Practical Guide To Quantitative Finance Interview
User Manual:
Open the PDF directly: View PDF 
.
Page Count: 212
| Download | |
| Open PDF In Browser | View PDF |
First Edition
Xinfeng Zhou
Edited by Brett Jiu
To the memory of my sister, Xinli Zhou
©Copyright 2008 by Xinfeng Zhou, http://www.quantfinanceinterviews.com
All right reserved.
No part of this book may be reproduced or transmitted in any form or by any means,
electronic or mechanical, including photocopying, recording or by any information
storage and retrieval system, without the written permission of the Publisher, except
where permitted by law.
Table of Contents
Chapter 1 General Principles ......................................................................................... 1
1. Build a broad knowledge base ....................................................................................................... 1
2. Practice your interview skills ......................................................................................................... 1
3. Listen carefully ............. ................................................................................................................. 2
4. Speak your mind ............................................................................................................................ 2
5. Make reasonable assumptions ...................... .................................................................................. 2
Chapter 2 Brain Teasers ................................................................................................. 3
2.1 Problem Simplification ............................................................................................ 3
Screwy pirates ................... ................................................................................................................. 3
Tiger and sheep .................................................................................................................................. 4
2.2 Logic Reasoning ............................................................ .......................................... 5
River crossing .................................................................................................................................... 5
Birthday problem ............. .................................................................................................................. 5
Card game .......................................................................................................................................... 6
Burning ropes ..................................................................................................................................... 7
Defective ball ..................................................................................................................................... 7
Trailing zeros ..................................................................................................................................... 9
Horse race .......................................................................................................................................... 9
Infinite sequence ................ .............................................................................................................. 10
2.3 Thinking Out of the Box ...................... .................................................................. 10
Box packing ..................................................................................................................................... 10
Calendar cubes ................................................................................................................................. 11
Door to offer .................................................... ................................................................................ 12
Message delivery ............................................................................................................................. 13
Last ball ........................................................... ................................................................................ 13
Light switches .................................................. ................................................................................ 14
Quant salary ..................................................................................................................................... 15
2.4 Application of Symmetry ...................................................................................... 15
Coin piles ......................................................................................................................................... 15
Mislabeled bags ............................................................................................................................... 16
Wise men ......................................................................................................................................... 17
2.5 Series Summation .................................................................................................. 17
Clock pieces ..................................................................................................................................... 18
Missing integers ............................................................................................................................... 18
Counterfeit coins I ........................................... ................................................................................ 19
2.6 The Pigeon Hole Principle ..................................................................................... 20
Matching socks ................................................................................................................................ 21
Handshakes ...................................................................................................................................... 21
Have we met before? ....................................... ................................................................................ 21
Ants on a square ............................................................................................................................... 22
Counterfeit coins II .......................................................................................................................... 22
Contents
2.7 Modular Arithmetic ................................................................................................ 23
Prisoner problem .............................................................................................................................. 24
Division by 9 ................................................................................................................................... 25
Chameleon colors ............................................................................................................................ 26
2.8 Math Induction ....................................................................................................... 27
Coin split problem ........................................... ................................................................................ 27
Chocolate bar problem ..................................................................................................................... 28
Race track ........................................................ ................................................................................ 29
2.9 Proof by Contradiction ........................................................................................... 31
Irrational number ............................................................................................................................. 31
Rainbow hats ................................................................................................................................... 31
Chapter 3 Calculus and Linear Algebra ...................................................................... 33
3.1 Limits and Derivatives ........................................................................................... 33
Basics of derivatives ........................................ ................................................................................ 33
Maximum and minimum ................................................................................................................. 34
L'Hospital's rule .............................................................................................................................. 35
3.2 Integration ........................................... ................................................................... 36
Basics of integration ........................................................................................................................ 36
Applications of integration .............................................................................................................. 38
Expected value using integration ..................................................................................................... 40
3.3 Partial Derivatives and Multiple Integrals ............................................................ .40
3.4 Important Calculus Methods .................................................................. ............... .41
Taylor's series .................................................................................................................................. 41
Newton's method ............................................................................................................................. 44
Lagrange multipliers ........................................................................................................................ 45
3.5 Ordinary Differential Equations ....... .................... ....................... ...... .....................46
Separable differential equations ....................................................................................................... 4 7
First-order linear differential equations ........................................................................................... 4 7
Homogeneous linear equations ......................... ............................................................................... 48
Nonhomogeneous linear equations .................. ............................... ................................................. 49
3.6 Linear Algebra ....................................................................................................... 50
Vectors ............................................................. ................................................................................ 50
QR decomposition ........................................................................................................................... 52
Determinant, eigenvalue and eigenvector ........................................................................................ 53
Positive semidefinite/definite matrix ............................................................................................... 56
LU decomposition and Cholesky decomposition ............................................................................ 57
Chapter 4 Probability Theory ....................................................................................... 59
4.1 Basic Probability Definitions and Set Operations .................................................. 59
Coin toss game ................................................. ................................................................................ 61
Card game ......................................................... ............................................................................... 61
Drunk passenger .............................................................................................................................. 62
ii
A Practical Guide To Quantitative Finance Interviews
N points on a circle .......................................................................................................................... 63
4.2 Combinatorial Analysis ......................................................................................... 64
Poker hands ...................................................................................................................................... 65
Hopping rabbit ................................................................................................................................. 66
Screwy pirates 2 ............................................................................................................ ................... 67
Chess tournament. .............................................. .............................................................................. 68
Application letters ......................................................................................................... ................... 69
Birthday problem ............................................................................................................................. 71
I OOth digit ..................................................................................................................... ................... 71
Cubic of integer ............................................................................................................................... 72
4.3 Conditional Probability and Bayes' formula .......................................... ............... 72
Boys and girls .............................................................................................................. .................... 73
All-girl world? .............................................................................................................. ................... 74
Unfair coin .................................................................................... ................................ ................... 74
Fair probability from an unfair coin ................................................................................................. 75
Dart game ......................................................................................................................................... 75
Birthday line .................................................................................................................................... 76
Dice order ........................................................................................................................................ 78
Monty Hall problem ....................................................................... .............................. .................... 78
Amoeba population ......................................................................... ................................................. 79
Candies in a jar ................................................................................................................................ 79
Coin toss game ............................................................................... .................................................. 80
Russian roulette series...................................................................................................................... 81
Aces ............................................................................................... .................................................. 82
Gambler's ruin problem ................................................................................................................... 83
Basketball scores ........................................................................................................... ................... 84
Cars on road ................................................................................................................. .................... 85
4.4 Discrete and Continuous Distributions .................................................................. 86
Meeting probability .......................................................................................................................... 88
Probability of triangle ...................................................................................................................... 89
Property of Poisson process ............................................................................................................. 90
Moments of normal distribution ................................................... ................................. .................. 91
4.5 Expected Value, Variance & Covariance .............................................. ................ 92
Connecting noodles ........................................................................ .................................................. 93
Optimal hedge ratio ...................................................................................................... ................... 94
Dice game ........................................................ .............................. ............................... ................... 94
Card game ........................................................ ................................................................................ 95
Sum of random variables ............................................................................................. .................... 95
Coupon collection ........................................................................................................ .................... 97
Joint default probability ................................................................. ............................... ................... 98
4.6 Order Statistics ...................................................................................................... 99
Expected value of max and min ....................................................................................................... 99
Correlation of max and min ........................................................................................... ................ 100
Random ants .................................................................................................................................. l 02
Chapter 5 Stochastic Process and Stochastic Calculus ............................................ 105
iii
Contents
5 .1 Markov Chain ........................................................................................... ........... 105
Gambler's ruin problem ................................................................................................................. 107
Dice question ................................................................................................................................. 108
Coin triplets ................................................................................................................................... l 09
Color balls ........................................................ .............................................................................. 113
5.2 Martingale and Random walk .............................................................................. 115
Drunk man .................................................................................... ................................ ................. 116
Dice game ...................................................................................................................................... 117
Ticket line .................................................................................................................... .................. 117
Coin sequence ................................................................................................................................ 119
5.3 Dynamic Programming ........................................................................................ 121
Dynamic programming (DP) algorithm ......................................................................................... 122
Dice game ...................................................................................................................................... 123
World series ..................................................................................... .............................................. 123
Dynamic dice game ....................................................................................................................... 126
Dynamic card game ....................................................................................................................... 127
5.4 Brownian Motion and Stochastic Calculus .......................................................... 129
Brownian motion ........................................................................................................................... 129
Stopping time/ first passage time ................................................................................................... 131
Ito's lemrna .................................................................................................................................... 135
Chapter 6 Finance ........................................................................................................ 137
6.1. Option Pricing ..................................................................................................... 137
Price direction of options ............................................................................................................... 137
Put-call parity ................................................................................................................................ 138
American v.s. European options .................................................................................................... 139
Black-Scholes-Merton differential equation .................................................................................. 142
Black-Scholes formula ................................................................................................................... 143
6.2. The Greeks .......................................................................................................... 149
Delta .............................................................................................................................................. 149
Gamma......................................................................................................................... .................. 152
Theta .............................................................................................................................................. 154
Vega ............................................................................................................................................... 156
6.3. Option Portfolios and Exotic Options ................................................................. 158
Bull spread ..................................................................................................................................... 159
Straddle .......................................................................................................................................... 159
Binary options ................................................................................................................ ................ 160
Exchange options ........................................................................................................................... 161
6.4. Other Finance Questions ................................................... .................................. 163
Portfolio optimization ................................................. ................................................................... 163
Value at risk ................................................................................................................................... 164
Duration and convexity .................................................................................................................. 165
Forward and futures ....................................................................................................................... 167
Interest rate models ........................................................................................................ ................ 168
IV
A Practical Guide To Quantitative Finance Interviews
Chapter 7 Algorithms and Numerical Methods ....................................................... 171
7.1. Algorithms .......................................... ... ............................................................. 171
Number swap ................................................................................................................................. 172
Unique elements ............................................................................................................................ 173
Homer's algorithm ........................................................................................................ ................. 174
Moving average ............................................................................................................................. 174
Sorting algorithm ........................................................................................................................... 174
Random permutation ...................................................................................................... ................ 176
Search algorithm ............................................................................................................................ 177
Fibonacci numbers ......................................................................................................................... 179
Maximum contiguous subarray ...................................................................................................... 180
7.2. The Power of Two ................................................................................. ............. 182
Power of 2? .................................................................................................................................... 182
Multiplication by 7........................................... .............................................................................. 182
Probability simulation .................................................................................................................... 182
Poisonous wine .............................................................................................................................. 183
7.3 Numerical Methods ............................................................................................. 184
Monte Carlo simulation ................................................................................................................. 184
Finite difference method ................................... ........................................................... .................. 189
v
Preface
This book will prepare you for quantitative finance interviews by helping you zero in on
the key concepts that are frequently tested in such interviews. In this book we analyze
solutions to more than 200 real interview problems and provide valuable insights into
how to ace quantitative interviews. The book covers a variety of topics that you are
likely to encounter in quantitative interviews: brain teasers, calculus, linear algebra,
probability, stochastic processes and stochastic calculus, finance and programming.
Professionals and students seeking to pursue a career in quantitative finance or related
quantitative fields will benefit most from thoroughly reading this book. In recent years,
we have seen a dramatic surge in demand for talents with strong quantitative skills from
investment banks, investment management firms, hedge funds, financial software
vendors and financial consulting companies. As a result, quant, an umbrella description
that encompasses quantitative analysts, quantitative researchers, quantitative strategists,
quantitative traders, and quantitative developers, has become an attractive career choice.
Dozens of financial engineering or computational finance programs have been
established in the last few years to educate professionals for quantitative finance jobs.
Graduates with backgrounds in finance, mathematics, physics, computer sciences, and
various engineering majors are contending for quant jobs as well. Naturally, the
competition is fierce. To be a successful candidate, you have to distinguish yourself
from many other excellent applicants.
In general, a successful candidate for a quantitative finance position is expected to have
a strong mathematics background (in probability, statistics, stochastic calculus, etc.),
solid programming skills and basic to intermediate-level finance knowledge. Most
candidates find quantitative interviews, or at least some interview problems, challenging.
Quantitative interviews cover a broad range of mathematics, finance and programming
topics that the candidates may have never used or even encountered in their daily work
or study. Moreover, most interview problems require strong problem-solving skills,
beyond reciting formulas or doing simple calculations. A successful candidate needs a
combination of knowledge and problem-solving skills in order to excel in quantitative
interviews. This is precisely what this book provides!
This book addresses these aspects by reviewing the necessary finance and mathematical
concepts that serve as tools to structure and solve interview problems. Since it includes
most of the topics used by quantitative interviewers, it presupposes some basic
preparation in mathematics, statistics, finance, and programming.
I also strongly recommend that you try to solve each problem on your own first before
reading the answer. Working out solutions on your own will help you improve your
problem-solving skills and help you quickly identify common approaches to tackling
quantitative problems.
Needless to say, you are likely to encounter some problems in interviews that are similar
to or exactly the same as the problems in this book. After all, the book covers many
essential quantitative topics using real interview problems. However, the goal of the
book is not to teach you how to game the system by remembering the answers! In fact,
just memorizing answers may not help much in your interview process. Unless you truly
understand the underlying concepts and can analyze the problems yourself, you will fail
to elaborate on the solutions and will be ill-equipped to answer many other problems
that use similar concepts. (Besides, many experienced quantitative interviewers are good
at catching those who have simply memorized "canned" answers.)
This is exactly the reason why I make significant effort to review essential concepts, to
present solution strategies, and to analyze the solutions in detail instead of simply
providing answers to problems. Furthermore, although the building blocks can be
learned, how one analyzes problems and implements these concepts usually makes a big
difference-and these are the skills you can acquire through practice, practice and
practice.
I realize that there may be better methods to solve some of the problems presented in
this book. It is entirely possible that despite my best efforts some inadvertent errors may
have crept in. Please email me at xinfeng@quantfinanceinterviews.com if you have a
better approach to solving some of these problems or find errors. I will be grateful for
your feedback and will post corrections and your constructive feedback on the book's
companion website http://www.quantfinanceinterviews.com. The website is a joint
venture with my editor, Brett Jiu. You will also find some extra interview problems
with answers that we have gathered.
I sincerely hope that you enjoy solving these problems and are successful in your
interviews.
Xinfeng Zhou
Notations
v
for each/for every/for all
3
there exists
therefore
whenever A is true, B is also true
s.t.
such that
a /\b
the minimum of a and b
avb
the maximum of a and b
n
TIX;
X1 XX2 X·· · XXn
i=I
factorial of nonnegative integer n, n ! =
n!
n
f1 i
( 0 ! = I)
i=I
x%y
modulo operation
<1>
empty set
Jf(x)dx
indefinite integral of f(x)
r
f(x)dx
definite integral of f(x) from a to b
max(x, 0)
N(µ, a 2 )
normal distribution with mean µ and variance a 2
cdf
pd/
cumulative density function
probability density function
Chapter 1 General Principles
Let us begin this book by exploring five general principles that will be extremely helpful
in your interview process. From my experience on both sides of the interview table,
these general guidelines will better prepare you for job interviews and will likely make
you a successful candidate.
1. Build a broad knowledge base
The length and the style of quant interviews differ from firm to firm. Landing a quant
job may mean enduring hours of bombardment with brain teaser, calculus, linear algebra,
probability theory, statistics, derivative pricing, or programming problems. To be a
successful candidate, you need to have broad knowledge in mathematics, finance and
programmmg.
Will all these topics be relevant for your future quant job? Probably not. Each specific
quant position often requires only limited knowledge in these domains. General problem
solving skills may make more difference than specific knowledge. Then why are
quantitative interviews so comprehensive? There are at least two reasons for this:
The first reason is that interviewers often have diverse backgrounds. Each interviewer
has his or her own favorite topics that are often related to his or her own educational
background or work experience. As a result, the topics you will be tested on are likely to
be very broad. The second reason is more fundamental. Your problem solving skills-a
crucial requirement for any quant job-is often positively correlated to the breadth of
your knowledge. A basic understanding of a broad range of topics often helps you better
analyze problems, explore alternative approaches, and come up with efficient solutions.
Besides, your responsibility may not be restricted to your own projects. You will be
expected to contribute as a member of a bigger team. Having broad knowledge will help
you contribute to the team's success as well.
The key here is "basic understanding." Interviewers do not expect you to be an expert on
a specific subject-unless it happens to be your PhD thesis. The knowledge used in
interviews, although broad, covers mainly essential concepts. This is exactly the reason
why most of the books I refer to in the following chapters have the word "introduction"
or "first" in the title. If I am allowed to give only one suggestion to a candidate, it will be
know the basics very well.
2. Practice your interview skills
The interview process starts long before you step into an interview room. In a sense, the
success or failure of your interview is often determined before the first question is asked.
Your solutions to interview problems may fail to reflect your true intelligence and
General Principles
knowledge if you are unprepared. Although a complete review of quant interview
problems is impossible and unnecessary, practice does improve your interview skills.
Furthermore, many of the behavioral, technical and resume-related questions can be
anticipated. So prepare yourself for potential questions long before you enter an
interview room.
3. Listen carefully
You should be an active listener in interviews so that you understand the problems well
before you attempt to answer them. If any aspect of a problem is not clear to you,
politely ask for clarification. If the problem is more than a couple of sentences, jot down
the key words to help you remember all the information. For complex problems,
interviewers often give away some clues when they explain the problem. Even the
assumptions they give may include some information as to how to approach the problem.
So listen carefully and make sure you get the necessary information.
4. Speak your mind
When you analyze a problem and explore different ways to solve it, never do it silently.
Clearly demonstrate your analysis and write down the important steps involved if
necessary. This conveys your intelligence to the interviewer and shows that you are
methodical and thorough. In case that you go astray, the interaction will also give your
interviewer the opportunity to correct the course and provide you with some hints.
Speaking your mind does not mean explaining every tiny detail. If some conclusions are
obvious to you, simply state the conclusion without the trivial details. More often than
not, the interviewer uses a problem to test a specific concept/approach. You should focus
on demonstrating your understanding of the key concept/approach instead of dwelling
on less relevant details.
5. Make reasonable assumptions
In real job settings, you are unlikely to have all the necessary information or data you'd
prefer to have before you build a model and make a decision. In interviews,
interviewers may not give you all the necessary assumptions either. So it is up to you to
make reasonable assumptions. The keyword here is reasonable. Explain your
assumptions to the interviewer so that you will get immediate feedback. To solve
quantitative problems, it is crucial that you can quickly make reasonable assumptions
and design appropriate frameworks to solve problems based on the assumptions.
We are now ready to review basic concepts in quantitative finance subject areas and
have fun solving real-world interview problems!
2
Chapter 2 Brain Teasers
In this chapter, we cover problems that only require common sense, logic, reasoning, and
basic-no more than high school level-math knowledge to solve. In a sense, they are
real brain teasers as opposed to mathematical problems in disguise. Although these brain
teasers do not require specific math knowledge, they are no less difficult than other
quantitative interview problems. Some of these problems test your analytical and general
problem-solving skills; some require you to think out of the box; while others ask you to
solve the problems using fundamental math techniques in a creative way. In this chapter,
we review some interview problems to explain the general themes of brain teasers that
you are likely to encounter in quantitative interviews.
2. 1 Problem Simplification
If the original problem is so complex that you cannot come up with an immediate
solution, try to identify a simplified version of the problem and start with it. Usually you
can start with the simplest sub-problem and gradually increase the complexity. You do
not need to have a defined plan at the beginning. Just try to solve the simplest cases and
analyze your reasoning. More often than not, you will find a pattern that will guide you
through the whole problem.
Screwy pirates
Five pirates looted a chest full of 100 gold coins. Being a bunch of democratic pirates,
they agree on the following method to divide the loot:
The most senior pirate will propose a distribution of the coins. All pirates, including the
most senior pirate, will then vote. If at least 50% of the pirates (3 pirates in this case)
accept the proposal, the gold is divided as proposed. If not, the most senior pirate will be
fed to shark and the process starts over with the next most senior pirate ... The process is
repeated until a plan is approved. You can assume that all pirates are perfectly rational:
they want to stay alive first and to get as much gold as possible second. Finally, being
blood-thirsty pirates, they want to have fewer pirates on the boat if given a choice
between otherwise equal outcomes.
How will the gold coins be divided in the end?
Solution: If you have not studied game theory or dynamic programming, this strategy
problem may appear to be daunting. If the problem with 5 pirates seems complex, we
can always start with a simplified version of the problem by reducing the number of
pirates. Since the solution to I-pirate case is trivial, let's start with 2 pirates. The senior
Brain Teasers
pirate (labeled as 2) can claim all the gold since he will always get 50% of the votes
from himself and pirate 1 is left with nothing.
Let's add a more senior pirate, 3. He knows that if his plan is voted down, pirate 1 will
get nothing. But if he offers private 1 nothing, pirate 1 will be happy to kill him. So
pirate 3 will offer private 1 one coin and keep the remaining 99 coins, in which strategy
the plan will have 2 votes from pirate 1 and 3.
If pirate 4 is added, he knows that if his plan is voted down, pirate 2 will get nothing. So
pirate 2 will settle for one coin if pirate 4 offers one. So pirate 4 should offer pirate 2 one
coin and keep the remaining 99 coins and his plan will be approved with 50% of the
votes from pirate 2 and 4.
Now we finally come to the 5-pirate case. He knows that if his plan is voted down, both
pirate 3 and pirate 1 will get nothing. So he only needs to offer pirate 1 and pirate 3 one
coin each to get their votes and keep the remaining 98 coins. If he divides the coins this
way, he will have three out of the five votes: from pirates 1 and 3 as well as himself.
Once we start with a simplified version and add complexity to it, the answer becomes
obvious. Actually after the case n = 5, a clear pattern has emerged and we do not need to
stop at 5 pirates. For any 2n + 1 pirate case (n should be less than 99 though), the most
senior pirate will offer pirates 1, 3, · · ·, and 2n -1 each one coin and keep the rest for
himself.
Tiger and sheep
One hundred tigers and one sheep are put on a magic island that only has grass. Tigers
can eat grass, but they would rather eat sheep. Assume: A. Each time only one tiger can
eat one sheep, and that tiger itself will become a sheep after it eats the sheep. B. All
tigers are smart and perfectly rational and they want to survive. So will the sheep be
eaten?
Solution: 100 is a large number, so again let's start with a simplified version of the
problem. If there is only 1 tiger ( n = 1 ), surely it will eat the sheep since it does not need
to worry about being eaten. How about 2 tigers? Since both tigers are perfectly rational,
either tiger probably would do some thinking as to what will happen if it eats the sheep.
Either tiger is probably thinking: if I eat the sheep, I will become a sheep; and then I will
be eaten by the other tiger. So to guarantee the highest likelihood of survival, neither
tiger will eat the sheep.
If there are 3 tigers, the sheep will be eaten since each tiger will realize that once it
changes to a sheep, there will be 2 tigers left and it will not be eaten. So the first tiger
that thinks this through will eat the sheep. If there are 4 tigers, each tiger will understand
4
A Practical Guide To Quantitative Finance Interviews
that if it eats the sheep, it will tum to a sheep. Since there are 3 other tigers, it will be
eaten. So to guarantee the highest likelihood of survival, no tiger will eat the sheep.
Following the same logic, we can naturally show that if the number of tigers is even, the
sheep will not be eaten. If the number is odd, the sheep will be eaten. For the case
n = l 00, the sheep will not be eaten.
2.2 Logic Reasoning
River crossing
Four people, A, B, C and D need to get across a river. The only way to cross the river is
by an old bridge, which holds at most 2 people at a time. Being dark, they can't cross the
bridge without a torch, of which they only have one. So each pair can only walk at the
speed of the slower person. They need to get all of them across to the other side as
quickly as possible. A is the slowest and takes 10 minutes to cross; B takes 5 minutes; C
takes 2 minutes; and D takes 1 minute.
What is the minimum time to get all of them across to the other side? 1
Solution: The key point is to realize that the l 0-minute person should go with the 5-
minute person and this should not happen in the first crossing, otherwise one of them
have to go back. So C and D should go across first (2 min); then send D back (lmin); A
and B go across ( 10 min); send C back (2min); C and D go across again (2 min).
It takes 17 minutes in total. Alternatively, we can send C back first and then D back in
the second round, which takes 17 minutes as well.
Birthday problem
You and your colleagues know that your boss A's birthday is one of the following 10
dates:
Mar 4, Mar 5, Mar 8
Jun 4, Jun 7
Sep 1, Sep 5
Dec 1, Dec 2, Dec 8
A told you only the month of his birthday, and told your colleague Conly the day. After
that, you first said: "I don't know A's birthday; C doesn't know it either." After hearing
1
Hint: The key is to realize that A and B should get across the bridge together.
5
Brain Teasers
what you said, C replied: "I didn't know A's birthday, but now I know it." You smiled
and said: "Now I know it, too." After looking at the 10 dates and hearing your comments,
your administrative assistant wrote down A's birthday without asking any questions. So
what did the assistant write?
Solution: Don't let the "he said, she said" part confuses you. Just interpret the logic
behind each individual's comments and try your best to derive useful information from
these comments.
Let D be the day of the month of A's birthday, we have De{l,2,4,5,7,8} . If the
birthday is on a unique day, C will know the A's birthday immediately. Among possible
Ds, 2 and 7 are unique days. Considering that you are sure that C does not know A's
birthday, you must infer that the day the C was told of is not 2 or 7. Conclusion: the
month is not June or December. (If the month had been June, the day C was told of may
have been 2; ifthe month had been December, the day C was told of may have been 7.)
Now C knows that the month must be either March or September. He immediately
figures out A's birthday, which means the day must be unique in the March and
September list. It means A's birthday cannot be Mar 5, or Sep 5. Conclusion: the
birthday must be Mar 4, Mar 8 or Sep 1.
Among these three possibilities left, Mar 4 and Mar 8 have the same month. So if the
month you have is March, you still cannot figure out A's birthday. Since you can figure
out A's birthday, A's birthday must be Sep 1. Hence, the assistant must have written Sep
1.
Card game
A casino offers a card game using a normal deck of 52 cards. The rule is that you tum
over two cards each time. For each pair, if both are black, they go to the dealer's pile; if
both are red, they go to your pile; if one black and one red, they are discarded. The
process is repeated until you two go through all 52 cards. If you have more cards in your
pile, you win $100; otherwise (including ties) you get nothing. The casino allows you to
negotiate the price you want to pay for the game. How much would you be willing to
pay to play this game?2
Solution: This surely is an insidious casino. No matter how the cards are arranged, you
and the dealer will always have the same number of cards in your piles. Why? Because
each pair of discarded cards have one black card and one red card, so equal number of
Hint: Try to approach the problem using symmetry. Each discarded pair has one black and one red card.
What does that tell you as to the number of black and red cards in the rest two piles?
2
6
A Practical Guide To Quantitative Finance Interviews
red and black cards are discarded. As a result, the number of red cards left for you and
the number of black cards left for the dealer are always the same. The dealer always
wins! So we should not pay anything to play the game.
Burning ropes
You have two ropes, each of which takes I hour to bum. But either rope has different
densities at different points, so there's no guarantee of consistency in the time it takes
different sections within the rope to bum. How do you use these two ropes to measure 45
minutes?
Solution: This is a classic brain teaser question. For a rope that takes x minutes to bum,
if you light both ends of the rope simultaneously, it takes x I 2 minutes to bum. So we
should light both ends of the first rope and light one end of the second rope. 30 minutes
later, the first rope will get completely burned, while that second rope now becomes a
30-min rope. At that moment, we can light the second rope at the other end (with the
first end still burning), and when it is burned out, the total time is exactly 45 minutes.
Defective ball
You have 12 identical balls. One of the balls is heavier OR lighter than the rest (you
don't know which). Using just a balance that can only show you which side of the tray is
heavier, how can you determine which ball is the defective one with 3 measurements?3
Solution: This weighing problem is another classic brain teaser and is still being asked
by many interviewers. The total number of balls often ranges from 8 to more than 100.
Here we use n = 12 to show the fundamental approach. The key is to separate the
original group (as well as any intermediate subgroups) into three sets instead of two. The
reason is that the comparison of the first two groups always gives information about the
third group.
Considering that the solution is wordy to explain, I draw a tree diagram in Figure 2.1 to
show the approach in detail. Label the balls 1 through 12 and separate them to three
groups with 4 balls each. Weigh balls 1, 2, 3, 4 against balls 5, 6, 7, 8. Then we go on to
explore two possible scenarios: two groups balance, as expressed using an "=" sign, or 1,
3
Hint: First do it for 9 identical balls and use only 2 measurements, knowing that one is heavier than the
rest.
7
Brain Teasers
2, 3, 4 are lighter than 5, 6, 7, 8, as expressed using an "<" sign. There is no need to
explain the scenario that 1, 2, 3, 4 are heavier than 5, 6, 7, 8. (Why?4)
If the two groups balance, this immediately tells us that the defective ball is in 9, 10, 11
and 12, and it is either lighter (l) or heavier (H) than other balls. Then we take 9, 10 and
11 from group 3 and compare balls 9, 10 with 8, 11. Here we have already figured out
that 8 is a normal ball. If 9, IO are lighter, it must mean either 9 or 10 isl or 11 is H. In
which case, we just compare 9 with 10. If 9 is lighter, 9 is the defective one and it is L; if
9 and 10 balance, then 11 must be defective and H; If 9 is heavier, 10 is the defective
one and it is L. lf9, IO and 8, 11 balance, 12 is the defective one. lf9, 10 is heavier, than
either 9 or 10 is H, or 11 is L.
You can easily follow the tree in Figure 2.1 for further analysis and it is clear from the
tree that all possible scenarios can be resolved in 3 measurements.
I /2/3/4 L or 5/617/8 H
l/2L or 6H
IL
6H
4L or 7/8H
2L
8H
4L
7H
9/ I0/11/ 12 Lor H
5H or 3L
3L
9/IOLor l IH
5H
9L
l IH
IOL
12Lorl211
12H
12L
9/IOfl or 11 L
IOH
I IL
9H
Figure 2.1 Tree diagram to identify the defective ball in 12 balls
In general if you have the information as to whether the defective ball is heavier or
Here is where the symmetry idea comes in. Nothing makes the I, 2, 3, 4 or 5, 6, 7, 8 labels special. If I, 2,
3, 4 are heavier than 5, 6, 7, 8, let's just exchange the labels of these two groups. Again we have the case
of I, 2, 3, 4 being lighter than 5, 6, 7, 8.
4
8
A Practical Guide To Quantitative Finance Interviews
lighter, you can identify the defective ball among up to 3n balls using no more than n
measurements since each weighing reduces the problem size by 2/3. If you have no
information as to whether the defective ball is heavier or lighter, you can identify the
defective ball among up to (3n - 3) I 2 balls using no more than n measurements.
Trailing zeros
How many trailing zeros are there in 100! (factorial of 100)?
Solution: This is an easy problem. We know that each pair of 2 and 5 will give a trailing
zero. If we perform prime number decomposition on all the numbers in 100!, it is
obvious that the frequency of 2 will far outnumber of the frequency of 5. So the
frequency of 5 determines the number of trailing zeros. Among numbers 1, 2, · · ·, 99, and
100, 20 numbers are divisible by 5 ( 5, 10, · · ·, 100 ). Among these 20 numbers, 4 are
divisible by 52 ( 25, 50, 75, 100 ). So the total frequency of 5 is 24 and there are 24
trailing zeros.
Horse race
There are 25 horses, each of which runs at a constant speed that is different from the
other horses'. Since the track only has 5 lanes, each race can have at most 5 horses. If
you need to find the 3 fastest horses, what is the minimum number of races needed to
identify them?
Solution: This problem tests your basic analytical skills. To find the 3 fastest horses,
surely all horses need to be tested. So a natural first step is to divide the horses to 5
groups (with horses 1-5, 6-10, 11-15, 16-20, 21-25 in each group). After 5 races, we will
have the order within each group, let's assume the order follows the order of numbers
(e.g., 6 is the fastest and IO is the slowest in the 6-10 group)5. That means 1, 6, 11, 16
and 21 are the fastest within each group.
Surely the last two horses within each group are eliminated. What else can we infer? We
know that within each group, if the fastest horse ranks 5th or 4th among 25 horses, then
all horses in that group cannot be in top 3; if it ranks the 3rd, no other horse in that group
can be in the top 3; if it ranks the 2nd, then one other horse in that group may be in top 3;
if it ranks the first, then two other horses in that group may be in top 3.
5 Such an assumption does not affect the generality of the solution. If the order is not as described, just
change the labels of the horses.
9
Brain Teasers
So let's race horses 1, 6, 11, 16 and 21. Again without loss of generality, let's assume
the order is 1, 6, 11, 16 and 21. Then we immediately know that horses 4-5, 8-10, 12-15,
16-20 and 21-25 are eliminated. Since 1 is fastest among all the horses, 1 is in. We need
to determine which two among horses 2, 3, 6, 7 and 11 are in top 3, which only takes one
extra race.
So all together we need 7 races (in 3 rounds) to identify the 3 fastest horses.
Infinite sequence
If x /\ x /\ x /\ x /\ x · · · = 2 , where x /\ y = xY, what is x ?
Solution: This problem appears to be difficult, but a simple analysis will give an elegant
solution. What do we have from the original equation?
limx/\x/\x/\x/\x···=2<=>limx/\x/\x/\x/\x···=2.
n terms
In other words,
as
n-1 terms
adding or minus one x /\ should yield the same result.
so x /\ x /\ x /\ x /\ x · · · = x /\ (x /\ x /\ x /\ x · · ·) = x /\ 2 = 2
x
= J2.
2.3 Thinking Out of the Box
Box packing
Can you pack 53 bricks of dimensions 1x1x4 into a 6 x 6 x 6 box?
Solution: This is a nice problem extended from a popular chess board problem. In that
problem, you have a 8 x 8 chess board with two small squares at the opposite diagonal
comers removed. You have many bricks with dimension 1x 2. Can you pack 31 bricks
into the remaining 62 squares? (An alternative question is whether you can cover all 62
squares using bricks without any bricks overlapping with each other or sticking out of
the board, which requires a similar analysis.)
A real chess board figure surely helps the visualization. As shown in Figure 2.2, when a
chess board is filled with alternative black and white squares, both squares at the
opposite diagonal comers have the same color. If you put a 1x 2 brick on the board, it
will always cover one black square and one white square. Let's say it's the two black
comer squares were removed, then the rest of the board can fit at most 30 bricks since
we only have 30 black squares left (and each brick requires one black square). So to
pack 31 bricks is out of the question. To cover all 62 squares without overlapping or
overreaching, we must have exactly 3 I bricks. Yet we have proved that 31 bricks cannot
10
A Practical Guide To Quantitative Finance Interviews
fit in the 62 squares left, so you cannot find a way to fill in all 62 squares without
overlapping or overreaching.
Removed
Removed +Figure 2.2 Chess board with alternative black and white squares
Just as any good trading strategy, if more and more people get to know it and replicate it,
the effectiveness of such a strategy will disappear. As the chess board problem becomes
popular, many interviewees simply commit it to memory (after all, it's easy to remember
the answer). So some ingenious interviewer came up with the newer version to test your
thinking process, or at least your ability to extend your knowledge to new problems.
lfwe look at the total volume in this 30 problem, 53 bricks have a volume of 212, which
is smaller then the box's volume 216. Yet we can show it is impossible to pack all the
bricks into the box using a similar approach as the chess board problem. Let's imagine
that the 6 x 6 x 6 box is actually comprised of small 2 x 2 x 2 cubes. There should be 27
small cubes. Similar to the chess board (but in 30), imagine that we have black cubes
and white cubes alternates-it does take a little 30 visualization. So we have either 14
black cubes & 13 white cubes or 13 black cubes & 14 white cubes. For any 1x1x4 brick
that we pack into the box, half ( 1x1x2) of it must be in a black 2 x 2 x 2 cube and the
other half must be in a white 2 x 2 x 2 cube. The problem is that each 2 x 2 x 2 cube can
only be used by 4 of the 1x1 x 4 bricks. So for the color with 13 cubes, be it black or
white, we can only use them for 52 1x1x4 tubes. There is no way to place the 53th
brick. So we cannot pack 53 bricks of dimensions 1x1x4 into a 6 x 6 x 6 box.
Calendar cubes
You just had two dice custom-made. Instead of numbers 1 - 6, you place single-digit
numbers on the faces of each dice so that every morning you can arrange the dice in a
way as to make the two front faces show the current day of the month. You must use
both dice (in other words, days 1 - 9 must be shown as 01 - 09), but you can switch the
11
Brain Teasers
order of the dice if you want. What numbers do you have to put on the six faces of each
of the two dice to achieve that?
Solution: The days of a month include 11 and 22, so both dice must have 1 and 2. To
express single-digit days, we need to have at least a 0 in one dice. Let's put a 0 in dice
one first. Considering that we need to express all single digit days and dice two cannot
have all the digits from 1 - 9, it's necessary to have a 0 in dice two as well in order to
express all single-digit days.
So far we have assigned the following numbers:
!
Dice one
Dice two
I
I:
If we can assign all the rest of digits 3, 4, 5, 6, 7, 8, and 9 to the rest of the faces, the
problem is solved. But there are 7 digits left. What can we do? Here's where you need to
think out of the box. We can use a 6 as a 9 since they will never be needed at the same
time! So, simply put 3, 4, and 5 on one dice and 6, 7, and 8 on the other dice, and the
final numbers on the two dice are:
Dice one
1
2
0
3
4
5
Dice two
1
2
0
6
7
8
Door to offer
You are facing two doors. One leads to your job offer and the other leads to exit. In front
of either door is a guard. One guard always tells lies and the other always tells the truth.
You can only ask one guard one yes/no question. Assuming you do want to get the job
offer, what question will you ask?
Solution: This is another classic brain teaser (maybe a little out-of-date in my opinion).
One popular answer is to ask one guard: "Would the other guard say that you are
guarding the door to the offer?" If he answers yes, choose the other door; if he answers
no, choose the door this guard is standing in front of.
There are two possible scenarios:
I . Truth teller guards the door to offer; Liar guards the door to exit.
2. Truth teller guards the door to exit; Liar guards the door to offer.
If we ask a guard a direct question such as "Are you guarding the door to the offer?" For
scenario I, both guards will answer yes; for scenario 2, both guards will answer no. So a
12
A Practical Guide To Quantitative Finance Interviews
direct question does not help us solve the problem. The key is to involve both guards in
the questions as the popular answer does. For scenario 1, if we happen to choose the
truth teller, he will answer no since the liar will say no; if we happen to choose the liar
guard, he will answer yes since the truth teller will say no. For scenario 2, if we happen
to choose the truth teller, he will answer yes since the liar will say yes; if we happen to
choose the liar guard, he will answer no since the truth teller with say yes. So for both
scenarios, if the answer is no, we choose that door; if the answer is yes, we choose the
other door.
Message delivery
You need to communicate with your colleague in Greenwich via a messenger service.
Your documents are sent in a padlock box. Unfortunately the messenger service is not
secure, so anything inside an unlocked box will be lost (including any locks you place
inside the box) during the delivery. The high-security padlocks you and your colleague
each use have only one key which the person placing the lock owns. How can you
securely send a document to your colleague?6
Solution: If you have a document to deliver, clearly you cannot deliver it in an unlocked
box. So the first step is to deliver it to Greenwich in a locked box. Since you are the
person who has the key to that lock, your colleague cannot open the box to get the
document. Somehow you need to remove the lock before he can get the document,
which means the box should be sent back to you before your colleague can get the
document.
So what can he do before he sends back the box? He can place a second lock on the box,
which he has the key to! Once the box is back to you, you remove your own lock and
send the box back to your colleague. He opens his own lock and gets the document.
Last ball
A bag has 20 blue balls and 14 red balls. Each time you randomly take two balls out.
(Assume each ball in the bag has equal probability of being taken). You do not put these
two balls back. Instead, if both balls have the same color, you add a blue ball to the bag;
if they have different colors, you add a red ball to the bag. Assume that you have an
unlimited supply of blue and red balls, if you keep on repeating this process, what will
be the color of the last ball left in the bag?7 What if the bag has 20 blue balls and l 3 red
balls instead?
6
7
Hint: You can have more than one lock on the box.
Hint: Consider the changes in the number ofred and blue balls after each step.
13
Brain Teasers
Solution: Once you understand the hint, this problem should be an easy one. Let (B, R)
represent the number of blue balls and red balls in the bag. We can take a look what will
happen after two balls are taken out.
Both balls are blue: (B, R)
(B-1, R)
Both balls are red: (B, R)
(B + 1, R-2)
One red and one blue: (B, R)
(B -1, R)
Notice that R either stays the same or decreases by 2, so the number of red balls will
never become odd if we begin with 14 red balls. We also know that the total number of
balls decreases by one each time until only one ball is left. Combining the information
we have, the last ball must be a blue one. Similarly, when we start with odd number of
red balls, the final ball must be a red one.
Light switches
There is a light bulb inside a room and four switches outside. All switches are currently
at off state and only one switch controls the light bulb. You may tum any number of
switches on or off any number of times you want. How many times do you need to go
into the room to figure out which switch controls the light bulb?
Solution: You may have seen the classical version of this problem with 3 light bulbs
inside the room and 3 switches outside. Although this problem is slightly modified, the
approach is exact the same. Whether the light is on and off is binary, which only allows
us to distinguish two switches. If we have another binary factor, there are 2 x 2 = 4
possible combinations of scenarios, so we can distinguish 4 switches. Besides light, a
light bulb also emits heat and becomes hot after the bulb has been lit for some time. So
we can use the on/off and cold/hot combination to decide which one of the four switches
controls the light.
