Features and training data

Prevent the model from learning information it will not have at serving time.

Understand the problem

Define event-time cutoffs, labeling windows, and feature freshness. Split data according to deployment conditions. Fit preprocessing on training data only and track feature lineage.

Trade-offs and pitfalls

Very fresh features improve responsiveness but increase serving cost and operational coupling.

Check your understanding

Describe temporal leakage in a random train/test split.

Your study notes