Feature engineering without leakage

Every feature needs a clear timestamp: the latest information available when a decision is made. Rolling values should exclude the unfinished or future observation according to the execution convention. Lagging a feature is often safer than assuming a bar close was available at its open.

Write the timestamp rule into the feature specification and test it with a tiny hand-checked example.

Normalisation, imputation, feature selection, dimensionality reduction, and target encoding can leak distribution information if they are fitted on the full dataset. Fit each transformation on the training segment and apply the frozen object to validation and test segments.

A feature does not need a perfect economic theory, but you should be able to explain what it measures, when it is available, and why it might relate to the decision. Features that appear only after a large search deserve extra holdout testing.

When a feature disappears in live data, the pipeline should fail safely instead of silently substituting a different value.