Tum on switches 1 and 2; move on to solve some other puzzles or do whatever you like
for a while; tum off switch 2 and turn on switch 3; get into the room quickly, touch the
bulb and observe whether the light is on or off.
The light bulb is on and hot - switch 1 controls the light;
The light bulb is off and hot - switch 2 controls the light;
The light bulb is on and cold - switch 3 controls the light;
The light bulb is off and cold - switch 4 controls the light.
14
A Practical Guide To Quantitative Finance Interviews
Quant salary
Eight quants from different banks are getting together for drinks. They are all interested
in knowing the average salary of the group. Nevertheless, being cautious and humble
individuals, everyone prefers not to disclose his or her own salary to the group. Can you
come up with a strategy for the quants to calculate the average salary without knowing
other people's salaries?
Solution: This is a light-hearted problem and has more than one answer. One approach is
for the first quant to choose a random number, adds it to his/her salary and gives it to the
second quant. The second quant will add his/her own salary to the result and give it to
the third quant; ... ; the eighth quant will add his/her own salary to the result and give it
back to the first quant. Then the first quant will deduct the "random" number from the
total and divide the "real" total by 8 to yield the average salary.
You may be wondering whether this strategy has any use except being a good brain
teaser to test interviewees. It does have applications in practice. For example, a third
party data provider collect fund holding position data (securities owned by a fund and
the number of shares) from all participating firms and then distribute the information
back to participants. Surely most participants do not want others to figure out what they
are holding. If each position in the fund has the same fund ID every day, it's easy to
reverse-engineer the fund from the holdings and to replicate the strategy. So different
random numbers (or more exactly pseudo-random numbers since the provider knows
what number is added to the fund ID of each position and complicated algorithm is
involved to make the mapping one to one) are added to the fund ID of each position in
the funds before distribution. As a result, the positions in the same fund appear to have
different fund IDs. That prevents participants from re-constructing other funds. Using
this approach, the participants can share market information and remain anonymous at
the same time.
2.4 Application of Symmetry
Coin piles
Suppose that you are blind-folded in a room and are told that there are 1000 coins on the
floor. 980 of the coins have tails up and the other 20 coins have heads up. Can you
separate the coins into two piles so to guarantee both piles have equal number of heads?
Assume that you cannot tell a coin's side by touching it, but you are allowed to tum over
any number of coins.
Solution: Let's say that we separate the 1000 coins into two piles with n coins in one pile
and 1000 - n coins in the other. If there are m coins in the first pile with heads up, there
15
Brain Teasers
must be 20 - m coins in the second pile with heads up. We also know that there are
n - m coins in the first pile with tails up. We clearly cannot guarantee that m = 10 by
simply adjusting n.
What other options do we have? We can tum over coins if we want to. Since we have no
way of knowing what a coin's side is, it won't guarantee anything if we selectively flip
coins. However, if we flip all the coins in the first pile, all heads become tails and all
tails become heads. As a result, it will have n-m heads and m tails (symmetry). So, to
start, we need to make the number of tails in the original first pile equal to the number of
heads in the second pile; in other words, to make n - m = 20- m. n = 20 makes the
equation hold. If we take 20 coins at random and tum them all over, the number of heads
among these turned-over 20 coins should be the same as the number of heads among the
other 980 coins.
Mislabeled bags
You are given three bags of fruits. One has apples in it; one has oranges in it; and one
has a mix of apples and oranges in it. Each bag has a label on it (apple, orange or mix).
Unfortunately, your manager tells you that ALL bags are mislabeled. Develop a strategy
to identify the bags by taking out minimum number of fruits? You can take any number
of fruits from any bags. 8
Solution: The key here is to use the fact that ALL bags are mislabeled. For example, a
bag labeled with apple must contain either oranges only or a mix of oranges and apples.
Let's look at the labels: orange, apple, mix (orange+ apple). Have you realized that the
orange label and the apple label are symmetric? If not, let me explain it in detail: If you
apple), then
pick a fruit from the bag with the orange label and it's an apple (orange
the bag is either all apples or a mix. If you pick a fruit from the bag with the apple label
orange), then the bag is either an orange bag or a mix.
and it's an orange (apple
Symmetric labels are not exciting and are unlikely to be the correct approach. So let's try
the bag with the mix label and get one fruit from it. If the fruit we get is an orange, then
we know that bag is actually orange (It cannot be a mix of oranges and apples since we
know the bag's label is wrong). Since the bag with the apple label cannot be apple only,
it must be the mix bag. And the bag with the orange label must be the apple bag.
Similarly, for the case that apples are in the bag with the mix label, we can figure out all
the bags using one single pick.
The problem struck me as a word game when I first saw it. But it does test a candidate's attention to
details besides his or her logic reasoning skills.
8
16
A Practical Guide To Quantitative Finance Interviews
Wise men
A sultan has captured 50 wise men. He has a glass currently standing bottom down.
Every minute he calls one of the wise men who can choose either to tum it over (set it
upside down or bottom down) or to do nothing. The wise men will be called randomly,
possibly for an infinite number of times. When someone called to the sultan correctly
states that all wise men have already been called to the sultan at least once, everyone
goes free. But if his statement is wrong, the sultan puts everyone to death. The wise men
are allowed to communicate only once before they get imprisoned into separate rooms
(one per room). Design a strategy that lets the wise men go free.
Solution: For the strategy to work, one wise man, let's call him the spokesman, will state
that every one has been called. What does that tell us? 1. All the other 49 wise men are
equivalent (symmetric). 2. The spokesman is different from the other 49 men. So
naturally those 49 equivalent wise men should act in the same way and the spokesman
should act differently.
Here is one of such strategies: Every one of the 49 (equivalent) wise men should flip the
glass upside down the first time that he sees the glass bottom down. He does nothing if
the glass is already upside down or he has flipped the glass once. The spokesman should
flip the glass bottom down each time he sees the glass upside down and he should do
nothing if the glass is already bottom down. After he does the 49th flip, which means all
the other 49 wise men have been called, he can declare that all the wise men have been
called.
2.5 Series Summation
Here is a famous story about the legendary mathematician/physicist Gauss: When he
was a child, his teacher gave the children a boring assignment to add the numbers from 1
to 100. To the amazement of the teacher, Gauss turned in his answer in less than a
minute. Here is his approach:
100
:Ln= 1
n=I
100
+ 2+ ··· + 99+ 100
+
+
+
+
:Ln = 100+ 99+···+
2 + 1
! !
! ! 101x100 :Ln =100xl01
2Ln = 101+101+···+101+101=
--2
100
n=I
n =I
17
Brain Teasers
This approach can be generalized to any integer N:
f
n=I
n = N ( N + 1)
2
The summation formula for consecutive squares may not be as intuitive:
fn =
2
N(N +1)(2N +1)
= N 3 + N 2 + N.
6
n=I
3
2
6
N
But if we correctly guess that
Ln
2
= aN 3 + bN 2 + cN + d and apply the initial
n=I
conditions
N=l
l=a+b+c+d
N=2
5 = 8a + 4b + 2c + d
we will have the solution that a= 1/3, b = 112, c = 116, d = 0. We can then easily show
that the same equation applies to all N by induction.
Clock pieces
A clock (numbered 1 - 12 clockwise) fell off the wall and broke into three pieces. You
find that the sums of the numbers on each piece are equal. What are the numbers on each
piece? (No strange-shaped piece is allowed.)
Solution: Using the summation equation,
L n = 12x13
2
12
78. So the numbers on each
n=I
piece must sum up to 26. Some interviewees mistakenly assume that the numbers on
each piece have to be continuous because no strange-shaped piece is allowed. It' s easy to
see that 5, 6, 7 and 8 add up to 26. Then the interviewees' thinking gets stuck because
they cannot find more consecutive numbers that add up to 26.
Such an assumption is not correct since 12 and 1 are continuous on a clock. Once that
wrong assumption is removed, it becomes clear that 12+1=13 and 11+2=13. So the
second piece is 11, 12, 1 and 2; the third piece is 3, 4, 9 and 10.
Missing integers
Suppose we have 98 distinct integers from I to I 00. What is a good way to find out the
two missing integers (within [ l, 100])?
18
A Practical Guide To Quantitative Finance Interviews
Solution: Denote the missing integers as x and y, and the existing ones are z 1 , • • ·, z98 •
Applying the summation equations, we have
100
98
100x101
n=I
i=I
2
L n = x + y + L z; => x + y =
98
- L, z;
i=I
Using these two equations, we can easily solve x and y . If you implement this strategy
using a computer program, it is apparent that the algorithm has a complexity of O(n) for
two missing integers in 1 to n.
Counterfeit coins I
There are 10 bags with 100 identical coins in each bag. In all bags but one, each coin
weighs 10 grams. However, all the coins in the counterfeit bag weigh either 9 or 11
grams. Can you find the counterfeit bag in only one weighing, using a digital scale that
tells the exact weight? 9
Solution: Yes, we can identify the counterfeit bag using one measurement. Take 1 coin
out of the first bag, 2 out of the second bag, 3 out the third bag, · · ·, and 10 coins out of
10
the tenth bag. All together, there are
Ln = 55 coins. If there were no counterfeit coins,
i=I
they should weigh 550 grams. Let's assume the i-th bag is the counterfeit bag, there will
be i counterfeit coins, so the final weight will be 550 ± i. Since i is distinct for each bag,
we can identify the counterfeit coin bag as well as whether the counterfeit coins are
lighter or heavier than the real coins using 550 ± i.
This is not the only answer: we can choose other numbers of coins from each bag as long
as they are all different numbers.
Glass balls
You are holding two glass balls in a 100-story building. If a ball is thrown out of the
window, it will not break if the floor number is less than X, and it will always break if
9 Hint: In order to find the counterfeit coin bag in one weighing, the number of coins from each bag must
be different. If we use the same number of coins from two bags, symmetry will prevent you from
distinguish these two bags if one is the counterfeit coin bag.
19
Brain Teasers
the floor number is equal to or greater than X. You would like to determine X. What is
the strategy that will minimize the number of drops for the worst case scenario? 10
Solution: Suppose that we have a strategy with a maximum of N throws. For the first
throw of ball one, we can try the N-th floor. If the ball breaks, we can start to try the
second ball from the first floor and increase the floor number by one until the second
ball breaks. At most, there are N -1 floors to test. So a maximum of N throws are
enough to cover all possibilities. If the first ball thrown out of N-th floor does not break,
we have N -1 throws left. This time we can only increase the floor number by N -1 for
the first ball since the second ball can only cover N - 2 floors if the first ball breaks. If
the first ball thrown out of (2N-1 )th floor does not break, we have N - 2 throws left. So
we can only increase the floor number by N - 2 for the first ball since the second ball
can only cover N - 3 floors if the first ball breaks ...
Using such logic, we can see that the number of floors that these two balls can cover
with a maximum of N throws is N + (N -1) + .. · + 1 = N ( N + 1) I 2 . In order to cover 100
stories, we need to have N(N + 1)/ 2:?: 100. Taking the smallest integer, we have N = 14.
Basically, we start the first ball on the 14th floor, if the ball breaks, we can use the
second ball to try floors l, 2, · · ·, 13 with a maximum throws of 14 (when the 13th or the
14th floor is X). If the first ball does not break, we will try the first ball on the
14+(14-1)=27th floor. If it breaks, we can use the second ball to cover floors
15, 16, .. ·, 26 with a total maximum throws of 14 as well...
2.6 The Pigeon Hole Principle
Here is the basic version of the Pigeon Hole Principle: if you have fewer pigeon holes
than pigeons and you put every pigeon in a pigeon hole, then at least one pigeon hole has
more than one pigeon. Basically it says that if you have n holes and more than n+ 1
pigeons, at least 2 pigeons have to share one of the holes. The generalized version is that
if you have n holes and at least mn + 1 pigeons, at least m +I pigeons have to share one
of the holes. These simple and intuitive ideas are surprisingly useful in many problems.
Here we will use some examples to show their applications.
10 Hint: Assume we design a strategy with N maximum throws. If the first ball is thrown once, the second
ball can cover N - I floors; if the first ball is thrown twice, the second ball can cover N - 2 floors ...
20
A Practical Guide To Quantitative Finance Interviews
Matching socks
Your drawer contains 2 red socks, 20 yellow socks and 31 blue socks. Being a busy and
absent-minded MIT student, you just randomly grab a number of socks out of the draw
and try to find a matching pair. Assume each sock has equal probability of being
selected, what is the minimum number of socks you need to grab in order to guarantee a
pair of socks of the same color?
Solution: This question is just a variation of the even simpler version of two-color-socks
problem, in which case you only need 3. When you have 3 colors (3 pigeon holes), by
the Pigeon Hole Principle, you will need to have 3 + 1 = 4 socks (4 pigeons) to guarantee
that at least two socks have the same color (2 pigeons share a hole).
Handshakes
You are invited to a welcome party with 25 fellow team members. Each of the fellow
members shakes hands with you to welcome you. Since a number of people in the room
haven't met each other, there's a lot of random handshaking among others as well. If you
don't know the total number of handshakes, can you say with certainty that there are at
least two people present who shook hands with exactly the same number of people?
Solution: There are 26 people at the party and each shakes hands with from I-since
everyone shakes hands with you-to 25 people. In other words, there are 26 pigeons and
25 holes. As a result, at least two people must have shaken hands with exactly the same
number of people.
Have we met before?
Show me that, ifthere are 6 people at a party, then either at least 3 people met each other
before the party, or at least 3 people were strangers before the party.
Solution: This question appears to be a complex one and interviewees often get puzzled
by what the interviewer exactly wants. But once you start to analyze possible scenarios,
the answer becomes obvious.
Let's say that you are the 6th person at the party. Then by generalized Pigeon Hole
Principle (Do we even need that for such an intuitive conclusion?), among the remaining
5 people, we conclude that either at least 3 people met you or at least 3 people did not
meet you. Now let's explore these two mutually exclusive and collectively exhaustive
scenarios:
Case 1: Suppose that at least 3 people have met you before.
21
Brain Teasers
If two people in this group met each other, you and the pair (3 people) met each other. If
no pair among these people met each other, then these people ( 2 3 people) did not meet
each other. In either sub-case, the conclusion holds.
Case 2: Suppose at least 3 people have not met you before.
If two people in this group did not meet each other, you and the pair (3 people) did not
meet each other. If all pairs among these people knew each other, then these people ( 2 3
people) met each other. Again, in either sub-case, the conclusion holds.
Ants on a square
There are 51 ants on a square with side length of 1. If you have a glass with a radius of
1/7, can you put your glass at a position on the square to guarantee that the glass
encompasses at least 3 ants? 11
Solution: To guarantee that the glass encompasses at least 3 ants, we can separate the
square into 25 smaller areas. Applying the generalized Pigeon Hole Principle, we can
show that at least one of the areas must have at least 3 ants. So we only need to make
sure that the glass is large enough to cover any of the 25 smaller areas. Simply separate
the area into 5 x 5 smaller squares with side length of 115 each will do since a circle with
radius of 117 can cover a square 12 with side length 1/5.
Counterfeit coins II
There are 5 bags with l 00 coins in each bag. A coin can weigh 9 grams, l 0 grams or 11
grams. Each bag contains coins of equal weight, but we do not know what type of coins
a bag contains. You have a digital scale (the kind that tells the exact weight). How many
times do you need to use the scale to determine which type of coin each bag contains? 13
Solution: If the answer for 5 bags is not obvious, let's start with the simplest version of
the problem-I bag. We only need to take one coin to weigh it. Now we can move on to
2 bags. How many coins do we need to take from bag 2 in order to determine the coin
types of bag l and bag 2? Considering that there are three possible types for bag 1, we
will need three coins from bag 2; two coins won't do. For notation simplicity, let's
change the number/weight for three types to - I, 0 and l (by removing the mean 10). If
Hint: Separate the square into 25 smaller areas; then at least one area has 3 ants in it.
A circle with radius r can cover a square with side length up to ,/2 rand ,/2 ;::: 1.414.
13 Hint: Start with a simpler problem. What if you have two bags of coins instead of 5, how many coins do
you need from each bag to find the type of coins in either bag? What is the minimum difference in coin
numbers? Then how about three bags?
11
12
22
A Practical Guide To Quantitative Finance Interviews
we only use 2 coins from bag 2, the final sum for 1 coin from bag 1 and 2 coins from
bag 2 ranges from -3 to 3 (7 pigeon holes). At the same time we have 9 ( 3 x 3) possible
combinations for the weights of coins in bag I and bag 2 (9 pigeons). So at least two
combinations will yield the same final sum (9>7, so at least two pigeons need to share
one hole), and we can not distinguish them. If we use 3 coins from bag 2, then the sum
ranges from -4 to 4, which is possible to cover all 9 combinations. The following table
exactly shows that all possible combinations yield different sums:
Sum
1 coin, bag 1
N
-I
0
I
0.0
=
.,;
-I
-4
-3
-2
·c;
0
-I
0
I
I
2
3
4
c
u
t")
Cl and C2 represent the weights ofcoins from bag 1 and 2 respectively.
Then how about 3 bags? We are going to have 33 = 27 possible combinations. Surely an
indicator ranging from -13 to 13 will cover it and we will need 9 coins from bag 3. The
possible combinations are shown in the following table:
Sum
C2 =-1
t")
0.0
C2=0
C2=1
-I
0
I
-I
0
I
-I
0
I
=.;;
-I
-13
-12
-II
-IO
-9
-8
-7
-6
-5
·c;
0
-4
-3
-2
-I
0
I
2
3
4
I
5
6
7
8
9
10
11
12
13
c
u
O'I
CJ, C2, and CJ represent the weights ofcoins from bag 1, 2, and 3 respectively.
Following this logic, it is easy to see that we will need 27 coins from bag 4 and 81 coins
from bag 5. So the answer is to take 1, 3, 9, 27 and 81 coins from bags I, 2, 3, 4, and 5,
respectively, to determine which type of coins each bag contains using a single weighing.
2. 7 Modular Arithmetic
The modulo operation---denoted as x%y or x mod y-finds the remainder of division of
number x by another number y. For simpicility, we only consider the case where y is a
positive integer. For example, 5%3 = 2. An intuitive property of modulo operation is
23
Brain Teasers
that if x1%y = x2 %y, then ( x, -x2 )%y = 0. From this property we can also show that
x%y, (x+I)%y, ···, and (x+ y-1)%y are all different numbers.
Prisoner problem
One hundred prisoners are given the chance to be set free tomorrow. They are all told
that each will be given a red or blue hat to wear. Each prisoner can see everyone else's
hat but not his own. The hat colors are assigned randomly and once the hats are placed
on top of each prisoner's head they cannot communicate with one another in any form, or
else they are immediately executed. The prisoners will be called out in random order and
the prisoner called out will guess the color of his hat. Each prisoner declares the color of
his hat so that everyone else can hear it. If a prisoner guesses correctly the color of his
hat, he is set free immediately; otherwise he is executed.
They are given the night to come up with a strategy among themselves to save as many
prisoners as possible. What is the best strategy they can adopt and how many prisoners
can they guarantee to save? 14
Solution: At least 99 prisoners can be saved.
The key lies in the first prisoner who can see everyone else's hat. He declares his hat to
be red if the number of red hats he sees is odd. Otherwise he declares his hat to be blue.
He will have a 1/2 chance of having guessed correctly. Everyone else is able to deduce
his own hat color combining the knowledge whether the number of red hats is odd
among 99 prisoners (excluding the first) and the color of the other 98 prisoners
(excluding the first and himself). For example, if the number of red hats is odd among
the other 99 prisoners. A prisoner wearing a red hat will see even number of red hats in
the other 98 prisoners (excluding the first and himself) and deduce that he is wearing a
red hat.
The two-color case is easy, isn't it? What if there are 3 possible hat colors: red, blue, and
white? What is the best strategy they can adopt and how many prisoners can they
guarantee to save? 15
Solution: The answer is still that at least 99 prisoners will be saved. The difference is
that the first prisoner now only has 1/3 chance of survival. Let's use the following
scoring system: red=O, green= I, and blue=2. The first prisoner counts the total score for
14 Hint: The first prisoner can see the number of red and blue hats of all other 99 prisoners. One color has
odd number of counts and the other has even number of counts.
15 Hint: That a number is odd simply means x%2 = 1 . Here we have 3 colors, so you may want to consider
x%3 instead.
24
A Practical Guide To Quantitative Finance Interviews
the rest of 99 prisoners and calculates s%3. If the remainder is 0, he announces red; if
the remainder is 1, green; 2, blue. He has 1/3 chance of living, but all the rest of the
prisoners can determine his own score (color) from the remainder. Let's consider a
prisoner i among 99 prisoners (excluding the first prisoner). He can calculate the total
score (x) of all other 98 prisoners. Since (x + 0)%3, ( x + 1) %3, and ( x + 2) %3 are all
different, so from the remainder that the first prisoner gives (for the 99 prisoners
including i), he can determine his own score (color). For example, if prisoner i sees that
there are 32 red, 29 green and 37 blue in those 98 prisoners (excluding the first and
himself). The total score of those 98 prisoners is 103. If the first prisoner announces that
the remainder is 2 (green), then prisoner i knows his own color is green (1) since
only 104%3 = 2 among 103, 104 and 105.
Theoretically, a similar strategy can be extended to any number of colors. Surely that
requires all prisoners to have exceptional memory and calculation capability.
Division by 9
Given an arbitrary integer, come up with a rule to decide whether it is divisible by 9 and
prove it.
Solution: Hopefully you still remember the rules from your high school math class. Add
up all the digits of the integer. If the sum is divisible by 9, then the integer is divisible by
9; otherwise the integer is not divisible by 9. But how do we prove it?
Let's express the original integer as a =a) on + an_, 10n-I + . .. +a, 101 +Go. Basically we
State that if an+ an-I + · · · + a 1 + a 0 = 9x ( X is a integer), then the a is divisible by 9 as
well. The proof is straightforward:
For any a=a)On+an_,10n-i+···+a,IO'+a0 , let b=a-(an+an_,+···+a,+a0 ). We
have b=an(lOn -l)+an_,(lon-i -l)+···+a,(10 1 -l)=a-9x, which is divisible by 9
since all (1 Ok -1), k = 1,- · ·,n are divisible by 9. Because both band 9x are divisible by 9,
a= b + 9x must be divisible by 9 as well.
(Similarly you can also show that a= (-lY an + (-lf- 1an-i + · · · + (-1) 1a,+ a 0
necessary and sufficient condition for a to be divisible by 11.)
= 1 lx
is the
25
Brain Teasers
Chameleon colors
A remote island has three types of chameleons with the following population: 13 red
chameleons, 15 green chameleons and 17 blue chameleons. Each time two chameleons
with different colors meet, they would change their color to the third color. For example,
if a green chameleon meets a red chameleon, they both change their color to blue. Is it
ever possible for all chameleons to become the same color? Why or why not? 16
Solution: It is not possible for all chameleons to become the same color. There are
several approaches to proving this conclusion. Here we discuss two of them.
Approach 1. Since the numbers 13, 15 and 17 are "large" numbers, we can simplify the
problem to 0, 2 and 4 for three colors. (To see this, you need to realize that if
combination (m + 1, n + 1, p + 1) can be converted to the same color, combination
(m,n,p) can be converted to the same color as well.) Can a combination (0,2,4) be
converted to a combination (0,0,6)? The answer is NO, as shown in Figure 2.3:
(0, 2, 4)
(I, 2,
30
Figure 2.3 chameleon color combination transitions from (0, 2, 4)
Actually combination (1, 2, 3) is equivalent to combination (0, 1, 2), which can only be
converted to another (0,1,2) but will never reach (0,0,3).
Approach 2. A different, and more fundamental approach, is to realize that in order for
all the chameleons to become the same color, at certain intermediate stage, two colors
must have the same number. To see this, just imagine the stage before a final stage. It
must has the combination (1,1,x). For chameleons of two different colors to have the
same number, their module of 3 must be the same as well. We start with 15 =3x,
13 = 3y + l, and 17 = 3z + 2 chameleon, when two chameleons of different colors meet,
we will have three possible scenarios:
16
Hint: consider the numbers in module of 3.
26
A Practical Guide To Quantitative Finance Interviews
(3x + 2,3y,3z + 1) = (3x',3y'+ 1,3z'+ 2),
one ymeetsonez
(3x,3y+ l,3z + 2) => { (3(x-1) + 2,3(y + 1),3z + 1) = (3x',3y'+ 1,3z'+ 2), onexmeetsone z
(3(x-1) +2,3y,3(z + 1) + 1) =(3x',3y'+ 1,3z'+ 2), onexmeetsone y
So the pattern is preserved and we will never get two colors to have the same module of
3. In other words, we cannot make two colors have the same number. As a result, the
chameleons cannot become the same color. Essentially, the relative change of any pair of
colors after two chameleons meet is either 0 or 3. In order for all the chameleons to
become one color, at least one pair's difference must be a multiple of 3.
2.8 Math Induction
Induction is one of the most powerful and commonly-used proof techniques in
mathematics, especially discrete mathematics. Many problems that involve integers can
be solved using induction. The general steps for proof by induction are the following:
• State that the proof uses induction and define an appropriate predicate P(n).
• Prove the base case P(l), or any other smallest number n for the predicate to be true.
• Prove that P(n) implies P(n + 1) for every integer n. Alternatively, in a strong
induction argument, you prove that P(l), P(2), ···, and P(n) together imply
P(n+ 1).
In most cases, the real difficulty lies not in the induction step, but to formulate the
problem as an induction problem and come up with the appropriate predicateP(n). The
simplified version of the problem can often help you identify P(n).
Coin split problem
You split 1000 coins into two piles and count the number of coins in each pile. If there
are x coins in pile one and y coins in pile two, you multiple x by y to get xy. Then you
split both piles further, repeat the same counting and multiplication process, and add the
new multiplication results to the original. For example, you split x to x, andx2 , y to y,
andy2 , then the sum is xy+x,x2 + y 1y 2 • The same process is repeated until you only
have piles of 1 stone each. What is the final sum? (The final 1's are not included in the
sum.) Prove that you always get the same answer no matter how the piles are divided.
27
Brain Teasers
Solution: Let n be the number of the coins and f(n) be the final sum. It is unlikely that
a solution will jump to our mind since the number n = 1000 is a large number. If you
aren't sure how to approach the problem, it never hurts to begin with the simplest cases
and try to find a pattern. For this problem, the base case has n = 2. Clearly the only split
is 1+ 1 and the final sum is I. When n = 3, the first split is 2 + 1 and we have xy = 2 and
the 2-coin pile will further give an extra multiplication result 1, so the final sum is 3.
This analysis also gives the hint that when n coins are split into x and n - x coins, the
total sum will be f(n)=x(n-x)+f(x)+f(n-x). 4 coins can be split into 2+2 or
3+1. For either case we can apply x(n-x)+ f(x)+ f(n-x) and yields the same final
sum6.
Claim: For n coins, independent of intermediate splits, the final sum is n(n - I) . 17
2
So how do we prove it? The answer should be clear to you: by strong induction. We
have proved the claim for the base cases n = 2, 3, 4. Assume the claim is true for
n = 2, · · ·, N - I coins, we need to prove that it holds for n = N coins as well. Again we
apply the equation f(n) = x(n-x)+ f(x)+ f(n-x). If N coins are split intox coins and
N - x coins, we have
f ( N) = x( N - x) + f ( x) + f ( N - x)
=x(N-x)+
N(N-1)
2
+
(N -x)(N-x-1)
2
So indeed it holds for n = N as well and f(n)
N(N-1)
=---
2
= n(n-l)
is true for any n
2
the conclusion to n=lOOO, we have /(n)=1000x999/2.
2. Applying
Chocolate bar problem
A chocolate bar has 6 rows and 8 columns (48 small 1x1 squares). You break it into
individual squares by making a number of breaks. Each time, break one rectangle into
two smaller rectangles. For example, in the first step you can break the 6x8 chocolate
bar into a 6 x 3 one and a 6 x 5 one. What is the total number of breaks needed in order
to break the chocolate bar into 48 small squares?
17
.f(2) = l, .f(3)- /(2) = 2 and /(4)-/(3) = 3 should give you enough hint to realize the pattern is
.f(n)=l+2+···+(n-l)=
28
n(n -1)
2
.
A Practical Guide To Quantitative Finance Interviews
Solution: Let m be the number of the rows of the chocolate bar and n be the number of
columns. Since there is nothing special for the case m = 6 and n = 8, we should find a
general solution for all m and n. Let's begin with the base case where m = 1 and n = 1.
The number of breaks needed is clearly 0. For m > 1 and n = 1, the number of breaks is
m -1; similarly for m = 1 and n > 1, the number of breaks is n -1. So for any m and n,
if we break the chocolate into m rows first, which takes m -1 breaks, and then break
each row into n small pieces, which takes m( n - l) breaks, the total number of breaks is
( m -1) + m(n -1) = mn - I. If we breaks it into n columns first and then break each
column into m small pieces, the total number of breaks is also mn -1. But is the total
number of breaks always mn - l for other sequences of breaks? Of course it is. We can
prove it using strong induction.
We have shown the number of breaks is mn -1 for base cases m 1, n = 1 and
m = 1, n;;::: 1. To prove it for a general m x n case, Jet's assume the statement is true for
cases where rows < m, columns n and rows m, columns < n. If the first break is
along a row and it is broken into two smaller pieces m x n1 and m x (n - n1), then the
total number of breaks is l+(mxn1 -l)+(mx(n-n,)-l)=mn-1. Here we use the
m, columns < n. Similarly, if it is broken into two pieces m1 x n and
results for
( m - m1) x n, the total number of breaks is 1 + ( m, x n -1) + ( ( m - m1) x n -1) = mn -1. So
the total number of breaks is always mn -1 in order to break the chocolate bar into
mx n small pieces. For the case m = 6 and n = 8, the number of breaks is 47.
Although induction is the standard approach used to solve this problem, there is actually
a simpler solution if you've noticed an important fact: the number of pieces always
increases by 1 with each break since it always breaks one piece into two. In the
beginning, we have a single piece. In the end, we will have mn pieces. So the number of
breaks must be mn -1.
Race track
Suppose that you are on a one-way circular race track. There are N gas cans randomly
placed on different locations of the track and the total sum of the gas in these cans is
enough for your car to run exactly one circle. Assume that your car has no gas in the gas
tank initially, but you can put your car at any location on the track and you can pick up
choose a starting
the gas cans along the way to fill in your gas tank. Can you
position on the track so that your car can complete the entire circle? 1
18
Hint: Start with N
=
l, 2 and solve the problem using induction.
29
Brain Teasers
Solution: If you get stuck as to how to solve the problem, again start with the simplest
cases (N = 1, 2) and consider using an induction approach. Without loss of generality,
let's assume that the circle has circumference of 1. For N = 1, the problem is trivial. Just
start at where the gas can is. For N = 2, The problem is still simple. Let's use a figure to
visualize the approach. As shown in Figure 2.4A, the amount of gas in can 1 and can 2,
expressed as the distance the car can travel, are x, and x2 respectively, so x, + x 2 = 1.
The corresponding segments are y, and y 2 , so y, + y 2
=1.
Since x, + x 2
=1
and
y, + y 2 = 1, we must have x, y, or x2 y 2 ( x, < y, and x 2 < Yi cannot both be true). If
x, y, , we can start at gas can 1, which has enough gas to reach gas can 2, and get more
gas from gas can 2 to finish the whole circle. Otherwise, we will just start at gas can 2
and pick up gas can 1 along the way to finish the whole circle.
Yt
A
Figure 2.4 Gas can locations on the cycle and segments between gas cans
The argument for N = 2 also gives us the hint for the induction step. Now we want to
show that if the statement holds for N = n, then the same statement also holds for
N
= n + 1.
As
shown
y, + y 2 + · · · + Yn+t = 1 for N
has
X;
m
Figure
2.48,
we
have
x1 + x 2 + ·· · + xn+i
= n + 1. So there must exist at least one i,
Y;· That means whenever the car reaches
X;,
it can reach
x;+i
= 1 and
that
with more gas
(For i = n + 1, it goes to i =I instead). In other words, we can actually "combine" x; and
to one gas can at the position of x; with an amount of gas X; + X;+i (and eliminate
the gas can i +I ). But such combination reduces the N = n +I problem to N = n, for
x;+i
30
A Practical Guide To Quantitative Finance Interviews
which the statement holds. So the statement also holds for N = n + 1. Hence we can
always choose a starting position on the track to complete the entire circle for any N.
There is also an alternative approach to this problem that provides a solution to the
starting point. Let's imagine that you have another car with enough gas to finish the
circle. You put that car at the position of a randomly chosen gas can and drive the car for
a full circle. Whenever you reach a gas can (including at the initial position), you
measure the amount of gas in your gas tank before you add the gas from the can to your
gas tank. After you finish the circle, read through your measurement records and find the
lowest measurement. The gas can position corresponding to the lowest measurement
should be your starting position if the car has no gas initially. (It may take some thinking
to fully understand this argument. I'd recommend that you again draw a figure and give
this argument some careful thoughts if you don't find the reasoning obvious.)
2.9 Proof by Contradiction
In a proof by contradiction or indirect proof, you show that if a proposition were false,
then some logical contradiction or absurdity would follow. Thus, the proposition must be
true.
Irrational number
Can you prove that J2 is an irrational number? A rational number is a number that can
be expressed as a ratio of two integers; otherwise it is irrational.
Solution: This is a classical example of proof by contradiction. If .J2 is not an irrational
number, it can be expressed as a ratio of two integers m and n. If m and n have any
common factor, we can remove it by dividing both m and n by the common factor. So in
the end, we will have a pair of m and n that have no common factors. (It is called
irreducible fraction.) Since m In = .J2, we have m2 = 2n 2 • So m2 must be an even
number and m must be an even number as well. Let's express m as 2x, where xis an
integer, since m is even. Then m2 = 4x 2 and we also have n2 = 2x 2 , which means n
must be even as well. But that both m and n are even contradicts the earlier statement
that m and n have no common factors. So .J2 must be an irrational number.
Rainbow hats
Seven prisoners are given the chance to be set free tomorrow. An executioner will put a
hat on each prisoner's head. Each hat can be one of the seven colors of the rainbow and
the hat colors are assigned completely at the executioner's discretion. Every prisoner can
31
Brain Teasers
see the hat colors of the other six prisoners, but not his own. They cannot communicate
with others in any form, or else they are immediately executed. Then each prisoner
writes down his guess of his own hat color. If at least one prisoner correctly guesses the
color of his hat, they all will be set free immediately; otherwise they will be executed.
They are given the night to come up with a strategy. Is there a strategy that they can
guarantee that they will be set free? 19
Solution: This problem is often perceived to be more difficult than the prisoner problem
in the modular arithmetic section. In the previous prisoner problem, the prisoners can
hear others' guesses. So one prisoner's declaration gives all the necessary information
other prisoners need. In this problem, prisoners won't know what others' guesses are. To
solve the problem, it does require an aha moment. The key to the aha moment is given
by the hint. Once you realize that if we code the colors to 0-6, (
tx,
J%7 must be
among 0, 1, 2, 3, 4, 5 or 6 as well. Then each prisoner i-let's label them as 0-6 as
well-should give a guess g; so that the sum of g; and the rest of 6 prisoners' hat color
codes will give a remainder of i when divided by 7, where g; is a unique number
between 0 and 6. For example, prisoner O's guess should make (g0 + IxkJ%7 = 0.
k;tO
This way, we can guarantee at least one of g;
= X;
for i = 0, I, 2, 3, 4, 5, 6.
We can easily prove this cone Iusion by contradiction. If g, * x,, then (
(since
(g; + L J%7 *
xk
i and g; and
kot1
0, l,2,3,4,5, and 6, then (
i
t,x,
X;
t
are both between 0 and 6). But if K
x, J % 7 * i
* X;
for all
J%7*0,1, 2,3,4,5,6, which is clearly impossible. So
at least one of g; must equal to x; . As a result, using this strategy, they are guaranteed
to be set free.
19
Hint: Let's assign the 7 colors of rainbow with code 0-6 and X; be the color code of prisoner i. Then
(t,x,
32
)%7 must be 0, 1, 2, 3, 4, 5 or 6. How many guesses can 7 prisoners make?
Chapter 3 Calculus and Linear Algebra
Calculus and linear algebra lay the foundation for many advanced math topics used in
quantitative finance. So be prepared to answer some calculus or linear algebra
problems-many of them may be incorporated into more complex problems-in
quantitative interviews. Since most of the tested calculus and linear algebra knowledge
is easy to grasp, the marginal benefit far outweighs the time you spend brushing up your
knowledge on key subjects. If your memory of calculus or linear algebra is a little rusty,
spend some time reviewing your college textbooks!
Needless to say, it is extremely difficult to condense any calculus/linear algebra books
into one chapter. Neither is it my intention to do so. This chapter focuses only on some
of the core concepts of calculus/linear algebra that are frequently occurring in
quantitative interviews. And unless necessary, it does so without covering the proof,
details or even caveats of these concepts. If you are not familiar with any of the concepts,
please refer to your favorite calculus/linear algebra books for details.
3. 1 Limits and Derivatives
Basics of derivatives
Let's begin with some basic definitions and equations used in limits and derivatives.
Although the notations may be different, you can find these materials in any calculus
textbook.
Derivative: Let y
= f(x), then
f '(x) = dy
dJC
= lim L\y =
d.1---tO
lim f(x +
f(x)
d.1---tO
The product rule: If u = u(x) and v = v(x) and their respective derivatives exist,
d(uv)
dv
du
---;;;- = u dx + v dx, (uv)' = u'v+ uv'
.
d
The quotient
rule: dx
.
The cham rule: If y
=
(u) =(
dv)fv 2 ,
v du
dx - u dx
f(u(x)) and u
The generalized power rule: dyn
dx
Some useful equations:
( U) = UV-2 UV
1
1
v
v
= u(x), then -ey = - dx
= nyn-t dy
dx
for "if n
du dx
O
1
Calculus and Linear Algebra
ln( ab) = ln a + ln b
n->oo
lim(l + xl = l + kx for any k
lim sinx = l
x
lim(ln x Ix') = 0 for any r > 0
!!_e"
dx
ex = lim(l +
= e" du
lim x' e-x = 0 for any r
x
)00
da" =(a" lna) du
dx
dx
dx
d
l du u'
-lnu=--=dx
u dx u
d .
d
.
d
2
-sm x = cos x, -cos x = - sm x, -tan x = sec x
dx
dx
dx
What is the derivative of y = ln x'"x ? 1
Solution: This is a good problem to test your knowledge of basic derivative formulasspecifically, the chain rule and the product rule.
Let u=lny=ln(lnx1"x)=lnxxln(lnx). Applying the chain rule and the product rule,
we have
du= d(ln y) = _!_ dy = d(ln x) x ln(lnx) + ln xx d(ln(lnx)) = ln(lnx) +
dx
dx
y dx
dx
dx
x
x ln x
. d(ln(ln x))
.
.
.
To denve
, we agam use the cham rule by settmg v = ln x:
dx
d(ln(ln x)) d(ln v) dv l I
I
----''-------'- =
= - x- =-dx
dv dx v x x ln x ·
:. _!_dy
y dx
x
xlnx
dy =y(ln(lnx)+l)=lnx'"x (ln(lnx)+l).
dx x
x
Maximum and minimum
Derivative f '(x) is essentially the slope of the tangent line to the curve y = f(x) and
the instantaneous rate of change (velocity) of y with respect to x. At point x = c, if
1
Hint: To calculate the derivative of functions with the format y =
f
(x)', it is common to take natural
logs on both sides and then take the derivative, since d(ln y) I dx =I I y x dy I dx.
34
A Practical Guide To Quantitative Finance Interviews
f '(c) > 0, f ( x) is an increasing function at c; if f '(c) < 0, f ( x) is a decreasing
function at c.
Local maximum or minimum: suppose that f(x) is differentiable at c and is defined
on an open interval containing c. If /(c) is either a local maximum value or a local
minimum value of f(x), then J '(c) = 0.
Second Derivative test: Suppose the secondary derivative of f(x), f "(x), is
continuous near c. If f '(c) = 0 and f"(c) > 0, then f(x) has a local minimum at c; if
f '(c) = 0 and /"(c) < 0, then .f(x) has a local maximum at c.
Without calculating the numerical results, can you tell me which number is larger, eli or
::re ?2
Solution: Let's take natural logs of eli and ::re. On the left side we have ::r In e, on the
right side we have e ln ::r. If eli >::re, eli > ::re <=> ::r x In e > ex In ::r <=> ln e > In ::r.
e
Is it true? That depends on whether f(x)
= lnx
x
::r
is an increasing or decreasing function
J '( x) = 1/ x x x2-
In x = 1- ln x ,
2
x
x
which is less than 0 when x > e (lnx > 1 ). In fact, f(x) has global maximum when
In e ln ::r
x = e for all x > 0. So > - - and eli > ::re .
. the denvattve
. . o f /( x ) , we have
tirom e to ::r. T akmg
e
::r
Alternative approach: If you are familiar with the Taylor's series, which we will discuss
ao 1
x x2 x3
in Section 3 .4, you can apply Taylor's series to ex : ex =
= 1+-+- +-+ · · · So
n=O n !
1! 2 ! 3 !
ex> 1+ x, Vx > 0. Let x = ::r I e-1, then eJr!e I e > ::r I e <=> eJr!e > ::r <=> eli >::re.
