Quant, FM, and Data Science Interview Compilation

Aaron Cao

Contents

Introduction for LSU Students 1

Good Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

Interview Problems 3
Miscellaneous Statistics Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

Additional Problems on Glassdoor 14

Introduction for LSU Students

There is a growing popularity of financial mathematics among the LSU undergraduate population.
Many of these students are Mathematics majors concentrating in Actuarial Sciences or Statistics,
but there are also quite a number in the Finance and Computer Science departments. The topics
and problems compiled in this handout will invariably be encountered by students in all of these
fields, and it is important that enthusiasts of the quant/FM/data science career path have a firm
grasp of the technical interview
Throughout my undergraduate years, I was able to compile a good number of popular and
frequently-asked interview problems from both my own interview experiences and from friends
in the field. The problems listed here primarily come from top-tier hedge funds, trading firms,
software companies, and even graduate schools. These are attractive career choices for many of
the aforementioned students, and along all of these career paths one will inevitably face a technical
interview. I hope that the content provided in this handout will help many generations of LSU
students to prepare for such a task.
Though the problems in this handout are by no means exhaustive, they do provide a good road
map for the general knowledge that is expected from professionals in the industry. I provide link
solutions and occasionally written solutions to every problem, so some research on the readers’ part
to fill in knowledge gaps will be necessary. For baseline proficiency I would recommend knowing
the undergraduate curriculum from at least the first two introductory statistics courses, the data
structures course, the mathematical statistics course, and the stochastics course.
On a closing note, I encourage any LSU reader to participate in the Problem Solving Seminar
and Putnam Competition during their undergraduate years. Beyond the specialized mathematics
and coding skills, the most important ability any person in this industry should have is the ability
to solve problems, and honing one’s problem-solving ability pays large dividends throughout life.

1
Good Resources
• Art of Problem Solving (https://artofproblemsolving.com/)

• Brilliant (https://brilliant.org/)

• 3Blue1Brown (http://www.3blue1brown.com/)

• Mark S. Joshi (https://en.wikipedia.org/wiki/Mark_S._Joshi)

• Xinfeng Zhou (https://www.biblio.com/xinfeng-zhou/author/1230731)

• Competitive Programming (https://cpbook.net/)

• Cracking the Coding Interview (http://www.crackingthecodinginterview.com/)

• DarthPrince (http://codeforces.com/blog/entry/15729)

• Inishan (https://codeforces.com/blog/entry/23054)

• Matei Zaharia (https://cs.stanford.edu/~matei/programming_contests/)

For questions and concerns feel free to contact me at [email protected].

2
Interview Problems
1. Discuss algorithms for parallel matrix multiplication.

• https://cse.buffalo.edu/faculty/miller/Courses/CSE633/Ortega-Fall-2012-CSE633.pdf
• https://www.tutorialspoint.com/parallel algorithm/matrix multiplication.htm
• https://www3.nd.edu/ zxu2/acms60212-40212/Lec-07-3.pdf

2. Design a risk and asset pricing model for tech startup equity.

• https://www.startups.co/articles/startup-equity-101
• https://www.investopedia.com/terms/c/capm.asp
• https://www.investopedia.com/articles/personal-finance/050515/how-calculate-beta-
private-company.asp
• https://www.financierworldwide.com/are-we-pricing-private-equity-risk-properly/s

3. Discuss ordinary least squares (OLS), maximum likelihood (MLE), and maximum a posteriori
(MAP) estimation.

• https://en.wikipedia.org/wiki/Ordinary least squares

• https://en.wikipedia.org/wiki/Proofs involving ordinary least squares
• https://en.wikipedia.org/wiki/Maximum likelihood estimation
• https://en.wikipedia.org/wiki/Maximum a posteriori estimation
• https://stats.stackexchange.com/questions/80424/sum-of-squared-difference-
and-gaussian-noise-model/80564#80564

4. A horizontal stick is one metre long. Fifty ants are placed in random positions on the stick,
pointing in random directions. The ants crawl head first along the stick, moving at one metre
per minute. If an ant reaches the end of the stick, it falls off. If two ants meet, they both
change direction. How long do you have to wait to be sure that all the ants have fallen off
the stick?

