How can I detect data leakage?

Look for evaluation results that are unexpectedly strong, a large decline from offline to production performance, or one column that dominates the model’s explanations.

Use Kumo’s column statistics, evaluation results, and Explainability to identify suspicious fields. For each field, confirm that its value would be available at the prediction anchor time in production.

Common leakage sources include a column that contains the future target, an event timestamp that precedes when the row’s information becomes known, or an entity attribute that is overwritten over time.

See Anchor Time and Data Leakage for the complete temporal-correctness workflow.