L-
L'Hospital's rule
Suppose that functions f(x) and g(x) are differentiable at x
Further suppose that lim/(a) = 0
x-+a
x-->a
and limg(a) = 0 or that limf(a)
x-7-a
Hint: Again consider taking natural logs on both sides;
monotonously increasing function.
2
and that limg'(a)-:;:. 0.
x -+a
In a > In b
±oo and
a > b since In x is a
35
Calculus and Linear Algebra
limg(a) ±oo, then lim f(x) = lim f '(x). L'Hospital's rule converts the limit from
x-+a
x-+a g(x) x-+a g '(x)
an indeterminate form to a determinate form.
What is the limit of ex I x 2 as x
oo, and what is the limit of x 2 In x as x
o+?
x
Solution: lim; is a typical example of L'Hospital's rule since lim ex = oo and
X-><0 X
X-+«>
2
limx = oo. Applying L'Hospital's rule, we have
x-+oo
lim f (x) = lim ex = lim f '(x) = lim !t.__.
x-+a g(x) x-+oo x 2 X-+00 g '(x) X-+OO 2x
The result still has the property that lim f (x) = lim ex = oo and lim g(x) = lim 2x = oo, so
x-+OO
X---?00
x-+oo
X-+OO
we can apply the L' Hospital's rule again:
lim f(x) =lime: = lim f '(x) = lim!t.__ = lim d(ex)/ dx =
= oo.
x-+oog(x) x...+oox
x-+oog'(x) x-+oo2x x-+ood(2x)/dx x-+«> 2
At first look, L 'Hospital's rule does not appear to be applicable to lim x 2 In x since it's
x-+0+
not in the format of lim f (x). However, we can rewrite the original limit as lim
x-+a g(x)
x-+0+ xand it becomes obvious that lim x-2 = oo and lim ln x = -oo. So we can now apply
X-+0+
L'Hospital's rule:
l/ x
.
. lnx 1. d(lnx)/ dx 1.
. x2 0
21nx= 11m-=
l1mx
11m-=
1m
=
1m
=
x
x-+o+ x-2 x-+o· d ( x -2 ) I dx x-+o+ -2 I x 3 x-+o+ -2
3. 2 Integration
Basics of integration
Again, let's begin with some basic definitions and equations used in integration.
If we can find a function F(x) with derivative f(x), then we call F(x) an
antiderivative of f (x) .
If f(x)=F'(x), [f(x)= [F'(x)dx=[F(x)t=F(b)-F(a)
36
A Practical Guide To Quantitative Finance Interviews
dFdx(x) -- f(x),
F( a ) =y"=> F( x ) =ya+
r/(
t )dt
Uk+I
The generalized power rule in reverse: Ju* du= - - + c (k
k+l
constant.
'* 1),
where c is any
Integration by substitution:
J/(g(x)) · g '(x)dx = Jf(u)du
with u = g(x), du= g '(x)dx
Substitution in definite integrals: r6 /(g(x)) · g '(x)dx = r f(u)du
l,
Jg(u)
Integration by parts: fudv
= uv- Jvdu
A. What is the integral ofln(x)?
Solution: This is an example of integration by parts. Let u = ln x and v = x, we have
d ( uv) = vdu + udv = (xx 1Ix )dx + ln xdx ,
: . flnxdx = xlnx- fdx = xlnx-x+c, where c is any constant.
B. What is the integral of sec(x) from x = 0 to x =
Jr I
6?
Solution: Clearly this problem is directly related to differentiation/integration of
trigonometric functions. Although there are derivative functions for all basic
trigonometric functions, we only need to remember two of them:
dx
dx
x =cos x,
=-sin x. The rest can be derived using the product rule or the quotient rule. For
example,
dsecx
dx
---=
d(l/ cosx)
sinx
=--=secxtanx
dx
cos 2 x
'
dtanx d(sinx/cosx) cos 2 x+sin 2 x
2
dx =
dx
=
cos 2 x
=sec x.
d(secx+ tanx)
------=sec x(sec x +tan x).
dx
37
Calculus and Linear Algebra
Since the (sec x + tan x) term occurs in the derivative, we also have
din Isecx+tanx I
dx
-------=
J
secx(secx+ tanx)
(secx+ tanx)
=secx
=> sec x =In Isec x + tan x I+c
and
r'
6
secx = ln(sec(JZ" I 6) + tan(JZ" I 6))-ln(sec(O) + tan(O)) =In(J3)
Applications of integration
A. Suppose that two cylinders each with radius 1 intersect at right angles and their
centers also intersect. What is the volume of the intersection?
Solution: This problem is an application of integration to volume calculation. For these
applied problems, the most difficult part is to correctly formulate the integration. The
general integration function to calculate 3D volume is V
=
r
A(z)dz where A(z) is the
I
cross-sectional area of the solid cut by a plane perpendicular to the z-axis at coordinate z.
The key here is to find the right expression for cross-sectional area A as a function of z.
Figure 3.1 gives us a clue. If you cut the intersection by a horizontal plane, the cut will
be a square with side-length
-(2z}2. Taking advantage of symmetry, we can
calculate the total volume as
An alternative approach requires even better 3D imagination. Let's imagine a sphere that
is inscribed inside both cylinders, so it is inscribed inside the intersection as well. The
sphere should have a radius of r I 2. At each cut perpendicular to the z-axis, the circle
from the sphere is inscribed in the square from the intersection as well. So
Acircte =
A.,quure· Since it's true for all z values, we have
V,phere
38
= -j- JZ"( 5")3 =
V;ntersection
=>
V;ntersection
= 16 / 3r 3 = 16 / 3·
A Practical Guide To Quantitative Finance Interviews
Figure 3.1 Interaction of two cylinders
B. The snow began to fall some time before noon at a constant rate. The city of
Cambridge sent out a snow plow at noon to clear Massachusetts Avenue from MIT to
Harvard. The plow removed snow at a constant volume per minute. At 1 pm, it had
moved 2 miles and at 2 pm, 3 miles. When did the snow begin to fall?
Solution: Let's denote noon as time 0 and assume snow began to fall T hours before
noon. The speed at which the plow moves is inversely related to the vertical crosssectional area of the snow: v = c1 I A(t), where vis the speed of the plow, c, is a constant
representing the volume of snow that the plow can remove every hour and A(t) is the
cross-sectional area of the snow. If t is defined as the time after noon, we also have
A(t) = c2 (t + T), where c2 is the rate of cross-sectional area increase per hour (since the
snow falls at a constant rate). So v =
c
c,
=
where
c2 (t+T) t+T
c
c=-'
Taking the
Cz
integration, we have
1c
--dt=cln(l+T)-clnT =cln (l+T)
- - = 2,
T+t
T
c
(2+T)
- - =3
1--dt=cln(2+T)-clnT=cln
T+t
T
i2
From these two equations, we get
39
Calculus and Linear Algebra
Overall, this question, although fairly straightforward, tests analytical skills, integration
knowledge and algebra knowledge.
Expected value using integration
Integration is used extensively to calculate the unconditional or conditional expected
value of continuous random variables. In Chapter 4, we will demonstrate its value in
probability and statistics. Here we just use one example to show its application:
If Xis a standard normal random variable, X - N(O, 1), what is E[X IX> O]?
Solution: SinceX - N(O, 1), the probability density function of x is f(x) =
and we have E[X Ix> O] =
r
rb
xf(x)dx =
x
Because d(-112x 2 )=-x and
Jk e-
112 x 2
e- 112 x 2 dx.
where c is an arbitrary constant, it is
obvious that we can use integration by substitution by letting u = -1I2x 2 • Replace
e- 112 x 2 with eu and xdx with -du, we have
r
2
I
x J2ii
e-1 12x dx -_
r
I
- .ffii
e udu -_ - &1
[
_
eu ]-- u = 0 and x = oo => u = -oo.
:. E[XIX>0]=1/J2;
3.3 Partial Derivatives and Multiple Integrals
Partial derivative: w=f(x, y)=> 88if (x0 ,y0 )= lim
x
&->0
=ix
a
.
• .
a2f
af
0 2J
a af
a Bf
Second order partial derivatives: = -(-), - - = - ( - ) = - ( - )
2
8x
ax ax
axay
ax By
By ax
The general chain rule: Suppose that w= f(xl'x 2 ,-··,x",) and that each of variables
xi' x 2 ,
• • ·,
x"' is a function of the variables ti' t 2 ,
• • ·,
tn. If all these functions have
·
fi1rst-order part1a
· I denvatives,
. .
Bx+
ax2· · · +aw
1
contmuous
then -aw =aw
--aw
-+
- -ax
" ' ior
Bt;
each i, I -5:: i
40
-5::
n.
ax, Bt;
Bx2 Bt;
Bx"' at;
A Practical Guide To Quantitative Finance Interviews
Changing Cartesian integrals into polar integrals: The variables in two-dimension
plane can be mapped into polar coordinates: x = r cos B, y = r sin B. Tthe integration in a
continuous polar region R is converted to
JfJ(x,y)dxdy
= JfJ(rcosB,rsinB)rdrdB.
R
Calculate
R
r
e-x212 dx.
Solution: Hopefully you happen to remember that the probability density function (pdf)
of the standard normal distribution is f(x)
=
v2Jr
e-x212 . By definition, we have
If you've forgotten the pdf of the standard normal distribution or if you are specifically
asked to prove (
v2tr
e-x 212dx =I, you will need to use polar integrals to solve the
problem:
( e-x212dx ( e-y212dy = (
=
( e-(x2+/i12dxdy =
r 1"
r 1"
e-r212rdrdB= -
= -[ e-r212
I[
r
e-(O)
n! xn +···
Taylor's series are often used to represent functions in power series terms. For example,
Taylor's series for three common transcendental functions, ex, sin x and cos x , at
x0 = 0 are
x
"' 1
x x2 x3
e =L-=1+-+-+-+···
n=O n !
1! 2 ! 3 !
'
•
smx=
x3 xs x1
L (-IYx2n+1
=x--+---+···
(2n+l)!
3! 5! 7!
'
oo
n=O
x2n
x2 x4 x6
L (-If
=1--+---+···
(2n)!
2! 4! 6!
oo
COSX=
n=O
The Taylor's series can also be expressed as the sum of the nth-degree Taylor
f "(x)
fn>(x)
0 (x-x ) 2 +· ·· +
0 (x-x f
polynomial T,,(x) = f(x 0 )+ J'(x0 )(x-x0 )+
and
0
0
2!
n!
a remainder Rn(x): f(x) = Tn(x) + Rn(x).
For some i between
IJC)
x
(n+l)!
IX-Xo r+I. Let Mbe the maximum of
i between x0 and x, we get constraint !Rn (x)I:::;;
Mx I X-X
0
ln+I
(n + l)!
A. What is i;?
Solution: The solution to this problem uses Euler's formula, e;o =cos e + i sine, which
can be proven using Taylor's series. Let's look at the proof. Applying Taylor's series to
e;e, cos e and sine, we have
ie
I!
iB
(ie) 2
2!
(ie) 3
3!
(ie) 4
4!
.e
I!
e2
2!
. e 3 e 4 . es
3! 4!
5!
e =1+-+--+--+--+···=1+1----1-+-+1-+···
e2 e4 e6
cos = 1- - + - - - + · · ·
2! 4! 6!
e
.
e 3 es e 1
sm e = e- - + - - - + · · ·
3! 5! 7!
42
. .
. e . e 3 • es . e 1
l sm e = l- -1-+ l - -1- + · · ·
I!
3!
5!
7!
A Practical Guide To Quantitative Finance Interviews
Combining these three series, it is apparent that
When 8 = :1r, the equation becomes
equation becomes
Hence,
eiJC 12
=
=cos e+ i sine.
cos :1r + i sin :1r = -1. When 8 = :1r I 2, the
=cos( Jr I 2) + i sin (Jr I 2) = i. 3 So In i
tn(i) = i ln i = i(i:1r I 2) =-Jr I 2
B. Prove (l + xf
eiJC
eiB
= In ( eiJCl 2 ) = i:1r I 2.
i; = e-JC 12 •
1+ nx for all x > -1 and for all integers n
2.
Solution: Let f (x) = (l + xf. It is clear that 1+ nx is the first two terms in the Taylor's
series of f(x) with x 0 = 0. So we can consider solving this problem using Taylor's
sen es.
For Xo = 0 we have (l +
xr = 1 for 'efn
2. The first and secondary derivatives off(x)
are f '(x) = n(l + xf-' and f "(x) = n(n-1)(1 + xy- 2 • Applying Taylor's series, we have
f(x)
= f(x 0 )+ f'(x 0 )(x-x0 )+ /"Ci) (x-x0 )2 = f(O)+ f'(O)x+
2!
= 1+ nx+ n(n-1)(1 +xr-
2
where x :::;; i :::;; 0 if x < 0 and x
Since x >-1 andn
x
f"(i) x 2
2!
i
0 if x > 0 .
2, we have n > 0, (n-1) > 0, (l+iy-2 > 0, x 2
Hence, n(n-l)(l+xr- 2 x 2
'
2
0.
and f(x)=(l+xY >l+nx.
If Taylor's series does not jump to your mind, the condition that n is an integer may give
you the hint that you can try the induction method. We can rephrase the problem as: for
every integer n 2, prove (1 + xY :2:: 1+ nx for x > -1 .
The base case: show (l + x y
since (I
1+ nx, Vx > -1 when n = 2, which can be easily proven
2x, Vx >-1.
The induction step: show that if (1 + xY :2:: 1+ nx, 'efx > -1 when n = k, the same
statement holds for
straightforward as well.
3
n = k + 1:
(1 + x)k+i :2:: 1+ (k + l)x, Vx > -1.
This
step
is
Clearly they satisfy equation ( e'"' )' = i' = e" = -1.
43
Calculus and Linear Algebra
(1 + x)k+i =(I+ x)k (1 + x)
(1 +loc)(l + x) = 1+(k + l)x+loc2 , 'fifx > -1
So the statement holds for all integers n 2 when x > -1.
Newton's method
Newton's method, also known as the Newton-Raphson method or the Newton-Fourier
method, is an iterative process for solving the equation f(x) = 0. It begins with an initial
value x0 and applies the iterative step xn+i
= xn - f(xn) to solve f(x) = 0 if Xpx2 ,. ..
f'(xJ
converge.4
Convergence of Newton's method is not guaranteed, especially when the starting point
is far away from the correct solution. For Newton's method to converge, it is often
necessary that the initial point is sufficiently close to the root; f(x) must be
differentiable around the root. When it does converge, the convergence rate is quadratic,
which means lxn+i -x11 ::;; c5<1, where x1 is the solution to f(x) = 0.
(xn -xi)
A. Solve x 2
= 37 to the third digit.
Solution: Let f(x) = x 2 -37, the original problem is equivalent to solving f(x)
x 0 = 6 is a natural initial guess. Applying Newton's method, we have
x1
=
x0
_
f(x0 )
f'(x 0 )
=
x0
_
-37
2x0
=
6 _ 36-37
2x6
=
=0.
6 •083 •
( 6.083 2 = 37.00289, which is very close to 37.)
If you do not remember Newton's method, you can directly apply Taylor's series for
function f(x) = J-; with f'(x) = +x- 112 :
/(37)
4
/(36)+ f '(36)(37-36) = 6+ I 112 = 6.083.
The iteration equation comes from the first-order Taylor's series:
f (x.
)
(( x ) z !( x ) + f '( x)(x -x ) = 0 =:>x =n x -f'(x)
•
44
1111
II
•
II
n•l
H
11•1
A Practical Guide To Quantitative Finance Interviews
Alternatively, we can use algebra since it is obvious that the solution should be slightly
higher than 6. We have (6 + y) 2 =
If we ignore the y 2 term,
which is small, then y = 0.083 and x = 6 + y = 6.083.
B. Could you explain some root-finding algorithms to solve f(x) = 0? Assume f(x) is
a differentiable function.
Solution: Besides Newton's method, the bisection method and the secant method are two
alternative methods for root-finding. 5
Bisection method is an intuitive root-finding algorithm. It starts with two initial values
a 0 and b0 such that f(a 0 ) < 0 and f(b0 ) > 0. Since f(x) is differentiable, there must be
an x between a0 and b0 that makes f(x) = 0. At each step, we check the sign of
f((an+bn)/2). If f((an+bn)/2)<0, we set bn+I =bn and an+I =(an+bn)/2; If
f((an+bJ!2)>0, we set an+I =an and bn+I =(an+bJ/2; If J((an+bJ!2)=0, or its
absolute value is within allowable error, the iteration stops and x = (an + bn) I 2. The
x -x
bisection method converges linearly, n+i
f <5<1, which means it is slower than
xn-xf
Newton's method. But once you find an a0 / b0 pair, convergence is guaranteed.
Secant method starts with two initial values x0 , x1 and applies the iterative step
xn+I
= xn -
xn -xn-I
f(xJ. It replaces the f '(xJ in Newton's method with a
f(xJ- f(xn-1)
linear approximation f ( xn) - f (xn-i) . Compared with Newton's method, it does not
xn -xn-1
require the calculation of derivative f'(xJ, which makes it valuable if f '(x) is difficult
to calculate. Its convergence rate is ( 1+JS)I2, which makes it faster than the bisection
method but slower than Newton's method. Similar to Newton's method, convergence is
not guaranteed if initial values are not close to the root.
Lagrange multipliers
The method of Lagrange multipliers is a common technique used to find local
maximums/minimums of a multivariate function with one or more constraints. 6
5 Newton's method is also used in optimization-including multi-dimensional optimization problems-to
find local minimums or maximums.
45
Calculus and Linear Algebra
Let f(x,, x2 ,
···,
xJ be a function of n variables x=(x" x 2 ,
vector Vf(x) = (
,
, ···, !. ).
···,
xJ with gradient
The necessary condition for maximizing or
minimizing f(x) subject to a set of k constraints
is that'\lf(x)+A,Vg,(x)+A..iVg2 (x)+···+A.k'\lgk(x)=O, where A,,···,A.k are called the
Lagrange multipliers.
What is the distance from the origin to the plane 2x + 3y + 4z = 12 ?
Solution: The distance (D) from the origin to a plane is the minimum distance between
the origin and points on the plane. Mathematically, the problem can be expressed as
min D 2 =J(x,y,z)=x2 +y2+z2
s.t. g(x,y,z)=2x+3y+4z-12=0
Applying the Lagrange multipliers, we have
= 2x+ 2A. = 0
ar +.IL aray = 2y+ 3.IL = 0
ay
araz +.IL OJaz = 2x + 4.IL = 0
qf
ax
+A. OJ
ax
2x+3y+4z-12 = 0
.IL=-24/29
x = 24 I 29 __..._ D
___,,
y = 36129
z = 48/29
=
48
(1i)2
29
+ (.J£)2
29
+ ( 29 )2 =
v29
In general, for a plane with equation ax+ by+ cz =d, the distance to the origin is
D=
ldl
.Ja2 +b2 +c2
3. 5 Ordinary Differential Equations
In this section, we cover four typical differential equation patterns that are commonly
seen in interviews.
The method of Lagrange multipliers is a special case of Karush-Kuhn-Tucker (KKT) conditions, which
reveals the necessary conditions for the solutions to constrained nonlinear optimization problems.
6
46
A Practical Guide To Quantitative Finance Interviews
Separable differential equations
A separable differential equation has the form dy = g(x)h(y). Since it is separable, we
dx
can express the original equation as dy = g(x)dx. Integrating both sides, we have the
h(y)
solution f dy = f g(x)dx.
h(y)
A. Solve ordinary differential equation y '+ 6xy = 0, y(O) = l
Solution: Let g(x) = -6x and h(y) = y, we have dy = -6xdx. Integrate both sides of
y
the equation: fdy = f-6xdx
y
lny=-3x 2 +c
y=e-3x 2 +", where c is a constant.
Plugging in the initial condition y(O) = 1, we have c = O and y = e-3x 2 •
B. Solve ordinary differential equation y' = x - Y .7
x+y
Solution: Unlike the last example, this equation is not separable in its current form. But
we can use a change of variable to tum it into a separable differential equation. Let
z = x + y, then the original differential equation is converted to
d ( z - x) = x - ( z - x)
dz - l = 2 x - l
zdz = 2xdx
dx
z
dx
z
(x+ y) 2 = z 2 = 2x2 +c
y 2 + 2xy-x2 = c
fzdz = f2xdx + c
First-order linear differential equations
A first-order differential linear equation has the form dy + P(x)y = Q(x). The standard
dx
approach to solving a first-order differential equation is to identify a suitable function
l(x), called an integrating factor, such that l(x)(y'+ P(x)y) = l(x)y'+ l(x)P(x)y
7
Hint: Introduce variable z = x + y.
47
Calculus and Linear Algebra
=(I(x)y)'; Then we have (I(x)y)' = l(x)Q(x) and we can integrate both sides to solve
J
for y: I(x)y = I(x)Q(x)dx => y
The integrating factor,
=
Jl(x)Q(x)dx
I(x)
.
l(x), must satisfy dl(x)
dx
= I(x)P(x),
which means l(x) is a
separable differential equation with general solution J(x) = efP 0.
Solution: This is a typical example of first-order linear equations with P(x)
I
Q(x) = - 2 • So l(x) = e
x
JP(x)dx
=e
Jo /x)dx = e"x = x
1
= _!_
x
and
1
and we have l(x)Q(x) =-.
x
:. l(x)(y'+ P(x)y) = ( xy )' = I(x)Q(x) =I Ix
J
Taking integration on both sides, xy = (l/ x)dx = lnx+c => y = lnx+c.
x
Plugging in y(l) =I, we get c =I and y = ln x + 1 .
x
Homogeneous linear equations
A homogenous linear equation is a second-order differential equation with the form
d2
d
dx
dx
c(x)y = 0.
It is easy to show that, if y, and y 2 are linearly independent solutions to the
homogeneous linear equation, then any y(x)=c,y,(x)+c2 y 2 (x), where c1 and c2 are
arbitrary constants, is a solution to the homogeneous linear equation as well.
When a, b and c (a -:t:. 0 ) are constants instead of functions of x,
linear equation has closed form solutions:
the homogenous
Let 'i and r2 be the roots of the characteristic equation ar 2 + br + c = O,9
8
The constant c is not needed in this case since it just scales both sides of the equation by a factor.
48
A Practical Guide To Quantitative Finance Interviews
2. If 1j and r2 are real and 1j
3. If
y
= r2 =r,
then the general solution is y
=c 1erx + c2xerx;
and r2 are complex numbers a± ip, then the general solution is
1j
= eax (c1 COS fJx + C2 sin fJx).
It is easy to verify that the general solutions indeed satisfy the homogeneous linear
solutions by taking the first and secondary derivatives of the general solutions.
What is the solution of ordinary differential equation y "+ y '+ y = 0?
Solution: In this specific case, we have a= b = c = 1 and b 2 -4ac = -3 < 0 , so we have
complex roots r = -1I2 ± f3 I 2i (a = -1I2, f3
differential equation is therefore
= f3 I 2 ), and the general solution to the
y = eax(c1 cospx+c2 sinpx) = e- 112 x {c1 cos(.J3I2x)+c2 sin(.J3 / 2x) ).
Nonhomogeneous linear equations
2
Unlike a homogenous linear equation a d -;' + b dy + cy = 0, a nonhomogeneous linear
dx
dx
equation a d 3: + b dy + cy = d(x) has no closed-form solution. But if we can find a
dx
dx
d2
dy
particular solution yP(x) for
+b dx +cy=d(x), then y=yP(x)+ y/x), where
2
a;,
Yi:(x) is the general solution of the homogeneous equation a d 3: +b dy +cy =0,
dx
dx
2
IS
a
general solution of the nonhomogeneous equation a d 3: + b dy + cy = d(x).
dx
dx
2
9
A quadrat1c
. equation
.
.
b y quadrat1c
.
ar ' + br + c = 0 has roots given
c
Ia r = -b +
- .Jb' - 4ac . You
2a
should either commit the formula to memory or be able to derive it using (r + b I 2a)' = (b' - 4ac) I 4a'.
49
Calculus and Linear Algebra
Although it may be difficult to identify a particular solution y P ( x) in general, in the
special case when d(x) is a simple polynomial, the particular solution is often a
polynomial of the same degree.
What is the solution of ODEs y "+ y '+ y
=
l and y "+ y '+ y
=
x?
Solution: In these ODEs, we again have a= b = c = 1 and b2 -4ac =-3 < 0, so we have
complex solutions r = -1I2 ± ..f3 I 2i (a = -1 I 2,
y = e- 112 x { c1 cos(
p = ..f3 I 2) and the general solution is
.J3 I 2x) +c sin(.J3 I 2x) ).
2
What is a particular solution for y "+ y '+ y
y"+ y'+ y=1 is
=
1? Clearly y
=
l is. So the solution to
To find a particular solution for y "+ y '+ y = x, Let y P (x) = mx + n, then we have
So the particular solution is x-1 and
the solution to y "+ y '+ y = x is
y = y P(x) + yg(x) = e- 112 x { c, cos(
.J3 I 2x) + c
2
sin(
.J3 I 2x)) +(x-1).
3. 6 Linear Algebra
Linear algebra is extensively used in applied quantitative finance because of its role in
statistics, optimization, Monte Carlo simulation, signal processing, etc. Not surprisingly,
it is also a comprehensive mathematical field that covers many topics. In this section, we
discuss several topics that have significant applications in statistics and numerical
methods.
Vectors
An n x l (column) vector is a one-dimensional array. It can represent the coordinates of
a point in the Rn (n-dimensional) Euclidean space.
50
A Practical Guide To Quantitative Finance Interviews
Inner product/dot product: the inner product (or dot product) of two Rn vectors x and
n
y is defined as
L
X;Y;
= xr y
i=I
Euclidean norm:
llxll
=ffx;
llx-yll
XTy
Then angle B between Rn vectors x and y has the property that cos B = llxll llYll . x and Y
are orthogonal if xr y = 0. The correlation coefficient of two random variables can be
viewed as the cosine of the angle between them in Euclidean space ( p =cos(}).
There are 3 random variables x, y and z. The correlation between x and y is 0.8 and the
correlation between x and z is 0.8. What is the maximum and minimum correlation
between y and z?
Solution: We can consider random variables x, y and z as vectors. Let B be the angle
between x and y, then we have cos B = Px,y = 0.8. Similarly the angle between x and z is
B as well. For y and z to have the maximum correlation, the angle between them needs
to be the smallest. In this case, the minimum angle is 0 (when vector y and z are in the
same direction) and the correlation is 1. For the minimum correlation, we want the
maximum angle between y and z, which is the case shown in Figure 3.2.
If you still remember some trigonometry,
all you need is that
cos(2B) = (cos8) 2 -(sin 8) 2
= 0.8 2 -0.62 = 0.28
Otherwise, you can solve the problem using
Pythagoras's Theorem:
'
0.8
0.8x1.2 =
y
z cos2B= .J1 2 -0.962 = 0.28
0.6
0.6
Figure 3.2 Minimum correlation and maximum angle between vectors y and z
51
Calculus and Linear Algebra
QR decomposition
QR decomposition: For each non-singular n x n matrix A, there is a unique pair of
orthogonal matrix Q and upper-triangular matrix R with positive diagonal elements such
that A =QR. io
QR decomposition is often used to solve linear systems Ax= b when A is a non-singular
matrix. Since Q is an orthogonal matrix, Q-' = QT and QRx = b
Rx= QT b. Because R
is an upper-triangular matrix, we can begin with xn (the equation is simply
Rn,nxn =(QTb)n ), and recursively calculate all
X;,
'Iii= n, n -1, ···,I.
If the programming language you are using does not have a function for the linear least
squares regression, how would you design an algorithm to do so?
Solution: The linear least squares regression is probably the most widely used statistical
analysis method. Let's go over a standard approach to solving linear least squares
regressions using matrices. A simple linear regression with n observations can be
expressed as
Y;
= /30x;,o + /31x;, 1 + · ·· + /3P_1x;,p-I + si'
term and
x;,t • • • ·, xi,p - I
'Iii= 1,···,n, where X;o
=1, 'Iii,
1s the intercept
are p -1 exogenous regressors.
The goal of the linear least squares regression is to find a set of f3 = [/30 , /3" ···,Pp-if
n
that makes
the smallest. Let's express the linear regression in matrix format:
i=I
Y=X/3+&, where Y=[Yi,Yz,-··,f;,f and &=[s"&2 ,-··,&nf are both nxl column
vectors; X is a n x p matrix with each column representing a regressor (including the
intercept) and each row representing an observation. Then the problem becomes
min f
/)
10
n
(/3)
=min
/)
L s;
2
i=l
= min(Y -X /3)r (Y - X /3)
/)
A nonsingular matrix Q is called an orthogonal matrix if Q ' = Q'. Q is orthogonal if and only if the
columns (and rows) of Q form an orthonormal set of vectors in R". The Gram-Schmidt
orthonormalization process (often improved to increase numerical stability) is often used for QR
decomposition. Please refer to a linear algebra textbook if you are interested in the Gram-Schmidt process.
52
A Practical Guide To Quantitative Finance Interviews
To minimize the function /(/3), taking the first derivative' 1 of f(/3) with respect to
we have /'(/3) = 2Xr (Y -X fl)= 0
/3,
(Xr X)/J = xrY, where (Xr X) is a p x p
symmetric matrix and xry is a p x I column vector.
Let A = (Xr X) and b = xrY, then the problem becomes Afl = b, which can be solved
using QR decomposition as we described.
Alternatively, if the programming language has a function for matrix inverse, we can
directly calculate fl as fl= (XT xr' XTY. 12
Since we are discussing linear regressions, it's worthwhile to point out the assumptions
behind the linear least squares regression (a common statistics question at interviews):
1. The relationship between Y and Xis linear: Y = X f3 + &.
2. E[&;]=O, Vi=l,-··,n.
3.
var(&;)=a 2 , i=l,-··,n (constant variance), andE[&;&J]=O,i;tj (uncorrelated
errors).
4. No perfect multicollinearity: p(x;,x) ;t ±1, i ;t j
correlation of regressors
5.
&
and
X;
X;
where p(xi'x1 )
is the
and xr
are independent.
Surely in practice, some of these assumptions are violated and the simple linear least
squares regression is no longer the best linear unbiased estimator (BLUE). Many
econometrics books dedicate significant chapters to addressing the effects of assumption
violations and corresponding remedies.
Determinant, eigenvalue and eigenvector
Determinant: Let A be an n x n matrix with elements {A;), where i, j = 1, · · ·, n. The
determinant of A is defined as a scalar: det(A) = Lfll(p)a1.p1a2.p2 ···an.p., where
p
p =(pl' p 2 ,
• ··,
Pn) is any permutation of (I, 2, · · ·, n); the sum is taken over all n!
possible permutations; and
11
To do that, you do need a little knowledge about matrix derivatives. Some of the important derivative
equations
for
vectors/matrices
are
oa' x = ox' a =a,
ox
ox
oAx
a;= A,
ox' Ax
T
----;--- = (A + A)x,
0 2 x' Ax
axax'
= 2A,
o(Ax + b)' C(Dx +e) =A' C(Dx +e) + D' C' (Ax+ b),
ax
12
The matrix inverse introduces large numerical error if the matrix is close to singular or badly scaled.
53
Calculus and Linear Algebra
lf/(p)= {
I, if p can be coverted to natural order by even number of exchanges
.
.
-1, 1f p can be coverted to natural order by odd number of exchanges
For example, determinants of 2 x 2 and 3 x 3 matrices can be calculated as
de{[: !])
=ad - be,
de{[: : rn
=
aei + bfg + cdh -ceg- ajh- bdi. 13
I
Determinant properties: det(Ar) = det(A), det(AB) = det(A)det(B), det(A- 1 ) = - det(A)
Eigenvalue: Let A be an n x n matrix. A real number A. is called an eigenvalue of A if
there exists a nonzero vector x in Rn such that Ax = Ax. Every nonzero vector x
satisfying this equation is called an eigenvector of A associated with the eigenvalue A..
Eigenvalues and eigenvectors are crucial concepts in a variety of subjects such as
ordinary differential equations, Markov chains, principal component analysis (PCA), etc.
The importance of determinant lies in its relationship to eigenvalues/eigenvectors. 14
The determinant of matrix A - Al, where I is an n x n identity matrix with ones on the
main diagonal and zeros elsewhere, is called the characteristic polynomial of A. The
equation det(A - Al)= 0 is called the characteristic equation of A. The eigenvalues of
A are the real roots of the characteristic equation of A. Using the characteristic equation,
we can also show that A,A.i ···An
= det(A)
n
n
and LA;= trace( A)= IA;,;·
A is diagonalizable if and only if it has linearly independent eigenvectors. 15 Let
A,, Ai, ···, An be the eigenvalues of A, x1, x2 ,
and X = [x1 Ix2 I··· I xn], then
• • ·,
xn be the corresponding eigenvectors.
X- 1AX=
In practice, determinant is usually not solved by the sum of all permutations because it is
computationally inefficient. LU decomposition and cofactors are often used to calculate determinants
instead.
14 Determinant can also be applied to matrix inverse and linear equations as well.
15 If all n eigenvalues are real and distinct, then the eigenvectors are independent and A is diagonalizable.
13
54
A Practical Guide To Quantitative Finance Interviews
If matrix A = [
what are the eigenvalues and eigenvectors of A?
Solution: This is a simple example of eigenvalues and eigenvectors. It can be solved
using three related approaches:
Approach A: Apply the definition of eigenvalues and eigenvectors directly.
Let.! be an eigenvalue and x = [ ::] be its corresponding eigenvector. By definition, we
have
So either A. = 3, in which case x1 = x2 (plug A. =3 into equation 2x1 + x2
.
.
.
normal.1zed eigenvector
1s
correspondmg
normalized eigenvector is [ 11
-1/v2
2x1 +x2
[I/ .Ji]
.J2 ,
= A.x1 )
and the
I/
. w h.1ch case the
or x 1 + x 2 = 0, m
and
A.= 1 (plug x2 = -x1 into equation
= A.x1 ).
Approach B: Use equationdet(A-A./) = 0.
det( A - A.I) = 0 :::::::> (2 - A. )(2 - A.) -1 = 0. Solving the equation, we have
A, = 1 and
Ai = 3.
Applying the eigenvalues to Ax= A.x, we can get the corresponding
eigenvectors.
n
Approach C: Use equations
n
A, ·Ai··· An = det(A) and LA; =trace( A)= LA;,;·
i=I
det(A) = 2x 2-1x1=3 and trace(A) = 2x2 = 4.
So we have A,xAi=3} :::::::> {A,=l • Again apply the eigenvalues to Ax = A.x, and we
A,+Ai=4
Ai=3
can get the corresponding eigenvectors.
55
Calculus and Linear Algebra
Positive semidefinite/definite matrix
When A is a symmetric n x n matrix, as in the cases of covariance and correlation
matrices, all the eigenvalues of A are real numbers. Furthermore, all eigenvectors that
belong to distinct eigenvalues of A are orthogonal.
Each of the following conditions is a necessary and sufficient condition to make a
symmetric matrix A positive semidefinite:
1. xr Ax 0 for any n x 1 vector x .
2. All eigenvalues of A are nonnegative.
3. All the upper left (or lower right) submatrices AK, K = 1, · · ·, n have nonnegative
determinants. 16
Covariance/correlation matrices must also be positive semidefinite. If there is no perfect
linear dependence among random variables, the covariance/correlation matrix must also
be positive definite. Each of the following conditions is a necessary and sufficient
condition to make a symmetric matrix A positive definite:
1. xr Ax > 0 for any nonzero n x 1 vector x .
2. All eigenvalues of A are positive.
3. All the upper left (or lower right) submatrices AK, K
determinants.
= 1, · ·., n
have positive
There are 3 random variables x, y and z. The correlation between x and y is 0.8 and the
correlation between x and z is 0.8. What is the maximum and minimum correlation
between y and z?
Solution: The problem can be solved using the positive semidefiniteness property of the
correlation matrix.
Let the correlation between y and z be p , then the correlation matrix for x, y and z is
0.8
p
0.8
p
0.81
p .
1
16 A necessary, but not sufficient, condition for matrix A to be positive semidifinite is that A has no
negative diagonal elements.
56
A Practical Guide To Quantitative Finance Interviews
0;8J)+o.8xdei([0;8
det(P)
= (1- p 2)-0.8 x (0.8-0.8p) + 0.8x (0.8p-0.8) = -0.28 + l .28p- p 2 ;;::: 0
So the maximum correlation between y and z is 1, the minimum is 0.28.
LU decomposition and Cholesky decomposition
Let A be a nonsingular n x n matrix. LU decomposition expresses A as the product of a
lower and upper triangular matrix: A= LU. 17
LU decomposition can be use to solve Ax= b and calculate the determinant of A:
n
LUx = b
n
Ux = y, Ly= b; det(A) = det(L)det(U) =II L;,;f] uj,j"
i=I
j=I
When A is a symmetric positive definite matrix, Cholesky decomposition expresses A
as A = RT R, where R is a unique upper-triangular matrix with positive diagonal entries.
Essentially, it is a LU decomposition with the property L =UT.
Cholesky decomposition is useful in Monte Carlo simulation to generate correlated
random variables as shown in the following problem:
How do you generate two N(O,l) (standard normal distribution) random variables with
correlation p if you have a random number generator for standard normal distribution?
Solution: Two _N(O,l) random variables xP x 2 with a correlation p can be generated
from independent N(O, 1) random variables z1 , z2 using the following equations:
X1
=Z1
X2 = PZ1 +
p2 Z2
It is easy to confirm that var( x1) = var( z1) = 1, var( x2 )
and cov(xpx2) = cov(z1,pz1+
= p 2 var( z1) + (1- p 2) var( z2) = 1,
p 2z2) = cov(z1,pz1) = p.
This approach is a basic example using Cholesky decomposition to generate correlated
random numbers. To generate correlated random variables that follow an-dimensional
17
LU decomposition occurs naturally in Gaussian elimination.
57
Calculus and Linear Algebra
multivariate
normal
= [µ,, µ 2 , • • ·, µn f
µ
distribution
X
= [X" X 2 ,-··,XJT
N(µ, I)
with
mean
and covariance matrix I (a n x n positive definite matrix) 18, we can
decompose the covariance matrix I into RT R and generate n independent N(O, 1)
random variables z" z2 ,
asX=µ+RTZ.
···,
zn. Let vector Z = [z,, z2 ,-··,znf, then X can be generated
19
Alternatively, X can also be generated using another important matrix decomposition
called singular value decomposition (SVD): For any n x p matrix X, there exists a
factorization of the form X = UD VT, where U and V are n x p and p x p orthogonal
matrices, with columns of U spanning the column space of X, and the columns of V
spanning the row space; D is a p x p diagonal matrix called the singular values of X.
For a positive definite covariance matrix, we have V = U and L = UDUT. Furthermore,
D is the diagonal matrix of eigenvalues A,, Ai, ···, A,n and U is the matrix of n
corresponding eigenvectors. Let D 112 be a diagonal matrix with diagonal elements
then it 1s clear that D = (D 112 ) 2 = (D 112 )(D112 f
and
Ji;, JI;., ···, Ji:,
I= UD 112 (UD 112 )T. Again, if we generate a vector of n independent N(O, 1) random
variables Z
18
=[z" z2 ,. • • ,zJ7', X can be generated as
X
Th e probb"l"
· o f mu1·
·
· ts
· f(x) = cxp(-'. where ct> is an empty set.
Q
Forany mutually exclusive events E,, £ 2 ,. • • E" , P(
E;) =
t
P(E;).
Random variable: A function that maps each outcome (ro) in the sample space (Q) into
the set of real numbers.
Let's use the rolling of a six-sided dice to explain these definitions and notations. A roll
of a dice has 6 possible outcomes (mapped to a random variable): 1, 2, 3, 4, 5, or 6. So
the sample space Q is {1,2,3,4,5,6} and the probability of each outcome is 116
(assuming a fair dice). We can define an event A representing the event that the outcome
is an odd number A = {1, 3, 5}, then the complement of A is Ac = {2, 4, 6} . Clearly
P( A) = P(l) + P(3) + P( 5) = 1I2. Let B be the event that the outcome is larger than 3:
B = {4, 5, 6}. Then the union is Au B = {1, 3, 4, 5, 6} and the intersection is
An B = {5}. One popular random variable called indicator variable (a binary dummy
variable) for event A is defined as the following:
IA
=
{
if X E {1, 3, 5}
. Basically I A = 1 when A occurs and I A = 0 if A'. occurs. The
0, if
{1, 3, 5}
1,
expected value of I A is £[I A]= P(A).
Now, time for some examples.
60
A Practical Guide To Quantitative Finance Interviews
Coin toss game
Two gamblers are playing a coin toss game. Gambler A has (n + 1) fair coins; B has n
fair coins. What is the probability that A will have more heads than B if both flip all their
coins?2
Solution: We have yet to cover all the powerful tools probability theory offers. What do
we have now? Outcomes, events, event probabilities, and surely our reasoning
capabilities! The one extra coin makes A different from B. If we remove a coin from A,
A and B will become symmetric. Not surprisingly, the symmetry will give us a lot of
nice properties. So let's remove the last coin of A and compare the number of heads in
A's first n coins with B's n coins. There are three possible outcomes:
£ 1 : A's n coins have more heads than B's n coins;
£ 2 : A's n coins have equal number of heads as B's n coins;
£ 3 : A's n coins have fewer heads than B's n coins.
By symmetry, the probability that A has more heads is equal to the probability that B has
more heads. So we have P(E1) = P(E3 ). Let's denoteP(E1) = P(E3 ) = xand P(E2 ) = y.
