Citadel·Software Engineer·Technical Phone Screen·Senior

Senior

May 2026

Summary

Citadel quant engineer interview, technical phone screen focused on factor research fundamentals. One meaty question on look-ahead bias that had more layers than I expected going in.

Questions Asked (1)

How do you figure out whether a quantitative factor is accidentally using future data? Walk through your diagnostic process.

Root Cause AnalysisTechnical Trade-offsProduct Analytics & Metrics

Author's notes

I started with the formula review angle, which was fine, but I underweighted how much they wanted to hear about point-in-time database joins specifically.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining what 'future data' means in the context of the factor and the prediction task, then systematically trace the factor's computation pipeline to identify any point where information from after the prediction timestamp could leak in. Use both logical reasoning and empirical tests to confirm or rule out leakage.

Pro tip: Mention that you always validate factors on a time-series split with a gap between train and test to catch subtle leakage, and that you keep a 'leakage checklist' for common pitfalls like using future-dated reference data or global normalization.

1. Define the prediction point and data availability

Clearly specify the exact timestamp at which the prediction is made and what data would be available at that time. This sets the boundary for what constitutes future data.

2. Trace the factor's computation pipeline

Map every data source, transformation, and aggregation step in the factor's calculation. Look for any operation that uses data with timestamps after the prediction point, including joins, window functions, or global statistics.

3. Check for common leakage patterns

Systematically inspect for typical leakage sources: using future values in rolling windows, normalizing with full-sample statistics, target encoding without time-awareness, or joining on future-dated reference tables.

4. Run empirical tests

Perform time-series cross-validation with a gap, compare factor performance on shuffled vs. chronological data, and check for suspiciously high predictive power. Also, simulate point-in-time data to see if the factor changes.

5. Validate with domain knowledge and documentation

Consult with domain experts and review data documentation to confirm that all inputs are indeed available at prediction time. Document any assumptions and edge cases.

Key Points to Mention

Point-in-time correctness: ensuring all data used is as-of the prediction timestamp.
Time-series cross-validation with a gap to prevent leakage from adjacent periods.
Common leakage sources: rolling windows with future data, global normalization, target encoding, and joins with future-dated tables.
Empirical tests: comparing performance on shuffled vs. chronological data, checking for unrealistic predictive accuracy.
Importance of data lineage and documentation to trace factor inputs.
Collaboration with domain experts to validate data availability assumptions.

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