Citibank·Data Scientist·Technical Phone Screen·Senior

Senior

May 2026

Summary

Interviewed for a Data Scientist role at Citibank. The technical screen leaned heavily into credit risk modeling fundamentals, specifically around logistic regression and its practical limits in a PD context. Not a lot of fluff, just one meaty question that took the whole conversation.

Questions Asked (1)

What are the limitations of logistic regression when used for probability of default modeling? Think about things like the linearity assumption in log-odds, handling interactions, multicollinearity, class imbalance, probability calibration, and the tradeoff between interpretability and flexibility.

Data ModelingTechnical Trade-offs

Author's notes

This question is broader than it sounds.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer by first acknowledging the strengths of logistic regression (interpretability, regulatory acceptance) and then systematically addressing its limitations in the context of probability of default modeling. For each limitation, briefly explain the issue, its impact on PD modeling, and potential workarounds or alternative approaches. Conclude by emphasizing the trade-off between interpretability and flexibility, and how to balance them in a regulated banking environment.

Pro tip: Tie each limitation to a business implication, such as regulatory compliance (e.g., Basel, SR 11-7) or model risk, to demonstrate that you understand the practical constraints in a bank like Citibank. Mention that while more flexible models exist, logistic regression remains a benchmark due to its transparency, and often a hybrid approach or careful feature engineering can mitigate many limitations.

1. Acknowledge strengths and context

Start by noting why logistic regression is widely used in PD modeling: interpretability, regulatory acceptance, and ease of implementation. This sets the stage for a balanced discussion.

2. Discuss linearity and additivity assumptions

Explain that logistic regression assumes linearity in the log-odds and additivity of features, which may not hold for complex financial data. Mention that non-linear relationships and interactions are not captured unless explicitly specified.

3. Address data-related challenges

Cover multicollinearity, which can inflate coefficient variance and reduce interpretability, and class imbalance, which can bias probability estimates. Discuss techniques like regularization, resampling, or using different evaluation metrics.

4. Examine probability calibration and performance metrics

Highlight that logistic regression may produce poorly calibrated probabilities if the model is misspecified or if there is class imbalance. Mention calibration techniques (Platt scaling, isotonic regression) and the importance of using proper scoring rules (e.g., Brier score, log loss).

5. Weigh interpretability vs. flexibility and conclude

Summarize the trade-off: logistic regression offers interpretability but may lack flexibility compared to machine learning models. Suggest that in regulated environments, one might use logistic regression as a baseline, employ feature engineering, or adopt explainable AI techniques for more complex models.

Key Points to Mention

Linearity in log-odds: logistic regression assumes a linear relationship between features and the log-odds of default, which may not capture non-linear patterns.
Handling interactions: interactions must be manually specified, which can be impractical with many features and may lead to overfitting if not careful.
Multicollinearity: high correlation among features can cause unstable coefficient estimates and reduce model interpretability; regularization (L1/L2) can help.
Class imbalance: PD datasets often have few defaults, leading to biased probability estimates; techniques like oversampling, undersampling, or using class weights can mitigate.
Probability calibration: logistic regression may not produce well-calibrated probabilities if assumptions are violated; calibration methods like Platt scaling or isotonic regression can improve.
Interpretability vs. flexibility: logistic regression is interpretable but less flexible; more complex models (e.g., gradient boosting) may improve performance but at the cost of interpretability and regulatory acceptance.

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