• https://math.stackexchange.com/questions/1036902/interesting-question-on-ants
• https://math.stackexchange.com/questions/1418351/random-ants-probability-question

5. Design a data structure for storing an integer-weighted probability distribution. Discuss

runtime tradeoffs for different designs.

• https://www.geeksforgeeks.org/random-number-generator-in-arbitrary-probability-
distribution-fashion/
• https://softwareengineering.stackexchange.com/questions/150616/get-weighted-
random-item
• https://cs.stackexchange.com/questions/59690/is-this-probability-distribution-
data-structure-already-discovered

3
 6. Given a signal, which is regularly sampled over time and is “noisy”, how can the noise
be reduced while minimizing the changes to the original signal? The standard method
is with a Fourier transform. What is the intuition for why it works? Can one optimize
hyperparameters for terms of the transform that sit inside the sum?

• https://exnumerus.blogspot.com/2011/12/how-to-remove-noise-from-signal-using.html
Solution. The simple answer is no. Sine and cosine are periodic, so one can not use convex
optimization techniques.
7. There are N lions and 1 sheep in a field. All the lions really want to eat the sheep, but the
problem is that if a lion eats a sheep, it becomes a sheep. A lion would rather stay a lion
than be eaten by another lion. There is no other way for a lion to die than to become a
sheep and then be eaten. When is it safe for any lion to eat?

• https://math.stackexchange.com/questions/937410/understanding-the-solution-of-a-
riddle-about-lions-and-sheep

8. What are some pros and cons of using daily returns data versus monthly returns data?

• https://stats.stackexchange.com/questions/124404/why-monthly-stock-returns-
instead-of-daily-returns-in-multiple-regressions
• https://www.fields.utoronto.ca/programs/scientific/09-10/bachelier/ talks/Sat/
Varley/bfs80groth.pdf
• https://www.quora.com/Should-I-use-daily-monthly-or-yearly-returns-in-portfolio-
variance-calculations-when-calculating-relevant-means-variances-exces-returns-covar

9. What is the geometric interpretation of regression?

• https://www.datasciencecentral.com/profiles/blogs/linear-regression-geometry
• https://www.youtube.com/watch?v=oWuhZuLOEFY
• https://www.youtube.com/watch?v=PbyP3goun2Y
• https://www.youtube.com/watch?v=444ZkgiHI3Q
• https://en.wikipedia.org/wiki/Regression analysis

10. Calculate the correlation of two vectors and write the code in Python.

• https://en.wikipedia.org/wiki/Pearson correlation coefficient

• https://en.wikipedia.org/wiki/Correlation and dependence
• https://stackoverflow.com/questions/19428029/how-to-get-correlation-of-
two-vectors-in-python

11. Design a research that uses dividend cut as a variable of interest.

• https://www.investopedia.com/ask/answers/06/dividendpaymentcut.asp
• https://www.investopedia.com/trading/dividends-interest-rates-effect-stock-options/

12. A population dies out with p = 0.2. It remains stable with p = 0.5. It doubles with p = 0.3.
What is the expected long term behavior of the population?

4
 • https://en.wikipedia.org/wiki/Markov chain
• https://en.wikipedia.org/wiki/Stochastic matrix
• https:// drive.google.com/file/d/1BydxM4mZNc4rUHF5VAZQDshY8xbfWzmL/
view?usp=sharing
Solution. This is a classic Markov chain problem. A transition probability matrix solves this
easily.