Since
L P( m) = 1, we have 2x + y = 1. For event £
1,
A will always have more heads
men
than B no matter what A's (n+l)th coin's side is; for event £ 3 , A will have no more
heads than B no matter what A's (n+l)th coin's side is. For event £ 2 , A's (n+l)th
coin does make a difference. If it's a head, which happens with probability 0.5, it will
make A have more heads than B. So the (n + l)th coin increases the probability that A
has more heads than B by 0.5 y and the total probability that A has more heads is
x+0.5y=x+0.5(1-2x)=0.5 when A has (n+l) coins.
Card game
A casino offers a simple card game. There are 52 cards in a deck with 4 cards for each
jack queen king ace
value 2, 3, 4, 5, 6, 7, 8, 9, 10, J, Q, K, A. Each time the cards are thoroughly shuffled
(so each card has equal probability of being selected). You pick up a card from the deck
and the dealer picks another one without replacement. If you have a larger number, you
win; if the numbers are equal or yours is smaller, the house wins-as in all other casinos,
the house always has better odds of winning. What is your probability of winning?
2 Hint: What are the possible results (events) if we compare the number of heads in A's first n coins with
B's n coins? By making the number of coins equal, we can take advantage of symmetry. For each event,
what will happen if A's last coin is a head? Or a tail?
61
Probability Theory
Solution: One answer to this problem is to consider all 13 different outcomes of your
card. The card can have a value 2, 3, ···,A and each has 1/13 of probability. With a
value of 2, the probability of winning is 0/51; with a value of 3, the probability of
winning is 4/51 (when the dealer picks a 2); ... ; with a value of A, the probability of
winning is 48/51 (when the dealer picks a 2, 3, · · ·, or K). So your probability of
. . .
wmnmg ts
1 ( 0 4
48)
4
4
12x13 8
iix 51+51+···+51 = 13x51x(O+l+···+l 2)=13x51 x 2 =17·
Although this is a straightforward solution and it elegantly uses the sum of an integer
sequence, it is not the most efficient way to solve the problem. If you have got the core
spirits of the coin tossing problem, you may approach the problem by considering three
different outcomes:
E, : Your card has a number larger than the dealer's;
£ 2 : Your card has a number equal to the dealer's;
£ 3 : Your card has a number lower than the dealer's.
Again by symmetry, P(E,) = P(E3 ). So we only need to figure out P(E2 ), the
probability that two cards have equal value. Let's say you have randomly selected a card.
Among the remaining 51 cards, only 3 cards will have the same value as your card. So
the probability that the two cards have equal value is 3/51. As a result, the probability
that you win is P( E,) =(1- P( E 2 )) I 2 = (1- 3 I 51) I 2 = 8I17.
Drunk passenger
A line of 100 airline passengers are waiting to board a plane. They each hold a ticket to
one of the 100 seats on that flight. For convenience, let's say that the n-th passenger in
line has a ticket for the seat number n. Being drunk, the first person in line picks a
random seat (equally likely for each seat). All of the other passengers are sober, and will
go to their proper seats unless it is already occupied; In that case, they will randomly
choose a free seat. You're person number 100. What is the probability that you end up
in your seat (i.e., seat #100) ?3
Solution: Let's consider seats #1 and #100. There are two possible outcomes:
Hint: If you are trying to use complicated conditional probability to solve the problem, go back and think
again. If you decide to start with a simpler version of the problem, starting with two passengers and
increasing the number of passengers to show a pattern by induction, you can solve the problem more
efficiently. But the problem is much simpler than that. Focus on events and symmetry and you will have
an intuitive answer.
3
62
A Practical Guide To Quantitative Finance Interviews
£ 1 : Seat # 1 is taken before # 100;
£ 2 : Seat # 100 is taken before # 1.
If any passenger takes seat # 100 before # 1 is taken, surely you will not end up in you
own seat. But if any passenger takes # 1 before # 100 is taken, you will definitely end up
in you own seat. By symmetry, either outcome has a probability of 0.5. So the
probability that you end up in your seat is 50%.
In case this over-simplified version of reasoning is not clear to you, consider the
following detailed explanation: If the drunk passenger takes #1 by chance, then it's clear
all the rest of the passengers will have the correct seats. If he takes # 100, then you will
not get your seat. The probabilities that he takes #1 or #100 are equal. Otherwise assume
that he takes the n-th seat, where n is a number between 2 and 99. Everyone between 2
and (n-1) will get his own seat. That means the n-th passenger essentially becomes the
new "drunk" guy with designated seat #1. If he chooses #1, all the rest of the passengers
will have the correct seats. If he takes # 100, then you will not get your seat. (The
probabilities that he takes # 1 or # 100 are again equal.) Otherwise he will just make
another passenger down the line the new "drunk" guy with designated seat # 1 and each
new "drunk" guy has equal probability of taking #1 or #100. Since at all jump points
there's an equal probability for the "drunk" guy to choose seat #1 or 100, by symmetry,
the probability that you, as the lOOth passenger, will seat in #100 is 0.5.
N points on a circle
Given N points drawn randomly on the circumference of a circle, what is the probability
that they are all within a semicircle?4
Solution: Let's start at one point and clockwise label the points as 1, 2, · · ·, N . The
probability that all the remaining N -1 points from 2 to N are in the clockwise
semicircle starting at point 1 (That is, if point 1 is at 12:00, points 2 to N are all
between 12:00 and 6:00) is 1I2N-i. Similarly the probability that a clockwise semicircle
starting at any point i, where i e {2, · · ·, N} contains all the other N -1 points is also
1/ 2N-I.
Claim: the events that all the other N -1 points are in the clockwise semicircle starting
at point i, i = 1, 2, · · ·, N are mutually exclusive. In other words, if we, starting at point i
and proceeding clockwise along the circle, sequentially encounters points i + 1, i + 2, · · ·,
N, 1, · · ·, i -1 in half a circle, then starting at any other point j, we cannot encounter all
4 Hint: Consider the events that starting from a point n, you can reach all the rest of the points on the circle
clockwise, n E {I, ... , N} in a semicircle. Are these events mutually exclusive?
63
Probability Theory
other points within a clockwise semicircle. Figure 4.1 clearly demonstrates this
conclusion. If starting at point i and proceeding clockwise along the circle, we
sequentially encounter points i + 1, i + 2, · ·., N, 1, · ·., i -1 within half a circle, the
clockwise arc between i -1 and i must be no less than half a circle. If we start at any
other point, in order to reach all other points clockwise, the clockwise arc between i -1
and i are always included. So we cannot reach all points within a clockwise semicircle
starting from any other points. Hence, all these events are mutually exclusive and we
have
P(QE} t.P(E;)=>P(QE} Nxl/2"-' N12"-'
The same argument can be extended to any arcs that have a length less than half a circle.
If the ratio of the arc length to the circumference of the circle is x ( x 1I2 ), then the
probability of all N points fitting into the arc is N x xN-i.
, 2.
66
A Practical Guide To Quantitative Finance Interviews
Solution: Let's begin with the simplest cases and consider solving the problem for any
number of stairs using induction. For n =I , there is only one way and f (1) =I. For
n = 2, we can have one 2-stair hop or two I-stair hops. So f (2) = 2. For any n > 2,
there are always two possibilities for the last hop, either it's a I-stair hop or a 2-stair hop.
In the former case, the rabbit is at (n-1) before reaching n, and it has f(n -1) ways to
reach (n -1). In the latter case, the rabbit is at (n-2) before reaching n, and it has
f(n-2) ways to reach (n-2). So we have f(n)=f(n-2)+/(n-l). Using this
function we can calculate f(n) for n = 3, 4, · · · 6
Screwy pirates 2
Having peacefully divided the loot (in chapter 2), the pirate team goes on for more
looting and expands the group to 11 pirates. To protect their hard-won treasure, they
gather together to put all the loot in a safe. Still being a democratic bunch, they decide
that only a majority - any majority - of them
together can open the safe. So they
ask a locksmith to put a certain number of locks on the safe. To access the treasure,
every lock needs to be opened. Each lock can have multiple keys; but each key only
opens one lock. The locksmith can give more than one key to each pirate.
What is the smallest number of locks needed? And how many keys must each pirate
carry?7
Solution: This problem is a good example of the application of combinatorial analysis in
information sharing and cryptography. A general version of the problem was explained
in a 1979 paper "How to Share a Secret" by Adi Shamir. Let's randomly select 5 pirates
from the I I-member group; there must be a lock that none of them has the key to. Yet
any of the other 6 pirates must have the key to this lock since any 6 pirates can open all
locks. In other words, we must have a "special" lock to which none of the 5 selected
pirates has a key and the other 6 pirates all have keys. Such 5-pirate groups are randomly
selected. So for each combination of 5 pirates, there must be such a "special" lock. The
minimum number of locks needed is (I I) = __!__!_!_ = 462 locks. Each lock has 6 keys,
5
5!6!
which are given to a unique 6-member subgroup. So each pirate must have
462x6
- - - = 252 keys. That's surely a lot of locks to put on a safe and a lot of keys for
11
each pirate to carry.
You may have recognized that the sequence is a sequence of Fibonacci numbers.
Hint: every subgroup of 6 pirates should have the same key to a unique lock that the other 5 pirates do
not have.
6
7
67
Probability Theory
Chess tournament
A chess tournament has 2n players with skills 1 > 2 > · · · >2". It is organized as a
knockout tournament, so that after each round only the winner proceeds to the next
round. Except for the final, opponents in each round are drawn at random. Let's also
assume that when two players meet in a game, the player with better skills always wins.
What's the probability that players 1 and 2 will meet in the final? 8
Solution: There are at least two approaches to solve the problem. The standard approach
applies multiplication rule based on conditional probability, while a counting approach
is far more efficient. (We will cover conditional probability in detail in the next section.)
Let's begin with the conditional probability approach, which is easier to grasp. Since
there are 2n players, the tournament will have n rounds (including the final). For round
1, players 2,3,-··,2n each have 2n1_ 1 probability to be 1's rival, so the probability that
. round 1 is
. 2n -2
. on that 1 and 2 do not
1 and 2 do not meet m
- = 2x(2n-l -1) . Cond"ition
2n -1
2n -1
meet in round 1, 2n-i players proceed to the 2nd round and the conditional probability
2n-l - 2 2 X (2n- 2 -1)
that 1 and 2 will not meet in round 2 is
= 2n-1 -1 • We can repeat the same
2n-1 -1
process until the (n - l)th round, in which there are 22 (= 2n I 2n-2 ) players left and the
conditional probability that 1 and 2 will not meet in round ( n -1) is
2 2 -2 2x(2 2- 1 -l)
=---22-1
22 -1
Let E1 be the event that 1 and 2 do not meet in round 1;
E2 be the event that 1 and 2 do not meet in rounds 1and2;
En-i be the event that 1 and 2 do not meet in round 1, 2, · · ·, n -1.
Apply the multiplication rule, we have
P(l and 2 meet in the nth game)= P(E1) x P(E2 I E 1) x · · · x P(En-i I E1E2 • • • En_ 2 )
2 X (2n-l - J) 2 X (2n- 2 -J)
2 X (2 2-l -1)
2n-I
=
X
X···X
=-2n - l
2n-I - l
22 - )
2n -1
Hint: Consider separating the players to two 2•-' subgroups. What will happen if player 1 and 2 in the
same group? Or not in the same group?
8
68
A Practical Guide To Quantitative Finance Interviews
Now let's move on to the counting approach. Figure 4.2A is the general case of what
happens in the final. Player 1 always wins, so he will be in the final. From the figure, it
is obvious that 2n players are separated to two 2n-i -player subgroups and each group
will have one player reaching the final. As shown in Figure 4.28, for player 2 to reach
the final, he/she must be in a different subgroup from 1. Since any of the remaining
players in 2, 3, · · ·, 2n are likely to be one of the (2n-i - 1) players in the same subgroup
as player 1 or one of the 2n-i players in the subgroup different from player 1, the
probability that 2 is in a different subgroup from 1 and that I and 2 will meet in the final
is simply
Clearly, the counting approach provides not only a simpler solution but
2n -1
also more insight to the problem.
nth round
(n-l)th round
General Case
1 & 2 in the Final
I
1
t
t
+
+
I
?
/\
/\
1 +
?
? +
?
nth round
(n-l)thround
I
2
/\
/\
1 +
?
2 +
?
2n-I players 2n-I players
2n-I players 2n-I players
A
B
Figure 4.2A The general case of separating 2" players into 2"- 1-player subgroups;
4.28 The special case with players 1 and 2 in different groups
Application letters
You're sending job applications to 5 firms: Morgan Stanley, Lehman Brothers, UBS,
Goldman Sachs, and Merrill Lynch. You have 5 envelopes on the table neatly typed with
names and addresses of people at these 5 firms. You even have 5 cover letters
personalized to each of these firms. Your 3-year-old tried to be helpful and stuffed each
cover letter into each of the envelopes for you. Unfortunately she randomly put letters
69
Probability Theory
into envelopes without realizing that the letters are personalized. What is the probability
that all 5 cover letters are mailed to the wrong firms? 9
Solution: This problem is a classic example for the Inclusion-Exclusion Principle. In fact,
a more general case is an example in Ross' textbook First Course in Probability.
Let's denote by E;, i = l,·· ·,5 the event that the i-th letter has the correct envelope. Then
is the probability that at least one letter has the correct envelope and
P(
1-P(
is the probability that all letters have the wrong envelopes. P(
can
be calculated using the Inclusion-Exclusion Principle:
t.P(E,)-
P(E1E2 ···E,)
It's obvious thatP(E;) =_!_, Vi= 1,-··,5. So i:P(E;) = 1.
5
i=I
P(E;,E;) is the event that both letter i, and letter i 2 have the correct envelope. The
probability that i1 has the correct envelope is 1I5; Conditioned on that i1 has the correct
envelope, the probability that i2 has the correct envelope is 1I4 (there are only 4
envelopes left). So P(E,.E; )=_!_x-1-= ( 5 - 2)!.
5 5-1
5!
I
2
J
5!
There are ( 5 =
members of P(E;,E;) in LP(E;,E;), so we have
2
2!(5-2)!
i1 0, then P(A IB) = P(AB) is the fraction
P(B)
of B outcomes that are also A outcomes.
11
Hint: The last two digits of x 3 only depend on the last two digits of x.
72
A Practical Guide To Quantitative Finance Interviews
Law of total orobabilitv: for any mutually exclusive events {F;} , i = 1, 2, · · ·, n, whose
n
union is the entire sample space ( F; n F1 =, Vi -:1; j;
LJ F; = n ), we have
i =I
n
P(E)
=
P(EF;)+P(EF2 ) + ··· +P(EFn) =LP(£ IF;)P(F;)
i=I
= P(E IF;)P(F;) + P(E IF;_)P(F2 ) + · ·· + P(E IFn)P(F,,)
Independent events: P(EF) = P(E)P(F) => P(EFc) = P(E)P(Fc).
Independence is a symmetric relation: X is independent of Y <=> Y is independent of X.
8 ayes ' F ormuIa: P(F.J I £) --
P(E I F. )P(F.)
n
J
J
L P(E IF; )P(F;)
if F;, i
=
1, · ·., n, are mutually
i=l
exclusive events whose union is the entire sample space.
As the following examples will demonstrate, not all conditional probability problems
have intuitive solutions. Many demand logical analysis instead.
Boys and girls
Part A. A company is holding a dinner for working mothers with at least one son. Ms.
Jackson, a mother with two children, is invited. What is the probability that both
children are boys?
Solution: The sample space of two children is given by 0= {(b,b),(b,g),(g,b),(g,g)}
(e.g., (g, b) means the older child is a girl and the younger child a boy), and each
outcome has the same probability. Since Ms. Jackson is invited, she has at least one son.
Let B be the event that at least one of the children is a boy and A be the event that both
children are boys, we have
P(A I B) = P(A nB) =
P( {(b,b)})
= _11_4 = _!_
P(B)
P( {(b,b),(b,g),(g,b)}) 3 I 4 3
Part B. Your new colleague, Ms. Parker is known to have two children. If you see her
walking with one of her children and that child is a boy, what is the probability that both
children are boys?
73
Probability Theory
Solution: the other child is equally likely to be a boy or a girl (independent of the boy
you've seen), so the probability that both children are boys is 1/2.
Notice the subtle difference between part A and part B. In part A, the problem essentially
asks given there is at least one boy in two children, what is the conditional probability
that both children are boys. Part B asks that given one child is a boy, what is the
conditional probability that the other child is also a boy. For both parts, we need to
assume that each child is equal likely to be a boy or a girl.
All-girl world?
In a primitive society, every couple prefers to have a baby girl. There is a 50% chance
that each child they have is a girl, and the genders of their children are mutually
independent. If each couple insists on having more children until they get a girl and once
they have a girl they will stop having more children, what will eventually happen to the
fraction of girls in this society?
Solution: It was surprising that many interviewees-include many who studied
probability-have the misconception that there will be more girls. Do not let the word
"prefer" and a wrong intuition misguide you. The fraction of baby girls are driven by
nature, or at least the X and Y chromosomes, not by the couples' preference. You only
need to look at the key information: 50% and independence. Every new-born child has
equal probability of being a boy or a girl regardless of the gender of any other children.
So the fraction of girls born is always 50% and the fractions of girls in the society will
stay stable at 50%.
Unfair coin
You are given 1000 coins. Among them, 1 coin has heads on both sides. The other 999
coins are fair coins. You randomly choose a coin and toss it 10 times. Each time, the
coin turns up heads. What is the probability that the coin you choose is the unfair one?
Solution: This is a classic conditional probability question that uses Bayes' theorem. Let
A be the event that the chosen coin is the unfair one, then A" is the event that the chosen
coin is a fair one. Let B be the event that all ten tosses tum up heads. Apply Bayes'
theorem we have P( A IB) = P( B I A)P( A) =
P(B)
P( B I A)P( A)
P(B I A)P(A) + P(B I A")P(A")
The priors are P( A) = 1/1000 and P( A") = 99911000. If the coin is unfair, it always
turns up heads, so P(B I A)= 1. If the coin is fair, each time it has 1/2 probability turning
74
A Practical Guide To Quantitative Finance interviews
up heads. So P(B I A")= (1I2) 10 = 111024. Plug in all the available information and we
have the answer:
P(AIB)-
P(BIA)P(A)
P(B I A)P(A) + P(B I A' )P(A')
=
1/lOOOxl
1/1000x1+999/1000x1I1024
Fair probability from an unfair coin
If you have an unfair coin, which may bias toward either heads or tails at an unknown
probability, can you generate even odds using this coin?
Solution: Unlike fair coins, we clearly can not generate even odds with one toss using an
unfair coin. How about using 2 tosses? Let PH be the probability the coin will yield
head, and Pr be the probability the coin will yield tails ( pH + Pr = 1 ). Consider two
independent tosses. We have four possible outcomes HH, HT, TH and TT with
probabilities P(HH) = PHPH, P(HT) = PHPr> P(TH) =Pr PH' and P(TT) =Pr Pr.
So we have P(HT) =P(TH). By assigning HT to winning and TH to losing, we can
generate even odds. 12
Dart game
Jason throws two darts at a dartboard, aiming for the center. The second dart lands
farther from the center than the first. If Jason throws a third dart aiming for the center,
what is the probability that the third throw is farther from the center than the first?
Assume Jason's skillfulness is constant.
Solution: A standard answer directly applies the conditional probability by enumerating
all possible outcomes. If we rank the three darts' results from the best (A) to the worst
(C), there are 6 possible outcomes with equal probability:
12 I should point out that this simple approach is not the most efficient approach since I am disregarding
the cases HH and TT. When the coin has high bias (one side is far more likely than the other side to occur),
the method may take many runs to generate one useful result. For more complex algorithm that increasing
efficiency, please refer to Tree Algorithms for Unbiased Coin Tossing with a Biased Coin by Quentin F.
Stout and Bette L. Warren, Annals of Probability 12 ( 1984), pp. 212-222.
75
Probability Theory
Outcome
1
2
3
4
5
6
1st throw
A
B
A
c
B
c
2nd throw
B
A
c
A
c
B
3rd throw
c
c
B
B
A
A
The information from the first two throws eliminates outcomes 2, 4 and 6. Conditioned
on outcomes 1, 3, and 5, the outcomes that the 3rd throw is worse than the 1st throw are
outcomes 1 and 3. So there is 2/3 probability that the third throw is farther from the
center than the first.
This approach surely is reasonable. Nevertheless, it is not an efficient approach. When
the number of darts is small, we can easily enumerate all outcomes. What if it is a more
complex version of the original problem:
Jason throws n ( n 2 5) darts at a dartboard, aiming for the center. Each subsequent dart
is farther from the center than the first dart. If Jason throws the (n + l)th dart, what is the
probability that it is also farther from the center than his first?
This question is equivalent to a simple question: what is the probability that the (n + 1)th
throw is not the best among all (n + 1) throws? Since the 1st throw is the best among the
first n throws, essentially I am saying the event that the (n + l)th throw is the best of all
( n + 1) throws (let's call it An+i ) is independent of the event that the 1st throw is the best
of the first n throws (let's call it A1 ). In fact, An+i is independent of the order of the first
n throws. Are these two events really independent? The answer is a resounding yes. If it
is not obvious to you that An+i is independent of the order of the first n throws, let's look
at it another way: the order of the first n throws is independent of
An+i •
Surely this claim
is conspicuous. But independence is symmetric! Since the probability of
1/(n+1), the probability that (n + l)th throw is not the best is n /(n + 1) . 13
An+i
is
For the original version, three darts are thrown independently, they each have a 1/3
chance of being the best throw. As long as the third dart is not the best throw, it will be
worse than the first dart. Therefore the answer is 2/3.
Birthday line
At a movie theater, a whimsical manager announces that she will give a free ticket to the
first person in line whose birthday is the same as someone who has already bought a
ticket. You are given the opportunity to choose any position in line. Assuming that you
13
Here you can again use symmetry argument: each throw is equally likely to be the best.
76
A Practical Guide To Quantitative Finance Interviews
don't know anyone else's birthday and all birthdays are distributed randomly throughout
the year (assuming 365 dals in a year), what position in line gives you the largest chance
of getting the free ticket? 1
Solution: If you have solved the problem that no two people have the same birthday in
an n-people group, this new problem is just a small extension. Assume that you choose
to be the n-th person in line. In order for you to get the free ticket, all of the first n -1
individuals in line must have different birthdays and your birthday needs to be the same
as one of those n -1 individuals.
p(n) = p(first n-1 peoplehavenosamebirthday) x p(yours among those n -1 birthdays)
365 x 364 x · · · (365 - n + 2)
365n-i
n -1
365
It is intuitive to argue that when n is small, increasing n will increase your chance of
getting the free ticket since the increase of p (yours among those n -1 birthdays) is
more significant than the decrease in p(first n-1 peoplehavenosamebirthday). So
when n is small, we have P(n) > P(n-1). As n increases, gradually the negative impact
of p (first n -1 people have no same birthday) will catch up and at a certain point we
will have P(n+l)P(n-1)
and P(n) > P(n + 1).
P(n - l) = 365 x 364 x ... x 365-(n -3) x n - 2
365 365
365
365
P(n)= 365 x 364 X···x 365-(n-2) x n-1
365 365
365
365
P(n+l)= 365 x 364 X···X 365-(n-2) x 365-(n-1) x_!!_
365 365
365
365
365
n-21
Hence,
P(n)>P(n-1)=> 365-(n-2) xn-1
-->-365
365
365
P(n) > P(n + 1)
n-1 > 365-(n-l) x--"365
365
365
2
=>
n - 3n - 363 < 0
n' -n-365 > 0 }
=> n = 20
You should be the 20th person in line.
14 Hint: If you are the n-th person in line, to get the free ticket, the first (n-I) people in line must not have
the same birthday and you must have the same birthday as one of them.
77
Probability Theory
Dice order
We throw 3 dice one by one. What is the probability that we obtain 3 points in strictly
increasing order? 15
Solution: To have 3 points in strictly increasing order, first all three points must be
different numbers. Conditioned on three different numbers, the probability of strictly
increasing order is simply l I 3 ! = l I 6 (one specific sequence out of all possible
permutations). So we have
P = P(different numbers in all three throws) x P(increasing orderl3 different numbers)
= (Ix2-x-i)x_J__
= 5 I 54
6 6
6
Monty Hall problem
Monty Hall problem is a probability puzzle based on an old American show Let's Make
a Deal. The problem is named after the show's host. Suppose you're on the show now,
and you're given the choice of 3 doors. Behind one door is a car; behind the other two,
goats. You don't know ahead of time what is behind each of the doors.
You pick one of the doors and announce it. As soon as you pick the door, Monty opens
one of the other two doors that he knows has a goat behind it. Then he gives you the
option to either keep your original choice or switch to the third door. Should you switch?
What is the probability of winning a car if you switch?
Solution: If you don't switch, whether you win or not is independent of Monty's action
of showing you a goat, so your probability of winning is 1/3. What if you switch? Many
would argue that since there are only two doors left after Monty shows a door with goat,
the probability of winning is 1/2. But is this argument correct?
If you look at the problem from a different perspective, the answer becomes clear. Using
a switching strategy, you win the car if and only if you originally pick a door with a goat,
which has a probability of 2/3 (You pick a door with a goat, Monty shows a door with
another goat, so the one you switch to must have a car behind it). If you originally
picked the door with the car, which has a probability of 113, you will lose by switching.
So your probability of winning by switching is actually 2/3.
15 Hint: To obtain 3 points in strictly increasing order, the 3 points must be different. For 3 different points
in a sequence, strictly increasing order is one of the possible permutations.
78
A Practical Guide To Quantitative Finance Interviews
Amoeba population
There is a one amoeba in a pond. After every minute the amoeba may die, stay the same,
split into two or split into three with equal probability. All its offspring, if it has any, will
behave the same (and independent of other amoebas). What is the probability the
amoeba population will die out?
Solution: This is just another standard conditional probability problem once you realize
we need to derive the probability conditioned on what happens to the amoeba one
minute later. Let P(E) be the probability that the amoeba population will die out and
apply the law of total probability conditioned on what happens to the amoeba one
minute later:
For the original amoeba, as stated in the question, there are four possible mutually
exclusive events each with probability 1/4. Let's denote F; as the event the amoeba dies;
F2 as the event that it stays the same;
F; as the event that it splits into two;
F4 as the
event that it splits into three. For event F;, P( E I F;) = I since no amoeba is left.
P(E IF2 )
= P(E)
since the state is the same as the beginning. For F;, there are two
amoebas; either behaves the same as the original one. The total amoeba population will
die only if both amoebas die out. Since they are independent, the probability that they
both will die out is P(E) 2 • Similarly we have P(F4 ) = P(E)3. Plug in all the numbers,
the equation becomes P(E)= 1/4xl+I/4 x P(E)+1/4xP(E) 2 +1/4xP(E) 3 • Solve
this equation with the restriction 0 < P(E) p = 6111. So you should
choose to be the second player and have 5I11 probability of losing.
If instead of one bullet, two bullets are randomly put in the chamber. Your opponent
played the first and he was alive after the first trigger pull. You are given the option
whether to spin the barrel. Should you spin the barrel?
81
Probability Theory
Solution: if you spin the barrel, the probability that you will lose in this round is 2/6. If
you don't spin the barrel, there are only 5 chambers left and your probability oflosing in
this round (conditioned on that your opponent survived) is 2/5. So you should spin the
barrel.
What if the two bullets are randomly put in two consecutive positions? If your opponent
survived his first round, should you spin the barrel?
Solution: Now we have to condition our probability on the fact that the positions of the
two bullets are consecutive. As shown in Figure 4.3, let's label the empty chambers as 1,
2, 3 and 4; label the ones with bullets 5 and 6. Since your opponent survived the first
round, the possible position he encountered is 1, 2, 3 or 4 with equal probability. With
1/4 chance, the next one is a bullet (the position was 4). So if you don't spin, the chance
of survival is 3/4. If you spin the barrel, each position has equal probability of being
chosen, and your chance of survival is only 2/3. So you should not spin the barrel.
0)
Figure 4.3 Russian roulette with two consecutive bullets.
Aces
Fifty-two cards are randomly distributed to 4 players with each player getting 13 cards.
What is the probability that each of them will have an ace?
Solution: The problem can be answered using standard counting methods. To distribute
52 cards to 4 players with 13 cards each has
82
52 !
permutations. If each player
13!13!13!13!
A Practical Guide To Quantitative Finance Interviews
needs to have one ace, we can distribute the aces first, which has 4 ! ways. Then we
48!
distribute the rest 48 cards to 4 players with 12 cards each, which has - - - - 12!l2!l2!l2 !
permutations. So the probability that each of them will have an Ace is
4 !x
48!
52!
=52x39x26xQ,
7
12!12!12!12! 13!13!13!13! 52 51 50 49
The logic becomes clearer if we use a conditional probability approach. Let's begin with
any one of the four aces; it has probability 52 I 52 =I of belonging to a pile. The second
ace can be any of the remaining 51 cards, among which 39 belong to a pile different
from the first ace. So the probability that the second ace is not in the pile of the first ace
is 39 I 51 . Now there are 50 cards left, among which 26 belong to the other two piles. So
the conditional probability that the third ace is in one of the other 2 piles given the first
two aces are already in different piles is 26 I 50. Similarly, the conditional probability
that the fourth ace is in the pile different from the first three aces given that the first
three aces are in different piles is 13 I 49 . So the probability that each pile has an ace is
Ix 39 x 26
51 50
49
Gambler's ruin problem
A gambler starts with an initial fortune of i dollars. On each successive game, the
gambler wins $1 with probability p, 0 < p < I, or loses $1 with probability q = 1- p . He
will stop if he either accumulates N dollars or loses all his money. What is the
probability that he will end up with N dollars?
Solution: This is a classic textbook probability problem called the Gambler's Ruin
Problem. Interestingly, it is still widely used in quantitative interviews.
From any initial state i (the dollars the gambler has), 0:::;; i:::;; N, let P; be the probability
that the gambler's fortune will reach N instead of 0. The next state is either i +I with
probability p or i -1 with probability q. So we have
=
P, pP,., +qP,_, => P,., -P,
=; (P,-P,_, ) =(;
We also have the boundary probabilities
So starting from
J
(P,_, -P,_, )
Po = 0 and
Pi, we can successively evaluate
=... =(; )' (J:-P, )
PN = I .
as an expression of
Pi :
83
Probability Theory
Extending this expression to PN, we have
PN
q
= 1 =[ 1+ -
(qJN-1] Pi ={1-(q/p)N
Pi,ifq/p*l
1- q Ip
+ ... + -
P
P
if q Ip = 1
NI>i,
l-q/p
zif I *1
{l-(qlpYp zif *112
=>Pi= { 1-(q/p)N'
q p
=>P;= 1-(q/p)N 1' p
l!N,
ifqlp=l
i/N,
if p=l/2
Basketball scores
A basketball player is taking 100 free throws. She scores one point if the ball passes
through the hoop and zero point if she misses. She has scored on her first throw and
missed on her second. For each of the following throw the probability of her scoring is
the fraction of throws she has made so far. For example, if she has scored 23 points after
the 40th throw, the probability that she will score in the 41 th throw is 23/40. After 100
throws (including the first and the second), what is the probability that she scores exactly
50 baskets? 18
Solution: Let (n,k), 1 k
throws and
n, be the event that the player scores k baskets after n
P,,,k = P( (n,k) ) . The solution is surprisingly simple if we use an induction
approach starting with n = 3. The third throw has 1/2 probability of scoring. So we have
= 1I2 and
= 1I2. For the case when n = 4, let's apply the law of total
probability
18 Hint: Again, do not let the number 100 scares you. Start with smallest n, solve the problem; try to find a
pattern by increasing n; and prove the pattern using induction.
84
A Practical Guide To Quantitative Finance Interviews
P((4,l) I (3,1)) x
2 1
1 1
= 3x2+ Ox 2 = 3
+ P((4,l) I (3,2))
1 1 1 1 1
= P((4,2) I(3,1)) x P..31 + P((4,2) I (3,2)) x
= -x-+-x- = '
.
'3232 3
1 2 1 1
= P((4,3) I(3,1))
= 0 x-+-x- = 1 + P((4,3) I (3,2)) x
.
.
'
2 3 2 3
= - 1-, Vk = 1, 2, · · ·, n - 1 , and give the hint that the law of
n-1
total probability can be used in the induction step.
The results indicate that P
k
Induction
that Pk= - 1-, Vk = 1, 2,···, n-1, we
n,
n-1
n,
step:
given
1
=1
(n+l)-1 n
probability:
=
P,,+i,k
= P( miss I(n, k)) P,,,k + P( score I (n, k-1)) P,,,k-I
=(l-:)
+
to
prove
Vk = 1, 2, .. ·, n. To show it, simply apply the law of total
p
n+l.k
need
=;
The equation is also applicable to the P,,+ 1,1 and
P,,+i,n•
although in these cases k - I = 0
n
1-, Vk=l,2, .. .,n-land Vn"?.2.
and (1-kn)=o, respectively. So we have P*=n,
Hence,
n-I
Pioo 50 = 1199 .
Cars on road
If the probability of observing at least one car on a highway during any 20-minute time
interval is 609/625, then what is the probability of observing at least one car during any
5-minute time interval? Assume that the probability of seeing a car at any moment is
uniform (constant) for the entire 20 minutes.
Solution: We can break down the 20-minute interval into a sequence of 4 non-
overlapping 5-minute intervals. Because of constant default probability (of observing a
car), the probability of observing a car in any 5-minute interval is constant. Let's denote
the probability to be p, then the probability that in any 5-minute interval we do not
observe a car is I - p .
85
Probability Theory
The probability that we do not observe any car in all four of such independent 5-minute
intervals is (1- p ) 4 = 1-609 I 625 = 16 I 625, which gives p = 3 I 5.
4.4 Discrete and Continuous Distributions
In this section, we review a variety of distribution functions for random variables that
are widely used in quantitative modeling. Although it may not be necessary to memorize
the properties of these distributions, having an intuitive understanding of the
distributions and having the ability to quickly derive important properties are valuable
skills in practice. As usual, let's begin with the theories:
Common function of random variables
Table 4.1 summarizes how the basic properties of discrete and continuous random
variables are defined or calculated. These are the basics you should commit to memory.
Random variable ill
Discrete
Continuous 19
Cumulative distribution function/cdf
F(a) = P{X
F(a) = [f(x)dx
Probability mass function /pmf
pmf:
p(x) = P{X = x}
pdf: f(x) =-F(x)
dx
Probabil!!Y_ dens!!Y_ function
Expected value/ E[ X]
Expected value of g(X)/ E[g(X)]
L
xp(x)
(xf(x)dx
g(x)p(x)
(g(x)f(x)dx
x:p(x)>O
L
d
x:p(x)>O
Variance of XI var( X)
E[(X -E[X]) 2 ] = E[X 2 ]-(E[X]) 2
Standard deviation of XI std(X)
.Jvar(X)
Table 4.1 Basic properties of discrete and continuous random variables
Discrete random variables
Table 4.2 includes some of the most widely-used discrete distributions. Discrete uniform
random variable represents the occurrence of a value between number a and b when all
values in the set {a, a+l,- · · , b} have equal probability. Binomial random variable
represents the number of successes in a sequence of n experiments when each trial is
19 For continuous random variables, P(X = x)
86
= 0, \Ix E (-00,00), so
P{X
x} = P{X < x}.
A Practical Guide To Quantitative Finance Interviews
independently a success with probability p . Poisson random variable represents the
number of events occurring in a fixed period of time with the expected number of
occurrences At when events occur with a known average rate A and are independent of
the time since the last event. Geometric random variable represents the trial number (n)
to get the first success when each trial is independently a success with probability p.
Negative Binomial random variable represents the trial number to get to the r-th success
when each trial is independently a success with probability p .
E[X]
var(X)
--
b+a
2
(b-a+l) 2 -1
12
np
np(l- p)
A.t
At
1
p
1- p
r
p
r(l- p)
Name
Probability mass function (pmf)
Uniform
P(x) =
Binomial
P(x)
Poisson
P(x)=
Geometric
P(x) = (1- py-' p, x = 1,2, ...
-
Negative
Binomial
P(x)= ( x-1) pr(l-py-r, x=r,r+l,-··
r-1
-
1
b-a+l
=(:
,
x = a,a+ 1,.··,b
)p'(I- Pr-',
e_,i, (A.t}"
x!
,
x
= 0,1,. · .,n
x = 0, 1, .. . 20
p2
p2
Table 4.2 Probability mass function, expected value and variance of discrete random
variables
Continuous random variables
Table 4.3 includes some of the commonly encountered continuous distributions.
Uniform distribution describes a random variable uniformly distributed over the interval
[a, b]. Because of the central limit theorem, normal distribution/Gaussian distribution is
by far the most popular continuous distribution. Exponential distribution models the
arrival time of an event if it has a constant arrival rate A.. Gamma distribution with
parameters (a, A.) often arises, in practice, as the distribution of the amount of time one
has to wait until a total of n events occur. Beta distributions are used to model events
Here we use the product of arrival rate A. and time t to define the parameter (expected value) since it is
the definition used in many Poisson process studies.
20
87
Probability Theory
that are constrained within a defined interval. By adjusting the shape parameters a and
it can model different shapes of probability density functions.2 1
Name
Probability density function (pdf)
E[X]
var(X)
Uniform
--
1
aS.xS.b
b-a '
--
b+a
2
(b-a) 2
Normal
Exponential
Gamma
Beta
.[i;a
e
12
-(x-µ)2
1
A
p,
ez;r-
'
x
E
(-oo, oo)
-AX
'
A.e-A.x (A-xt-1
r(a)
,
r(a)=
r(a + fi) xa-1 (1- x)p-1
r(a)r(p)
'
r
e-yya-i
O 1- y => y > 112
x + (1- y) > y - x => y < 1I2 + x
(y- x) + (1- y) > x => x < 112
The feasible area is shown in Figure 4.5. The
case for x < y is the left gray triangle. Using
112
symmetry, we can see that the case for x > y is
X
1
Figure 4.5 Distribution of cuts X and Y
the right gray triangle.
22
Hint: Let the first cut point be x, the second one bey, use the figure to show the distribution of x and y.
89
Probability Theory
The total shadowed area represents the region where 3 segments can form a triangle,
which is 1/4 of the square. So the probability is 114.
Property of Poisson process
You are waiting for a bus at a bus station. The buses arrive at the station according to a
Poisson process with an average arrival time of 10 minutes (.A = 0.1 I min). If the buses
have been running for a long time and you arrive at the bus station at a random time,
what is your expected waiting time? On average, how many minutes ago did the last bus
leave?
Solution: Considering the importance of jump-diffusion processes in derivative pricing
and the role of Poisson processes in studying jump processes, let's elaborate more on
exponential random variables and the Poison process. Exponential distribution is widely
used to model the time interval between independent events that happen at a constant
1
. 1 rate) 11.:
average rate (arr1va
/()
t =
{Ae-"'
0
(t 0) . The expecte d amva
. 1 time
.
. 1/ /L1
is
(t < 0)
and the variance is 1IA 2 • Using integration, we can calculate the cdf of an exponential
distribution to be F(t) =P(r s; t) = l-e-'1 and P(r > t) =e_,,, where r is the random
variable for arrival time. One unique property of exponential distribution is
memorylessness: P { r > s + t I r > s} = P( r > t}. 23 That means if we have waited for s
time units, the extra waiting time has the same distribution as the waiting time when we
start at time 0.
When the arrivals of a series of events each independently follow an exponential
distribution with arrival rate A., the number of arrivals between time 0 and t can be
-A.I At'
modeled as a Poisson process P(N(t) = x) = e
, x = 0, 1, · · · 24 The expected
x!
number of arrivals is At and the variance is also At. Because of the memoryless nature
of exponential distribution, the number of arrivals between time s and t is also a Poisson
e-.l(t-s) ( A(t-S)
process
r
x!
Taking advantage of the memoryless property of exponential distribution, we know that
the expected waiting time is 1I A = 10 min. If you look back in time, the memoryless
property stills applies. So on average, the last bus arrived 10 minutes ago as well.
23
P{r > s+t Ir> s} = e-.<(s+i> I e--<.• = e_,., = P(x > t}
24
More rigorously, N(t) is defined as a right-continuous function.
90
A Practical Guide To Quantitative Finance Interviews
This is another example that your intuition may misguide you. You may be wondering
that if the last bus on average arrived 10 minutes ago and the next bus on average will
arrive 10 minutes later, shouldn't the average arrival time be 20 minutes instead of 10?
The explanation to the apparent discrepancy is that when you arrive at a random time,
you are more likely to arrive in a long time interval between two bus arrivals than in a
short one. For example, if one interval between two bus arrivals is 30 minutes and
another is 5 minutes, you are more likely to arrive at a time during that 30-minute
interval rather than 5-minute interval. In fact, if you arrive at a random time, the
E[X 2 ]
expected residual life (the time for the next bus to arrive) is
for a general
2E[X]
distribution. 25
Moments of normal distribution
If X follows standard normal distribution ( X - N (0, 1) ), what is £[ xn] for n = 1, 2, 3
and4?
Solution: The first to fourth moments of the standard normal distribution are essentially
the mean, the variance, the skewness and the kurtosis. So you probably have
remembered that the answers are 0, 1, 0 (no skewness), and 3, respectively.
Standard normal distribution has pdf f(x)
=
e-x212 • Using simple symmetry we
'\/27r
have E[xn] = ( xn
e-x 212 dx = 0 when n is odd. For n = 2, integration by parts are
'\/27r
often used. To solve E[Xn] for any integer n, an approach using moment generating
functions may be a better choice. Moment generating functions are defined as
Le'.. p(x),
M(t) = E[e 1x]
={
if x is discrete
x
( e'x f(x)dx,
if x is continuous
Sequentially taking derivative of M(t), we get one frequently-used property of M(t):
!
