Time series and forecasting · insight-0237

Fit normalisation statistics at each cutoff, never on the full time series

Situation
A notebook standardises the whole series with its mean and standard deviation, then runs a rolling-origin backtest.
Action
Put scaling, imputation and any target transformation inside the per-cutoff pipeline so each origin fits them on its own training window only.
Reason
Preprocessing fitted on data that includes the test period leaks test information; for time series the leak is information from the future.
Boundary
Neural forecasters, ML forecasts with scaled inputs, PCA on sensor panels.

Symptom or risk

Every training window has seen the future mean and spread, so level shifts after the cutoff are partly anticipated and the backtest is optimistic.

Fixed physical transformations with no fitted parameters (unit conversion, log) carry no leak.

Refitting preprocessing per cutoff costs compute; cache by cutoff.

Reusing a scaler fitted in a separate exploratory notebook.

Fitting preprocessing on the full dataset before splitting causes data leakage; preprocessing should be fitted on training data only, for example inside a Pipeline.
Common pitfalls and recommended practices (scikit-learn User Guide, section 12)