13. Given a matrix of correlations, write an algorithm to cluster the stocks with correlation equal
to a certain value.

• https://www.cs.princeton.edu/sites/default/files/uploads/karina marvin.pdf
• https://arxiv.org/pdf/1511.07945.pdf
• http://www.diva-portal.org/smash/get/diva2:196577/FULLTEXT01.pdf
• https://en.wikipedia.org/wiki/Correlation clustering
• https://quant.stackexchange.com/questions/2263/how-to-cluster-stocks-and-
construct-an-affinity-matrix
• https://en.wikipedia.org/wiki/K-nearest neighbors algorithm
Solution. A basic idea is to use the k-nearest-neighbors algorithm.

14. Discuss correlation matrices with panel data.

• https://stackoverflow.com/questions/47886162/correlation-matrix-for-panel-data
• https://www.researchgate.net/post/correlation matrix for variables in panel data
• https://www.statalist.org/forums/forum/general-stata-discussion/general/1432235-
correlation-matrix-in-panel-data-model

15. Discuss problems with multiple inheritance in Java.

• http://www.lambdafaq.org/what-about-the-diamond-problem/
• https://www.geeksforgeeks.org/java-and-multiple-inheritance/
• https://www.journaldev.com/1775/multiple-inheritance-in-java
• https://javapapers.com/core-java/why-multiple-inheritance-is-not-supported-in-java/

16. What are the differences between Python lists and Java arrays?

• https://www.pythoncentral.io/the-difference-between-a-list-and-an-array/
• https://stackoverflow.com/questions/27769511/python-list-vs-java-array-efficiency
• https://stackoverflow.com/questions/33978318/arraylist-in-java-vs-list-in-python
• https://www.quora.com/How-do-I-think-about-an-array-in-Java-or-a-list-in-Python

17. If someone comes up to you with a new factor, how would you consider incorporating it into
an existing factor model?

• https://www.investopedia.com/terms/m/multifactor-model.asp

5
 • https://en.wikipedia.org/wiki/Multiple factor models
• https://ocw.mit.edu/courses/mathematics/18-s096-topics-in-mathematics-with-
applications-in-finance-fall-2013/lecture-notes/MIT18 S096F13 lecnote15.pdf
• https://faculty.washington.edu/ezivot/research/factormodellecture handout.pdf
• https://web.stanford.edu/ wfsharpe/mia/fac/mia fac3.htm

18. Design an algorithm to play Connect 4.

• https://connect4.gamesolver.org/
• http://blog.gamesolver.org/solving-connect-four/01-introduction/
• http://web.mit.edu/sp.268/www/2010/connectFourSlides.pdf
• https://roadtolarissa.com/connect-4-ai-how-it-works/

19. Write a recursive function to compute the number of partitions of a natural number.

• https://stackoverflow.com/questions/14053885/integer-partition-
algorithm-and-recursion
• This is also a classic generating function problem.
• https://www.overleaf.com/read/sbtmxdddtnzy

20. Given some regression filters, talk about their upsides and downside versus principal compo-
nent analysis and other dimensionality reduction techniques.

• https://en.wikipedia.org/wiki/Principal component analysis

• https://en.wikipedia.org/wiki/Dimensionality reduction
• https://dsp.stackexchange.com/questions/19962/linear-regression-filter-properties
• http://faculty.chicagobooth.edu/bryan.kelly/research/pdf/forecasting theory.pdf

21. Design a neural network, hidden Markov model, or state machine to solve the knight’s tour
problem.

• https://dmitrybrant.com/knights-tour
• https://math.stackexchange.com/questions/87991/knights-tour-as-a-neural-network
• http://www.jamesphoughton.com/2013/09/14/knights-hidden-path-0-
hidden-markov.html
• http://stanford.edu/ cpiech/cs221/handouts/practiceMidterms.html
• https://community.computingatschool.org.uk/files/6118/original.pdf
• https://scholarworks.sjsu.edu/cgi/viewcontent.cgi?referer=https://www.google.com/
&httpsredir=1&article=8383&context=etd theses
• https://en.wikipedia.org/wiki/Finite-state machine

22. Given a dictionary of words and a string, generate all valid anagrams.

• https://stackoverflow.com/questions/20680145/best-algorithm-to-find-anagram-
of-word-from-dictonary

6
 • https://stackoverflow.com/questions/25298200/given-a-dictionary-and-a-list-of-letters-
find-all-valid-words-that-can-be-built

23. You have 12 balls that appear identical. However, one is a different weight from the others
(could be either lighter or heavier). You also have a balance scale. With only three weighs
on the scale, devise a method to find the odd ball and determine if it is heavier or lighter.