=!
M'(t) =
E[e 1x] = E[Xe 1x]
M"(t)
E[Xe'x] = E[X 2e1x]=>M"(O)
25
=> M'(O) = E[X],
= E[X 2 ],
The residual life is explained in Chapter 3 of"Discrete Stochastic Process" by Robert G. Gallager.
91
Probability Theory
and Mn(O) = E[Xn], Vn
l in general.
We can use this property to solve E[Xn] for X - N(O, 1).
distribution M(t)=E[e1x]= (e 1x
(.}z; e- O So Indeed when h =paA, the hedge portfolio has the minimum
Oh 2 S
•
O"s
variance.
Dice game
Suppose that you roll a dice. For each roll, you are paid the face value. If a roll gives 4, 5
or 6, you can roll the dice again. Once you get I, 2 or 3, the game stops. What is the
expected payoff of this game?
Solution: This is an example of the law of total expectation. Clearly your payoff will be
different depending on the outcome of first roll. Let E[X] be your expected payoff and
Y be the outcome of your first throw. You have 1/2 chance to get YE {I, 2, 3}, in which
case the expected value is the expected face value 2, so E[X I YE {I, 2,3}] = 2; you have
94
A Practical Guide To Quantitative Finance Interviews
112 chance to get YE {4,5, 6}, in which case you get expected face value 5 and extra
throw(s). The extra throw(s) essentially means you start the game again and have an
extra expected value E[ X] . So we have E[ X IY E ( 4, 5, 6)] =5 + E[ X]. Apply the law of
total expectation, we have E[X] = E[E[X IY]] =
2 +tx(5+ E[X]) => E[X] = 7 .27
+x
Card game
What is the expected number of cards that need to be turned over in a regular 52-card
deck in order to see the first ace?
Solution: There are 4 aces and 48 other cards. Let's label them as card l, 2, · · ·, 48. Let
if card i is turned over before 4 aces
otherwise
x- { 0,1,
i -
The total number of cards that need to be turned over in order to see the first ace is
x
=
1+
a
a
i=I
i=l
L xi' so we have E[ X] = 1+ LE[ XJ
As shown in the following sequence,
each card i is equally likely to be in one of the five regions separated by 4 aces:
1A2A3A4A5
So the probability that card i appears before all 4 aces is 115, and we have E[ Xi] = 1I5.
48
Therefore, E[X] = 1+LE[ Xi]= 1+48/ 5 = 10.6.
i=I
This is just a special case for random ordering of m ordinary cards and n special cards.
The expected position of the first special card is 1 +
f E[Xi] = 1 + __!!!__
.
n+
i=l
1
Sum of random variables
Assume that XI, X2, .. ·, and xn are independent and identically-distributed (IID)
random variables with uniform distribution between 0 and 1. What is the probability
28
that Sn = X I + X 2 + · · · + X n <
- 1?•
You will also see that the problem can be solved using Wald's equality in Chapter 5.
Hint: start with the simplest case where n =l, 2, and 3. Try to find a general formula and prove it using
induction.
27
28
95
Probability Theory
Solution: This problem is a rather difficult one. The general principle to start with the
simplest cases and try to find a pattern will again help you approach the problem; even
though it may not give you the final answer. When n = 1, P(S1 ::; l) is 1. As shown in
Figure 4.6, when n = 2, the probability that X, + X 2
::;
1 is just the area under
X 1 +X2 ::;1 within the square with side length 1 (a triangle). So P(S2
::;
1) = 1/2. When
n = 3, the probability becomes the tetrahedron ABCD under the plane X 1 + X 2 + X 3 ::; 1
within the cube with side length 1. The volume of tetrahedron ABCD is 1/6. 29 So
P(S3 ::; 1) =I I 6. Now we can guess that the solution is 1 In!. To prove it, let's again
resort
to
induction.
Assume
P(Sn::; 1) = 11 n!.
We
need
to
prove
that
P(Sn+i :s;l)=l/(n+l)!.
A ,
\
\
\
\
\
\
\
\
\
\
\
.
"' ..
\
\
I
\
I
B
'
I
I
,.
Jv·'-<" =::: :::::::_-_-__
0
x,
n=3
Figure 4.6 Probability that Sn s: 1 when n = 2 or n= 3.
Here we can use probability by conditioning. Condition on the value of X n+i, we have
P(Sn+i ::;])= lf(Xn+i)P(Sn ::;I-Xn+JdXn+P where f(Xn+i) is the probability density
function of xn+I' so f ( xn+I) = 1. But how do we calculate P(Sn ::; 1- xn+I)? The cases
of n = 2 and n = 3 have provided us with some clue. For Sn ::; 1-Xn+i instead of Sn ::; 1,
we essentially need to shrink every dimension of the n-dimensional simplex 30 from 1 to
29
You can derive it by integration: f. A(z)dz = f.1 /2z 2dz =I /6, where A(z) is the cross-sectional area.
30
An n-Simplex is the n-dimensional analog of a triangle.
96
A Practical Guide To Quantitative Finance Interviews
1- x n+i · So its volume should be (1- x n+I
n!
r
instead of _!_. Plugging in these results,
n!
11
We have P(sn+I PAn =-J3i7.
For the minimum probability, we have 0.65-.J0.21I2pA 8 = 0.5 => PAn = .J3i7.
In this problem, do not start with P(A or B defaults)=0.65-.J0.2112pA 8 and try to set
p A8
=±1
to calculate the maximum and minimum probability since the correlation
cannot be ±I. The range of correlation is restricted to [
-.J3i7, ..J377] .
4. 6 Order Statistics
LetXbe a random variable with cumulative distribution function Fx(x). We can derive
the distribution function for the minimum I:,
= min( xi' x 2' ... ' x n) and for the maximum
Zn= max(XpX2 ,-··,Xn) of n IID random variables with cdf Fx(x)as
P(I:,
x) = (P(X
x)Y =>I-Frn (x) = (1-Fx(x)Y =>frn (x) = efx (x)(I-Fx(x)y-i
P(Zn
x) = (P(X
x)Y => Fz" (x) = (Fx(x)Y => fz" (x) = nfx(x)(Fx (x)y-i
Expected value of max and min
Let XPX2 ,-··,Xn be 110 random variables with uniform distribution between 0 and 1.
What are the cumulative distribution function, the probability density function and
expected value of Zn= max(XpX2 ,-··,Xn)? What are the cumulative distribution
function, the probability density function and expected value of Yn = min(X1,X2 , • • ·, Xn)?
Solution: This is a direct test of textbook knowledge. For uniform distribution on [O, 1],
Fx(x) = x and fx(x) = 1. Applying Fx(x) and fx(x) to Zn= max(X1,X2 , .. ·,Xn) we
have
P(Zn
x) = (P(X
x)Y => F2" (x) = (Fx(x)f
=> fz" (x) = nfx(x)(Fx(x)y-1
= Xn
= nxn-1
99
Probability Theory
and E[Zn]=
n [x
1xfz (x)dx= 1nxndx=n+l
n
P(Y,,
;?:
n+I ]'
=n
-.
0
n+l
= 1-(1- Fx (x)f =1-(1- xY
=>fr" (x) = nfx (x)(l-Fx (x)f-1 = n(l-xy-'
x) =(P(X;?: x)Y =>Fr" (x)
and E[Y,,]= rnx(l-xf-'dx= rn(l-y)yn-ldx=[yn]I __n_[Yn+i]' =-1- .
0
0
.b
.b
n+l
n+l
Correlation of max and min
Let X 1 and X 2 be IID random variables with uniform distribution between 0 and 1,
Y = min(XpX2 ) and Z
Z $ z for any y, z
E
= max(X"X2 ).
What is the probability of Y;::: y given that
[O, l]? What is the correlation of Y and Z?
Solution: This problem is another demonstration that a figure is worth a thousand words.
As shown in Figure 4.7, the probability that Z z is simply the square with side length
z. So
Since Z=max(X1,X2 ) and Y=min(X1,X2 ), we must have
Y $ Z for any pair of X, and X 2 • So if y > z, P(Y y I Z z) = 0. For y z, that X,
and X 2 satisfies Y;:::y and Z$z is the square with vertices (y,y),(z,y),(z,z), and
(y,z), which has an area (z- y) 2 • So P(Y;::: y n Z
P(Y;::: y I Z $ z) = {
$
z) = (z - y) 2 • Hence
(z - y) 2 I z 2 , if 0 $ z $ 1 and 0 $ y
0,
$
otherwise
Now let's move on to calculate the correlation of Y and Z.
corr(Y,Z) =
100
cov(Y,Z) =
std(Y) xstd(Z)
E[fZ]- E[Y]E[Z]
E[Y ]-E[Y]2 x
J
2
z
A Practical Guide To Quantitative Finance Interviews
I
z
y
0
y
z
I
Figure 4.7 Distribution of X1, X2, their maximum and minimum.
1-=.!., E[Z]=2-=3..
Using previous problem's conclusions, we have E[Y]=2+1 3
2+1 3
From the pdfs of Y and Z, fr. (x) = n(l- xt-1 = 2(1-x) and / 2 (z) = nzn-i = 2z, we can
also get E[Yn2 ] = r' 2(1- y)y 2dy = 3__3_ = _.!._ and E[Z;] = r' 2z3dz = 3.., which give us the
1
3 4 6
1
4
1
2 ( -2 ) 2 = 1
1 ( -1) 2 =-and
variances: var(Y)=E[Y 2 ]-E[Y]2 =-var(Z)=-6
3
18
4
3
18
To calculate E[fZ], we can use E[.YZ] =
!r
33
To solve this equation, we
need f(y,z). Let's again go back to Figure 4.7. From the figure we can see that when
0:::; z:::; 1 and 0:::; y:::; z, F(y,z) is the shadowed area with probability
F(y,z) = P(Y:::; ynZ:::; z) = P(Z:::; z)-P(Y
:.f(y,z) =
;az
F(y,z) = 2 and E[.YZ] =
!r
ynZ:::; z) = z 2 -(z- y) 2 = 2zy- y 2
=
r
=
!
z 3dz =
You may have noticed that var(Y) = var(Z) and wonder whether it is a coincidence for n = 2. It is
actually true for all integer n. You may want to think about why that is true without resorting to
calculation. Hint: var(x) = var(l - x) for any random variable x.
33
101
Probability Theory
An alternative and simpler approach to calculate E[fZ] is again to take advantage of
symmetry. Notice that no matter x1 ::::: x2 or x1 > x 2 , we always have yz = x1x 2
( z = max(x1,x2 ) and y = min(xpx2 ) ).
:.E[fZ]=
11
1 1
2 2
1
4
X1X2dx1dx2 =E[X1]E[X2]=-X-=-.
1
cov(Y Z)
'.J
Hence cov(Y, Z) = E[rz]- E[Y]E[Z] =- and corr(Y, Z) =
36
var(Y) x var(Z)
=
1
2
Sanity check: That Y and Z have positive autocorrelation make sense since when Y
becomes large, Z tends to become large as well ( Z Y ).
Random ants
500 ants are randomly put on a 1-foot string (independent uniform distribution for each
ant between 0 and 1). Each ant randomly moves toward one end of the string (equal
probability to the left or right) at constant speed of 1 foot/minute until it falls off at one
end of the string. Also assume that the size of the ant is infinitely small. When two ants
collide head-on, they both immediately change directions and keep on moving at I
foot/min. What is the expected time for all ants to fall off the string?34
Solution: This problem is often perceived to be a difficult one. The following
components contribute to the complexity of the problem: The ants are randomly located;
each ant can go either direction; an ant needs to change direction when it meets another
ant. To solve the problem, let's tackle these components.
When two ants collide head-on, both immediately change directions. What does it mean?
The following diagram illustrates the key point:
Before collision:
After
switch
When an ant A collides with another ant B, both switch direction. But if we exchange the
ants' labels, it's like that the collision never happens. A continues to move to the right
and B moves to the left. Since the labels are randomly assigned anyway, collisions make
no difference to the result. So we can assume that when two ants meet, each just keeps
on going in its original direction. What about the random direction that each ant chooses?
Once the collision is removed, we can use symmetry to argue that it makes no difference
which direction that an ant goes either. That means if an ant is put at the x-th foot, the
34
Hint: If we switch the label of two ants that collide with each other, it's like that the collision never
happened.
102
A Practical Guide To Quantitative Finance Interviews
expected value for it to fall off is just x min. If it goes in the other direction, simply set x
to 1- x. So the original problem is equivalent to the following:
What is the expected value of the maximum of 500 IID random variables with uniform
distribution between 0 and 1?
Clearly the answer is 499 min, which is the expected time for all ants to fall off the
500
string.
103
Chapter 5 Stochastic Process and Stochastic Calculus
In this chapter, we cover a few topics-Markov chain, random walk and martingale,
dynamic programming-that are often not included in introductory probability courses.
Unlike basic probability theory, these tools may not be considered to be standard
requirements for quantitative researchers/analysts. But a good understanding of these
topics can simplify your answers to many interview problems and give you an edge in
the interview process. Besides, once you learn the basics, you'll find many interview
problems turning into fun-to-solve math puzzles.
5. 1 Markov Chain
A Markov chain is a sequence of random variables X 0 ,X1 ,. • .,xn, ... with the Markov
property that given the present state, the future states and the past states are independent:
P{Xn+i=jlXn=i,Xn_1 =in_1,···,X0 =i0 }=pij=P{Xn+i=JIXn=i} for all n, i0 ,
in-1' i, and j, where i, j
E
{l, 2, .. ., M} represent the state space S = {s" s 2 ,
.. .,
···,
sM} of
x.
In other words, once the current state is known, past history has no bearing on the future.
For a homogenous Markov chain, the transition probability from state i to state j does
not depend on n. 1 A Markov chain with M states can be completely described by an
M x M transition matrix P and the initial probabilities P(X0 ) •
Transition matrix: P={Pu}=
P11
P1 2
P1M
P 21
P22
P2M
'
PM1
PM2
where pij
IS
the transition
PMM
probability from state i to state j.
M
Initialprobabilities: P(X0 )=(P(X0 =l), P(X0 =2), .. ., P(X0 =M)), LP(X0 =i)=l.
i:I
The probability of a path: P(X, = i,' X2 = i2 ... ' xn =in I Xo = io) = pioi1 pi1i2 .. . P;n_lin
Transition graph: A transition graph is often used to express the transition matrix
graphically. The transition graph is more intuitive than the matrix, and it emphasizes
1 In this chapter, we only consider finite-state homogenous Markov chains (i.e., transition probabilities do
not change over time).
Stochastic Process and Stochastic Calculus
possible and impossible transitions. Figure 5.1 shows the transition graph and the
transition matrix of a Markov chain with four states:
0.5
1
o.s
[f
;
<
os
0.25
>
P=
0.4
2
3
4
'-·
0
0.5
0
0.5
l
0.5
0
0
0
II
0.25 0.25
2
0.4
0.4
0.2
3
0
0
1
4
0.4
Figure 5.1 Transition graph and transition matrix of the Play
Classification of states
State j is accessible from state i if there is a directed path in the transition graph from i to
j (::In such that Pt>> 0). Let TiJ = min(n :Xn = j I X 0 = i), then P(TiJ < oo) > 0) if and
only if state j is accessible from state i. States i and j communicate if i is accessible
from} and} is accessible from i. In Figure 5.1, state 3 and 1 communicate. State 4 is
accessible form state l, but they do not communicate since state 1 is not accessible from
state 4.
We say that state i is recurrent if for every state j that is accessible from i, i is also
accessible from j (VJ, P(TiJ < oo) > 0 => P(TiJ < oo) = 1). A state is called transient if it is
not recurrent ( ::3}, P(TiJ < oo) > 0 and P(I';j < oo) < 1 ).
In Figure 5.1, only state 4 is
recurrent. States 1, 2 and 3 are all transient since 4 is accessible from 1/2/3, but 1/2/3 are
not accessible from 4.
Absorbing Markov chains: A state i is called absorbing if it is impossible to leave this
state ( P;; =l, piJ = 0, VJ * i ). A Markov chain is absorbing if it has at least one absorbing
state and if from every state it is possible to go to an absorbing state. In Figure 5. I, state
4 is an absorbing state. The corresponding Markov chain is an absorbing Markov chain.
Equations for absorption probability: The probability to reach a specific absorbing
state s, a1,. • ·, aM, are unique solutions to equations a_, =I, a; = 0 for all absorbing
M
state(s) i * s, and a;= Iajpii for all transient states i. These equations can be easily
106
A Practical Guide To Quantitative Finance Interviews
derived using the law of total probability by conditioning the absorption probabilities on
the next state.
Equations for the expected time to absorption: The expected times to absorption,
µ 1,. • ·, µ M, are unique solutions to the equations µ; = 0 for all absorbing state( s) i and
m
µ;
= 1+I, pijµj for all transient states i. These equations can be easily derived using the
j=l
law of total expectation by conditioning the expected times to absorption on the next
state. The number 1 is added since it takes one step to reach the next state.
Gambler's ruin problem
Player M has $1 and player N has $2. Each game gives the winner $1 from the other. As
a better player, M wins 2/3 of the games. They play until one of them is bankrupt. What
is the probability that M wins?
Solution: The most difficult part of Markov chain problems often lies in how to choose
the right state space and define the transition probabilities Py 's, Vi, j. This problem has
fairly straightforward states. You can define the state space as the combination of the
money that player M has ($m) and the money that player N has ($n):
{(m,n)} = {(3,0),(2,1),(1,2),(0,3)}. (Neither m nor n can be negative since the whole
game stops when one of them goes bankrupt.) Since the sum of the dollars of both
players is always $3, we can actually simplify the state space using only m:
{m} = {0,1,2,3}.
I
The transition graph and the corresponding transition matrix are shown in Figure 5.2.
I
1/3
2/3
l
l/3
r£J
l
2
2/3
3
Po.o
P={Jt}= [ P1.o
P2.o
P2.o
Po.1
P1.1
P2.1
P2.1
Po.2
P1.2
P2.2
P2.2
Po,31
l
I
P1,J = K
P2.J
0
P2.J
0
0
0
1
K
0
0
0
x
l
Figure 5.2 Transition matrix and transition graph for Gambler's ruin problem
The initial state is X 0 = 1 (M has $1 at the beginning). At state 1, the next state is 0 (M
loses a game) with probability 1/3 and 2 (M wins a game) with probability 2/3. So
Pi.o =113 and p 1,2 =213. Similarly we can get p 2 ,1 =113 and p 2 ,3 =213. Both state 3
(Mwins the whole game) and state 0 (Mloses the whole game) are absorbing states.
To calculate the probability that M reaches absorbing state 3, we can apply absorption
probability equations:
107
Stochastic Process and Stochastic Calculus
3
3
a3 = 1, a 0 = 0 , and a 1 = """'
£..J p, '1.a 1., a 2 = """'
£..J p 2 ,.f.a .I.
j=O
j=O
Plugging in the transition probabilities using either the transition graph or transition
{ a 1 = 417
.
a 1 = 1/ 3 x 0 + 2 I 3 x a 2 }
matnx, we have
a 2 = 1/ 3 x a 1 + 2 I 3 x 1
a 2 = 617
So, starting from $1, player M has 417 probability of winning.
Dice question
Two players bet on roll(s) of the total of two standard six-face dice. Player A bets that a
sum of 12 will occur first. Player B bets that two consecutive 7s will occur first. The
players keep rolling the dice and record the sums until one player wins. What is the
probability that A will win?
Solution: Many of the simple Markov chain problems can be solved using pure
conditional probability argument. It is not surprising considering that Markov chain is
defined as conditional probability:
P{Xn+1 = 1· 1 X n
= i' X n-1 = in-1' • • • ' X 0 = i0 } = p y.. = P{Xn+1 = 1· 1 X n = i} ·
So let's first solve the problem using conditional probability arguments. Let P(A) be the
probability that A wins. Conditioning P(A) on the first throw's sum F, which has three
possible outcomes F = 12, F = 7 and F {7, 12}, we have
P(A) = P(A IF= 12)P(F = 12) + P(A IF= 7)P(F = 7)+ P(A
{7,12})P(F {7, 12})
Then we tackle each component on the right hand side. Using simple permutation, we
Also it
can easily see that P(F=12)=1136, P(F=7)=6136,
is obvious that P(AIF=12)=1 and
(The game essentially
starts over again.) To calculate P(A IF= 7), we need to further condition on the second
throw's total, which again has three possible outcomes: E = 12, E = 7, and E {7,12}.
P(A IF= 7) = P(A IF= 7,E = 12)P(E = 12 IF= 7)+P(A IF= 7,E = 7)P(E = 71F=7)
+ P(A IF= 7, E
{7,12} )P(E
{7, 12} IF= 7)
= P(A IF= 7,E = 12)x1/36+ P(A IF =7,E = 7)x6/36
= 1x1I36+0x 6136 + P(A)x 29 I 36 = l /36+ 29 I 36P(A)
Here the second equation relies on the independence between the second and the first
rolls. If F = 7 and E = 12, A wins; if F = 7 and E = 7, A loses; if F = 7 and
108
A Practical Guide To Quantitative Finance Interviews
E !l {7,12}, the game essentially starts over again. Now we have all the necessarily
information for P( A) . Plugging it into the original equation, we have
P(A) = P(A IF= I 2)P(F = 12) + P(A IF= 7)P(F = 7) + P(A IF !l {7, 12} )P(F !l {7, 12})
=Ix 1/36 + 6136x (1/36 + 29 /36P(A)) + 29 /36P(A)
Solving the equation, we get P( A) = 7I13 .
This approach, although logically solid, is not intuitively appealing. Now let's try a
Markov chain approach. Again the key part is to choose the right state space and define
the transition probabilities. It is apparent that we have two absorbing states, 12 (A wins)
and 7-7 (B wins), at least two transient states, S (starting state) and 7 (one 7 occurs, yet
no 12 or 7-7 occurred). Do we need any other states? Theoretically, you can have other
states. In fact, you can use all combination of the outcomes of one roll and two
consecutive rolls as states to construct a transition matrix and you will get the same final
result. Nevertheless, we want to consolidate as many equivalent states as possible. As
we just discussed in the conditional probability approach, if no 12 has occurred and the
most recent roll did not yield 7, we essentially go back to the initial starting state S. So
all we need are states S, 7, 7-7 and 12. The transition graph and probability to reach state
12 are shown in Figure 5.3.
6/36
7
cfil)
1
Probability to absorption state 12
a12 =I, a1-1 = 0
}
as= l/36 x I +6/36 x a7 + 29/36x as
a7
7 /13
=l / 36xl+6/36x0+29/36xas
Figure 5.3 Transition graph and probability to absorption for dice rolls
Here the transition probability is again derived from conditional probability arguments.
Yet the transition graph makes the process crystal clear.
Coin triplets
Part A. If you keep on tossing a fair coin, what is the expected number of tosses such
that you can have HHH (heads heads heads) in a row? What is the expected number of
tosses to have THH (tails heads heads) in a row?
Solution: The most difficult part of Markov chain is, again, to choose the right state
space. For the HHH sequence, the state space is straightforward. We only need four
states: S (for the starting state when no coin is tossed or whenever a T turns up before
HHH), H, HH, and HHH. The transition graph is
109
Stochastic Process and Stochastic Calculus
At state S, after a coin toss, the state will stay at S when the toss gives a T. If the toss
gives an H, the state becomes H. At state H, it has 1/2 probability goes back to state S if
the next toss is T; otherwise, it goes to state HH. At state HH, it also has 112 probability
goes back to state S if the next toss is T; otherwise, it reaches the absorbing state HHH.
So we have the following transition probabilities: Ps.s
PH.HH =
+,
PHH,s
= +,
PHH,HHH =
+, and PHHH ,HHH = 1.
= t,
Ps.H
= t,
PH .s
= t,
We are interested in the expected number of tosses to get HHH, which is the expected
time to absorption starting from state S. Applying the standard equations for the
expected time to absorption, we have
µs=l+± µs+tµ11
µH =I +!µs +t µHH
µHH
= l+tµs +t µHHll
µHHH =0
So from the starting state, the expected number of tosses to get HHH is 14.
Similarly for expected time to reach THH, we can construct the following transition
graph and estimate the corresponding expected time to absorption:
112
An
1/21\)
\__W
µs
1/2
µr
= 1+I µs +I µr
= 1+fµr +f µrH
= 1+ I µT + I µTHI/
µTHH = 0
µ711
µs =8
µT =4
=2
µ71111 = 0
µTH
So from the starting state S, the expected number of tosses to get THH is 8.
Part B. Keep flipping a fair coin until either HHH or THH occurs in the sequence. What
is the probability that you get an HHH subsequence before THH? 2
Hint: This problem does not require the drawing of a Markov chain. Just think about the relationship
between an HHH pattern and a THH pattern. How can we get an HHH sequence before a THH sequence?
2
110
A Practical Guide To Quantitative Finance Interviews
Solution: Let's try a standard Markov chain approach. Again the focus is on choosing
the right state space. In this case, we begin with starting state S. We only need ordered
subsequences of either HHH or THH. After one coin is flipped, we have either state Tor
H. After two flips, we have states TH and HH. We do not need TT (which is equivalent
to T for this problem) or HT (which is also equivalent to T as well). For three coin
sequences, we only need THH and HHH states, which are both absorbing states. Using
these states, we can build the following transition graph:
1/2
in
112
Cf)
({)
A(\•
@J
®
112
Figure 5.4 Transition graph of coin tosses to reach HHH or THH
We want to get the probability to reach absorbing state HHH from the starting state S.
Applying the equations for absorption probability, we have
aHHH = 1, arHH
=0
ar =0,aTH =0
as =tar +taH
- I
as-8
ar =tar +taTH, aH =tar +faHH
aH
aTH =tar +taTHH•aHH =tar +taHHH
- I
aHH -2
= .l4
So the probability that we end up with the HHH pattern is 1/8.
This problem actually has a special feature that renders the calculation unnecessary. You
may have noticed that ar = 0. Once a tail occurs, we will always get THH before HHH.
The reason is that the last two coins in THH is HH, which is the first two coins in
sequence HHH. In fact, the only way that the sequence reaches state HHH before THH
is that we get three consecutive Hs in the beginning. Otherwise, we always have a T
before the first HH sequence and always end in THH first. So if we don't start the coin
flipping sequence with HHH, which has a probability of 118, we will always have THH
beforeHHH.
Part C. (Difficult) Let's add more fun to the triplet game. Instead of fixed triplets for the
two players, the new game allows both to choose their own triplets. Player 1 chooses a
triplet first and announces it; then player 2 chooses a different triplet. The players again
toss the coins until one of the two triplet sequences appears. The player whose chosen
triplet appears first wins the game.
111
Stochastic Process and Stochastic Calculus
If both player 1 and player 2 are perfectly rational and both want to maximize their
probability of winning, would you go first (as player 1)? If you go second, what is your
probability of winning?3
Solution: A common misconception is that there is always a best sequence that beats
other sequences. This misconception is often founded on a wrong assumption that these
sequences are transitive: if sequence A has a higher probability occurring before
sequence B and sequence B has a higher probability occurring before sequence C, then
sequence A has a higher probability occurring before sequence C. In reality, such
transitivity does not exist for this game. No matter what sequence player 1 chooses,
player 2 can always choose another sequence with more than l/2 probability of winning.
The key, as we have indicated in Part B, is to choose the last two coins of the sequence
as the first two coins of player l's sequence. We can compile the following table for
each pair of sequences:
2's winning
Probability HHH THH
Plaver 1
HTH HHT TTH
THT
HTT
TTT
l/8
215
l/2
3/10
5/12
215
112
HHH
I
THH
7/8
v
1/2
3/4
l/3
1/2
l/2
315
HTH
315
l/2
11
l/3
3/8
l/2
l/2
7/12
"" HHT
1/2
1/4
2/3
11
1/2
5/8
2/3
7110
=::=TTH
7/10
2/3
5/8
l/2
v
2/3
l/4
112
TUT
7/12
112
1/2
3/8
1/3
I
l/2
315
HTT
315
l/2
1/2
l/3
3/4
1/2
11
7/8
TTT
1/2
215
5112
3/10
l/2
215
1/8
I
M
;;...
Table 5.1 Player 2's winning probability with different coin sequence pairs
As shown in Table 5.1 (you can confirm the results yourself), no matter what player l's
choices are, player 2 can always choose a sequence to have better odds of winning. The
best sequences that player 2 can choose in response to 1's choices are highlighted in
bold. In order to maximize his odds of winning, player I should choose among HTH,
HTT, THH and THT. Even in these cases, player 2 has 2/3 probability of winning.
3 This problem is a difficult one. Interested reader may find the following paper helpful: "Waiting Time
and Expected Wailing Time-Paradoxical Situations" by V. C. Hombas, The American Statistician, Vol. 51,
No. 2 (May, 1997). pp. 130-133. In this section, we will only discuss the intuition.
112
A Practical Guide To Quantitative Finance Interviews
Color balls
A box contains n balls of n different colors. Each time, you randomly select a pair of
balls, repaint the first to match the second, and put the pair back into the box. What is
the expected number of steps until all balls in the box are of the same color? (Very
difficult)
Solution: Let Nn be the number of steps needed to make all balls the same color, and let
F;, i =1, 2, · · ·, n, be the event that all balls have color i in the end. Applying the law of
total expectation, we have
Since all the colors are symmetric (i.e., they should have equivalent properties), we have
P[F;] = PIF2J = ... = P[Fn] = l In and E[Nn] = E[Nn IF;]= E[Nn I F2] = E[Nn I Fn]. That
means we can assume that all the balls have color l in the end and use E[Nn IF;] to
represent E[Nn].
So how do we calculate E[Nn IF;]? Not surprisingly, use a Markov chain. Since we
only consider event F;, color l is different from other colors and colors 2, · · ·, n
become equivalent. In other words, any pairs of balls that have no color l ball involved
are equivalent and any pairs with a color l ball and a ball of another color are equivalent
if the order is the same as well. So we only need to use the number of balls that have
color l as the states. Figure 5.5 shows the transition graph.
Figure 5.5 Transition graph for all n balls to become color 1
Staten is the only absorbing state. Notice that there is no state 0, otherwise it will never
reach F; . In fact, all the transition probability is conditioned on F; as well, which makes
the transition probability
I F; higher than the unconditional probability
P;,;+i
and
For example, Pi,o IF; = 0 and Pi,o = 1/ n. (Without
conditioning, each ball is likely to be the second ball, so color l has l In probability of
being the second ball.) Using the conditional transition probability, the problem
essentially becomes expected time to absorption with system equations:
P;,;-i
IF; lower than
P;,;+i
P;,;-i ·
E[N; IF;]= l + E[N;-1 IF;] x I';,;-1 IF; +E[N; I F;]x P;,; IF;+ E[N;+1 IF;]x I';,;+1 IF;·
113
Stochastic Process and Stochastic Calculus
IF;, let's rewrite the probability as P(xk+i = i -11 xk = i, F; ),
Vk = 0, 1, .. ,, to make the derivation step clearer:
To calculate
P( x
k+1
P;,;-i
. II x =l'F)
=1k
' i
P(xk=i,xk+ 1 =i-1,F;)
P(xk = i,F;)
P(F;
lxk+I
=i-1,xk =i)xP(xk+I =i-llxk =i)xP(xk =i)
P(F; Ixk
_ P(F;
lxk+I
= i) x P(xk = i)
=i-l)xP(xk+I =i-llxk =i)
P(F; Ixk
= i)
i-1 i(n-i)
x---= n n(n-l)_(n-i)x(i-1)
i/n
n(n-1)
The first equation is simply the definition of conditional probability; the second equation
is the application of Bayes' theorem; the third equation applies the Markov property. To
derive P(F; Ixk = i) = i In, we again need to use symmetry. We have shown that if all
the balls have different colors, then we have P[ F;] = P[ F2 ] = .. · = P[ Fn] = I/ n. What is
the probability of ending in a given color, labeled as c, if i of the balls are of color c? It
is simply i In. To see that, we can label the color of each of the i balls of color c as
cJ, j = 1, .. ·,i (even though they are in fact the same color). Now it's obvious that all
balls will end with color cJ with probability 1In. The probability for c is the sum of
probabilities of c/s, which gives the result i In.
Similarly we have P(F; Ixk+I = i-1) = (i-1)/ n. For P(xk+I = i-1 I xk = i), we use a
basic counting method. There are n(n -1) possible permutations to choose 2 balls out of
n balls. In order for one color 1 ball to change color, the second ball must be color 1,
which has i choices; the first ball needs to be another color, which has (n-i) choices.
So P( xk+I
= z. - l I xk = z.) = i(n -i) .
n(n -1)
Applying the same principles, we can get
.
.
(n-i)x2i
.
.
(n-i)x(i+l)
P(xk+i =1lxk =1,F;)=
, P(xk+i =1+llxk =1,F;)=
.
n(n-1)
n(n-1)
Plugging into E[ N; IF;] and simplifying E[ N; I F;] as Z;, we have
(n-i) x 2ix Z;
114
= n(n-1) + (n-i)(i + l)Z;+i + (n-i)(i- l)Z;_1 •
A Practical Guide To Quantitative Finance Interviews
Using these recursive system equations and the boundary condition Zn= 0, we can get
Z1=(n-l)2.4
5.2 Martingale and Random walk
Random walk: The process {Sn; n
1} is called a random walk if {X;;i
1} are IID
(identical and independently distributed) random variables and Sn= X, +···Xn, where
n = 1, 2, · · · The term comes from the fact that we can think of Sn as the position at time
n for a walker who makes successive random steps X,, X 2 , • • •
If X; takes values 1 and -1 with probabilities p and 1 - p respectively, Sn is called a
simple random walk with parameter p. Furthermore, if p = f, the process Sn ts a
symmetric random walk. For symmetric random walk, it's easy to show that
E[Sn] = 0 and var(Sn) = E[S;]-E[Sn] 2 = E[S;] = n .5
Symmetric random walk is the process that is most often tested in quantitative
interviews. The interview questions on random walk often revolve around finding the
first n for which Sn reaches a defined threshold a, or the probability that Sn reaches
a for any given value of n .
Martingale: a martingale { Zn;n
1} is a stochastic process with the properties that
E[I Zn I]< oo for all n and E[Zn+I IZn= zn,Zn-t = zn_ 1 ,-··,Z1 = z1 ] = zn. The property of a
martingale can be extended to E[ Zm; m > n I Zn= zn,Zn-I = zn-P · · ·,Z, = z1] = zn, which
means the conditional expected value of future Zm is the current value Zn. 6
A symmetric random walk is a martingale. From the definition of the symmetric random
walk we have Sn+i = {
Sn + 1 with probability 112
..
, so E[Sn+i ISn =sn,···,S1 =s1 ]=sn.
sn -1 wt.th probab1hty
1/2
Since E[S:+i -(n + 1)] = t[(Sn + 1)2 +(Sn -1)2]-(n + 1) = s;-n,
as well.
s; - n
is a martingale
Even this step is not straightforward. You need to plug in the ts and try a few cases starting with
i = n - I. The pattern will emerge and you can see that all the terms containing z._,,
Z2 cancel out.
4
z._,. ·· ·,
5 Induction
again can be used for its proof. Var(S,) = Var(Z,) =I. Induction step: If Var(S.) = n, then we
have Var(S.,,) = Var(S. + x.,,) = Var(S.) + Var(x.,,) = n +I since x•• , is independent of Sn.
Do not confuse a martingale process with a Markov process. A martingale does not need to be a Markov
process; a Markov process does not need to be a martingale process, either.
6
115
Stochastic Process and Stochastic Calculus
Stopping rule: For an experiment with a set of IID random variables X 1,X2 ,-··, a
stopping rule for {X;;i
1} is a positive integer-value random variable N (stopping time)
such that for each n > l, the event {N :5: n} is independent of Xn+i•Xn+ 2 ,··· . Basically it
says that whether to stop at n depends only on X 1,X2 ,·· ·,Xn (i.e., no look ahead).
Wald's Equality: Let N be a stopping rule for IID random variables XPX 2 ,-·· and let
SN = X, + X 2 + · · · + X N, then E[ SN] = E[ X]E[ N].
Since it is an important-yet relatively little known-theorem, let's briefly review its
proof. Let In be the indicator function of the event {N n}. So SN can be written as
SN= LX)n, where In =1 if
and In =0 if N:5:n-1.
n=I
From the definition of stopping rules, we know that In is independent of Xn, Xn+i • · · ·
(it only depends on XI' X 2 ,
···,
Xn_ 1 ). So E[X)n]=E[Xn]E[In]=E[X]E[In] and
E[SN l = E[t,x.1"] = t.E[ X/"] = t.E[ X]E[I.] = E[ xJt,E[ /"] = E[X]E[N] .7
A martingale stopped at a stopping time is a martingale.
Drunk man
A drunk man is at the 17th meter of a 100-meter-long bridge. He has a 50% probability
of staggering forward or backward one meter each step. What is the probability that he
will make it to the end of the bridge (the lOOth meter) before the beginning (the 0th
meter)? What is the expected number of steps he takes to reach either the beginning or
the end of the bridge?
Solution: The probability part of the problem-often appearing in different disguises-is
among the most popular martingale problems asked by quantitative interviewers.
Interestingly, few people use a clear-cut martingale argument. Most candidates either
use Markov chain with two absorbing states or treat it as a special version of the
gambler's ruin problem with p = 1I2. These approaches yield the correct results in the
end, yet a martingale argument is not only simpler but also illustrates the insight behind
the problem.
7 For detailed proof and applications of Wald's Equality, please refer to the book Disc:rete Stochastic
Processes by Robert G. Gallager.
116
A Practical Guide To Quantitative Finance Interviews
Let's set the current pos1t10n (the 17th meter) to O; then the problem becomes a
symmetric random walk that stops at either 83 or -17. We also know that both Sn and
- n are martingales. Since a martingale stopped at a stopping time is a martingale,
s
and
- N (where sN = xi + x 2 + ... + x N with N being the stopping time) are
martingales as well. Let pa be the probability that it stops at a = 83, p /3 be the
N
probability it stops at
-/3 = -17
( p /3
=1- pa),
and N be the stopping time. Then we
have
E[ SN] = pa X 83- (1- pa) X 17 = S 0
=0
-N] = E[pa x 83 2 +(1- Pa) x 17 2 ]-E[N] =
}
sg -0 = 0
{ p
=>
= 0.17
= 1441
Hence, the probability that he will make it to the end of the bridge (the 1OOth meter)
before reaching the beginning is 0.17, and the expected number of steps he takes to
reach either the beginning or the end of the bridge is 1441.
We can easily extend the solution to a general case: a symmetric random walk starting
from 0 that stops at either a (a > 0) or -/3 ( p > 0 ). The probability that it stops at a
instead of -/3 is Pa = f3 /(a+ jJ). The expected stopping time to reach either a or -/3
is
E[ N] = ajJ.
Dice game
Suppose that you roll a dice. For each roll, you are paid the face value. If a roll gives 4, 5
or 6, you can roll the dice again. If you get 1, 2 or 3, the game stops. What is the
expected payoff of this game?
Solution: In Chapter 4, we used the law of total expectation to solve the problem. A
simpler approach-requiring more knowledge-is to apply Wald's Equality since the
problem has clear stopping rules. For each roll, the process has 1/2 probability of
stopping. So the stopping time N follows a geometric distribution with p = 1I2 and we
have E[N] =II p = 2. For each roll, the expected face value is E[X] = 7 I 2. The total
expected payoff is E[SN] = E[X]E[N] = 7 /2x 2 = 7.
Ticket line
At a theater ticket office, 2n people are waiting to buy tickets. n of them have only $5
bills and the other n people have only $10 bills. The ticket seller has no change to start
117
Stochastic Process and Stochastic Calculus
with. If each person buys one $5 ticket, what is the probability that all people will be
able to buy their tickets without having to change positions?
Solution: This problem is often considered to be a difficult one. Although many can
correctly formulate the problem, few can solve the problem using the reflection
principle. 8 This problem is one of the many cases where a broad knowledge makes a
difference.
Assign + 1 to the n people with $5 bills and -1 to the n people with $10 bills. Consider
the process as a walk. Let (a,b) represent that after a steps, the walk ends at b. So we
start at (0,0) and reaches (2n,O)after 2n steps. For these 2n steps, we need to choose n
steps as +l, so there are ( 2nJ = 2 n! possible paths. We are interested in the paths that
n
n!n!
have the property b 0, VO< a< 2n steps. It's easier to calculate the number of
complement paths that reach b = -1, 30 O, ),
S,
=
O;
II
= h.
Denote N n (a,b) as the
as the number possible paths from (0,a) to
in other words, N:'(a,b) are the paths that contain
( k, 0), :30 < k < n. The reflection principle says that if a, h > 0, then
(a, h) = N,, (-a, h ).
The proof is
intuitive: for each path (0, a) to (k, 0), there is a one-to-one corresponding path from (0, -a) to (k, 0).
118
A Practical Guide To Quantitative Finance Interviews
b
-1 ------------------2
Figure 5.6 Reflected paths: the dashed line is the reflection of the solid line
after it reaches -1
Coin sequence
Assume that you have a fair coin. What is the expected number of coin tosses to get n
heads in a row?
Solution: Let E[f(n)] be the expected number of coin tosses to get n heads in a row. In
the Markov chain section, we discussed the case where n = 3 (to get the pattern HHH).
