← Citibank Interview Insights

Citibank·Data Scientist·Technical Phone Screen·Senior

Senior
May 2026

Summary

Citibank data scientist interview that went pretty deep into regression theory and model diagnostics. Not a casual screen, they clearly wanted someone who could reason through the math and not just name-drop techniques.

Questions Asked (3)

Q1

What are the core OLS assumptions required for estimates to be unbiased, efficient, and consistent? For each one, name a diagnostic test and a remedy if the assumption is violated.

Technical Trade-offsData Modeling
Author's notes

I knew most of these but fumbled the normality one.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around the Gauss-Markov theorem, systematically walking through each OLS assumption with its property guarantee (unbiasedness, efficiency, consistency), a named diagnostic test, and a concrete remedy. This demonstrates both theoretical depth and practical modeling experience, which is critical for a quantitative role at a financial institution like Citibank.

Pro tip: Distinguish between assumptions needed for unbiasedness (linearity, exogeneity) versus efficiency (homoscedasticity, no autocorrelation) versus consistency — this precision signals graduate-level econometric fluency and sets you apart from candidates who just recite a memorized list.

1. Anchor with the Gauss-Markov Theorem

Open by stating that OLS is BLUE (Best Linear Unbiased Estimator) under the Gauss-Markov assumptions, and briefly note that consistency requires additional conditions. This frames the entire answer and shows you understand the 'why' behind the assumptions.

2. Cover Unbiasedness Assumptions

Address linearity in parameters, random sampling, and strict exogeneity (E[ε|X]=0), explaining that violations — such as omitted variable bias or endogeneity — lead to biased and inconsistent estimates. Mention the Ramsey RESET test for misspecification and IV/2SLS as a remedy for endogeneity.

3. Cover Efficiency Assumptions

Discuss homoscedasticity and no autocorrelation, which are required for OLS to be efficient (minimum variance). Name the Breusch-Pagan or White test for heteroscedasticity and the Durbin-Watson or Breusch-Godfrey test for autocorrelation, with remedies of robust standard errors (HC/HAC) or GLS/WLS.

4. Address No Perfect Multicollinearity

Explain that perfect multicollinearity makes OLS estimates indeterminate, while high multicollinearity inflates standard errors. Mention the Variance Inflation Factor (VIF) as the diagnostic and remedies such as ridge regression, dropping correlated predictors, or PCA.

5. Connect to Real-World Modeling Decisions

Briefly tie the assumptions to practical consequences in a financial context — for example, autocorrelation is common in time-series credit or market data, making HAC standard errors or ARIMA-based residual modeling essential. This shows you can apply theory to Citibank's actual data environment.

Key Points to Mention

  • Gauss-Markov theorem and the BLUE property — distinguishing unbiasedness, efficiency, and consistency as separate guarantees
  • Exogeneity / no endogeneity: Hausman test for endogeneity, IV or 2SLS as remedy
  • Homoscedasticity: Breusch-Pagan / White test, remedies include WLS or heteroscedasticity-robust (HC3) standard errors
  • No autocorrelation: Durbin-Watson / Breusch-Godfrey test, remedies include HAC (Newey-West) standard errors or GLS
  • No perfect multicollinearity: VIF diagnostic, remedies include ridge regression or dimensionality reduction via PCA
  • Normality of errors is NOT required for unbiasedness or efficiency — only for finite-sample inference — demonstrating this nuance shows advanced understanding

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

Q2

You have n=10,000 observations where the error variance is proportional to x1 squared and two predictors x2 and x3 are correlated at 0.98. Walk through exactly how you'd validate the model, refit it, and compare results including what happens to your standard errors, confidence intervals, and hypothesis tests.

Data ModelingTechnical Trade-offsRoot Cause Analysis
Author's notes

This is where the interview got real.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around two distinct problems — heteroscedasticity (error variance proportional to x1²) and multicollinearity (x2/x3 correlation of 0.98) — and address each with diagnosis, remediation, and impact assessment. Walk through the diagnostic tests first, then the refitting strategy, and finally quantify the downstream effects on inference metrics like standard errors, CIs, and p-values. This demonstrates systematic thinking and deep statistical fluency, which Citibank's risk-sensitive environment demands.

Pro tip: Explicitly distinguish between how heteroscedasticity inflates/deflates standard errors asymmetrically across observations versus how multicollinearity inflates standard errors uniformly for the correlated predictors — interviewers at financial firms love when candidates show they understand these are orthogonal problems requiring separate fixes, not a single combined solution.

1. Diagnose Both Issues Formally

Run a Breusch-Pagan or White test for heteroscedasticity and plot residuals vs. x1² to confirm the variance structure. Compute the Variance Inflation Factor (VIF) for x2 and x3 — a correlation of 0.98 will yield VIFs well above 10, confirming severe multicollinearity.

2. Remediate Heteroscedasticity via WLS

Since the error variance is proportional to x1², use Weighted Least Squares with weights w_i = 1/x1_i² to stabilize variance across observations. Alternatively, apply a log or Box-Cox transformation on the response if theoretically justified, then verify residual homoscedasticity post-refit.