• https://www.mathsisfun.com/pool balls solution.html

24. You are given two unfair coins. You flip both of them and ones comes up heads 23 of the
time while the other comes up heads 31 of the time. Given you had a uniform prior on the
bias before flipping, what is the probability that the first coin is more biased than the second
coin?

• https://math.stackexchange.com/questions/1114093/why-would-a-uniform-
prior-distribution-give-a-different-result-than-a-purely-fre
• https://www.probabilisticworld.com/calculating-coin-bias-bayes-theorem/
• https://stats.stackexchange.com/questions/291955/bayesian-biased-prior-formula
• https://math.stackexchange.com/questions/1689448/
statistical-testing-of-a-biased-coin

25. What is the expected number of draws from a standard deck until you see an ace?

• https://math.stackexchange.com/questions/1138853/
expected-number-of-cards-you-should-turn-before-finding-an-ace

26. You are given two eggs, and access to a 100-story building. Both eggs are identical. The
aim is to find out the highest floor from which an egg will not break when dropped out of a
window from that floor. If an egg is dropped and does not break, it is undamaged and can
be dropped again. However, once an egg is broken, thats it for that egg. Generalize for any
number of eggs and floors and code the problem with dynamic programming.

• https://www.geeksforgeeks.org/egg-dropping-puzzle-dp-11/
• http://datagenetics.com/blog/july22012/index.html

27. Aaron samples from the Uniform(0,1) distribution. Then Brooke repeatedly samples from
the same distribution until she obtains a number higher than Aaron’s. How many samples
is she expected to make?
Solution. Call Aaron’s number a (with associated random variable A) and let B be a random
variable that represents Brooke’s sample at any given point. We know that P (B < a) = a,
P (B > a) = 1 − a, and of course P (B = a) = 0, as it is a continuous distribution. We now
find the conditional expectation from the geometric distribution. Since the probability of
1
getting a number higher than Aaron’s is 1 − a, we expect 1−a samples. Note that this is the
conditional expectation of the number of samples given the event A = a, not the final answer.
We now find the unconditional expectation of the number of samples. To do this, let N be
a random variable that represents the unconditional number of samples before exceeding A.

7
 1
We have calculated E[N |A = a] = 1−a . To find E[N ] explicitly, we must invoke the Law of
1
Iterated Expectation. This can be done by noting E[N |A] = 1−A . Thus,
 
1
E[N ] = E[E[N |A]] = E .
1−A

By the “law of the unconscious statistician,” we have

  Z 1
1 1
E = dx = ∞.
1−A 0 1−x

Thank you to the poster: https://stats.stackexchange.com/questions/487200/expected-number-of-uniform-draws-to-exceed-a-

first-uniform-draw
and to the email from Yakov Dlougach.

28. You are given n unit vectors in n-dimensional space. Find a vector that forms the same angle
to all of them.
Solution. We first present a less efficient solution. Consider the (n − 1)-sphere that intersects
the end of all of the vectors. Find the center of this sphere and then solve n equations in n
variables.
Now for a more clever solution. Call our desired vector w and the vectors vi for i = 1, . . . , n.
Note that w · (vi − vj ) = 0 for all i 6= j. This is because w makes the same angle to all
vi ’s. Thus, w is orthogonal to span(vi − v1 ) for all 2 ≤ i ≤ n. Compute this subspace you
are done, and it can be done in O(n) time with the Gram-Schmidt process to find a fully
orthogonal vector.
29. You have a strategy with supposed Sharpe ratio 8. After n days it has lost money. What
does n have to be before you reject the hypothesis that the Sharpe is 8?

