← Citadel Interview Insights

Citadel·Data Scientist·Technical Phone Screen·Senior

SeniorPrefer not to say
May 2026

Summary

Citadel data scientist interview that went deep on probability and statistics, the kind of session where you realize pretty fast whether your math fundamentals are actually solid or just surface-level. Four questions, all technical, no behavioral fluff.

Questions Asked (4)

Q1

If three random variables X, Y, Z all have the same pairwise correlation ρ, what is the minimum value ρ can take?

Data ModelingAlgorithms & Data Structures
Author's notes

This one tripped me up more than I expected.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Recognize that the correlation matrix must be positive semi-definite (PSD). Construct the 3x3 correlation matrix with all off-diagonal entries equal to ρ, then find the minimum ρ such that all eigenvalues are non-negative. The answer is ρ ≥ -1/2.

Pro tip: Mention that this bound generalizes to n variables with ρ ≥ -1/(n-1), and that it's a common trick in quant interviews to test understanding of covariance matrices.

1. Set up the correlation matrix

Write the 3x3 matrix with 1s on the diagonal and ρ on all off-diagonals. This matrix must be positive semi-definite (PSD) for any valid set of random variables.

2. Check eigenvalues

Compute the eigenvalues of the matrix. For a matrix with constant off-diagonal ρ, the eigenvalues are 1+2ρ (multiplicity 1) and 1-ρ (multiplicity 2).

3. Apply PSD condition

Require all eigenvalues to be non-negative: 1+2ρ ≥ 0 and 1-ρ ≥ 0. The second inequality gives ρ ≤ 1, which is always true for correlations. The first gives ρ ≥ -1/2.

4. Verify attainability

Show that ρ = -1/2 is achievable by constructing three random variables with the given correlation. For example, let X, Y, Z be such that X+Y+Z=0 and each has variance 1, then pairwise correlation is -1/2.

Key Points to Mention

  • Correlation matrix must be positive semi-definite (PSD).
  • For a 3x3 matrix with all off-diagonals equal to ρ, eigenvalues are 1+2ρ and 1-ρ (twice).
  • PSD condition requires 1+2ρ ≥ 0, so ρ ≥ -1/2.
  • The bound is tight: ρ = -1/2 is attainable (e.g., with variables summing to zero).
  • Generalization: for n variables, minimum ρ is -1/(n-1).
  • This is a common test of understanding covariance matrices in quant interviews.

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

Q2

In a simple linear regression of Y on X, if you know R² and the slope β(y|x), can you derive the slope β(x|y) from the reverse regression of X on Y?

Data ModelingTechnical Trade-offs
Author's notes

I knew the relationship existed but blanked on the exact form mid-answer.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Derive the relationship between the two slopes using the definitions of simple linear regression and the coefficient of determination. Show that β(x|y) = r² / β(y|x), where r² = R², and explain the intuition behind this inverse relationship.

Pro tip: Emphasize that this relationship holds only in simple linear regression with an intercept; in multiple regression, the reverse regression slopes are not simply related to R² and the forward slopes. This shows you understand the limitations and can avoid overgeneralizing.

1. Recall definitions

State the simple linear regression model for Y on X: Y = α + βX + ε, and for X on Y: X = γ + δY + ν. Note that β = Cov(X,Y)/Var(X) and δ = Cov(X,Y)/Var(Y).

2. Express R² in terms of variances and covariance

Use the fact that R² = [Cov(X,Y)]² / (Var(X)Var(Y)) for simple linear regression with an intercept.

3. Relate the slopes

Multiply β and δ: β * δ = [Cov(X,Y)/Var(X)] * [Cov(X,Y)/Var(Y)] = [Cov(X,Y)]² / (Var(X)Var(Y)) = R². Therefore, δ = R² / β.

4. Discuss implications and limitations

Explain that this relationship holds only for simple linear regression with an intercept. In multiple regression, the reverse regression coefficients are not simply related to R² and the forward coefficients.

Key Points to Mention

  • Definition of simple linear regression slopes: β(y|x) = Cov(X,Y)/Var(X), β(x|y) = Cov(X,Y)/Var(Y).
  • Coefficient of determination R² = [Cov(X,Y)]² / (Var(X)Var(Y)) for simple linear regression with intercept.
  • Derivation: β(y|x) * β(x|y) = R², so β(x|y) = R² / β(y|x).
  • The relationship is symmetric in the sense that the product of slopes equals R².
  • This relationship does not generalize to multiple regression; reverse regression is not simply the inverse of forward regression.
  • Assumption of an intercept term is crucial; without it, the relationship may not hold.

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

Q3

X and Y are independent Uniform(0,1). Compute the covariance between max(X, Y) and min(X, Y).