For any integer n, we can consider an induction approach. Using the Markov chain
approach, we can easy get that E[/(l)] = 2, E[/(2)] = 6 and E[/(3)] = 14. A natural
guess for the general formula is that £[/(n)] = 2n+I -2. As always, let's prove the
formula using induction. We have shown the formula is true for n =1,2,3. So we only
need to prove that if E[f(n)] = 2n+I -2, E[f(n + 1)] = 2n+ 2 -2. The following diagram
shows how to prove that the equation holds for E[f(n + 1)]:
__ _.
P=l/2
P=l/2
The state before (n + 1) heads in a row (denoted as (n + l)H) must be n heads in a row
(denoted as nH ). It takes an expected E[f ( n)] = 2n+I - 2 tosses to reach nH.
Conditioned on state nH, there is a 1/2 probability it will go to (n+l)H (the new toss
yields H) and the process stops. There is also a 1/2 probability that it will go to the
119
Stochastic Process and Stochastic Calculus
starting state 0 (the new toss yields 1) and we need another expected E[f(n + 1)] tosses
to reach (n+ l)H. So we have
E[f(n + 1)] = E[F(n)] +txl +
+x E[f(n+ 1)]
E[/(n+l)] = 2x E[F(n)]+ 2 =
r+ 2 -2
General Martingale approach: Let's use HH · · · Hn to explain a general approach for the
expected time to get any coin sequence by exploring the stopping times of martingales. 9
Imagine a gambler has $1 to bet on a sequence of n heads ( HH · · · Hn ) in a fair game
with the following rule: Bets are placed on up to n consecutive games (tosses) and each
time the gambler bets all his money (unless he goes bankrupt). For example, if H
appears at the first game, he will have $2 and he will put all $2 into the second game. He
stops playing either when he loses a game or when he wins n games in a roll, in which
case he collects $2n (with probability I I 2n ). Now let's imagine, instead of one gambler,
before each toss a new gambler joins the game and bets on the same sequence of n heads
with a bankroll of $1 as well. After the i-th game, i gamblers have participated in the
game and the total amount of money they have put in the game should be $i. Since each
game is fair, the expected value of their total bankroll is $i as well. In other words, if we
denote x; as the amount of money all the participating gamblers have after the i-th game,
then (x; -i) is a martingale.
Now, let's add a stopping rule: the whole game will stop if one of the gamblers becomes
the first to get n heads in a roll. A martingale stopped at a stopping time is a martingale.
So we still have E[(x; -i)] = 0. If the sequence stops after the i-th toss ( i n ), the
(i- n + 1)-th player is the (first) player who gets n heads in a roll with payoff 2n. So all
the (i - n) players before him went bankrupt; the {i- n + 2) -th player gets ( n -1) heads
in a roll with payoff 2n-J ; ... ; the i-th player gets one head with payoff 2. So the total
payoff is fixed and X; = 2n + 2n-I + · · · + i = r+I - 2 .
Hence, E[(x; - i)] = 2n+I
-
2- E[i] =0
E[i] = 2n+I -2.
This approach can be applied to any coin sequences-as well as dice sequences or any
sequences with arbitrary number of elements. For example, let's consider the sequence
HHTTHH. We can again use a stopped martingale process for sequence HHTTHH. The
gamblers join the game one by one before each toss to bet on the same sequence
HHTTHH until one gambler becomes the first to get the sequence HHITHH. If the
sequence stops after the i-th toss, the {i- 5)th gambler gets the HHTTHH with payoff
9 If you prefer more details about the approach, please refer to "A Martingale Approach to the Study of
Occurrence of Sequence Patterns in Repeated Experiments" by Shuo-Yen Robert Li, The Annals of
Probability, Vol. 8, No. 6 (Dec., 1980), pp. 1171-1176.
120
A Practical Guide To Quantitative Finance Interviews
26 . All the (i - 6) players before him went bankrupt; the (i -4)th player loses in the
second toss (HT); the (i-3)th player and the (i - 2)th player lose in the first toss (D;
the (i - l)th player gets sequence HH with payoff 22 and the i-th player gets H with
payoff 2.
Hence, E[(x; -i)] = 26 + 22 + 2 1 - E[i] = 0
E[i] = 70.
5.3 Dynamic Programming
Dynamic Programming refers to a collection of general methods developed to solve
sequential, or multi-stage, decision problems. 10 It is an extremely versatile tool with
applications in fields such as finance, supply chain management and airline scheduling.
Although theoretically simple, mastering dynamic programming algorithms requires
extensive mathematical prerequisites and rigorous logic. As a result, it is often perceived
to be one of the most difficult graduate level courses.
Fortunately, the dynamic programming problems you are likely to encounter in
interviews-although you often may not recognize them as such-are rudimentary
problems. So in this section we will focus on the basic logic used in dynamic
programming and apply it to several interview problems. Hopefully the solutions to
these examples will convey the gist and the power of dynamic programming.
A discrete-time dynamic programming model includes two inherent components:
1. The underlying discrete-time dynamic system
A dynamic programming problem can always be divided into stages with a decision
required at each stage. Each stage has a number of states associated with it. The decision
at one stage transforms the current state into a state in the next stage (at some stages and
states, the decision may be trivial if there is only one choice).
Assume that the problem has N + 1stages (time periods). Following the convention, we
label these stages as 0, 1, · · ·, N -1, N. At any stage k, 0 k N -1, the state transition
can be expressed as xk+i = f (xk, uk, wk), where xk is the state of system at stage k; 11 uk
is the decision selected at stage k; w* is a random parameter (also called disturbance).
This section barely scratches the surface of dynamic programming. For up-to-date dynamic
programming topics, I'd recommend the book Dynamic Programming and Optimal Control by Professor
Dimitri P. Bertsekas.
11 In general, xk can incorporate all past relevant information. In our discussion, we only consider the
present information by assuming Markov property.
'0
121
Stochastic Process and Stochastic Calculus
Basically the state of next stage xk+i is determined as a function of the current state xk,
current decision uk (the choice we make at stage k from the available options) and the
random variable wk (the probability distribution of wk often depends on xk and uk ).
2. A cost (or profit) function that is additive over time.
Except for the last stage (N), which has a cost/profit g N( x N) depending only on x N, the
costs at all other stages gk(xk,uk, wk) can depend on xk, uk, and wk. So the total
N-1
cost/profit is gN(xN)+ Lgk(xk,uk,wk)}.
The goal of optimization is to select strategies/policies for the decision sequences
tr*= {u0 *,-··,uN-i *} that minimize expected cost (or maximize expected profit):
N-1
J,...(x0)=minE{gN(xN)+ Lgk(xk'uk'wk)}.
,..
k=O
Dynamic programming (DP) algorithm
The dynamic programming algorithm relies on an idea called the Principle of
Optimality: If tr*= {u0 *,-··,uN _1*} is the optimal policy for the original dynamic
programming problem, then the tail policy tr;*= {u; *,-· ·,uN-i *} must be optimal for the
NI
tail subproblem E {g N(xN) + L gk (xk, uk, wk)} .
k-i
NI
DP algorithm: To solve the basic problem J,...(x0 )=minE{gN(xN)+ Lgk(xk,uk,wk)},
,..
k-0
start with J N(xN) = gN(xN), and go backwards minimizing cost-to-go function Jk(xk):
Jk(xk)= min E{gk(xk,uk , wk)+Jk+i(f(xk'uk,wk))},k=O,-··,N-1. Then the J 0 (x0 )
11,cU,(.r,) ..-,
generated from this algorithm is the expected optimal cost.
Although the algorithm looks complicated, the intuition is straightforward. For dynamic
programming problems, we should start with optimal policy for every possible state of
the final stage (which has the highest amount of information and least amount of
uncertainty) first and then work backward towards earlier stages by applying the tail
policies and cost-to-go functions until you reach the initial stage.
Now let's use several examples to show how the DP algorithm is applied.
122
A Practical Guide To Quantitative Finance Interviews
Dice game
You can roll a 6-side dice up to 3 times. After the first or the second roll, if you get a
number x, you can decide either to get x dollars or to choose to continue rolling. But
once you decide to continue, you forgo the number you just rolled. If you get to the third
roll, you'll just get x dollars if the third number is x and the game stops. What is the
game worth and what is your strategy?
Solution: This is a simple dynamic programming strategy game. As all dynamic
programming questions, the key is to start with the final stage and work backwards. For
this question, it is the stage where you have forgone the first two rolls. It becomes a
simple dice game with one roll. Face values 1, 2, 3, 4, 5, and 6 each have a 1/6
probability and your expected payoff is $3.5.
Now let's go back one step. Imagine that you are at the point after the second roll, for
which you can choose either to have a third roll with an expected payoff of $3.5 or keep
the current face value. Surely you will keep the face value if it is larger than 3.5; in other
words, when you get 4, 5 or 6, you stop rolling. When you get 1, 2 or 3, you keep rolling.
So your expected payoff before the second roll is 3I6x3.5+1I6 x ( 4 + 5 + 6) = $4.25.
Now let's go back one step further. Imagine that you are at the point after the first roll,
for which you can choose either to have a second roll with expected payoff $4.25 (when
face value is 1, 2, 3 or 4) or keep the current face value. Surely you will keep the face
value if it is larger than 4.25; In other words, when you get 5 or 6, you stop rolling. So
your expected payoff before the first roll is 4 I 6 x 4.25 + 1I6 x (5 + 6) = $14 I 3 .
This backward approach----called tail policy in dynamic programming-gives us the
strategy and also the expected value of the game at the initial stage, $14/3.
World series
The Boston Red Sox and the Colorado Rockies are playing in the World Series finals. In
case you are not familiar with the World Series, there are a maximum of 7 games and
the first team that wins 4 games claims the championship. You have $100 dollars to
place a double-or-nothing bet on the Red Sox.
Unfortunately, you can only bet on each individual game, not the series as a whole. How
much should you bet on each game so that if the Red Sox wins the whole series, you win
exactly $100, and if Red Sox loses, you lose exactly $100?
Solution: Let (i,j) represents the state that the Red Sox has won i games and the
Rockies has wonj games, and let f (i, j) be our net payoff, which can be negative when
we lose money, at state (i,j). From the rules of the game, we know that there may be
between 4 and 7 games in total. We need to decide on a strategy so that whenever the
123
Stochastic Process and Stochastic Calculus
series is over, our final net payoff is either +100---when Red Sox wins the
championship--or -100-when Red Sox loses. In other words, the state space of the
final stage includes {(4,0), (4,1), (4,2), (4,3)} with payoff f(i,})=100 and
{(0,4), (1,4), (2,4), (3,4)} with payoff f(i,j) = -100. As all dynamic programming
questions, the key is to start with the final stage and work backwards-even though in
this case the number of stages is not fixed. For each state (i, j), if we bet $y on the Red
Sox for the next game, we will have (f(i, j) + y) ifthe Red Sox wins and the state goes
to (i + 1, }), or (f (i, })- y) if the Red Sox loses and the state goes to (i, j +I). So
clearly we have
y}
f(i + 1, J) = f (i, J) +
f(i, j+l)=f(i, j)-y
{f (i, J) =(JU+ 1, J) + f(i, J + 1)) I 2.
y=(f(i+l, j)-f(i, }+1))12
For example, we have /(3, 3) = /( 4, 3)+ /( 3, 4) = lOO-IOO =0. Let's set up a table
2
2
with the columns representing i and the rows representing j. Now we have all the
information to fill in /(4, 0), /(4, I), /(4, 3), /(4, 2), f(O, 4), f (1, 4), /(2, 4),
/(3, 4), as well as /(3,3). Similarly we can also fill in all f(i,j) for the states where
i =3 or j =3 as shown in Figure 5.7. Going further backward, we can fill in the net
payoffs at every possible state. Using equation y=(f(i+l, j)-f(i, }+1))12, we can
also calculate the bet we need to place at each state, which is essentially our strategy.
If you are not accustomed to the table format, Figure 5.8 redraws it as a binomial tree, a
format you should be familiar with. If you consider that the boundary conditions are
/(4, 0), /(4, I), /(4, 3), /(4, 2), f(O, 4), /(1, 4), /(2, 4), and /(3, 4), the
underlying asset either increases by 1 or decrease by I after each step, and there is no
interest, then the problem becomes a simple binomial tree problem and the bet we place
each time is the delta in dynamic hedging. In fact, both European options and American
options can be solved numerically using dynamic programming approaches.
124
A Practical Guide To Quantitative Finance Interviews
Red Sox
wins
"'
:.:2
CJ
=
=
'Cl
"'=
cs=
u
0
1
2
Red Sox
3
4
wins
0
100
1
100
2
100
3
1·t-100
4
Q
-100 -100 -100 -100
"'
0
=
=
'Cl
=
"'
cs=
1
:.:2
CJ
u
0
1
"'
:.:2
=
=
cs"'=
'Cl
u
0
0
0
1
2
4
2
00
3
4
-100 -100 -100 -100
Red Sox
3
4
bets
62.5 87.5 100
1
-31.3
2
-62.5 -37.5
0
50
100
3
-87.5 -75
-50
0
100
4
-100 -100 -100 -100
0
3
00
Red Sox
wins
2
37.5 75
100
0
1
2
3
12.5
4
"'
0
25
=
=
'Cl
"'=
cs=
1
37.5 37.5 25
:.:2
CJ
u
2
3
25 37.5 50
12.5 25
50
50
100
4
Figure 5. 7 Payoffs and bets at different states
100
(4,0)
<
87.5
(3,0)
62.5
(2,0)
31.25
0
(0,0)
(1,0)
-31.25
(0,1)
100
(4,1)
75
(3, I)
50
(3,2)
37.5
(2,1)
0
(I, I)
100
(4,2)
0
(2,2)
-37.5
(1,2)
-62.5
(0,2)
100
0
(4,3)
(3,3)
-100
(3,4)
-50
(2,3)
-100
(2,4)
-75
(1,3)
-87.5
(0,3)
<
-100
(1,4)
-100
(0,4)
Figure 5.8 Payoff at different states expressed in a binomial tree
125
Stochastic Process and Stochastic Calculus
Dynamic dice game
A casino comes up with a fancy dice game. It allows you to roll a dice as many times as
you want unless a 6 appears. After each roll, if 1 appears, you will win $1; if 2 appears,
you will win $2; ... ; if 5 appears, you win $5; but if 6 appears all the moneys you have
won in the game is lost and the game stops. After each roll, if the dice number is 1-5,
you can decide whether to keep the money or keep on rolling. How much are you
willing to pay to play the game (if you are risk neutral)? 12
Solution: Assuming that we have accumulated n dollars, the decision to have another
roll or not depends on the expected profit versus expected loss. If we decide to have an
extra roll, our expected payoff will become
1
l
1
1
1
1
5
-(n+ 1)+-(n+2)+-(n +3)+-(n+ 4)+-(n+5)+-x 0 = -n+ 2.5.
6
6
6
6
6
6
6
n + 2.5 > n, which means that we should
6
keep rolling if the money is no more than $14. Considering that we will stop rolling
when n 15, the maximum payoff of the game is $19 (the dice rolls a 5 after reaching
the state n=l4 ). We then have the following: /(19)=19, /(18)=18, /(17)=17,
/(16) = 16, and /(15) = 15. When n:::;; 14, we will keep on rolling, so
We have another roll if the expected payoff
t
E[f(n) In:::;; 14] = _!.. E[f(n + i)]. Using this equation, we can calculate the value for
6 i;I
E[f(n)] recursively for all n = 14, 13, ···, 0. The results are summarized in Table 5.2.
Since E[/(O)] = 6.15, we are willing to pay at most $6.15 for the game.
18
17
16
15
14
13
12
II
10
Elf0ll 19.00
18.00
17.00
16.00
15.00
14.17
13.36
12.59
11.85
11.16
n
8
7
6
5
4
3
2
I
0
9.91
9.34
8.80
8.29
7.81
7.36
6.93
6.53
6.15
n
19
9
Elt02J. 10.52
Table 5.2 Expected payoff of the game when the player has accumulated n dollars
12 Hint: If you decide to have another roll, the expected amount you have after the roll should be higher
than the amount before the roll. As the number of dollars increases, you risk losing more money if a 6
appears. So when the amount of dollar reaches a certain number, you should stop rolling.
126
A Practical Guide To Quantitative Finance Interviews
Dynamic card game
A casino offers yet another card game with the standard 52 cards (26 red, 26 black). The
cards are thoroughly shuffled and the dealer draws cards one by one. (Drawn cards are
not returned to the deck.) You can ask the dealer to stop at any time you like. For each
red card drawn, you win $1; for each black card drawn, you lose $1. What is the optimal
stopping rule in terms of maximizing expected payoff and how much are you willing to
pay for this game?
Solution: It is another problem perceived to be difficult by many interviewees. Yet it is a
simple dynamic programming problem. Let (b, r) represent the number of black and red
cards left in the deck, respectively. By symmetry, we have
red cards drawn - black cards drawn = black cards left - red cards left = b - r
At each (b, r), we face the decision whether to stop or keep on playing. If we ask the
dealer to stop at (b, r), the payoff is b-r. If we keep on going, there is _b_
b+r
probability that the next card will be black-in which case the state changes to
(b-1, r)-and _r_ probability that the next card will be red-in which case the state
b+r
changes to (b, r-1). We will stop if and only if the expected payoff of drawing more
cards is less than b - r. That also gives us the system equation:
E[f(b,r)] = max(b-r, _b_E[f(b-l,r)]+-r-[f(b,r-1)]). 13
b+r
b+r
As shown in Figure 5.9 (next page), using the boundary conditions /(0, r) =0,
f(b, O)=b, 'Vb, r=O, 1, ···, 26, and the system equation for E[f(b, r)], we can
recursively calculate E[f(b, r)] for all pairs of band r.
The expected payoff at the beginning of the game is E [! (26, 26)] = $2.62.
13 You probably have recognized this system equation as the one for American options. Essentially you
decide whether you want to exercise the option at state (b, r).
127
Stochastic Process and Stochastic Calculus
Number of Black Cards Left
f(b,r)
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
1
0
0.50 1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
2
0
0.33 0.67 1.20 2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
3
0
0.25 0.50 0.85 1.34 2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
14
0
0.20 0.40 0.66 1.00 1.44 2.07 3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
5
0
0.17 0.33 0.54 0.79 1.12 1.55 2.15 3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
I&
0
0.14 0.29 0.45 0.66 0.91 1.23 1.66 2.23 3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
17
0
0.13 0.25 0.39 0.56 0.76 1.01 1.34 1.75 2.30 3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
8
0
0.11 0.22 0.35 0.49 0.66 0.86 1.11 1.43 1.84 2.36 3.05 4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
0
0.10 0.20 0.31 0.43 0.58 0.75 0.95 1.21 1.52 1.92 2.43 3.10 4
5
6
7
8
9
10
11
12
13
14
15
16
17
10
0
0.09 0.18 0.28 0.39 0.52 0.66 0.83 1.04 1.30 1.61 2.00 2.50 3.15 4
5
6
7
8
9
10
11
12
13
14
15
16
'E
11
0
0.08 0.17 0.26 0.35 0.46 0.59 0.74 0.91 1.12 1.38 1.69 2.08 2.57 3.20 4
5
6
7
8
9
10
11
12
13
14
15
(.)
12
0
0.08 0.15 0.24 0.32 0.42 0.54 0.66 0.81 0.99 1.20 1.46 1.77 2.15 2.63 3.24 4
5
6
7
8
9
10
11
12
13
14
13
0
0.07 0.14 0.22 0.30 0.39 0.49 0.60 0.73 0.89 1.06 1.28 1.53 1.84 2.22 2.70 3.28 4.03 5
6
7
8
9
10
11
12
13
14
0
0.07 0.13 0.20 0.28 0.36 0.45 0.55 0.67 0.80 0.95 1.13 1.35 1.60 1.91 2.29 2.75 3.33 4.06 5
6
7
8
9
10
11
12
..a 15
0
0.06 0.13 0.19 0.26 0.33 0.42 0.51 0.61 0.73 0.86 1.02 1.20 1.42 1.67 1.98 2.36 2.81 3.38 4.09 5
6
7
8
9
10
11
:I 16
0
0.06 0.12 0.18 0.24 0.31 0.39 0.47 0.57 0.67 0.79 0.93 1.08 1.27 1.48 1.74 2.05 2.42 2.87 3.43 4.13 5
6
7
8
9
10
17
0
006 0.11 0.17 0.23 029 0.36 0.44 0.53 0.62 0.73 0.85 0.99 1.15 1.33 1.55 1.81 2.11 2.48 2.93 3.48 4.16 5
6
7
8
9
18
0
0.05 0.11 0 .16 0.22 0.28 0.34 0.41 0.49 0.58 0.67 0.78 0.90 1.04 1.21 1.39 1.61 1.87 2.17 2.54 2.99 3.53 4.19 5
6
7
8
19
0
0.05 0.10 0.15 0.20 0.26 0.32 0.39 0.46 0.54 0.63 0.73 0.84 0.96 1.10 1.26 1.45 1.67 1.93 2.24 2.60 3.04 3.57 4.22 5.01 6
7
120
0
0.05 0.10 0.14 0.19 0.25 0.31 0.37 0.43 0.51 0.59 0.68 0.78 0.89 1.01 1.16 1.32 1.51 1.73 1.99 2.30 2.66 3.09 3.62 4.25 5.03 6
121
0
0.05 0.09 0.14 0.19 0.24 0.29 0.35 0.41 0.48 0.55 0.63 0.72 0.83 0.94 1.07 1.21 1.38 1.57 1.79 2.05 2.35 2.72 3.1 5 3.66 4.28 5.05
22
0
0.04 0.09 0.13 0.18 0.23 0.28 0.33 0.39 0.45 0.52 0.60 0.68 0.77 0 .87 0.99 1.12 1.26 1.43 1.62 1.85 2.11 2.41 2.77 3.20 3.71 4.32
23
0
0.04 0.08 0.13 0.17 0.22 0.26 0.32 0.37 0.43 0.49 0.56 0.64 0.72 0 .82 0.92 1.04 1.17 1.32 1.48 1.68 1.90 2.16 2.47 2.82 3.25 3.75
24
0
0.04 0.08 0.12 0.16 0.21 0.25 0.30 0.35 0.41 0.47 0.53 0.60 0.68 0.77 0.86 0.97 1.08 1.22 1.37 1.54 1.73 1.96 2.22 2.52 2.88 3.30
j25
0
0.04 0.08 0.12 0.16 0.20 0.24 0.29 0.34 0.39 0.45 0.51 0.57 0.64 0 .72 0.81 0 .90 1.01 1.13 1.26 1.42 1.59 1.78 2.01 2.27 2.57 2.93
0
0.04 0.07 0.11 0.15 0.19 0.23 0.28 0.32 0.37 0.43 0.48 0.54 0.61 0.68 0.76 0.85 0.95 1.06 1.18 1.31 1.46 1.64 1.83 2.06 2.32 2.62
=
_, 19
CD
Ill
"
"O
GI
a::
....
0
...CD
E
z
Figure 5.9 Expected payoffs at different states (b, r)
128
A Practical Guide To Quantitative Finance Interviews
5.4 Brownian Motion and Stochastic Calculus
In this section, we briefly go over some problems for stochastic calculus, the counterpart
of stochastic processes in continuous space. Since the basic definitions and theorems of
Brownian motion and stochastic calculus are directly used as interview problems, we'll
simply integrate them into the problems instead of starting with an overview of
definitions and theorems.
Brownian motion
A. Define and enumerate some properties of a Brownian motion? 1
Solution: This is the most basic Brownian motion question. Interestingly, part of the
definition, such as W(O) = 0, and some properties are so obvious that we often fail to
recite all the details.
A continuous stochastic process W(t), t
0, is a Brownian motion if
•
W(O)=O;
•
The increments of the process W(t,)-W(O), W(t2 )-W(t1 ),
t,
•
Each
t2
::;; • • •
of these
···,
W(tJ-W(tn_ 1),
tn are independent;
increments
is
normally
distributed
with
distribution
W(t;+ 1)-W(t;) - N(O, f;+i -t;).
Some of the important properties of Brownian motion are the following: continuous (no
E[W(t) 2 ] = t;
W(t)- N(O,t);
martingale property
jumps); E[W(t)] = O;
E[W(t+s)IW(t)]=W(t); cov(W(s),W(t))=s,
'v'O 0 and B2 < 0?
Solution: A standard solution takes advantage of the fact that B, - N(O, I), and B2 - B,
is independent of B" which is again a normal distribution: B2 - B,
B, = x > 0 , then for B2 < 0, we must have B2
-
N(O, 1). If
B, < -x.
P(B, >0,B2 0,B2 -B, <-B,)
=
=
rJi;
r
e-x212dx
(Ji; e-y212dy = r[
r7/4Jr
.l.i12,, 2tr
e-(x'+/)12dxdy
7 !4tr-3!2tr[-e-r2 12]00
2tr
0
=_!_
8
But do we really need the integration step? If we fully take advantage of the facts that B,
and B2 -B, are two 110 N(O, I), the answer is no. Using conditional probability and
independence, we can reformulate the equation as
P(B,
>O,B2 < 0) = P(B, >O)P(B2 -B, < O)P(I B2 -B, l>I B, I)
=l/2xl/2xl/2=118
2
W(s)-N(O,s). So E[exp{'1.W(s)}] is the moment generating function of normal random variable
N(O,s).
130
A Practical Guide To Quantitative Finance Interviews
This approach is better demonstrated in Figure 5.10. When we have B1 > 0 and
B2 -B1 < -B,, which accounts for 1/8 of the density volume. (All 8 regions separated by
x =0, y = 0, y =x, and y =-x have the same density volume by symmetry.)
,_
0.15
I
·u;
c
Q)
0
0.1
' '-.
/ '"-,
f
I
0.05 -
l
"
Figure 5.10 Probability density graph of (81' 8r81)
Stopping time/ first passage time
A. What is the mean of the stopping time for a Brownian motion to reach either -1 or 1?
Solution: As we have discussed, B} -t is martingale. It can be proved by applying Ito's
lemma:
d(Bl2 -t) = B(B12 -t) dBi+ o(B,2 -t) dt+! 82(B12 2-t) dt = 2BldBI -dt+ dt = 2B,dB,.
at
2
So d(B12 -t) has no drift term and is a martingale. Let T=min{t; B1 =1 or -1}. At
continuous time and space, the following property still applies: A martingale stopped at
131
Stochastic Process and Stochastic Calculus
Bi - T is a martingale and E[Bi - TJ= Bg - 0 = 0.
The probability that B, hits 1 or -1 is 1, so Bi= 1 => E[T] = E[ Bi]= I.
a stopping time is a martingale! So
B. Let W (t) be a standard Wiener process and rx ( x > 0) be the first passage time to
level x ( rx = min{t; W(t) = x} ). What is the probability density function of rx and the
expected value of rx ?
Solution: This is a textbook problem that is elegantly solved using the reflection
principle, so we will simply summarize the explanation. For any Wiener process paths
that reach x before t ( rx t ), they have equal probability ending above x or below x at
time t,
The explanation lies in the reflection
principle. As shown in Figure 5.11, for each path that reaches x before t and is at a level
y above x at time t, we can switch the sign of any move starting from rx and the
reflected path will end at 2x- y that is below x at time t. For a standard Wiener process
(Brownian motion), both paths have equal probability.
P(rx
t, W(t):?. x) + P(rx
t) = P(rx
= 2P(W(t):?. x) = 2
:.
m
t, W(t)
r
x) = 2P(rx
t, W(t):?. x)
e-w2 121 dw
,. ...;2;rt
r:
"2m
Jx1v1
"2;r
Take the derivative with respect to t, we have
r ( ) dP{rx
t =
'
dt
Jr
t}
= dP{rx
Ji) = 2 N '( X I vtr.) X-f
x _
t} d(x I
dt
d(x/ Ji)
312
2
::::>
xe-x2121
,
t.&i
\.I
vX
O
> .
From part A, it's easy to show that the expected stopping time to reach either a (a> 0)
or -/3 ( f3 > 0) is again E[ N] = af3. The expected first passage time to level x is
3
lf we define
M(I)
=max W(s), then P(r :s; t) if and only if M(t)?. x. Taking the derivative of P(r, :s; t)
with respect to x, we can derive the probability density function of M(t).
132
A Practical Guide To Quantitative Finance Interviews
essentially the expected stopping time to reach either x or -oo and E [ rx] =xx oo = oo.
Although we have P( rx
oo) = 2- 2N(x
1, the expected value of rx is oo !
W(t)y _________________________________________________,
I
I
I
•
/i
I
I
I
I
I
I
----------r-----------------------.
2x-y
I
I
.......
rx
t
Figure 5.11 Sample path of a standard Weiner process and its reflected path
= dW(t).
If X starts at 0,
what is the probability that X hits 3 before hitting -5? What if X has drift m, i.e.
dX(t) = mdt + dW(t)?
C. Suppose that Xis a Brownian motion with no drift, i.e. dX(t)
Solution: A Brownian motion is a martingale. Let p 3 be the probability that the
Brownian motion hits 3 before -5. Since a martingale stopped at a stopping time is a
+ (-5)(1= 0
= 5 I 8 . Similar to random walk, if we
martingale, we have
have stopping boundaries (a> 0) and -P (p > 0 ), the probability that it stops at a
instead of -P is Pa = P/(a+ p). The expected stopping time to reach either a or -P
is again
E[ N] = ap.
When X has drift m, the process is no longer a martingale. Let P(t, x) be the probability
that the process hits 3 before hitting -5 when X = x at time t. Although Xis no longer a
133
Stochastic Process and Stochastic Calculus
martingale process, it is still a Markov process. So P(t,x) = P(x) is actually independent
oft. Applying the Feynman-Kac equation4 , we have
mP,.(x)+l/2Pxx(x)=O for -5 1
=> P(O)=c +c = - - - c, +c2e10m =O
C2=ll(e-6m_elOm)
I
2 e'°m-e-6m
A different and simpler approach takes advantage of the exponential martingale:
Z(t)=exp{A.W(t)-tA. 21}. Since W(t)=X(t)-mt, X(t)-mt is a Brownian motion as
J
well. Applying the exponential martingale, we have E [exp ( A.(X - mt)-t A. 2t) =I for
any constant .A.. To remove the terms including time t, we can set A.= -2m and the
equation becomes E[ exp(-2mX)] =I. Since a martingale stopped at a stopping time is
a martingale, we have
exp(-2mx3)+ (1-
elOm -1
-6m.
e -e
= 1=> iom
D. Suppose that Xis a generalized Weiner process dX = dt + dW(t), where W(t) is a
Brownian motion. What is the probability that Xever reaches -1?
Solution: To solve this problem, we again can use the equation
E[ exp(-2mX)J = 1
from the previous problem with m = l. It may not be obvious since we only have one
apparent boundary, -1. To apply the stopping time, we also need a corresponding
positive boundary. To address this problem, we can simply use +oo as the positive
boundary and the equation becomes
4
Let X be an Ito process given by equation dX (I) = /J(t, X)dt + y( t, X )dW and
Define function V (I. x) = E[f ( X,. ) I X,
= x]. then
f
(x) be a function of X.
V (I. x) is a martingale process that satisfies the partial
av
av 1 ,
a1 v
differential equation - + /J(t,x)-+-y· (1,x)- = 0 and terminal condition V(T, x) = f(x) for all
a1
as 2
as'
x.
134
A Practical Guide To Quantitative Finance Interviews
Ito's lemma
Ito's lemma is the stochastic counterpart of the chain rule in ordinary calculus. Let X(t)
be an Ito process satisfying dX(t)=fi(t,X)dt+y(t,X)dW(t), and f(X(t),t) be a
twice-differentiable function of X(t) and t. Then f(X(t),t) is an Ito process satisfying
.
of
of 1
02 f
Dnft rate=-+ fi(t,X)-+-y 2 (t,X)2
ot
ox 2
ox
A. Let B, be a Brownian motion and z, = .JiB,. What is the mean and variance of Z,? Is
z,
a martingale process?
Solution: As a Brownian motion, B,
a constant at t, Z,
t
=.JiB,
N(O, t), which is symmetric about 0. Since
Ji
is
is symmetric about 0 and has mean 0 and variance
x var(B,) = t 2 • More exactly, Z,
N(O, t 2 )
.
Although Z, has unconditional expected value 0, it is not a martingale. Applying Ito's
r:
az
az
0 2z
_
r:
lemma to Z, = viB,, we have dZ, = -1 dB, + -1 dt+tx--; dt =tt 112 B1dt+vtdB1 •
aB,
at
oB,
For all the cases that B,
:;e 0,
which has probability 1, the drift term
zero. 5 Hence, the process
Z, = .JiB, is not a martingale process.
tr
112
B,dt is not
B. Let W(t) be a Brownian motion. Is W(t) 3 a martingale process?
5
A generalized Wiener process dx = a(x, t)dt + b(x, t)dW(t) is a martingale process if and only if the
drift term has coefficient a(x,t) = 0.
135
Stochastic Process and Stochastic Calculus
Solution: Applying Ito's lemma to f(W(t), t)=W(t) 3 , we have
8/
81 =3W(t)2
8W(t)
'
82/
ar=O, 8 W(t)2 =6W(t), and df(W(t), t)=3W(t)dt+3W(t) 2 dW(t). So again for the
cases W(t) :;t 0, which has probability 1, the drift term is not zero. Hence, W(t) 3 is not a
martingale process.
136
Chapter 6 Finance
It used to be common for candidates with no finance knowledge to get hired into
quantitative finance positions. Although this still happens for candidates with specialized
knowledge that is in high demand, it's more likely that you are required, or at least
expected, to have a basic grasp of topics in finance. So you should expect to answer
some finance questions and be judged on your answers.
Besides classic textbooks,' there are a few interview books in the market to help you
prepare for finance interviews.2 If you want to get prepared for general finance problems,
you may want to read a finance interview book to get a feel for what types of questions
are asked. The focus of this chapter is more on the intuitions and mathematics behind
derivative pricing instead of basic finance knowledge. Derivative problems are popular
choices in quantitative interviews-even for divisions that are not directly related to
derivative markets-because these problems are complex enough to test your
understanding of quantitative finance.
6. 1. Option Pricing
Let's begin with some notations that we will use in the following sections.
T: maturity date; t: the current time; r = T - t : time to maturity; S : stock price at time t;
r: continuous risk-free interest rate; y: continuous dividend yield; a: annualized asset
volatility; c: price of a European call; p: price of a European put; C: price of an
American call; P: price of an American put; D: present value, at t, of future dividends; K:
strike price; PV: present value at t.
Price direction of options
How do vanilla European/American option prices change when S, K, r , a, r, or D
changes?
Solution: The payoff of a call is max(S - K, 0) and the payoff of a put is max(K - S, 0).
A European option can only be exercised at the expiration time, while an American
option can be exercised at any time before maturity. Intuitively we can figure out that
the price of a European/American call should decrease when the strike price increases
1 For basic finance theory and financial market knowledge, I recommend Investments by Zvi Bodie, Alex
Kane and Alan J. Marcus. For derivatives, Options, Futures and Other Derivatives by John C. Hull is a
classic. If you want to gain a deeper understanding of stochastic calculus and derivative pricing, I'd
recommend Stochastic Calculus for Finance (Volumes I and II) by Steven E. Shreve.
2 For example, Vault Guide to Finance Interviews and Vault Guide to Advanced and Quantitative
Finance Interviews.
Finance
since a call with a higher strike has no higher-and sometimes lower-payoff than a call
with a lower strike. Using similar analyses, we summarize the effect of changing market
conditions on an option's value in Table 6.1.
The impact of time to maturity on the price of a European call/put is uncertain. If there is
a large dividend payoff between two different maturity dates, a European call with
shorter maturity that expires before the ex-dividend date may be worth more than a call
with longer maturity. For deep in-the-money European puts, the one with shorter
maturity is worth more since it can be exercised earlier (time value of the money).
Variable
Euro ean call
American call American Put
Stock nee
Strike nee
Time to maturi
Volatilit
Risk-free rate
Dividends
Table 6.1 Impact of S, K,
r , a, r, and Don option prices
r: increase; t : decrease;?: increase or decrease
It is also worth noting that Table 6.1 assumes that only one factor changes value while
all others stay the same, which in practice may not be realistic since some of the factors
are related. For example, a large decrease in interest rate often triggers a stock market
rally and increases the stock price, which has an opposite effect on option value.
Put-call parity
Put-call parity: c + K-rr = p + S - D, where the European call option and the European
put option have the same underlying security, the same maturity T and the same strike
price K. Since p 2:: 0, we can also derive boundaries for c, S - D - Ke-rr c S, from
the put-call parity.
For American options, the equality no longer holds and it becomes two inequalities:
S-D-K:::;
S-K-rr.
Can you write down the put-call parity for European options on non-dividend paying
stocks and prove it?
138
A Practical Guide To Quantitative Finance Interviews
Solution: The put-call parity for European options on non-dividend paying stocks is
c + K-rr = p + S. We can treat the left side of the equation as portfolio A-a call and a
zero-coupon bond with face value K-and the right side as portfolio B-a put and the
underlying stock, which is a protective put. Portfolio A has payoff
max(ST-K,O)+K=max(SroK) at maturity T; portfolio B has payoff
max(K-SroO)+ST =max(SroK) at T. Since both portfolios have the same payoff at T
and no payoff between t and T, the no-arbitrage argument3 dictates that they must have
the same value at t. Hence, c+ K-r• = p + S.
If we rearrange the put-call parity equation into c - p = S-K-r', it will give us different
insight. The portfolio on the left side of the equation-long a call and short a put-has
the payoff max( ST - K, 0)- max( K - Sro 0) = ST - K, which is the payoff of a forward
with delivery price K. A forward with delivery price K has present value S - K-rr. So
we again have the put-call parity c - p = S - K-r•. This expression shows that when the
strike price K = sr• (forward price), a call has the same value as put; when K < srr, a
call has higher value; and when K > sr•, a put has higher value.
American v.s. European options
A. Since American options can be exercised at any time before maturity, they are often
more valuable than European options with the same characteristics. But when the stock
pays no dividend, the theoretical price for an American call and European call should be
the same since it is never optimal to exercise the American call. Why should you never
exercise an American call on a non-dividend paying stock before maturity?
Solution: There are a number of solutions to this popular problem. We present three
arguments for the conclusion.
Argument I. If you exercises the call option, you will only get the intrinsic value of the
call S -K. The price of the American/European call also includes time value, which is
positive for a call on a non-dividend paying stock. So the investor is better off selling the
option than exercising it before maturity.
In fact, if we rearrange the put-call parity for European options, we have
c = S-K-r' + p = (S-K)+(K-K-r')+ p. The value of a European call on a nondividend paying stock includes three components: the first component is the intrinsic
value S-K; the second component is the time value of the strike (if you exercise now,
A set of transactions is an arbitrage opportunity if the initial investment '.S O; payoff 2': O; and at least one
of the inequalities is strict.
3
139
Finance
you pay K now instead of K at the maturity date, which is lower in present value); and
the third component is the value of the put, which is often considered to be a protection
against falling stock price. Clearly the second and the third components are both positive.
So the European call should be worth more than its intrinsic value. Considering that the
corresponding American call is worth at least as much as the European call, it is worth
more than its intrinsic value as well. As a result, it is not optimal to exercise the
American call before maturity.
Argument 2. Let's compare two different strategies. In strategy 1, we exercise the call
and receive cash S -K. Alternatively, we can keep the call,
option4 at time t (t <
short the underlying stock and lend K dollars with interest rate r (the cash proceedings
from the short sale, S, is larger than K). At the maturity date T, we exercise the call if it's
in the money, close the short position and close the lending. Table 6.2 shows the cash
flow of such a strategy:
n
It clearly shows that at time t, we have the same cash flow as exercising the call, S-K.
But at time T, we always have positive cash flow as well. So this strategy is clearly
better than exercising the call at time t. By keeping the call alive, the extra benefit can be
realized at maturity.
Cash flow
T
t
ST SK
Sr>K
Call
0
0
Sr-K
Short Stock
s
-Sr
-Sr
Lend Katt
-K
Kerr
Kerr
Total
S-K
Kerr -ST> 0
Kerr -K > 0
Table 6.2 Payoff of an alternative strategy without exercising the call
Argument 3. Let's use a mathematical argument relying on risk-neutral pricing and
Jensen's inequality-if/(X) is a convex function, 5 then
From
Figure 6.1, it's obvious that the payoff (if exercised when S > K ) of a call option
C(S) = (S - Kf is a convex function of stock price with property
4
We assume S > K in our discussion. Otherwise, the call surely should not be exercised.
5
A function /(X) is convex if
and only
f "(x) > 0, Vx, then /(X) is convex.
140
O P(1K).
For this specific problem, we should have 8/ 9xP(90)=8/9x9=8>P(80). Since the
put option with strike price $80 is currently price at 8, it is overpriced and we should
short it. The overall arbitrage portfolio is to short 9 units of put with K = $80 and long 8
units of put with K = 90. At time 0, the initial cash flow is 0. At the maturity date, we
have three possible scenarios:
90, payoff= 0 (No put is exercised.)
90 > ST
80, payoff= 8 x (90- ST) > 0 (Puts with K = 90 are exercised.)
ST < 80, payoff= 8 x (90-ST )-9 x (80- ST)= ST > 0 (All puts are exercised.)
The final payoff 0 with positive probability that payoff > 0. So it is clearly an
arbitrage opportunity.
Black-Scholes-Merton differential equation
Can you write down the Black-Scholes-Merton differential equation and briefly explain
how to derive it?
Solution: If the evolution of the stock price is a geometric Brownian motion,
dS = µSdt+(J"SdW(t), and the derivative V = V(S,t) is a function of Sand t, then
applying Ito's lemma yields:
dV=(av +µSav
at
as
2
as
as
dW(t), where W(t)isaBrownianmotion.