• https://www.investopedia.com/terms/s/sharperatio.asp
• https://stats.stackexchange.com/questions/155223/testing-sharpe-ratio-significance
• http://www.econ.uzh.ch/static/wp iew/iewwp320.pdf

30. What are some methods to prevent overfitting?

Solution. Add more data, use data augmentation, use models that generalize well, add regu-
larization, reduce model complexity, cross-validation, early stopping, ensemble methods such
as bootsrapping or bagging (not going to add links for all of these but one should be familiar
with all of them).
31. Given two sorted


arrays of integers of lengths m and n where m << n, find their intersection
m
in O n log n time.

• https://articles.leetcode.com/here-is-phone-screening-question-from/
• https://siderite.blogspot.com/2016/08/finding-intersection-of-two-large.html
• https://www.geeksforgeeks.org/union-and-intersection-of-two-sorted-arrays-2/

8
 n
32. Find the expected number of cycles of length greater than 2
in a random permutation of
{1, . . . , n}.

• https://math.stackexchange.com/questions/73550/the-limit-of-truncated-sums-of-harmonic-
series-lim-limits-k-to-infty-sum-n

33. Devise a way to uniformly sample from a disk. Then do it without square rooting.

• http://mathworld.wolfram.com/DiskPointPicking.html
• https://667-per-cm.net/2016/09/23/
uniform-sampling-of-a-disk-and-implications-for-sampling-the-internet/
• https://math.stackexchange.com/questions/927347/uniform-distribution-over-disk

34. Find the eigenvalues of an n × n matrix with n0 s on the diagonal and 1’s everywhere else

• https://math.stackexchange.com/questions/175228/
suppose-a-is-an-n-by-n-matrix-with-its-diagonal-entries-are-n-and-other-entries

35. Given i.i.d. random variables X, Y ∼ N (0, 1), find the conditional distribution of X given
that X + Y > 0. Prove this is a valid probability distribution.
Solution. The sum of i.i.d. normal variables is also normal. One can prove this with moment
generating functions. The distribution of X + Y is N (0 + 0, 1 + 1) = N (0, 2). Use this
distribution to find the conditional distribution.

36. Implement a stack with two queues. Then do it with one queue (the dumb way).
Solution. Move everything from one queue to the next except for the last element, then
return the last element. Continue doing this by alternating the queues. This can be done
with two loops. The dumb way with one queue is to try queue.push(queue.pop()) n − 1
times and then do queue.pop().

37. Efficiently find (and then program) a way to find a number that uses each of the digits
1, . . . , 9 exactly once and such that the number determined by the first k digits is divisible
by k for all k ∈ {1, . . . , 9}.

• http://mathforum.org/library/drmath/view/56742.html

38. Create a classifier for data using a support-vector machine.

• https://en.wikipedia.org/wiki/Support-vector machine
• https://blog.statsbot.co/support-vector-machines-tutorial-c1618e635e93
• http://web.mit.edu/zoya/www/SVM.pdf
• https://www.svm-tutorial.com/

39. Your friend claims he can tell the five colors of skittles apart by taste alone. The probability
of a skittle being any particular color is 15 . You give your friend 3 skittles and he gets 2
correct. Should you believe him? What if you give him 100 and he gets 40 correct?

• https://www.channelfireball.com/articles/magic-math-how-many-games-do-you-need-
for-statistical-significance-in-playtesting/

9
 Solution. This is a classic hypothesis testing problem. Test the null of p = 0.2 versus p > 0.2.
Then use the normal approximation from the binomial distribution or a binomial distribution
calculator.

40. Fast mental mathematics questions:

• What is 0.5382 − 0.332?

