EasyDeepLearn

How does time-based leakage happen and how do you prevent it?

hard

Answer

  • (1) Using future timestamps in features (fitting on aggregate features computed with post-event data).
  • (2) Random shuffle instead of time-based split.
  • (3) Target encoding across time (using future target information).
  • Fix: (1) time-based split (train ← past, test ← future).
  • (2) Purged CV with embargo (gap between train/test to avoid autocorrelation).
  • (3) Feature engineering only uses features known BEFORE prediction time (as-of-timestamps in feature store).
  • (4) Backtesting with rolling / expanding windows.
Check yourself — multiple choice
  • Random
  • Future info in features / random shuffle / target encoding across time; fix via time-based split + purged CV with embargo + as-of-timestamps + rolling backtest
  • Not real
  • Same as leakage

Time leakage: split by time + purged CV + as-of features + backtest.

#data-quality#pipeline

Practise MLOps & Data Quality

215 interview questions in this topic.

Related questions