Data ModelingAlgorithms & Data Structures
Author's notes

Probably the question I felt best about, mostly because I remembered that Cov(max, min) can be rewritten using Cov(X+Y, XY) type identities or by just computing E[max*min], E[max], E[min] separately.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Use the identity Cov(max, min) = E[max*min] - E[max]E[min]. Compute E[max] and E[min] using order statistics, and compute E[max*min] by recognizing that max*min = X*Y (since one is max and the other min). Then plug into the covariance formula.

Pro tip: Mention that max*min = X*Y is a key simplification that avoids integration. Also, note that the negative covariance makes intuitive sense: when the minimum is high, the maximum must be even higher, but the spread is constrained.

1. Set up the covariance formula

Write Cov(max, min) = E[max*min] - E[max]E[min]. This decomposes the problem into computing three expectations.

2. Compute E[max] and E[min]

For two independent Uniform(0,1), E[max] = 2/3 and E[min] = 1/3. Derive these using order statistics or integration.

3. Compute E[max*min]

Observe that max(X,Y)*min(X,Y) = X*Y. Since X and Y are independent, E[X*Y] = E[X]E[Y] = (1/2)*(1/2) = 1/4.

4. Calculate covariance

Plug values: Cov = 1/4 - (2/3)*(1/3) = 1/4 - 2/9 = (9 - 8)/36 = 1/36. Wait, check: 1/4 = 9/36, 2/9 = 8/36, so difference is 1/36. But is it positive? Actually, let's recompute: E[max]E[min] = (2/3)*(1/3)=2/9≈0.222, E[max*min]=0.25, so Cov=0.25-0.222=0.0278=1/36. That's positive. But intuitively, max and min should be negatively correlated? Let's check: For two uniforms, max and min are actually positively correlated? No, think: if min is high, max must be even higher, so they move together? Actually, if min is high, max is also high, so positive correlation. Yes, positive covariance. So 1/36 is correct.

5. Verify and interpret

Confirm the result and provide intuition: positive covariance because when the minimum is large, the maximum is also large, and vice versa.

Key Points to Mention

  • Covariance definition: Cov(U,V) = E[UV] - E[U]E[V]
  • Order statistics for uniform distribution: E[min] = 1/(n+1), E[max] = n/(n+1) for n=2
  • Independence of X and Y implies E[XY] = E[X]E[Y]
  • Key identity: max(X,Y)*min(X,Y) = X*Y
  • Final answer: 1/36
  • Intuitive interpretation: positive covariance because min and max tend to move together

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

Q4

Given a monotone function Y = g(X), how do you derive the pdf of X if you know the pdf of Y?

Data ModelingTechnical Trade-offs
Author's notes

Standard change-of-variables.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying that the function is monotone (either increasing or decreasing) and that the transformation is one-to-one. Then use the change-of-variables formula: f_X(x) = f_Y(g(x)) * |g'(x)|, explaining each component and noting the absolute value for the decreasing case. Finally, mention the importance of checking the support of X and verifying that the resulting pdf integrates to 1.

Pro tip: Emphasize that monotonicity ensures a one-to-one mapping, which simplifies the derivation; if the function were not monotone, you'd need to sum over all preimages. Also, mention that in practice, you often work with the CDF method for safety, especially when the transformation is not strictly monotone or when dealing with piecewise functions.

1. Clarify assumptions and notation

Confirm that g is monotone (increasing or decreasing) and differentiable, and that Y = g(X) is a one-to-one transformation. Define the pdfs f_X and f_Y and their supports.

2. Apply the change-of-variables formula

For monotone g, the pdf of X is f_X(x) = f_Y(g(x)) * |g'(x)|. Explain that the derivative term accounts for the scaling of probability mass under the transformation.

3. Handle increasing vs. decreasing cases

If g is increasing, g'(x) > 0, so |g'(x)| = g'(x). If g is decreasing, g'(x) < 0, so |g'(x)| = -g'(x). Emphasize that the absolute value ensures a positive pdf.

4. Determine the support of X

Find the range of X by inverting the transformation: X = g^{-1}(Y). The support of X is the image of the support of Y under g^{-1}. Ensure the pdf is zero outside this support.

5. Verify the result

Check that the derived pdf integrates to 1 over the support of X. Optionally, mention the CDF method as an alternative derivation: F_X(x) = P(X ≤ x) = P(g(X) ≤ g(x)) or P(g(X) ≥ g(x)) depending on monotonicity, then differentiate.

Key Points to Mention

  • Monotonicity ensures a one-to-one transformation, so no summation over multiple preimages is needed.
  • The change-of-variables formula: f_X(x) = f_Y(g(x)) * |g'(x)|.
  • The absolute value of the derivative is crucial for decreasing functions to keep the pdf positive.
  • The support of X must be derived by inverting the transformation and mapping the support of Y.
  • The CDF method (differentiating the CDF) is a robust alternative, especially for non-monotone or piecewise functions.
  • Always verify that the resulting pdf integrates to 1 over its support.

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