1
• Calculate 2
+ 14 + 1
16
.
• Calculate 23 · 21.
• https://www.intmath.com/blog/letters/intmath-newsletter-wus-squaring-trick-
patrickjmt-google-calculus-8858
• https://www.quora.com/Whats-a-math-trick-that-is-not-very-well-known

41. Discuss how you would model the acquaintance graph of the United States. Use this model
to guess the average degree of a vertex over this graph.

• https://www.princeton.edu/ mjs3/mccormick salganik zheng10.pdf

• http://www.stat.columbia.edu/ gelman/research/published/DiPreteetal.pdf

42. Given two data sets X and Y , we run two linear regressions to obtain y ∼ ax + b and
x ∼ cy + d. What are the bounds on ac?

• https://en.wikipedia.org/wiki/Simple linear regression

Solution. For simplicity one can assume that both data sets have mean 0, as they can always
be scaled to mean 0. Then we simply use the covariance definition of slope and finish with
the Cauchy-Schwarz inequality on the sums. The bounds will be between 0 and 1.

43. Give an example of two variables that are uncorrelated but dependent.

• https://stats.stackexchange.com/questions/85363/simple-examples-of-uncorrelated-
but-not-independent-x-and-y

44. Choose n − 1 points randomly on a line segment and break the segment at those points.
What is the probability that the resulting n segments form an n-gon?

• https://math.stackexchange.com/questions/2848881/the-probability-of-those-n-
broken-parts-of-sticks-to-form-a-closed-polygon

45. Estimation questions galore:

• Weight of an average adult giraffe,

• Number of US businesses that went bankrupt in 2014,
• Largest change of temperature over one day in the US,
• Weight of your phone if it was converted to solid gold,
• Average global life expectancy of a woman,
• 0.99100 ,
• Smallest n such that n! has 100 digits,

10
 • 33.8 ,
• xx = 1000 estimate x,
• ln(314).
Solution. The majority of these are not intended for one to obtain an exact answer, but for
discussion and mental evaluation. The numerical ones are a bit more interesting.
For exponents, it is generally a good idea to use logarithms and then observe behaviors at
small or large values. For instance, when estimating x = 0.99100 , one can perform ln(x) =
100 ln(0.99) = 100 ln(1 − 0.01) ≈ 100(−0.01). This is because for small x, ln(1 + x) is very
close to x. Thus, ln(x) ≈ −1 and x ≈ e−1 . This is very close to the actual answer.
For the n! problem, the number of base-10 digits of an integer m is blog10 (m)c + 1. We are
looking for the smallest n such that blog10 (n!)c ≥ 99. We can then change our base from 10
to e and use Stirling’s approximation to observe the growth of ln(n!).
For 33.8 , we will approach the problem two ways. Don’t fall into the trap of thinking it is
particularly close to 34 , as exponentiation increases the number very quickly.
4 4
3
1) We first note that 33.8 = 330.2 = √
5
3
= √81
5 . We can estimate fifth roots with a derivative
3
√
5 1
trick. Let y = 3. Let x be the nearest fifth power to 3, so x = 1 and we write y = (1 + 2) 5 .
1
Here we can consider ∆x = 2 and y = x 5 = 1. We are essentially taking the derivative of
1 4
y = x 5 which is y 0 = 5x− 5 so we can note that ∆y = ∆x4 = 25 = 0.4. Thus, y + ∆y =
√ 81
5x 5
1 + 0.4 = 1.4 so 5 3 ≈ 1.4 and 33.8 ≈ 1.4 ≈ 57.857. However, this is not a good estimate, as
our ∆x is twice as large x itself.
2) Observe the derivative of 3x directly, or 3x ln(3). Our estimate will be 34 − 34 ln(3).
To estimate ln(3), we can use a numerical method (https://math.stackexchange.com/
questions/1179348/estimate-ln3-using-taylor-expansion-up-to-3rd-order) and we
are done. This gives a much better approximation.
We use a similar method for xx . Note that 44 = 256 and 55 = 3125 so 4 < x < 5. Also note
the derivative is xx (ln(x) + 1). We essentially wish to solve for ∆x in 44 + 44 (ln(4) + 1)∆x =
1000. We can estimate ln(4) with the same method as the previous problem and we are done.
For ln(314) we can again use a numerical method to approximate, but for a large number it
can be tedious. Instead, note that ln does not grow particularly fast, and that 73 = 343 is
close to 314. Thus, ln(314) ≈ ln(73 ) = 3 ln(7). We can then approximate ln(7) much easier.

