Label Leakage Quant ML Interview Guide
Label leakage quant ML interview guide for timing, survivorship, feature construction, splits, diagnostics, and examples.
Candidates defending backtests and model validation choices.
Leakage means future information slipped in
Label leakage occurs when training features contain information that would not be available at prediction time or that directly encodes the target.
Timing is the first check
For each feature, ask when it became known, how it was revised, and whether it was computed using data after the decision point.
Concrete example
Using end-of-day volume to predict an intraday trade decision leaks information if that full-day volume was unknown when the decision was made.
Diagnostics can reveal leakage
Suspiciously high validation metrics, performance concentrated in impossible features, or sharp collapse under walk-forward tests should trigger investigation.
Common mistakes
Candidates often focus only on train/test separation. Leakage can live inside feature construction even when train and test rows are separated.
Practice the pattern
Use the LeetQuidity curriculum and calibration to turn this topic into a focused practice plan.