← DRW Interview Insights

DRW·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
May 2026

Summary

DRW data scientist interview that leaned heavily on probability and stats fundamentals, with a game theory curveball thrown in at the end. The math was dense and moved fast.

Questions Asked (4)

Q1

Given a finite Markov chain, how do you find its stationary distribution?

Algorithms & Data StructuresTechnical Trade-offs
Author's notes

Knew the answer conceptually but fumbled the formal setup under pressure.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the stationary distribution as a probability vector π satisfying πP = π, then explain how to solve this as a left eigenvector problem or linear system. Emphasize the conditions for existence and uniqueness (irreducibility and positive recurrence) and mention computational methods for large chains.

Pro tip: In practice, for large state spaces, direct linear algebra is infeasible; power iteration or Markov Chain Monte Carlo (MCMC) methods are preferred. Also, always check that the chain is irreducible and aperiodic to guarantee a unique stationary distribution.

1. Define stationary distribution

State that a stationary distribution π is a probability vector (sums to 1, non-negative) such that π = πP, where P is the transition matrix.

2. Check existence and uniqueness

Mention that for a finite irreducible chain, a unique stationary distribution exists. If the chain is periodic, the stationary distribution still exists but convergence may not occur.

3. Solve the linear system

Set up the system π(P - I) = 0 with the constraint sum(π) = 1. Solve using linear algebra (e.g., Gaussian elimination, eigen-decomposition).

4. Consider computational methods

For large chains, use iterative methods like power iteration (π_{k+1} = π_k P) until convergence, or solve the system using sparse linear solvers.

5. Alternative: detailed balance

If the chain is reversible, solve the detailed balance equations π_i P_{ij} = π_j P_{ji} for all i,j, which can simplify the computation.

Key Points to Mention

  • Stationary distribution satisfies π = πP and sums to 1.
  • Existence and uniqueness require irreducibility and positive recurrence (finite irreducible chain suffices).
  • Solving as a left eigenvector problem with eigenvalue 1.
  • Computational approaches: direct linear solve, power iteration, or MCMC for large state spaces.
  • Detailed balance condition for reversible chains.
  • Convergence to stationary distribution depends on aperiodicity and mixing time.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

State and derive the expected values of the exponential distribution and the Poisson distribution.

Algorithms & Data Structures
Author's notes

The exponential part was fine, integration by parts and you get 1/lambda.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clearly stating the expected values: for exponential distribution, E[X] = 1/λ (or 1/β depending on parameterization); for Poisson distribution, E[X] = λ. Then derive each using the definitions: for exponential, integrate x * f(x) from 0 to infinity; for Poisson, sum k * P(X=k) from k=0 to infinity. Show the key steps and mention the memoryless property and Poisson process connection.

Pro tip: Be explicit about the parameterization you're using (rate λ vs. scale β) to avoid confusion, and relate the two distributions via the Poisson process—this shows deeper understanding and is highly relevant for a data science role at a trading firm.

1. State the expected values

Clearly state that for an exponential distribution with rate λ, E[X] = 1/λ; for a Poisson distribution with rate λ, E[X] = λ. Mention that parameterization matters (e.g., scale θ = 1/λ gives E[X] = θ).

2. Derive exponential expectation

Write the PDF f(x) = λe^{-λx} for x ≥ 0. Compute E[X] = ∫_0^∞ x λe^{-λx} dx using integration by parts (or recognize it as the mean of a gamma distribution with shape 2). Show the steps: let u = x, dv = λe^{-λx} dx, etc., leading to 1/λ.

3. Derive Poisson expectation

Write the PMF P(X=k) = e^{-λ} λ^k / k!. Compute E[X] = Σ_{k=0}^∞ k * e^{-λ} λ^k / k!. Simplify by canceling k and shifting the index to get λ e^{-λ} Σ_{j=0}^∞ λ^j / j! = λ e^{-λ} e^{λ} = λ.

4. Connect the distributions

Explain that in a Poisson process with rate λ, the number of events in a unit interval is Poisson(λ) and the inter-arrival times are Exponential(λ). This connection reinforces both derivations and highlights practical relevance.

5. Summarize and mention properties

Summarize the results and note key properties: exponential is memoryless, Poisson has variance equal to mean, and both are fundamental in modeling waiting times and event counts.