46. You have $100 and are betting on a fair coin flip. You can bet any percentage of the $100.
If you win, you gain 1.2 times your bet (and your bet back), but if you lose, you lose your
bet. What is the optimal bet size to maximize long-run expected earnings?

• https://en.wikipedia.org/wiki/Kelly criterion
Solution. We approach this problem analytically. Let f (x, y) represent the maximum ex-
pected money the player can get starting with x red cards and y blue cards. Observing
simpler cases, we can see that f (n, 0) = 2n , as there are no blue cards and the player simply
doubles their money n times. Our goal is to find a recurrence for f (n, m) with this base case.
For the first card, the player bets a proportion p of their money on red. p can be negative
and this signifies a bet for blue. Then there are two cases: the flipped card is either red or
blue.

11
 Now let’s observe f (n, 1). The profit if red is thus (1 + p)f (n − 1, 1), as the player wins
and has 1 + p of what the player had before, then the process repeats sans one red card.
The profit if blue is (1 − p)f (n, 0) since the player gains nothing and the process repeats
sans one blue card. We have f (n, 1) = (1 + p)f (n − 1, 1) + (1 − p)f (n, 0) and in general
f (n, m) = (1 + p)f (n − 1, m) + (1 − p)f (n, m − 1).
2n+1
Plugging in f (n, 0) = 2n , we get that f (n, 1) = n+1
. Note that the proportion p does not
even matter! In fact, the general form is

2n+m
f (n, m) = n+m
.
n

Before we prove this, note how interesting it is that p does not matter. The only thing that
matters is that when there is only one color remaining the player bets all their money. In
fact, the other bets the player makes do not matter either.
n+m
Observe n+m . This is the number of possible orderings of the cards. The term 2n+m



n
n
represents picking a random ordering and betting all of the player’s money on it at every
1

stage. If the player wins, they obtain $2n+m and the probability that they win is n+m .
n
Obviously this strategy has high variance, but it can be modified.
The idea arises from observing the original recurrence f (n, m) = (1 + p)f (n − 1, m) + (1 −
p)f (n, m − 1). Note that setting p = n−m
m+n
gives our desired result of 0 variance, and we can
show that our expected profit stays the same. But this method is not particularly motivated.
Instead, let us change our betting method. Previously we picked a random ordering of cards
and bet on it, but we can also bet on all the ordering of cards equally and simultaneously.
n m
With n red and m blue, n+m of the orderings start with red and n+m start with blue. Betting
the difference of n−m
n+m
on red every time clearly has 0 variance, as exactly one ordering will
n+m 1

be correct every time. This correct bet will make $2 and n+m of your money will be
n
bet on it. This is exactly the result obtained from the Kelly criterion.

Miscellaneous Statistics Problems

1. X and Y are i.i.d. N (0, 1) random variables. You are given that X > 0 and Y > 0. What
is the probability that Y > X?
1
Solution. The conditions make this very easy, as by symmetry the answer is simply 2
.

2. X and Y are i.i.d. N (0, 1) random variables. What’s the probability that Y > 3X?
Solution. Rearrange and see that we want P (Y − 3X > 0). Note that the linear combination
of i.i.d. normal variables is normal, so Y − 3X ∼ N (0, 10). Thus, the probability is 12 by
symmetry.

