Work out data leakage risk check instantly with clear inputs, formula shown and shareable results.
Leakage inflates offline scores in three common ways: duplicate or near-duplicate rows straddling the split, features that could only be known after the outcome, and random splitting of time-ordered data so the model trains on the future. This check weights those three signals into one score so a suspiciously good validation result can be triaged before it reaches a stakeholder deck.
Composite risk score
score = min(60, 3 x overlap%) + min(30, 15 x post-outcome features) + 20 if chronological order is violated
It lets the model learn from records that happen after the ones it is scored on, which is impossible at inference time. Always split at a cut-off date.
The model effectively memorises a row and is then tested on its twin. Deduplicate on a business key and on content hashes before splitting.