Who this is for — Anyone doing ML or backtests on time series. Leakage is the silent error that turns a «perfect» model into live losses.
Data leakage occurs when information not available at decision time enters features, targets or preprocessing — look-ahead bias, shuffling time series, normalizing on the full dataset, joins with data revised ex post.
Canonical treatment: Data leakage and look-ahead bias — causality, preprocessing, audit and the boundary with survivorship bias.
In plain terms — The model «cheats» by seeing exam answers before responding — great on paper, useless live.
Common forms
| Type | Example |
|---|---|
| Look-ahead | Feature uses current bar close to enter at open |
| Target leakage | Target built with post-event data unknown at t |
| Global preprocessing | StandardScaler fit on train+test together |
| Survivorship | Universe = only stocks still listed today |
| Revisions | Macro «final» instead of first print |
In finance time is causal: splits must respect chronology (out-of-sample).
Prevention
- Features with explicit lags; entries on next bar open
- Walk-forward and purge between adjacent folds
- Pipeline audit: «what did I know at t?»
- Compare to naive baseline — huge jump? suspect leakage
Common mistake — Random `train_test_split` on OHLCV — mixes future and past, unrealistic metrics.
Example — Model predicts gap up: feature includes `high - open` on the same bar you enter at open → leakage, 90%+ accuracy, useless live.
Card
- Rule: no data from t+1 in decision at t.
- Test: shift features 1 bar — performance collapses? likely leakage.
- Related: Overfitting often masks leakage.
Sources
- scikit-learn, Common pitfalls and recommended practices - data leakage, train-test separation and pipeline controls.