The Black-Scholes-Merton differential equation is a partial differential equation that
.
av
at
av
as
1
2
a1v
as
should be satisfied by V: -+rS-+-(J" 2S 2 - 2 =rV.
To derive the Black-Scholes-Merton differential equation, we build a portfolio with two
components: long one unit of the derivative and short av unit of the underlying stock.
as
Then the portfolio has value
142
n =V-
av S and the change of n follows equation
as
A Practical Guide To Quantitative Finance Interviews
dTI = dV - av dS
as
av
av 1
a2 v
av
av
= (-+ µS-+-a 2S 2- 2 )dt +aS-dW(t)--(µSdt + aSdW(t))
at
as 2
as
as
as
=(av +_!_a2 s2 a 2 V )dt
as 2
at 2
It is apparent that this portfolio is risk-free since it has no diffusion term. It should have
risk-free rate of return as well: dTI
=
r(V - av S)dt. Combining these results we have
as
av 1 2 2 a2 v
av
av
av 1 2 2 a 1 v
(at + 2 a S as 2 )dt = r(V - as S)dt => at + rS as + 2 a S as2 = rV,
which is the Black-Scholes-Merton differential equation.
The Black-Scholes-Merton differential equation is a special case of the discounted
Feynman-Kac theorem. The discounted Feynman-Kac theorem builds the bridge
between stochastic differential equations and partial differential equations and applies to
all Ito processes in general:
Let X be an Ito process given by equation dX(t) = fl(t,X)dt + y(t,X)dW(t) and /(x)
be a function of X. Define function V(t,x) = E[e-r(T-t>f(Xr) I X 1 = x], then V(t,x) is a
martingale process that satisfies the partial differential equation
av
at
av
ax
1
2
a1 v
ax
- + f3(t,x)-+-y 2 (t,x)-2 = rV(t,x)
and boundary condition V(T,x) = f(x) for all x.
Under risk-neutral measure, dS=rSdt+aSdW(t). Let S=X, f3(t,X)=rS and
y(t, X) = aS, then the discounted Feynman-Kac equation becomes the Black-Scholes.
. 1 equation
av
av
a1v
M erton 11erentia
+ rs + -1a 2s2 -=rV.
at
as 2
as 2
Black-Scholes formula
The Black-Scholes formula for European calls and puts with continuous dividend yield y
1s:
143
Finance
d = ln(Se-Y' I K)+(r+ a 2 I 2)r = ln(S I K)+(r-y+a 2 I 2)r
w here
1
a../r
a../r
2 12)r -dr
d2_ln(SIK)+(r-y-a
r
- , avr
avr
N(x) is the cdf of the standard normal distribution and N '(x) is the pdf of the standard
normal distribution: N(x) = (
tb-
v2ff
e-y2 12 dy
and N'(x) =
tb-
v2K
e-x212 •
If the underlying asset is a futures contract, then yield y = r. If the underlying asset is a
foreign currency, then yield y = rl, where r1 is the foreign risk-free interest rate.
A. What are the assumptions behind the Black-Scholes formula?
Solution: The original Black-Scholes formula for European calls and puts consists of the
equations c = SN ( d,) - Ke-rr N ( d 2 ) and p = Ke-rr N (-d2 ) - SN ( -d, ), which require the
following assumptions:
1. The stock pays no dividends.
2. The risk-free interest rate is constant and known.
3. The stock price follows a geometric Brownian motion with constant driftµ and
volatility a: dS = µSdt+aSdW(t).
4. There are no transaction costs or taxes; the proceeds of short selling can be fully
invested.
5. All securities are perfectly divisible.
6. There are no risk-free arbitrage opportunities.
B. How can you derive the Black-Scholes formula for a European call on a non-dividend
paying stock using risk-neutral probability measure?
Solution: The Black-Scholes formula for a European call on a non-dividend paying
stock is
c = SN(d, )-Ke-r• N(d2 ), where d, = ln(S / K) +
a
144
CT
r
2 12
)r and d 2 = d, - a../r.
A Practical Guide To Quantitative Finance Interviews
Under the risk-neutral probability measure, the drift of stock price becomes the risk-free
interest rate r(t): dS = r(t)Sdt + aSdW(t). Risk-neutral measure allows the option to be
priced as the discounted value of its expected payoff with the risk-free interest rate:
V(t) = E [ e
- ( r(u)du
]
V(T) S(t) , 0::; t
T, where V(T) is the payoff at maturity T.
When r is constant, the formula can be further simplified as V(t)=e-rrE[V(T)IS(t)].
Under risk-neutral probabilities, dS = rSdt + CYSdW(t). Applying Ito's lemma, we get
d(ln(S)) = (r-CY2 /2)dt + adW(t) => ln ST - N(lnS + (r-CY2 I 2)r, a 2 r).
So ST= se K
otherwise
0,
a 2 I 2)r _ d
d
Se(r-cri 12)r+cr.fic > K =>c> ln(K IS)- (r- - 2 an
1
avr
E[V(T) Is]= E[ max( ST -K,O) Is]=
.c
1-e-&i 12 dc
(ser+crJT& -K)-
Jf;
di
1 e+·-.ficr)i 12dc - K [ - 1 e-ci/2dc
--
-- Se rr [
di
J2;
di
J2;
Let & = c-a.J;, then de =di, c = -d2 => & =-d2 -a.J; =-d, and we have
Serr [
1 -(c-Jrcr)i 12dc --Serr [
1 -&i/2dr;:;-e
r;:;-e
c
di "27!
d1 "27l
=
S err N (d ) ,
1
Ki
ld2 v2Jr
:. E[V(T)] =Serr N( d 1 )-KN(d2 ) and V(t) = e-rr E[V(T)] =SN ( d, )-Ke-rr N(d2 )
From the derivation process, it is also obvious that 1-N(-d2 )
neutral probability that the call option finishes in the money.
= N(d2 )
is the risk-
C. How do you derive the Black-Scholes formula for a European call option on a non-
dividend paying stock by solving the Black-Scholes-Merton differential equation?
145
Finance
Solution:
You can skip this problem if you don't have background in partial differential
equations (PDE). One approach to solving the problem is to convert the Black-ScholesMerton differential equation to a heat equation and then apply the boundary conditions
to the heat equation to derive the Black-Scholes formula.
av =-av av =av dy =_!_aV and
at af ' as ay ds s ay
a2v = av(av)= av(_!_ avJ=.=!. av+_!_ av(avJ=.=!. av +-1 a2v .6
as2 as as as s ay s 2 ay s as ay s 2 ay s 2 ay 2
av
av 1
a2v
The Black-Scholes-Merton differential equation - + rS-+-CT S ---rV = 0
at
as 2
as
2v
av+ ( r--CT
1 2)av
1 2a
can be converted to - -+-CT
---rV
= 0.
af
2
ay 2 ay
Let y=lnS
(S=eY)and f=T-t
'
then
2
2
2
2
Let
- the equation
. becomes--+
au ( r--CT
1 2) -+-CT
au I 2 -a1u2 =0.
u=er•v,
af
2
ay 2 ay
Finally, let
au= -au+ (r--CT
I 2)
af ar
au au
ay ax
x=y+(r2
and r=f, then - = - and
au wh.1ch transtiorms the equation
. to
-,
ax
. . I equation
. becomes a heat/d"ffu
. equation
. -au = -I CT 2 -a1u
So the ongma
1 s1on
2
ar
2
ax
.
For h eat
. -au = -I CT 2 -a1u2 , w here u = u(x, r ) 1s. a tiunction
. o f time
.
. bl ex,
equation
r and space vana
ar
ax
2
with boundary condition u(x,0) = u0 (x), the solution is
u(x, r) =
I
i
,-;;--!-o
v2JrrCT
00
7
u0 (1/f )exp ( - (x-1/1)2}
l/f.
2
2CT r
The Jog is taken to convert the geometric Brownian motion to an arithmetic Brownian motion; r = T - t
is used to convert the equation from a backward equation to a forward equation with initial condition at
r = 0 (the boundary condition at t = T
r = 0 ).
6
146
A Practical Guide To Quantitative Finance Interviews
For European calls, the boundary condition is u0 (ST)= max (ST - K, 0).
S=exp(x-(r-0.5o-2 )x-). When
X=lfl
and r=O, ST
(max(ell'-K,O)exp(
u(S,r)=u(x, r)=
27rTO"
=
27rTO"
Let c
=
Vf-X
o-Ji,
-l K
Vt- n '
I (ell'
!nK
then de=
2o-
T
=e'll.
2o-
T
r
rjdV'
dVf
,-- ,
2
exp[-(x-V1) J=e-c 212 and when
2o- 2 r
Cf.YT
_1n(KIS)-(r-o- 2 /2)r _ d
r
-- 2
o-vr
&-
Now, it's clear that the equation for u(S, r) is exactly the same as the equation for
E[V(T)IS] in question B. Hence, we have V(S,t)=e-r'u(S,r)=SN(d1 )-Ke-r'N(d2 )
as well.
D. Assume zero interest rate and a stock with current price at $1 that pays no dividend.
When the price hits level $H ( H > 1) for the first time you can exercise the option and
receive $1. What is this option worth to you today?
Solution: First let's use a brute-force approach to solve the problem by assuming that
the stock price follows a geometric Brownian motion under risk-neutral measure:
dS = rSdt + o-SdW(t). Since r = 0,
dS = aSdW(t) => d(ln S) = -to- 2 dt + o-dW(t).
When t = 0, we have S0 =1=>ln(S0 )=0.
7 The
Ou
1 a1 u
fundamental solution to heat equation - = - - with initial condition u0 (1/') =/(If/) is
OT 2 2
u(x,t)= [__..,p(x,
ax
=lfl)f(lf/)dl/', where p(x, =x/x0 =If/)= &exp{-(x-11')2 /2t}.
For detailed discussion about heat equation, please refer to The Mathematics of Financial Derivatives by
Paul Wilmott, Sam Howison, and Jeff Dewynne.
147
Finance
Hence, ln S = -to- 21 + o-W(t) => ln S + to- 2t = W(t) is a Brownian motion.
O'"
Whenever S reaches $H, the payoff is $1. Because the interest rate is 0, the discounted
payoff is also $1 under risk-neutral measure. So the value of the option is the probability
that S ever reaches $H, which is equivalent to the probability that In S ever reaches
lnH. Again we can apply the exponential martingale Z(t) = exp{A.W(t)-tA. 2 1} as we
2
didinChapter5: E[Z(t)]=E [ exp { A. 1nS+1-oo-2 t -fA. 2t }] =l.
To remove the terms including time t, we can set A.= a and the equation becomes
E[ exp(lnS)] =I. The Let P be the probability that lnS ever reaches lnH (using -oo
as the negative boundary for stopping time), we have
Pexp(lnH)+(l-P)exp(-oo) = Px H = 1=>P=1/ H.
So the probability that Sever reaches $His l/H and the price of the option should be
$1/H. Notice that S is a martingale under the risk-neutral measure; 8 but In S has a
negative drift. The reason is that In S follows a (symmetrical) normal distribution, but S
itself follows a lognormal distribution, which is positively skewed. As T oo, although
the expected value of Sr is I, the probability that Sr 1 actually approaches 0.
It is simpler to use a no-arbitrage argument to derive the price. In order to pay $1 when
the stock price hits $H, we need to buy 1IH shares of the stock (at $1 IH). So the option
should be worth no more than $1 IH. Yet if the option price C is less than $1/H
( C < 1I H => CH < 1 ), we can buy an option by borrowing C shares of the stock. The
initial investment is 0. Once the stock price hits $H, we will excise the option and return
the stock by buying C shares at price $H, which gives payoff I - CH > 0. That means we
have no initial investment, yet we have possible positive future payoff, which is
contradictory to the no arbitrage argument. So the price cannot be less than $1 IH. Hence,
the price is exactly $1/H.
E. Assume a non-dividend paying stock follows a geometric Brownian motion. What is
the value of a contract that at maturity T pays the inverse of the stock price observed at
the maturity?
Once we recognize that S is a martingale under the risk neutral measure, we do not need the assumption
that S follows a geometric Brownian motion. S has two boundaries for stopping: 0 and H. The boundary
conditions are /(0) = o and /(//) = 1. Using the martingale, the probability that it will ever reaches His
8
P x II + (I - P) x 0
148
= S,,
= I =>
P = I I H.
A Practical Guide To Quantitative Finance Interviews
Solution: Under risk-neutral measure dS = rSdt + aSdW(t). Apply Ito's lemma to
2
1
V=-:
S
J
av +av- +1 a- v-2 a 2 S2 dt+-aSdW(t)
av
dV = ( -rS
as
a1
2 as
as
123 a 2 S 2 ) dt-I2 aSdW(t)=(-r+a 2 )Vdt-aVdW(t)
= (-I-2 rS+0+-S
2S
S
So V follows a geometric Brownian motion as well and we can apply Ito's lemma to
lnV:
d(ln V)
2
v
v
(-r+-a
1 2 ) dt-adW(t).
= (-(-r+
a 2 ) + 0 -1-v-2 a 2 Jdt +-adW(t)=
v
2v
v
2
Discounting the payoff by e-rr, we have
V= e-rr E [VT] = i, e-2
2
rr+ a r.
6. 2. The Greeks
All Greeks are first-order or second-order partial derivatives of the option price with
respect to different underlying factors, which are used to measure the risks-as well as
potential returns-of the financial derivative. The following Greeks for a derivative fare
routinely used by financial institutions:
82 f
Bf
Bf
Bf
Gamma: r = - · Theta: 0=-· Vega: v = - · Rho: p=as '
8S 2 '
Bt '
Ba '
Br
Bf
Delta
For a European call with dividend yieldy: d
= e-yr N(d1)
For a European put with dividend yieldy: d =-e-yr[l-N(d1)]
A. What is the delta of a European call option on a non-dividend paying stock? How do
you derive the delta?
Solution: The delta of a European call on a non-dividend paying stock has a clean
expression: d = N(d1 ). For the derivation, though, many make the mistake by treating
149
Finance
N(d1)and N(d2 ) as constants in the call pricing formula c=SN(d1 )-Ke-r'N(d2 )and
simply taking the partial derivative on S to yield N(d1 ). The derivation step is actually
more complex than that since both N(d1) and N(d2 ) are functions of S through d1 and
d 2 • Sothecorrectpartialderivativeis
ac =N(d )+Sxj_N(d )-Ke-r• j_N(d
1
as
1
as
as
2 ).
Take the partial derivative with respect to S for N (d 1 ) and N (d 2 ) 9 :
j_N(d)=N'(d)j_d =-l-e-d?12x
as
I
I
as
.J2;
I
I
= I
e-d,212
Sa.f; Sa.J27ri
j_N(d )=N'(d )j_d =-l-e-d?t2x
as
2
2
=
as
1
Sa.J2;;
a
have
as
e
2
.J2;
-d12 / 2 u./Td1 -u2 r/ 2
e
e
I
= I
e-(d1-u..fT)212
sa.J; sa.J2;rr
1
=
-d12 / 2
a
as
ac
ac
last two components of- cancel out and - = N(d1).
as
as
So we
s
K
e
S rr
x-e
K
_ a
r' -N(d2 )=0.
as
Hence, the
B. What is your estimate of the delta of an at-the-money call on a stock without dividend?
What will happen to delta as the at-the-money option approaches maturity date?
Solution: For an at-the-money European call, the stock price equals the strike price.
(r + a 2 I 2)r
r
a r
· ·
=
c
=(-+-)vr >0 and 8.=N(d1 )>0.5. As shown m Figure
avr
a 2
6.2, all at-the-money call options indeed have !-:.. > 0.5 and the longer the maturity, the
higher the!-:... As T-t-;O,
which is also
2
shown in Figure 6.2 ( T = I 0 days). The same argument is true for calls on stock with
continuous dividend rate y if r > y .
a
Figure 6.2 also shows that when Sis large (S >> K ), !-:.. approaches I. Furthermore, the
shorter the maturity, the faster the delta approaches 1. On the other hand, if Sis small
( S << K ), !-:.. approaches 0 and the shorter the maturity, the faster the delta approaches 0.
9
d
2
150
= d -a/;=> N '(d ) = §_er N '(d)
1
2
K
1 •
Bd2 =ad,
as
as
A Practical Guide To Quantitative Finance Interviews
Delta of Call Options
---..
,......
.
..
················
,,
......
1
--.= 10days
0.9
0.8
- - - - - t=
1 month
•••••••••• t=
3 months
/
I
, .·•
.•
•••
I .•
I .•
I •••
,..·
0.7
I
',' ...··......··
I•"
0.6
s
Q)
0.5
-·-·-·-·-·-·-·-·-·-·-·-·-·-·-·-·-·-·"
.·•
..···I
•" I
.• I
••• I
•"
I
.•
I
•••
I
••
I
••
I
•••
I
•"
I
.•
I
0
0.4
0.3
0.2
0.1
o····"···
75
...·· ,'
..
·
,
...·· ,,'
..
..·
,
...
···
·····
•"
I
__,,;"'
80
85
90
95
100
105
110
115
120
125
Spot Price
Figure 6.2 Variation of delta of a European call option with respect to S and
T. K = 100, r = 0.05, a= 0.25.
C. You just entered a long position for a European call option on GM stock and decide
to dynamically hedge the position to eliminate the risk from the fluctuation of GM stock
price. How will you hedge the call option? If after your hedge, the price of GM has a
sudden increase, how will you rebalance your hedging position?
Jf
Solution: Since d 1 = ln(S / K) + (r
+ a 2 12 )' and /). = e-yr N(d1) is a monotonously
a r
increasing function of d 1, we have S t
d 1 t => /). t .
One hedging method is delta hedging, for which we
e-yr N(d1) shares of stock
for each unit of call option to make the portfolio delta-neutral. Since /). shares of GM
stock costs more than one unit of GM option, we also need to invest cash (if the option
price exactly follows the Black-Scholes formula, we need to lend $Ke-rr N(d2 ) for each
151
Finance
unit of option) in the money market. If there is a sudden increase in S, d 1 increases and
increases as well. That means we need to short more stock and lend more cash
(Ke-rr N(d2 ) also increases).
The delta hedge only replicates the value and the slope of the option. To hedge the
curvature of the option, we will need to hedge gamma as well.
D. Can you estimate the value of an at-the-money call on a non-dividend paying stock?
Assume the interest rate is low and the call has short maturity.
Solution:
When
environment, r
S=K,
we
0 and e-rr
We also have N(d, )-N(d2 )
r
where d 2 =(a -
a,-
2 )vr
have c=S(N(d1 )-e-rrN(d2 )). In a
1, so c
low-interest
S(N(d,)-N(d2 )).
= f"i
e- 112 x2dx,
12 v27r
r
ar
and d 1 =(a +2)vr .
For a small r, a typical a for stocks(< 40% per year) and a short maturity(< 3 months),
both d 2 and d 1 are close to 0. For example, if r = 0.03, a= 0.3, and T =I I 6 year, then
d 2 --
-
0 •02 an d e -I/id? -- 0.98 .
In practice, this approximation is used by some volatility traders to estimate the implied
volatility of an at-the-money option.
(The approximation e- 112 x 2
1 causes a small overestimation since e- 112 x 2 < 1; but the
approximation -e-rr K -K causes a small underestimation. To some extent, the two
opposite effects cancel out and the overall approximation is fairly accurate.)
Gamma
For a European call/put with dividend yield y:
r = N '(d,
S0 a
152
T
A Practical Guide To Quantitative Finance Interviews
What happens to the gamma of an at-the-money European option when it approaches its
maturity?
Solution: From the put-call parity, it is obvious that a call and a put with identical
characteristics have the same gamma (since r =0 for both the cash position and the
underlying stock). Taking the partial derivative of the /1 of a call option with respect to
N'(d )e-y•
S We haver=
I
'
Sa/; '
So for plain vanilla call and put options, gamma is always positive.
Figure 6.3 shows that gamma is high when options are at the money, which is the stock
price region that /1 changes rapidly with S. If S << K or S >> K (deep in the money or
out of the money), gamma approaches 0 since /1 stays constant at I or 0.
The gamma of options with shorter maturities approaches 0 much faster than options
with longer maturities as S moves away from K. So for deep in-the-money or deep outof-the-money options, longer maturity means higher gamma. In contrast, if the stock
prices are close to the strike price (at the money) as the maturity nears, the slope of delta
for an at-the-money call becomes steeper and steeper. So for options close to the strike
price, shorter-term options have higher gammas.
As r 0, an at-the-money call/put has r --+ oo ( /1 becomes a step function). This can
be shown from the formula of gamma for a European call/put with no dividend,
r
=
N'(d1).
Sa/;.
When S = K, d1 = lim(!.._ + a)/;
T-40 a
2
0
lim N '(d1)
T-40
yet the denominator has a limit limSa/;--+ 0, so r
T--+0
V 2Jr
•
The numerator is I I .J2;;
oo. In other words, When t = T,
delta becomes a step function. This phenomenon makes hedging at-the-money options
difficult when t
T since delta is extremely sensitive to changes in S.
153
Finance
Gamma of Call/Put Options
0.1
0.09
0.08
-.=
- - - - - t=
.......... t=
10days
1 month
3 months
0.07
0.06
ro
E
E 0.05
ro
C)
0.04
..,
..
···I
......··· / ,
0.03
0.02
•••
,.•
....·· ,,,
..
··
·····
......,,
o--..a.:;;-;,...
___
,o'
I
I
I
I
I
I
I
I
I
0.01
75
80
85
90
95
100
105
110
115
120
125
Spot Price
Figure 6.3 Variation of gamma of a European call option with respect to Sand T.
K = 100, r= 0.05, a= 0.25.
Theta
For a European call option: 0
=-
For a European put option: 0
=
J;
SN'(d )ae-yr
2 r
J";
SN'(d )ae-yr
2 r
+ ySe-yr N ( d1) - rKe-rr N (d 2 )
ySe-yr N(-di) + rKe-rrN(-d2 )
When there is no dividend, the theta for a European call option is simplified to
0
=-
S<>K,
and
A Practical Guide To Quantitative Finance Interviews
Hence,
When
0 has large negative value and the
smaller the r, the more negative the 0.
Theta of Call Options
·····..........
-5
........
....
_.,,,,
-10
...•-;'
_.,--- ....
,,,,,,,...··············
///
-15
.I:.
I- -20
-25 -
- - 't
= 10 days
- - - - - 't=
......... • 't=
1 month
3 months
-30
-35
75
80
85
90
95
100
105
Spot Price
110
115
120
125
Figure 6.4 Variation of theta of a European call option with respect to Sand
T. K = 100, a= 0.25, r= 0.05
A. When will a European option have positive theta?
Solution: For American options as well as European calls on non-dividend paying assets,
theta is always negative. But for deep in-the-money European puts, their values may
increase as t approaches T if all other factors remain the same, so they may have positive
theta.
A put option on a non-dividend paying asset has 0
put option is deep in-the-money (S< 0. That's also the reason why it can be optimal to exercise a deep in-themoney American put before maturity.
For deep in-the-money European call options with high dividend yield, the theta can be
positive as well. If a call option with high dividend yield is deep in-the-money ( S >> K ),
so the component ySe-Y'N(d1 ) can make 0 positive.
B. You just entered a long position for a call option on GM and hedged the position by
shorting GM shares to make the portfolio delta neutral. If there is an immediate increase
or decrease in GM's stock price, what will happen to the value of your portfolio? Is it an
arbitrage opportunity? Assume that GM does not pay dividends.
Solution: A position in the underlying asset has zero gamma. So the portfolio is deltaneutral and long gamma. Therefore, either an immediate increase or decrease in the GM
stock price will increase the portfolio value. The convexity (positive gamma) enhances
returns when there is a large move in the stock price in either direction.
Nevertheless, it is not an arbitrage opportunity. It is a trade-off between gamma and
theta instead. From the Black-Scholes-Merton differential equation, the portfolio V
.
. av
av l
a2 v =
1
satisfies the equation -+rS-+-o- 2S 2 - +-o- 2S 2r = rV. For a delta2
at
as 2
as
2
neutral portfolio, we have E> + _!_ a 2s 2 r = r V. This indicates that gamma and theta often
2
have opposite signs. For example, when an at-the-money call approaches maturity,
gamma is large and positive, so theta is large and negative. Our delta neutral portfolio
has positive gamma and negative theta. That means if the price does not move, the
passage of time will result in a lower portfolio value unless we rebalance. So the
portfolio does not provide an arbitrage opportunity.
Vega
For European options: v =
ac = ap =Se-Y' ./;N '( d
aa aa
1)
At-the-money options are most sensitive to volatility change, so they have higher vegas
than either in-the-money or out-of-the-money options. The vegas of all options decrease
as time to expiration becomes shorter ( J-:;
0 ) since a long-term option is more
sensitive to change in volatility.
A. Explain implied volatility and volatility smile. What is the implication of volatility
156
A Practical Guide To Quantitative Finance Interviews
smile for the Black-Scholes pricing model?
Solution: Implied volatility is the volatility that makes the model option price equal to
the market option price. Volatility smile describes the relationship between the implied
volatility of the options and the strike prices for a given asset. For currency options,
implied volatilities tend to be higher for in-the-money and out-of-the-money options
than for at-the-money options. For equity, volatility often decreases as the strike price
increases (also called volatility skew). The Black-Scholes model assumes that the asset
price follows a lognormal distribution with constant volatility. In reality, volatilities are
neither constant nor deterministic. In fact, the volatility is a stochastic process itself.
Furthermore, there may be jumps in asset prices.
B. You have to price a European call option either with a constant volatility 30% or by
drawing volatility from a random distribution with a mean of 30%. Which option would
be more expensive?
Solution: Many would simply argue that stochastic volatility makes the stock price
more volatile, so the call price is more valuable when the volatility is drawn from a
random distribution. Mathematically, the underlying argument is that the price of a
European call option is a convex function of volatility and as a result
e ( E[ al) ::;; E [ e( a)], where a is the random variable representing volatility and e is the
call option price. Is the underlying argument correct? It's correct in most, but not all,
cases. If the call price e is always a convex function of a, then
Ba
0. Be is the
Ba
Vega of the option. For a European call option,
v = -Be = Sv,TN ,(d1 ) =
Ba
'\/2tr
exp (-d12 I 2 ) .
The secondary partial derivative
B2e =
exp(-d12 I 2) d1d2
Ba 2 .&
a
=
Ba
is called Volga. For a European call option,
v d1d2.
a
v is always positive. For most out-of-the-money call options, both d1 and d 2 are
negative; for most in-the-money call options, both d1 and d 2 are positive. So d 1d 2 > 0
in most cases and e is a convex function of a when d 1d 2 > 0. But theoretically, we can
have conditions that d1 > 0 and d2 < 0 and
Ba
< 0 when the option is close to being
157
Finance
at-the-money. So the function is not always convex. In those cases, the option with
constant volatility may have a higher value.
C. The Black-Scholes formula for non-dividend paying stocks assumes that the stock
follows a geometric Brownian motion. Now assume that you don't know the stochastic
process followed by the stock price, but you have the European call prices for all
(continuous) strike prices K. Can you determine the risk-neutral probability density
function of the stock price at time T?
Solution: The payoff a European call at its maturity date is Max( ST - K, 0). Therefore
i (s -K)fsr (s)ds, where
under risk-neutral measure, we have c = e-rr JK
fsr (s)
is the
probability density function of ST under the risk-neutral probability measure. Taking the
first and second derivatives of c with respect to K, 10 we have
ac
aK
= e-rr
F (s-K)fs (s)ds
aKk
r
=e-rr f'a(s-K) f. (s)ds-e-rr(K-K)xl
JK aK
Sr
=e-rr J; -fsr (s)ds
and 82 c
aK 2
=
aK aK
= e-rT
F -f. (s)ds = e-rT f. (K).
aK JK
Sr
Sr
Hence the risk-neutral probability density function is f.s (K) =err 82 c2 .
r
aK
6.3. Option Portfolios and Exotic Options
In addition to the pricing and properties of vanilla European and American options, you
may be expected to be familiar with the construction and payoff of basic option-based
trading strategies---covered call, protective put, bull/bear spread, butterfly spread,
straddle, etc. Furthermore, if you are applying for a derivatives-related position, you
JO To calculate the derivatives requires the Leibniz integral rule, a fonnula for differentiating a definite
integral whose limits are functions of the differential variable:
a
af(x, z)
ab
aa
f(x,z)dx =
·
dx+ f(b(z),z)--f(a(z),z)az •(zl
(z)
az
az
az
r(z)
158
!(z)
A Practical Guide To Quantitative Finance Interviews
should also have a good understanding of pricing and hedging of some of the common
exotic derivatives-binary option, barrier option, Asian option, chooser option, etc.
Bull spread
What are the price boundaries for a bull call spread?
Solution: A bull call spread is a portfolio with two options: long a call c1 with strike K 1
and short a call c2 with strike K 2 (K1 < K 2 ). The cash flow of a bull spread is
summarized in table 6.3.
Cash flow
Time 0
Long
c1
Short
c2
Total
Matur!!Y_T
ST::; Kl
Kl< ST O';, both options are undervalued. When the market prices converge to the prices
with the realized volatility, both the call and the put will become more valuable.
Although initially a straddle with an at-the-money call and an at-the-money put ( K = S )
has a delta close to 0, as the stock price moves away from the strike price, the delta is no
longer close to 0 and the investor is exposed to stock price movements. So a straddle is
not a pure bet on stock volatility. For a pure bet on volatility, it is better to use volatility
swaps or variance swaps. 11 For example, a variance swap pays Nx(O'; -Kvar), where N
is the notional value, O'; is the realized variance and Kvar is the strike for the variance.
Binary options
What is the price of a binary (cash-or-nothing digital) European call option on a nondividend paying stock if the stock price follows a geometric Brownian motion? How
would you hedge a cash-or-nothing call option and what's the limitation of your hedging
strategy?
Solution: A cash-or-nothing call option with strike price K pays $1 if the asset price is
above the strike price at the maturity date, otherwise it pays nothing. The price of the
option is c8 = e r• N(d2 ) if the underlying asset is a non-dividend paying stock. As we
have discussed in the derivation of the Black-Scholes formula, N(d2 ) is the probability
that a vanilla call option finishes in the money under the risk-neutral measure. So its
discounted value is e-r• N(d2 ).
Theoretically, a cash-or-nothing call option can be hedged using the standard delta
hedging strategy. Since = ac8 = e-r• N '( d 2 ) 1 , a long position in a cash-or-
as
nothing call option can be hedged by shorting e-r• N '( d 2 )
shares (and a risk-free
SO' r
money market position). Such a hedge works well when the difference between S and K
is large and r is not close to 0. But when the option is approaching maturity T ( r
0)
11 For detailed discussion about volatility swaps, please refer to the paper "More Than You Ever Wanted
to Know about Volatility Swaps" by Kresimir Demeterfi, et al. The paper shows that a variance swap can
be approximated by a portfolio of straddles with proper weights inversely proportional to Ilk.
160
A Practical Guide To Quantitative Finance Interviews
and the stock price Sis close to K, "'1. is extremely volatile 12 and small changes in the
stock price cause very large changes in "'1.. In these cases, it is practically impossible to
hedge a cash-or-nothing call option by delta hedging.
We can also approximate a digital option using a bull spread with two calls. If call
options are available for all strike prices and there are no transaction costs, we can long
I I 2& call options with strike price K - & and short I I 2c call options with strike price
K + &. The payoff of the bull spread is the same as the digital call option if ST K - &
(both have payoff 0) or ST K + & (both have payoff $1 ). When K - & < ST < K + &,
their payoffs are different. Nevertheless, if we set &
0, such a strategy will exactly
replicate the digital call. So it provides another way of hedging a digital call option. This
hedging strategy suffers its own drawback. In practice, not all strike prices are traded in
the market. Even if all strike prices were traded in the market, the number of options
needed for hedging, I I 2&, will be large in order to keep c small.
Exchange options
How would you price an exchange call option that pays max ( ST,i - ST, 2 ,
0) at maturity.
Assume that S1 and S2 are non-dividend paying stocks and both follow geometric
Brownian motions with correlation p.
Solution: The solution to this problem uses change of numeraire. Numeraire means a
unit of measurement. When we express the price of an asset, we usually use the local
currency as the numeraire. But for modeling purposes, it is often easier to use a different
asset as the numeraire. The only requirement for a numeraire is that it must always be
positive.
The payoff of the exchange option depends on both ST I (price of SI at maturity date
and
sT,2
(price of s2 at
n
n, so it appears that we need two geometric Brownian motions:
dS1 = µ 1S1dt + o-1S 1dW,, 1
dS2 = µ 1 S2 dt + o-2S2dW,, 2
Yet if we use S1 as the numeraire, we can convert the problem to just one geometric
Brownian motion. The final payoff is max ( sT,2 - sT,I • 0) =
sT,I
-1, oJ. When
max (
T ,I
1
12
S
K and r
0 => ln(S I K)
0 => d 1
(r I a+ 0.5a)J";
0 => !!..
-n
e I
v2tr Savr
oo.
161
Finance
SI and s2 are geometrical Browian motions,
f = s 2 is a geometric Brownian motion as
SI
well. One intuitive explanation is that both In S1 and ln S2 follow normal distributions,
so ln f
=ln S2 -
ln S 1 follows a normal distribution as well and f follows a lognormal
f = S2
distribution. More rigorously, we can apply the Ito's lemma to
:
SI
dif= af dS + af dS +1._82f(dS)2+1._a21(dS )2+ a2f dSdS
1
2
asI 1 as2 2 2 asI2
asIas2 1 2
2 as22
s2
s2
s2 d
s2 d
2s 2d
s 2d
=-µI -dt0"1 -dW,.I + µ2 t + 0'2 W,.2 + 0'1 t- P0"10'2 t
SI
SI
SI
SI
SI
SI
= (µ2 - µ1 + 0"12 - P0"10"2) fdt - aJdW,,1 + D"2fdW,,2
= (µ2 - µ1+0-12 - P0"10"2 )fdt +
To make
and
s
___!2:_
-2pa10"2 +a; x fdW,,3
l
f =-S2
.
a martmgale,
set µ 2 - µ 1+ a 12 - pa1a 2 =0 and we have E-[ST2
-·SI
ST.I
=-S2 ,
SI
is a martingale under the new measure. The value of the exchange option using
S11
( ;,:: -1, 0
S, as the numeraire is C,
with underlying asset price S
=S
So
2 ,
SI
J
J
which is just the value ofa call option
strike price K
its
value
=1,
interest rate r
=0,
and volatility
where
IS
I
_ In ( S2 I S1 ) +
d _d _ r
h
d1 .Jr
and 2 - 1 av r. T e payoff of the exchange option
O's r
expressed in local currency is S 1Cs = S2 N(d1)-S1N(d2 ).
162
A Practical Guide To Quantitative Finance Interviews
6.4. Other Finance Questions
Besides option pricing problems, a variety of other quantitative finance problems are
tested in quantitative interviews as well. Many of these problems tend to be positionspecific. For example, if you are applying for a risk management job, prepare to answer
questions about VaR; for fixed-income jobs, get ready to answer questions about interest
rate models. As I explained in Chapter 1, it always helps if you grasp the basic
knowledge before the interview. In this section, we use several examples to show some
typical interview problems.
Portfolio optimization
You are constructing a simple portfolio using two stocks A and B. Both have the same
expected return of 12%. The standard deviation of A's return is 20% and the standard
deviation of B's return is 30%; the correlation of their returns is 50%. How will you
allocate your investment between these two stocks to minimize the risk of your portfolio?
Solution: Portfolio optimization has always been a crucial topic for investment
management firms. Harry Markowitz's mean-variance portfolio theory is by far the most
well-known and well-studied portfolio optimization model. The essence of the meanvariance portfolio theory assumes that investors prefer (I) higher expected returns for a
given level of standard deviation/variance and (2) lower standard deviations/variances
for a given level of expected return. Portfolios that provide the minimum standard
deviation for a given expected return are termed efficient portfolios. The expected return
and the variance of a portfolio with N assets can be expressed as
T
µP =w1µ1 +w2µ2 + ·· ·+wNµN =w µ
N
var(rP) = La}w;2 + Z:aifwiwJ
i=l
= wr L: w
i*}
where wi' Vi= 1,- · ·, N , is the weight of the i-th asset in the portfolio; µ;,Vi = I,···, N, is
the expected return of the i-th asset;
0";2
is the variance of i-th asset's return;
aif = pifa;a1 is the covariance of the returns of the i-th and thej-th assets and pif is their
correlation; w is an N x I column vector of w; 's; µ is an N x 1 column vector ofµ; 's;
L is the covariance matrix of the returns of N assets, an N x N matrix.
Since the optimal portfolio minimizes the variance of the return for a given level of
expected return, the efficient portfolio can be formulated as the following optimization
problem:
163
Finance
min wTLw
w
where e is an N x 1 vector with all elements equal to 1. 13
,
S .(. WTµ
= µ P, WT e = 1
For this specific problem, the expected returns are 12% for both stocks. So µP is always
12% no matter what wA and w8 ( wA + w8 = 1) are. The variance of the portfolio is
22
222
var(rP) =o-AwA
+a8 w8 + PA.BaAa8 wAw8
+a;(l-wA)2 +2pA,BO"AO"BWA(l-wA)
Taking the derivative of var(rP) with respect to wA and setting it to zero, we have
avar(rp)
2
- - - = 2aAwA
awA
2
-20-8(1-wA) +2pA.BO"AaB(l-wA)-2pA .BO"AO"BWA = 0
So we should invest 617 of the money in stock A and 1/7 in stock B.
Value at risk
Briefly explain what VaR is. What is the potential drawback of using VaR to measure
the risk of derivatives?
Solution: Value at Risk (VaR) and stress test---or more general scenario analysis-are
two important aspects of risk management. In the Financial Risk Manager Handbook, 14
VaR is defined as the following: VAR is the maximum loss over a target horizon such
that there is a low, pre-specified probability that the actual loss will be larger.
Given a confidence level a
a= [
VaR
(0, 1), the
E
VaR can be implicitly defined as
xf(x)dx, where x is the dollar profit (loss) and
f
(x) is its probability density
function. In practice, a is often set to 95% or 99%. VaR is an extremely popular choice
in financial risk management since it summarizes the risk to a single dollar number.
13
The optimal weights have closed form solution w* =AL 1e + y:L- 1µ, where
A.=
C-µ B
P
D
r
=
µA-8
I'
'
A
I
= e IL- e > 0,
B
= e IL
I
µ' c
= µ 'L
I
µ > 0, D
= Ac -
2
B .
D
Financial Ri.'>k Manager Handbook by Phillippe Jorion is a comprehensive book covering different
aspects of risk management. A classic book for VaR is Value at Risk, also by Philippe Jori on.
14
164
A Practical Guide To Quantitative Finance Interviews
Mathematically, it is simply the (negative) first or fifth percentile of the profit
distribution.
As a percentile-based measure on the profit distribution, VaR does not depend on the
shape of the tails before (and after) probability 1- a, so it does not describe the loss on
the left tail. When the profit/loss distribution is far from a normal distribution, as in the
cases of many derivatives, the tail portion has a large impact on the risk, and VaR often
does not reflect the real risk. 15 For example, let's consider a short position in a credit
default swap. The underlying asset is bond A with a $1 M notional value. Further assume
that A has a 3% default probability and the loss given default is 100% (no recovery).
Clearly we are facing the credit risk of bond A. Yet if we use 95% confidence level,
VaR(A) = 0 since the probability of default is less than 5%.
Furthermore, VaR is not sub-additive and is not a coherent measure of risk, which
means that when we combine two positions A and B to form a portfolio C, we do not
always have VaR(C) VaR(A)+ VaR(B). For example, if we add a short position in a
credit default swap on bond B with a $IM notional value. B also has a 3% default
probability independent of A and the loss given default is 100%. Again we have
VaR(B) = 0. When A and B form a portfolio C, the probability that at least one bond will
default becomes
So VaR(C)=$1M>VaR(A)+VaR(B).
Lack of sub-additivity directly contradicts the intuitive idea that diversification reduces
risk. So it is a theoretical drawback of VaR.
(Sub-additivity is one property of a coherent risk measure. A risk measure p(X) is
considered coherent if the following conditions holds: p(X + Y) p(X) + p(Y);
p(aX)=ap(X), Va>O;
if
and p(X+k)=p(X)-k for any
constant k. It is defined in Coherent Measure ofRisk by Artzner, P., et al., Mathematical
Finance, 9 (3):203-228. Conditional VaR is a coherent risk measure.)
Duration and convexity
The duration of a bond is defined as D
= __!_ dP,
p dy
where P is the price of the bond and y
is yield to maturity. The convexity of a bond is defined as C
Taylor's expansion, M
p
2
when
=
_!_ d 2
p dy
Applying
is small, M
p
For a fixed-rate bond with coupon rate c and time-to-maturity T:
15
Stress test is often used as a complement to VaR by estimating the tail risk.
165
Finance
Another important concept is dollar duration: $D = - dP = P x D. Many market
dy
participants use a concept called DVO 1: D VO 1 = -
dP
, which measures the
10,000x cry
price change when the yield changes by one basis point. For some bond derivatives,
such as swaps, dollar duration is especially important. A swap may have value P = 0, in
which case dollar duration is more meaningful than duration.
