← Citibank Interview Insights

Citibank·Data Scientist·Technical Phone Screen·Senior

Senior
May 2026

Summary

Interviewed for a Data Scientist role at Citibank, technical screen focused heavily on credit risk modeling, specifically LGD. One question but it had a lot of moving parts and I don't think I covered all of them as cleanly as I wanted.

Questions Asked (1)

Q1

How would you handle missing values in a dataset for LGD modeling? Walk through multiple imputation, model-based imputation, business rule fills, indicator variables, and when you'd actually just leave missingness in the model explicitly.

Data ModelingTechnical Trade-offs
Author's notes

This one sprawled.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing missingness as a business and regulatory issue in LGD modeling, not just a technical one, then systematically compare imputation methods against model performance and interpretability. Emphasize that the choice depends on missingness mechanism (MCAR/MAR/MNAR), data volume, and whether missingness itself carries predictive signal. Conclude with a validation strategy that tests each approach against out-of-time and out-of-sample data.

Pro tip: In credit risk, missingness often encodes default or workout status—e.g., missing recovery data for defaults still in progress—so always ask whether the missingness is informative before imputing. Documenting the missingness mechanism and its business cause will impress risk reviewers more than any single imputation technique.

1. Diagnose the missingness

Quantify missing rates by variable and segment, and classify the mechanism as MCAR, MAR, or MNAR using tests and business knowledge. Check whether missingness correlates with the target (LGD) or with default status.

2. Evaluate imputation options

Compare multiple imputation (MICE), model-based imputation (e.g., regression or random forest), and business rule fills (e.g., regulatory floors, peer averages) on bias, variance, and downstream LGD model performance.

3. Engineer missingness indicators

Create binary indicator variables for missingness where it may be predictive, and test whether they add lift in the LGD model. This preserves information that imputation might erase.

4. Decide when to leave missingness explicit

If missingness is MNAR and informative, consider leaving it as a separate category or using a model that handles missingness natively (e.g., XGBoost with missing direction). Validate that this does not violate regulatory interpretability requirements.

5. Validate and document

Backtest each approach on out-of-time samples, compare LGD predictions and rank-ordering, and document the rationale for regulators. Ensure the chosen method is stable and reproducible.

Key Points to Mention

  • Missingness mechanism (MCAR, MAR, MNAR) and its implications for LGD
  • Multiple imputation (MICE) and its ability to reflect uncertainty
  • Model-based imputation (e.g., regression, random forest) and risk of overfitting
  • Business rule fills (e.g., regulatory floors, peer group averages) and their conservatism
  • Missingness indicator variables to capture informative missingness
  • When to leave missingness explicit (e.g., tree-based models, MNAR) and regulatory considerations

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