3. X and Y are i.i.d. N (0, 1) random variables. You are given that Y > 0. What is the
probability that Y > 3X?
Solution. The key is that N (0, 1)2 is cyclically symmetric. When plotting the distributions,
the pdf will be cyclically symmetric about the origin. Then one can perform a geometric
probability calculation to obtain an answer in terms of arctan.

12
4. How can one use the normal distribution to sample points uniformly from a disk? How can
one use a uniform disk to sample points from a normal distribution?

• https://en.wikipedia.org/wiki/Box%E2%80%93Muller transform
Solution. For the first part, sample x and y from N (0, 1). The ordered pair (x, y) can then
be normalized. The points are cyclically symmetric, as probability of a point is proportional
x2 +y 2
to e− 2 so for fixed x2 + y 2 all points should have the same probability. For the second
part, use the Box-Muller transformation.

5. I flip 10,000 identical coins and 5200 come up heads. Are my coins fair?
Solution. Another classic hypothesis testing question. Test the null p = 0.5 against the
alternative p > 0.5. Use a binomial distribution calculator or the normal approximation to
finish. The coins have very low probability of being fair.

6. Derive the analytical solution for linear regression β = (X T X)−1 X T y.

• https://towardsdatascience.com/analytical-solution-of-linear-regression-a0e870b038d5
Solution. The goal is to minimize the cost function J(β) = (y − Xβ)T (y − Xβ). Expand and
differentiate with respect to β.

7. I take n samples from a distribution. Why is the canonical best estimator for the mean the
sample mean? Discuss estimators for the variance and standard deviation. What are the
estimators that minimize bias? Are they different from the ones that minimize MSE?

• https://en.wikipedia.org/wiki/Mean squared error

• https://en.m.wikipedia.org/wiki/Bessel%27s correction
Solution. The expected value of the sample mean is the mean. Furthermore, the MSE of the
sample mean is always equal to the variance of the sample mean. Estimators that minimize
bias are often different from the ones that minimize MSE. For particular estimators such as
the one given by Bessel’s correction, one may need to invoke the Central Limit Theorem.

8. I have three random variables X, Y , and Z with pairwise correlations all equal to r. Whar are
the bounds on r? What are the bounds on corr(X, Z) if corr(X, Y ) = a and corr(Y, Z) = b?

• https://math.stackexchange.com/questions/284877/
correlation-between-three-variables-question
Solution. The Cauchy-Schwarz inequality can give us the answer with the classic “correlations
are cosines” idea. However, this particular problem can also be solved with the correlation
matrix, as correlation matrices are positive-semidefinite. Our particular 3 × 3 correlation
matrix has 1’s along the diagonal and r’s everywhere else. The eigenvalues are − 12 and 1
which correspond to the minimum and maximum, respectively.

9. Suppose one has two covariance matrices A and B. Is AB also a covariance matrix? What
if AB = BA?

• https://math.stackexchange.com/questions/982797/prove-that-the-product-of-two-
positive-semidefinite-and-symmetric-matrices-has-n

13
 Solution. The essential argument is that AB must be symmetric for it to be a covariance
matrix. Thus the answer to the first part is no. AB = BA satisfies symmetry, so the problem
boils down to whether it is positive-semidefinite. One can establish that AB is similar to a
positive semidefinite matrix and therefore must be positive semidefinite

10. What happens to the coefficient of determination (R2 ) when more independent variables are
added to a regression model?
Solution. Having more covariates will in general give a better fit, however, this does not
necessarily mean a better model (in terms of generalization). Thus, model comparison should
be carried out at the end in terms of how the model explains the data (e.g., likelihood, R2 ,
etc), and how simple the model is (i.e., Occam’s razor). Overfitting can also be potentially
induced, so one can look at the BIC model selection criterion where having too many variables
is penalized.

Additional Problems on Glassdoor

https://www.glassdoor.com/Interview/quant-interview-questions-SRCH KO0,5.htm

14