When n bonds with values P;, i = 1, ···, n, and Durations Di (convexities Ci) form a
portfolio, the duration of the portfolio is the value-weighted average of the durations of
n p
n p
n
the components: D = _!_Di ( C = _!_Ci), where P =LP;· The dollar duration of
Lp
i=I
Lp
i=I
i=I
n
the portfolio is simply the sum of the dollar durations of the components: $D = L$Di.
i=I
What are the price and duration of an inverse floater with face value $100 and annual
coupon rate 30% -3r that matures in 5 years? Assume that the coupons are paid
semiannually and the current yield curve is flat at 7.5%.
Solution: The key to solving basic fixed-income problems is cash flow replication. To
price a fixed-income security with exotic structures, if we can replicate its cash flow
using a portfolio of fundamental bond types such as fixed-rate coupon bonds (including
zero-coupon bonds) and floating-rate bonds, no-arbitrage arguments give us the
following conclusions:
Price of the exotic security= Price of the replicating portfolio
Dollar duration of the exotic security = Dollar duration of the replicating portfolio
To replicate the described inverse floater, we can use a portfolio constructed by shorting
3 floating rate bonds, which is worth $100 each, and longing 4 fixed-rate bonds with a
7. 5% annual coupon rate, which is worth $100 each as well. The coupon rate of a
floating-rate bond is adjusted every 0.5 years payable in arrear: the coupon rate paid at
t + 0.5 y is determined at t. The cash flows of both positions and the whole portfolio are
summarized in the following table. It is apparent that the total cash flows of the portfolio
are the same as the described inverse floater. So the price of the inverse float is the price
of the replicating portfolio: P;nverse = $100.
166
A Practical Guide To Quantitative Finance Interviews
YearO
Year 0.5
...
Short 3 floatingrate bonds
300
-150r0
...
Long 4 bonds with
7.5% COU_QOn rate
-400
15
...
15
Total
-100
15-150r0
...
30-300r.i
Cash flow
Year4.5
Year 5
-300-150r4.S
400+15
The dollar duration of the inverse floater is the same as the dollar duration of the
portfolio as well: $Dinverse = 4 x $Dfu:ed - 3 x $Dfloating. Since the yield curve is flat,
r0 = 7.5% and the floating-rate bond is always worth $103.75 (after the payment of
$3.75, the price of the floating-rate bond is $100) at year 0.5, and the dollar duration 16 is
103.75 0.5 - d(103.75/(l+y / 2)) _
$Djloating -0.5x
-48.19.
2 -lOOx
dy
(1+ y/2)
l+ y/2
c/ 2
lOO zr, w here Tis
· th e matunty
·
+
l=I (1+y I 2)'
(1+yI2)
of the bond. So the dollar duration of the fixed-rate bond is
· o f a fi1xed -rate bond is
· P=
The pnce
$D
fixed
So
= _ dP =
dy
$Dinverse
Dinverse
1
1+y/2
= 4 x $Dfu:ed -
('I.!._2 (1+y/2)'
c I 2 + lOOT J= 410.64.
(1+Y/2)
2T
l=I
3 x $Dfloating
=1498
and the duration of the inverse floater is
= $Dinverse / P;nverse =14.98.
Forward and futures
What's the difference between futures and forwards? If the price of the underlying asset
is strongly positively correlated with interest rates, and the interest rates are stochastic,
which one has higher price: futures or forwards? Why?
Solution: Futures contracts are exchange-traded standardized contracts; forward
contracts are over-the-counter agreements so they are more flexible. Futures contracts
are marked-to-market daily; forwards contacts are settled at the end of the contract term.
16
The initial duration of a floating rate bond is the same as the duration of a six-month zero coupon bond.
167
Finance
If the interest rate is deterministic, futures and forwards have the same theoretical price:
F = se b, the drift rate is negative; when R(t) < b, the drift rate is positive. So the
Vasicek model has the desirable property of mean-reverting towards long-term average
b. But with constant volatility, the interest rate has positive probability of being negative,
which is undesirable.
Cox-Ingersoll-Ross model: dR(t) =a( b-R(t)) dt
168
R(u) dW(t)
A Practical Guide To Quantitative Finance Interviews
The Cox-Ingersoll-Ross model keeps the mean-reversion property of the Vasicek model.
But the diffusion rate a)R(u) addresses the drawback of Vasicek model by
guaranteeing that the short rate is positive.
No-arbitrage short-rate models
Ho-Lee model: dr = B(t)dt+adz
The Ho-Lee model is the simplest no-arbitrage short-rate model where B(t) is a timedependent drift. B(t) is adjusted to make the model match the current rate curve.
Hull-White model: dR(t) =a( b(t)-R(t) )dt + adW(t)
The Hull-White model has a structure similar to the Vasicek model. The difference is
that b(t) is a time-dependent variable in the Hull-White model to make it fit the current
term structure.
169
Chapter 7 Algorithms and Numerical Methods
Although the percentage of time that a quant spends on programming varies with the job
function (e.g., quant analyst/researcher versus quant developer) and firm culture, a
typical quant generally devotes part of his or her time to implementing models through
programming. Therefore, programming skill test is often an inherent part of the
quantitative interview.
To a great extent, the programming problems asked in quantitative interviews are similar
to those asked in technology interviews. Not surprisingly, many of these problems are
platform- or language-specific. Although C++ and Java still dominate the market, we've
seen a growing diversification to other programming languages such as Matlab, SAS, SPlus, and R. Since there are many existing books and websites dedicated to technology
interviews, this chapter will not give a comprehensive review of programming problems.
Instead, it discusses some algorithm problems and numerical methods that are favorite
topics of quantitative interviews.
7. 1. Algorithms
In programming, the analysis of algorithm complexity often uses asymptotic analysis
that ignores machine-dependent constants and studies the running time T(n) -the
number of primitive operations such as addition, multiplication, and comparison-as the
number of inputs n - oo. 1
Three of the most important notations in algorithm complexity are big- 0 notation,
notation and 0 notation:
O(g(n)) =
all n
{
Q
/(n): there exist positive constants c and n0 such that 0 5: f(n) 5: cg(n) for
n0 } • It is the asymptotic upper bound of f (n ).
n(g(n)) = { f(n): there exist positive constants c and n0 such that 0 5: cg(n) 5: f(n) for
all n
n0 } • It is the asymptotic lower bound of f (n ).
E>(g(n))
=
{
/(n): there exist positive constants c1, c 2 , and n0 such that
c1g(n) 5: f(n) 5: c2 g(n) for all n
n0 }. It is the asymptotic tight bound of f(n).
Besides notations, it is also important to explain two concepts in algorithm complexity:
1 If you want to review basic algorithms, I highly recommend "Introduction to Algorithm" by Thomas H.
Cormen, Charles E. Leiserson, Ronald L. Rivest and Clifford Stein. It covers all the theories discussed in
this section and includes many algorithms frequently appearing in interviews.
Algorithms and Numerical Methods
Worst-case running time W(n): an upper bound on the running time for any n inputs.
Average-case running time A(n): the expected running time if the n inputs are
randomly selected.
For many algorithms, W(n) and A(n) have the same O(g(n)). But as we will discuss
in some problems, they may well be different and their relative importance often
depends on the specific problem at hand.
A problem with n inputs can often be split into a subproblems with n I b inputs in each
subproblem. This paradigm is commonly called divide-and-conquer. If it takes f(n)
primitive operations to divide the problem into subproblems and to merge the solutions
of the subproblems, the running time can be expressed as a recurrence equation
T(n) = aT(n I b) + f(n), where a 2:: 1, b > 1, and f(n) 0.
The master theorem is a valuable tool in finding the tight bound for recurrence
equation T(n)=aT(nlb)+f(n): If f(n)=O(n 10gba-c) for some constant e>O,
e( n 10gba), since f(n) grows slower than n 10gba. If f(n) =e( n logk n) for
some k 0, T(n) = e( n 10gha logk+I n), since /(n) and n g grow at similar rates. If
T(n) =
10g 6 a
10 6 a
f(n)
=n( n 10gha+c) for
c 0, and af(n I b)
cf(n) for some constant
= 0(/(n)), since f(n) grows faster than n 10gha.
Let's use binary search to show the application of the master theorem. To find an
element in an array, if the numbers in the array are sorted (a, a2 $ · · · an), we can use
binary search: The algorithm starts with alntlJ' If al nt 2J = x, the search stops. If
alnt 2J > x, we only need to search a"· .. ,aln 12_ 1J. If alnt 2J < x, we only need to search
alntl+IJ,. .. ,an. Each time we can reduce the number of elements to search by half after
making one comparison. So we have a
f(n)
= e( n10g21 log 0 n)
=
I, b
=
2, and
f (n) = I.
Hence,
and the binary search has complexity 0(1ogn).
Number swap
How do you swap two integers, i and j, without using additional storage space?
Solution: Comparison and swap are the basic operations for many algorithms. The most
common technique for swap uses a temporary variable, which unfortunately is forbidden
in this problem since the temporary variable requires additional storage space. A simple
172
A Practical Guide To Quantitative Finance Interviews
mathematic approach is to store the sum of i and} first, then extract i's value and assign
it to j and finally assign j's value to i. The implementation is shown in the following
code: 2
void swap(int &i I
int &j)
{
i
i + j;
//st o re the sum o f i and j
j
i - j;
//chang e j
to i's va lue
i
i - j;
/ / change i
to j's va lue
An alternative solution uses bitwise XOR (") function by taking advantage of the fact
that x " x = 0 and 0 " x = x:
void swap(int &i I
int &j) {
i
i
j;
j
j
i; // j
i " (j " i)
i
i
i
j ; Iii
(i " j)
j
A
i
Unique elements
If you are given a sorted array, can you write some code to extract the unique elements
from the array? For example, if the array is [1, 1, 3, 3, 3, 5, 5, 5, 9, 9, 9, 9], the unique
elements should be [1, 3, 5, 9].
Solution: Let a be an n-element sorted array with elements a 0
Whenever
we encounter a new element a; in the sorted array, its value is different from its
a1
···
an- I.
previous element (a; * aH ). Using this property we can easily extract the unique
elements. One implementation in C++ is shown as the following function: 3
template vector unique(T a[], int n)
{
vector vec; / / v ector used to avoid resi z i n g problem
vec.reserve(n); //re s erver to a vo id r ea ll oca tion
vec . push_back(a[O]);
for(int i=l; i(n)
primitive operations. The running time T(n) follows the following recursive function:
T(n) = {2T(n/ 2) + E>(n),
1,
if n > 1
if n = 1 ·
Applying the master theorem to T(n) with a= 2, b = 2, and f(n)
f(n)
= e( n10gba log 0 n ).
So T(n)
= E>(n logn).
= E>(n),
we have
For merge sort, A(n) and W(n) are the
same as T(n).
Quicksort: Quicksort is another recursive sorting method. It chooses one of the
elements, A;, from the sequence and compares all other values with it. Those elements
smaller than A; are put in a subarray to the left of A;; those elements larger than A; are
put in a subarray to the right of A;. The algorithm is then repeated on both subarrays
(and any subarrays from them) until all values are sorted.
175
Algorithms and Numerical Methods
In the worst case, quicksort requires the same number of comparisons as the insertion
sort. For example, if we always choose the first element in the array (subarray) and
compare all other elements with it, the worst case happens when A1, ···,An are already
sorted. In such cases, one of the subarray is empty and the other has n - 1 element. Each
step only reduces the subarray size by one. Hence, W ( n) = El (
t, i)
= 0(n2 ).
To estimate the average-case running time, let's assume that the initial ordering is
random so that each comparison is likely to be any pair of elements chosen from
A1 , ···,An. If we suspect that the original sequence of elements has a certain pattern, we
can always randomly permute the sequence first with complexity 0(n) as explained in
-
-
the next problem. Let AP and Aq be the pth and qth element ( 1 :::; p < q
n) in the final
-
-
-
sorted array. There are q - p + 1 numbers between AP and Aq. The probability that AP
-
-
-
.......
and Aq is compared is the probability that Aq is compared with AP before Ap+i• · ·, or
Aq-i
is compared with either AP or
Aq
(otherwise, AP and
Aq
are separated into
different subarrays and will not be compared), which happens with probability
P(p,q) =
2
q- p +I
(you can again use the symmetry argument to derive this probablity).
The total expected number of comparison is A(n) =
n
q-1
=
n
q-1 (
2
)
q- p+ 1
= 8(nlgn).
Although theoretically quicksort can be slower than merge sort in the worst cases, it is
often as fast as, if not faster than, merge sort.
Random permutation
A. If you have a random number generator that can generate random numbers from
either discrete or continuous uniform distributions, how do you shuffle a deck of 52
cards so that every permutation is equally likely?
Solution: A simple algorithm to permute n elements is random permutation by sorting. It
assigns a random number to each card and then sorts the cards in order of their assigned
random numbers. 4 By symmetry, every possible order (out of n! possible ordered
sequences) is equally likely. The complexity is determined by the sorting step, so the
If we use the continuous uniform distribution, theoretically any two random numbers have zero
probability of being equal.
4
176
A Practical Guide To Quantitative Finance Interviews
running time is E>(n log n). For a small n, such as n = 52 in a deck of cards, the
complexity E>(nlogn) is acceptable. For large n, we may want to use a faster algorithm
known as the Knuth shuffle. For n elements A[l], · · ·, A[n], the Knuth shuffle uses the
following loop to generate a random permutation:
for (i=l ton) swap(A[i], A[Random(i, n)]),
where Random(i, n) is a random number from the discrete uniform distribution between
i and n.
The Knuth shuffle has a complexity of E>(n) and an intuitive interpretation. In the first
step, each of the n cards has equal probability of being chosen as the first card since the
card number is chosen from the discrete uniform distribution between 1 and n; in the
second step, each of the remaining n - 1 cards elements has equal probability of being
chosen as the second card; and so on. So naturally each ordered sequence has 1In!
probability.
B. You have a file consisting of characters. The characters in the file can be read
sequentially, but the length of the file is unknown. How do you pick a character so that
every character in the file has equal probability of being chosen?
Solution: Let's start with picking the first character. If there is a second character, we
keep the first character with probability 1/2 and replace the pick with the second
character with probability 112. If there is a third character, we keep the pick (from the
first two characters) with probability 2/3 and replace the pick with the third character
with probability 1/3. The same process is continued until the final character. In other
words, let Cn be the character that we pick after we have scanned n characters and the
(n + l)th character exists, the probability of keeping the pick is _n_ and the probability
n+l
of switching to the (n + l)th character is - 1-. Using simple induction, we can easily
n+l
prove that each character has 1/ m probability of being chosen if there are m characters.
Search algorithm
A. Develop an algorithm to find both the minimum and the maximum of n numbers
using no more than 3n/2 comparisons.
Solution: For an unsorted array of n numbers, it takes n-1 comparisons to identify
either the minimum or the maximum of the array. However, it takes at most 3n/2
comparisons to identify both the minimum and the maximum. If we separate the
elements to n/2 pairs, compare the elements in each pair and put the smaller one in group
177
Algorithms and Numerical Methods
A and the larger one in group B. This step takes n I 2 comparisons. Since the minimum
of the whole array must be in group A and the maximum must be in group B, we only
need to find the minimum in A and the maximum in B, either of which takes n I 2 - 1
comparisons. So the total number of comparisons is at most 3n/2. 5
B. You are given an array of numbers. From the beginning of the array to some position,
all elements are zero; after that position, all elements are nonzero. If you don't know the
size of the array, how do you find the position of the first nonzero element?
Solution: We can start with the 1st element; if it is zero, we check the 2nd element; if the
2nd element is zero, we check the 4th element. .. The process is repeated until the ith
.
2; + 2i-l
step when the 2' th element is nonzero. Then we check the
th element. If it is
2
2; + 2i-l
zero, the search range is limited to the elements between the
th element and
2
the 2; th element; otherwise the search range is limited to the elements between the
.
2; + 2i-l
2'- 1th element and the
th element. .. Each time, we cut the range by half. This
2
method is basically a binary search. If the first nonzero element is at position n, the
algorithm complexity is E>(log n ).
C. You have a square grid of numbers. The numbers in each row increase from left to
right. The numbers in each column increase from top to bottom. Design an algorithm to
find a given number from the grid. What is the complexity of your algorithm?
Solution: Let A be an n x n matrix representing the grid of numbers and x be the number
we want to find in the grid. Begin the search with the last column from top to bottom:
A 1.n, · · ·, An.n· If the number is found, then stop the search. If An.n < x, x is not in the grid
and the search stops as well. If
A;,n
< x < A;+i.n• then we know that all the numbers in
rows l, · · ·, i are less than x and are eliminated as well. 6 Then we search the (i + l)th row
from right to left. If the number is found in the (i + l)th row, the search stops. If
A 1,;+i > x, x is not in the grid since all the number in rows i + l and above are larger than
x. If
A;+i.J+I
> x > Ai+l.J' we eliminate all the numbers in columns j + 1,- · -,n. Then we
can search along column from
A;+i.J
towards
An.J
until we find x (or x does not exist in
Slight adjustment needs to be made if n is odd, but the upper bound 3n/2 still applies.
" i can be 0, which means x < A1.., in which case we can search the first row from right to left.
5
178
A Practical Guide To Quantitative Finance Interviews
the grid) or a k that makes
from
Ak+ 1 .
,}
towards
Ak+i i ...
'
Ak.J
< x < Ak+l,J and then we search left along the row k + 1
Using this algorithm, the search takes at most 2n steps. So
its complexity is O(n).
Fibonacci numbers
Consider the following C++ program for producing Fibonacci numbers:
int Fibonacci(int n)
if
(n
<=
0)
return O;
else if (n==l)
return l;
else
return Fibonacci(n-l)+Fibonacci(n-2);
If for some large n, it takes 100 seconds to compute Fibonacci(n), how long will it take
to compute Fibonacci(n+ 1), to the nearest second? Is this algorithm efficient? How
would you calculate Fibonacci numbers?
Solution: This C++ function uses a rather inefficient recursive method to calculate
Fibonacci numbers. Fibonacci numbers are defined as the following recurrence:
Po = 0, F; =I, F,. = Fn-1 + Fn- 2' Vn 'C::. 2
F has closed-formed solution Fn =
n
(1+v'5f-(1-J5f
r;
, which can be easily proven
2nv5
using induction. From the function, it is clear that
T(O) =I, T(l) = 1, T(n) = T(n-1) + T(n-2) +I.
So the running time is a proportional to a sequence of Fibonacci numbers as well. For a
r;
large n, (1- v :J y
0, so
T(n+l)
T(n)
J5 +I .
2
.
If it takes I 00 seconds to compute
Fibonacci(n), the time to compute Fibonacci(n+l) is
v'5+1
2
seconds. 7
7 12 '< _, 12
. .
. The
A(n
x A n ) x A if n IS odd
[ F,,+,
F,,
Fn ]
=[
'
multiplication of two 2x2 matrices has complexity 0(1). So T(n)=T(n/2)+0(1).
Applying the master theorem, we have T(n) = 0(logn).
Maximum contiguous subarray
Suppose you have a one-dimensional array A with length n that contains both positive
and negative numbers. Design an algorithm to find the maximum sum of any contiguous
j
subarray A[i,j] of A: V(i,j) = LA[x], l-5: i 5: j
n.
x=-i
Solution: Almost all trading systems need such an algorithm to calculate maximum runup or maximum drawdown of either real trading books or simulated strategies. Therefore
this is a favorite algorithm question of interviewers, especially interviewers at hedge
funds and trading desks.
The most apparent algorithm is an O(n 2 ) algorithm that sequentially calculates the
V(i,j) 's from scratch using the following equations:
j
V(i,i) = A[i] when j = i and V(i,j)
= LA[x] = V(i,j-1) + A[j] when
j >i.
x-i
As the V(i,j) 's are calculated, we also keep track of the maximum of V(i,j) as well as
the corresponding subarray indices i andj.
180
A Practical Guide To Quantitative Finance Interviews
A more efficient approach uses the divide-and-conquer paradigm. Let' s define
i
T(i)
= LA[x]
and T(O) = 0, then V(i,j)
= T(j)-T(i-1),
Vl -5, i -5, j -5, n . Clearly for
x=I
any fixed j, when T(i-1) is minimized, V(i,j) is maximized. So the maximum
subarray ending at} is Vmax =T(J)-Tmin where Tmin =min(T(l),- ·-,T(j-1)) . lfwe keep
track of and update Vmax and Tmin as j increases, we can develop the following O(n)
algorithm:
T = A[l]; Vmax = A[l]; Tmin =min(O,T)
For j
{ T
= 2 ton
=
T + A[j];
If T < Tmin , then Tmin = T;
}
Return Vmax;
The following is a corresponding C++ function that returns Vmax and indices i and j
given an array and its length:
double maxSubarray(double A[], int len, int &i, int &j)
double T=A[O], Vmax=A[O];
d o ubl e Tmin = min(O.O, T);
for(int k=l; k Vmax)
if
(T P; , we lose and the coin tossing
stops. If s; = P; , we continue to toss more coins. Some p values (e.g., I /3) are infinite
oo ). In these cases, the probability to
series when expressed as a binary number ( n
reach s; -:t:- P; is 1 as i increases. If the sequence is finite, (e.g., 114=0.0 I) and we reach
the final stage with s" = Pn , we lose (e.g., for 114, only the sequence 00 will be
classified as a win; all other three sequences 01, 10 and 11 are classified as a loss). Such
a simulation will give us probability p of winning.
Poisonous wine
You've got 1000 bottles of wines for a birthday party. Twenty hours before the party,
the winery sent you an urgent message that one bottle of wine was poisoned. You
happen to have IO lab mice that can be used to test whether a bottle of wine is poisonous.
The poison is so strong that any amount will kill a mouse in exactly 18 hours. But before
the death on the 18th hour, there are no other symptoms. Is there a sure way that you can
find the poisoned bottle using the 10 mice before the party?
Solution: If the mice can be tested sequentially to eliminate half of the bottles each time,
the problem becomes a simple binary search problem. Ten mice can identify the
poisonous bottle in up to 1024 bottles of wines. Unfortunately, since the symptom won't
show up until 18 hours later and we only have 20 hours, we cannot sequentially test the
mice. Nevertheless, the binary search idea still applies. All integers between 1 and 1000
can be expressed in 10-bit binary format. For example, bottle 1000 can be labeled as
1111101000 since 1000 = 29 + 2 8 + 2 7 + 2 6 + 2 5 + 2 3 .
Now let mouse I take a sip from every bottle that has a I in the first bit (the lowest bit
on the right); let mouse 2 take a sip from every bottle with a 1 in the second bit; ... ;and,
finally, let mouse 10 take a sip from every bottle with a 1 in the 10th bit (the highest bit).
Eighteen hours later, if we line up the mice from the highest to the lowest bit and treat a
live mouse as 0 and a dead mouse as I, we can easily back track the label of the
poisonous bottle. For example, if the 6th, 7th, and 9th mice are dead and all others are
alive, the line-up gives the sequence 0I01100000 and the label for the poisonous bottle
is 2 8 + 2 6 + 2 5 =352.
183
Algorithms and Numerical Methods
7.3 Numerical Methods
The prices of many financial instruments do not have closed-form analytical solutions.
The valuation of these financial instruments relies on a variety of numerical methods. In
this section, we discuss the application of Monte Carlo simulation and finite difference
methods.
Monte Carlo simulation
Monte Carlo simulation is a method for iteratively evaluating a deterministic model
using random numbers with appropriate probabilities as inputs. For derivative pricing, it
simulates a large number of price paths of the underlying assets with probability
corresponding to the underlying stochastic process (usually under risk-neutral measure),
calculates the discounted payoff of the derivative for each path, and averages the
discounted payoffs to yield the derivative price. The validity of Monte Carlo simulation
relies on the law of large numbers.
Monte-Carlo simulation can be used to estimate derivative prices if the payoffs only
depend on the final values of the underlying assets, and it can be adapted to estimate
prices if the payoffs are path-dependent as well. Nevertheless, it cannot be directly
applied to American options or any other derivatives with early exercise options.
A. Explain how you can use Monte Carlo simulation to price a European call option?
Solution: If we assume that stock price follows a geometric Brownian motion, we can
simulate possible stock price paths. We can split the time between t and T into N
equally-spaced time steps. 10 So D..t = T-t and t; =t+/)..txi, for i=0,1,2,···,N. We
N
then simulate the stock price paths under risk-neutral probability using equation
s. = s.t-1 e(r-u212 )+uJ6ic; '
I
where c.I 's are IID random variables from standard normal
distribution. Let's say that we simulate M paths and each one yields a stock price Sr.k'
where k = I, 2, · · ·, M, at maturity date T.
'° For European options, we can simply set N=l. But for more general options, especially the pathdependent ones, we want to have small time steps and therefore N should be large.
184
A Practical Guide To Quantitative Finance Interviews
The estimated price of the European call is the present value of the expected payoff,
M
:Lmax(Sr,k -K,0)
which can be calculated as C = e-r """k_.=1- - - - - M
B. How do you generate random variables that follow N(µ,a 2 ) (normal distribution
with mean µ and variance a 2 ) if your computer can only generate random variables
that follow continuous unifonn distribution between 0 and 1?
Solution: This is a great question to test the basic knowledge of random number
generation, the foundation of Monte Carlo simulation. The solution to this question can
be dissected to two steps:
1. Generate random variable of x N(O,l) from uniform random number generator
using inverse transform method and rejection method.
2. Scale x to µ + ax to generate the final random variables that follow N (µ, a
2) •
The second step is straightforward; the first step deserves some explanations. A popular
approach to generating random variables is the inverse transform method: For any
continuous random variable X with cumulative density function F ( U = F(X) ), the
random variable X can be defined as the inverse function of U: X = F-1(U), 0,:::;; U ,: :; 1.
It is obvious that X = F- 1(U) is a one-to-one function with 0,:::;; U,:::;; 1. So any
continuous random variable can be generated using the following process:
•
Generate a random number u from the standard uniform distribution.
•
Compute the value x such that u = F(x) as the random number from the
distribution described by F.
For this model to work, F- 1(U) must be computable. For standard normal distribution,
U =F(X) = (
l e -x
r::;-
212
dx. Th e
·
mverse
funchon
.
has no
. 1 so lut10n.
.
analyhca
v2tr
Theoretically, we can come up with the one-to-one mapping of X to U as the numeric
solution of ordinary differential equation F'(x)
integration method such as the Euler method. 11
the rejection method:
11
To integrate y
= f(x) =
tb-
e-x212 using numerical
v2tr
Yet this approach is less efficient than
= F(x) with first derivative y' = f (x) and a known initial value y
0
=
F(x0 ) , the Euler
method chooses a small step size h ( h can be positive or negative) to sequentially approximate y values:
185
Algorithms and Numerical Methods
Some random variables have pdf f(x), but no analytical solution for F- 1 (U). In these
cases, we can use a random variable with pdf g(y) and Y = c-'(U) to help generate
random variables with pdf /(x). Assume that Mis a constant such that
f
(y)
g(y)
M, 'Vy.
We can implement the following acceptance-rejection method:
•
Sampling step: Generate random variable y from g(y) and a random variable v
from standard uniform distribution [O, 1] .
•
Acceptance/rejection step: If v
f(y) , accept x
Mg(y)
= y;
otherwise, repeat the
sampling step. 12
An exponential random variable (g(x)=A-e--<..-) with A-=l has cdfu=G(x)=l-e-x.
So the inverse function has analytical solution x =- log(l - u) and a random variable
with exponential distribution can be conveniently simulated. For standard normal
distribution, f(x)
f(x)
g(x)
=
/2ex-x212
·{;'
=
...;2;r
<
e-·· 212 ,
v;
v;
/2e-(x-1)2 12+112
/2e1 12::::::::
1.32, 'VO< X < 00
So we can choose M = 1.32 and use the acceptance-rejection method to generate
x - N(O, 1) random variables and scale them to N(µ,a 2 ) random variables.
C. Can you explain a few variance reduction techniques to improve the efficiency of
Monte Carlo simulation?
Solution: Monte Carlo simulation, in its basic form, is the mean of IID random variables
-
estimator f
l
M
Since the expected value of each Y; is unbiased, the
M i=I
is unbiased as well. If Var(Y) =a and we generate IID f;, then
t;,.Y;,-··,YM: Y
Var(Y) =a I JM, where Mis the number of simulations. Not surprisingly, Monte Carlo
F(x0 + h) = F(x0 ) + f(x0 ) x h, F(x 0 + 2h) = F(x0 + h) + f(x 0 + h) x h, · · ·. The initial value of the cdfof
a standard normal can be F(O) = 0.5.
12
P(X
186
x)
ex:
f(y)
I' g(y)-·--dy
Mg(y)
x
=
M
I' f(y)dy => F(x)
'
=
P(X x)
P(X < oo)
=
I' f(y)dy
'
A Practical Guide To Quantitative Finance Interviews
simulation is computationally intensive if a is large. Thousands or even millions of
simulations are often required to get the desired accuracy. Depending on the specific
problems, a variety of methods have been applied to reduce variance.
Antithetic variable: For each series of e; 's, calculate its corresponding payoff
Y( e,, · · ·, eN) . Then reverse the sign of all e; 's and calculate the corresponding payoff
Y(-e,,-··,-eN). When Y(e,,···,eN) and Y(-e" .. ·,-eN) are negatively correlated, the
variance is reduced.
Moment matching: Specific samples of the random variable may not match the
population distribution well. We can draw a large set of samples first and then rescale
the samples to make the samples' moments (mean and variance are the most commonly
used) match the desired population moments.
Control variate: If we want to price a derivative X and there is a closely related
derivative Y that has an analytical solution, we can generate a series of random numbers
and use the same random sequences to price both X and Y to yield X and Y. Then X can
be estimated as X + (Y - i). Essentially we use (Y - f) to correct the estimation error of
x.
A
Importance sampling: To estimate the expected value of h(x) from distribution/(x),
instead of drawing x from distribution/(x), we can draw x from distribution g(x) and
use Monte Carlo simulation to estimate expected value of h(x)f(x):
g(x)
E
x
f<
(x)]·
[h(x)] = Jh(x)f(x)dx = Jh(x)f(x) g(x)dx = E x [ h(x)f
g(x)
g< >
g(x)
>
13
If h(x)f(x) has a smaller variance than h(x), then importance sampling can result in a
g(x)
more efficient estimator. This method is better explained using a deep out-of-the-money
option as an example. If we directly use risk-neutral f (ST) as the distribution, most of
the simulated paths will yield h(ST) = 0 and as a result the estimation variance will be
large. If we introduce a distribution g(ST) that has much wider span (fatter tail for ST),
more simulated paths will have positive h(ST). The scaling factor f(x) will keep the
g(x)
estimator unbiased, but the approach will have lower variance.
13
Importance sampling is essentially a variance reduction method using a change of measure.
187
Algorithms and Numerical Methods
Low-discrepancy sequence: Instead of using random samples, we can generate a
deterministic sequence of "random variable" that represents the distribution. Such lowdiscrepancy sequences may make the convergence rate 1IM.
D. If there is no closed-form pricing formula for an option, how would you estimate its
delta and gamma?
Solution: As we have discussed in problem A, the prices of options with or without
closed-form pricing formulas can be derived using Monte Carlo simulation. The same
methods can also be used to estimate delta and gamma by slightly changing the current
underlying price from S to S ±JS, where JS is a small positive value. Run Monte
Carlo simulation for all three starting prices S - JS, S and S + JS, we will get their
corresponding option prices /(S-JS), /(S) and /(S +JS).
.
Estimated delta:
!:!,.
/(S+JS)-/(S-JS)
= -Jf =------JS
2JS
.
d
(/(S +JS)- f(S) )-(/(S)- /(S -JS))
Estimate gamma: f' =
2
JS
To reduce variance, it's often better to use the same random number sequences to
estimate /(S-JS), /(S) and f(S +JS). 14
E. How do you use Monte Carlo simulation to estimate ;r?
Solution: Estimation of ;r is a classic example of Monte Carlo simulation. One standard
method to estimate ;r is to randomly select points in the unit square (x and y are
independent uniform random variables between 0 and 1) and determine the ratio of
points that are within the circle x 2 + y 2 s 1. For simplicity, we focus on the first quadrant.
As shown in Figure 7.1, any points within the circle satisfy the equation x;2 + l s 1. The
percentage of the points within the circle is proportional to its area:
p = Number of ( X;, y;)
within X;2 + y} s 1 = I/ 4;r =
Number of (xpy;) within the square
1x1
4
= 4p .
So we generate a large number of independent (x, y) points, estimate the ratio of the
points within the circle to the points in the square, and multiply the ratio by 4 to yield an
estimation of Jr. Figure 7.1 uses only 1000 points for illustration. With today's
14
The method may not work well if the payoff function is not continuous.
188
A Practical Guide To Quantitative Finance Interviews
computing power, we can easily generate millions of (x, y) pairs to estimate ff with
good precision. 1,000 simulations with 1,000,000 (x, y) points each using Matlab took
less than 1 minute on a laptop and gave an average estimation of ff as 3.1416 with
standard deviation 0.0015.
...,.-.. ,,,. .......
. ..
, .....
,.. .. ..' ..
•
0.9
•
&.'
•
!t. •
0.8
•
0.7
.,
;•
.......
, .....-._
., . ... .
•• ·<-.
••I
•
fL_t.
I
•••••• •
0.6
•• :f:·
•
>- 0.5
0.4
0.3
0.2
•
0.1 .
0
0
0.2
0.6
0.4
0.8
x
Figure 7 .1 A Monte Carlo simulation method to estimate rr
Finite difference method
The finite difference method is another popular numerical technique for derivative
pricing. It numerically solves a differential equation to estimate the price of a derivative
by discretizing the time and the price of the underlying security. We can convert the
Black-Scholes-Merton equation, a second order nonlinear partial differential equation, to
a heat diffusion equation (as we did in Chapter 6). This new equation, expressed as a
function of r (time to maturity) andx (a function of the price of the underlying security),
is a general differential equation for derivatives. The difference between various
derivatives lies in the boundary conditions. By building a grid of x and r and using the
boundary conditions, we can recursively calculate u at every x and r using finite
difference methods.
A. Can you briefly explain finite difference methods?
Solution: There are several version of finite difference methods used in practice. Let's
briefly go over the explicit difference method, the implicit difference method and the
189
Algorithms and Numerical Methods
Crank-Nicolson method. As shown in Figure 7.2, if we divide the range of r, [O,
T],
N discrete intervals with increment /1 r = TI N and divide the range of x, [ x0 , x J
] ,
into
into
J discrete intervals with increment l1x =(xJ -x0 ) I J, the timer and the space of x can be
expressed as a grid of rn, n = l, ···, N and x1 , j =I,··., J .
x
•
x.I.
n
un+/
j+/
un
Un.+/
u'!
.1- I
un+I
.I
.I
j-1
X;-I
...
Xo
0
Figure 7.2 Grid of rand x for finite different methods
The explicit difference method uses the forward difference at time rn and the second8u
order central difference at x1.: -
Br
Un+I -Un
1
11r
Un+I -2Un +Un_I
1 -
-
1
1
(11x) 2
1
82U
-
-
8x 2
•
Rearranging terms, we can express un+i
as a linear combination of un+"
unJ and un_
:
J
J
J 1
un+I
=
J
J
conditions
190
+(1-2a)unJ
u;, and
+ aun+I'
where
.I
a= 11t /(11x) 2 • Besides, we often have boundary
for all n =I, · ·., N; j = 0, · ·., J . Combining the boundary
A Practical Guide To Quantitative Finance Interviews
conditions and equation u;+' =au;_, +(1-2a)u; +au;+,, we can estimate all u;'s on
the grid.
The implicit difference method uses the backward difference at time tn+i and the
au
second-order central difference at x. : - :::::
J
ar
n+I
n
1
1
u . -u.
!l.r
The Crank-Nicolson method uses the central difference at time (t" +tn+ 1 ) 12 and the
second-order central difference at x 1 :
aU
a,
-::::::
Un+I
-U";
;
!l.r
} ( u";+I -2u"; +Un;-I
=-
2
(/l.x) 2
+
Un+I
-2Un+I
;+I
;
(/l.x) 2
+ Un+I
;-1
J:::::::-a2
U
8x 2
•
B. If you are solving a parabolic partial differential equation using the explicit finite
difference method, is it worse to have too many steps in the time dimension or too many
steps in the space dimension?
Solution:
The equation for
the explicit finite difference method is
u;+' =au;_,+(1-2a)u;+au;+" where a=l1t/(/l.x) 2 • For the explicit finite difference
method to be stable, we need to have 1- 2a > 0 l1t /(!l.x) 2 < 1I2. So a small l1t (i.e.,
many time steps) is desirable, but a small !l.x (too many space steps) may make
l1t /(/l.x) 2 > 1I2 and the results unstable. In that sense, it is worse to have too many steps
in the space dimension. In contrast, the implicit difference method is always stable and
convergent.
191
Index
absorbing Markov chain, 106
absorbing state, 113
absorption probability, 107
algorithm complexity, 171
analytical skills, 9
antithetic variable, 187
average-case running time, 172
Bayes' Formula, 73
binary option, 160
binomial theorem, 65, 71
bisection method, 45
bitwise XOR, 173
Black-Scholes formula, 143
Black-Scholes-Merton differential
equation, 142
boundary condition, 115
Brownian motion, 129
bull spread, 159
Cartesian integral, 41
chain rule, 33, 34
characteristic equation, 54
Cholesky decomposition, 57
coherent risk measure, 165
combination, 65
conditional probability, 68, 72, 75, 83
continuous distribution, 87
control variate, 187
convex function, 140
convexity, 165
correlation, 92
covariance,92
Cox-Ingersoll-Ross model, 168
Crank-Nicolson method, 191
cross-sectional area, 3 8
delta, 149
derivative, 33, 35
determinant, 53
diagonalizable, 54
discounted Feynman-Kac equation, 143
discrete distribution, 86
divide-and-conquer, 180
dollar duration, 166
duration, 165
dynamic programming, 121
dynamic programming algorithm, 122
eigenvalue, 54
eigenvector, 54
European put, 13 7
event, 60, 63
exchange option, 161
expected time to absorption, 110
expected times to absorption, 107
expected value, 86
explicit difference method, 190
exponential martingale, 129
Feynman-Kac equation, 134
Fibonacci numbers, 179
finite difference method, 189
first passage time, 131
first-order differential linear equation,
47
fixed-rate coupon bond, 166
floating-rate bond, 166
forwards, 167
futures, 167
Gamma, 149
general chain rule, 40
generalized power rule, 33
heat equation, 146
Ho-Lee model, 168
homogenous linear equation, 48
Homer's algorithm, 174
Hull-White model, 168
implicit difference method, 191
importance sampling, 187
Inclusion-Exclusion Principle, 65
independence, 73
induction, 27, 29
insertion sort, 175
integration, 36
193
integration by parts, 37
integration by substitution, 37, 40
interest rate model, 168
intersection, 60
inverse floater, 166
Ito's lemma, 13 5
Jensen's inequality, 140
jump-diffusion process, 90
L'Hospital's rule, 36
Lagrange multipliers, 45
law of total expectation, 93, 113
Law of total probability, 73
linear least squares, 52
linear regression, 53
logic, 6
low-discrepancy sequence, 188
LU decomposition, 57
Markov chain, 105
Markov property, 105, 114
mark-to-market, 168
martingale, 115
master theorem, 172
maximum, 35
maximum drawdown, 180
merge sort, 175
minimum, 35
module, 26
modulo operation, 23
moment generating function, 91
moment matching, 187
Monte Carlo simulation, 184
moving average, 174
multiplication rule:, 72
mutually exclusive, 60, 63
Newton's method, 44
Newton-Raphson method, 44
nonhomogeneous linear equation, 49
normal distribution, 91
numerical method, 184
order statistics, 99
orting algorithm, 174
out of the box, 3, 12
194
outcome, 59
partial derivative, 40
partial differential equations, 146
permutation, 65
Pigeon Hole Principle, 20, 21
Poison process, 90
Poisson process, 90
polar integral, 41
portfolio optimization, 163
positive definite, 56
positive semidefinite, 56
principle of counting, 64
Principle of Optimality, 122
probability density function, 41, 86
probability mass function, 86
probability space, 59
product rule, 34, 37
product rule:, 33
proof by contradiction, 31
put-call parity, 138
QR decomposition, 52
quicksort, 175
quotient rule, 33, 37
random permutation, 176
random variable, 60
random walk, 115
reflection principle, 118, 132
replicating portfolio, 166
Rho, 149
running time, 171
sample space, 59
secant method, 45
separable differential equation, 47
simplified version, 3, 4
singular value decomposition, 58
state space, 107
stopping rule, 116
straddle, 159
sub-additivity, 165
summation equation, 18
symmetric random walk, I 15
symmetry, 16
A Practical Guide To Quantitative Finance Interviews
system equation, 127
Taylor's series, 42, 43
Theta, 149
transition graph, I 05, I 09, 111
transition matrix, 105
union, 60
uropean call, 137
Value at Risk, 164
variance reduction, 186
Vasicek model, 168
vector, 51
Vega, 149
Volga, 157
Wald's Equality, 116
worst-case running time, 172
195
II
I I Ill llll 1111111111111111111111111111111
18555154R00119
Made in the USA
San Hernardino. C' A
19 January 2015
Source Exif Data:
File Type : PDF File Type Extension : pdf MIME Type : application/pdf PDF Version : 1.3 Linearized : No Page Count : 212 Title : Xinfeng Zhou-A Practical Guide To Quantitative Finance Interviews-CreateSpace Independent Publishing Platform (2008).djvu Author : j Producer : Mac OS X 10.12.4 Quartz PDFContext Create Date : 2017:04:28 03:31:26Z Modify Date : 2017:04:28 03:31:26ZEXIF Metadata provided by EXIF.tools