3. Address Multicollinearity Between x2 and x3

Evaluate whether x2 and x3 are theoretically redundant — if so, drop one based on domain knowledge or use Ridge Regression to shrink correlated coefficients. If both are theoretically necessary, consider creating a composite feature (e.g., PCA component) or applying regularization, and document the trade-off between interpretability and stability.

4. Refit and Compare Models Systematically

Refit using WLS (and Ridge if needed) and compare OLS vs. corrected model on: coefficient estimates, standard errors, 95% confidence intervals, and t-statistics/p-values. With n=10,000, expect OLS standard errors for x2 and x3 to be inflated by a factor of ~7x (≈1/√(1-0.98²)) due to multicollinearity, and heteroscedasticity-corrected SEs to shift asymmetrically based on x1 magnitude.

5. Validate and Communicate Trade-offs

Use cross-validation (e.g., k-fold) to confirm predictive performance hasn't degraded, and run a likelihood ratio test or compare AIC/BIC between OLS and WLS. Clearly articulate that hypothesis tests in the original OLS model were unreliable — some significant predictors may become insignificant and vice versa — which is critical in a financial modeling context where regulatory interpretability matters.

Key Points to Mention

  • Breusch-Pagan/White test for heteroscedasticity and VIF > 10 as the multicollinearity threshold, with the exact VIF formula showing ~7x SE inflation at r=0.98
  • WLS with weights 1/x1² as the theoretically correct fix for known variance structure, versus HC3 robust standard errors as a lighter-touch alternative that corrects inference without refitting
  • The distinction that heteroscedasticity makes OLS unbiased but inefficient (wrong SEs), while multicollinearity makes coefficients unstable and SEs inflated — both corrupt hypothesis tests but through different mechanisms
  • Ridge regression as a principled multicollinearity remedy that introduces bias to reduce variance, and the bias-variance trade-off this entails for a large n=10,000 dataset
  • Post-refit validation: checking that residuals vs. x1² are now homoscedastic, condition number of the design matrix has improved, and cross-validated RMSE is comparable or better
  • Regulatory and business implications at Citibank: unreliable p-values in the original model could lead to incorrect feature selection in credit risk or fraud models, making this remediation a compliance concern, not just a statistical one

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

Q3

Under what circumstances would you move away from OLS toward ridge regression, LASSO, or a GLM, and what's the reasoning behind each switch?

Technical Trade-offsData Modeling
Author's notes

Felt more comfortable here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer by first establishing OLS as the baseline and its core assumptions, then systematically walk through the specific data conditions or modeling goals that motivate each alternative method. Ground your reasoning in both statistical theory and practical business consequences, which is especially relevant in a financial context like Citibank where model interpretability and regulatory compliance matter.

Pro tip: Mention that in a banking/finance context, LASSO's automatic feature selection is valuable for regulatory model documentation and explainability, while ridge is preferred when you need to retain all predictors but control multicollinearity — this shows you understand real-world constraints beyond just model accuracy.

1. Anchor with OLS Assumptions

Briefly state when OLS is ideal — low-dimensional data, no severe multicollinearity, normally distributed errors, and homoscedasticity. This sets the baseline and shows you understand why you'd ever deviate from it.

2. Motivate Ridge Regression

Explain that ridge is appropriate when multicollinearity is present or when p is large relative to n, as the L2 penalty shrinks correlated coefficients together without zeroing them out, reducing variance at the cost of a small bias. Emphasize it retains all predictors, which is useful when all features are theoretically meaningful.

3. Motivate LASSO

Describe LASSO as the go-to when you suspect only a sparse subset of features are truly predictive, since the L1 penalty drives irrelevant coefficients exactly to zero, performing automatic feature selection. This is particularly valuable in high-dimensional settings like credit risk modeling with hundreds of potential predictors.

4. Motivate GLMs

Explain that GLMs are necessary when the response variable violates OLS's normality and constant-variance assumptions — for example, using logistic regression for binary default outcomes, Poisson regression for count data like transaction frequencies, or Gamma regression for skewed positive continuous outcomes like loan amounts.

5. Tie Back to Business Context

Close by noting that the choice also depends on non-statistical factors such as interpretability requirements, regulatory scrutiny (SR 11-7 model risk guidelines at banks), computational constraints, and whether prediction accuracy or inference is the primary goal.

Key Points to Mention

  • Bias-variance tradeoff: regularization introduces bias to reduce variance, improving out-of-sample performance in high-dimensional or multicollinear settings
  • L1 vs L2 penalty distinction: L1 (LASSO) produces sparse solutions via coefficient zeroing; L2 (ridge) shrinks all coefficients proportionally but retains them all
  • Multicollinearity as the primary trigger for ridge — VIF scores or condition number as diagnostic tools
  • GLM link functions and exponential family distributions — matching the error distribution to the response type (binary, count, skewed continuous)
  • Cross-validation for selecting the regularization hyperparameter lambda in ridge and LASSO
  • Elastic Net as a hybrid approach when both sparsity and handling of correlated predictors are needed simultaneously

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