Key Points to Mention

  • Parameterization: rate λ vs. scale θ = 1/λ for exponential; Poisson always uses rate λ.
  • Derivation of exponential expectation via integration by parts or gamma function.
  • Derivation of Poisson expectation via summation and index shift.
  • Connection to Poisson process: inter-arrival times are exponential, counts are Poisson.
  • Memoryless property of exponential distribution.
  • Variance of Poisson equals its mean (λ), and variance of exponential is 1/λ^2.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

For a non-singular square matrix A, express the sum of the eigenvalues of its inverse in terms of the eigenvalues of A, and explain the reasoning.

Algorithms & Data StructuresTechnical Trade-offs
Author's notes

Actually enjoyed this one.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by recalling that if λ is an eigenvalue of A, then 1/λ is an eigenvalue of A⁻¹, assuming A is invertible. Then the sum of eigenvalues of A⁻¹ is the sum of reciprocals of eigenvalues of A, which can be expressed as the trace of A⁻¹. Explain the reasoning using the eigenvalue definition and properties of inverses.

Pro tip: Mention that the sum of eigenvalues equals the trace, and for A⁻¹ it's the trace of A⁻¹, which is also the sum of reciprocals of eigenvalues of A. This shows you understand the connection between eigenvalues, trace, and matrix inverse.

1. Recall eigenvalue definition

State that for a non-singular matrix A, if Av = λv for a non-zero vector v, then λ is an eigenvalue of A.

2. Relate eigenvalues of A and A⁻¹

Apply A⁻¹ to both sides of Av = λv to get v = λ A⁻¹ v, which implies A⁻¹ v = (1/λ) v. Thus, 1/λ is an eigenvalue of A⁻¹.

3. Sum of eigenvalues

The sum of eigenvalues of A⁻¹ is the sum over all eigenvalues λ_i of A of 1/λ_i. This is also equal to the trace of A⁻¹.

4. Express in terms of A's eigenvalues

Conclude that if the eigenvalues of A are λ₁, λ₂, ..., λₙ, then the sum of eigenvalues of A⁻¹ is ∑_{i=1}^n 1/λ_i.

Key Points to Mention

  • Eigenvalues of A⁻¹ are reciprocals of eigenvalues of A.
  • The sum of eigenvalues equals the trace of the matrix.
  • Non-singularity ensures all eigenvalues are non-zero.
  • The characteristic polynomial relationship: det(A - λI) = 0 implies det(A⁻¹ - μI) = 0 with μ = 1/λ.
  • Trace of A⁻¹ is the sum of reciprocals of eigenvalues of A.
  • This holds for any square matrix over a field where eigenvalues exist (e.g., complex numbers).

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q4

In a specific two-player game theory scenario, calculate the probability of a particular outcome.

Algorithms & Data StructuresProduct Analytics & Metrics
Author's notes

This was the one that got me.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify the specific two-player game and the outcome in question, as the probability depends on the game's structure and strategies. Then, identify the appropriate solution concept (e.g., Nash equilibrium, mixed strategies) and compute the probability using probability theory and game-theoretic tools. Finally, validate the result with a simple simulation or edge cases.

Pro tip: At DRW, interviewers value clear assumptions and practical validation. State your assumptions explicitly and mention that you would verify the result with a quick Monte Carlo simulation, showing both theoretical rigor and practical implementation skills.

1. Clarify the Game and Outcome

Ask questions to understand the game rules, players' strategies, and the exact outcome whose probability is requested. Confirm whether strategies are pure or mixed, and if the game is simultaneous or sequential.

2. Identify Solution Concept

Determine the appropriate equilibrium concept (e.g., Nash equilibrium, dominant strategies) and whether the probability is under equilibrium play or a specific strategy profile.

3. Compute the Probability

Use probability theory and game-theoretic calculations to derive the probability. For mixed strategies, set up equations for indifference conditions and solve for probabilities.

4. Validate and Interpret

Check the result with a simple simulation or by testing edge cases. Interpret the probability in the context of the game and discuss any implications.

Key Points to Mention

  • Nash equilibrium and mixed strategies
  • Indifference principle for mixed strategies
  • Probability calculations (conditional probability, independence)
  • Monte Carlo simulation for validation
  • Assumptions about player rationality and information
  • Practical relevance to data science and trading (e.g., modeling strategic interactions)

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.