Guide · updated 2026-11-04
Avoiding leakage in data preparation
Fitting cleaning statistics on training data only, and other leakage traps.
What leaks
Medians, means, category lists or scaling learned from all rows include test information. So do features computed from the future relative to the prediction date.
Fit, then apply
Learn statistics on the training rows, store them, and apply the stored values to test and new data.
Check the timeline
For every feature, ask whether it was known at prediction time.