Feature Engineering

Most "amazing offline, useless in production" models are not modeling failures - they are leakage failures introduced during feature engineering. This skill builds features that improve cross-validated performance without smuggling the target into the inputs, and rejects features that only look good because they cheated.

Operating procedure

Leakage prevention comes first because every later step (encoding, aggregation, imputation) is a leakage opportunity; validation comes last because a feature that does not move the cross-validated metric does not ship.

Step 1: Gather inputs

Collect before writing transforms; label unknowns as guesses.

1. Prediction task and the exact prediction time: what is known at the moment the model must predict? Everything else is off-limits.
2. Target definition and how it was produced - features derived from the same process as the label are proxy-leakage suspects.
3. Data shape: row count, column list with dtypes, entity keys, and whether rows have a time dimension.
4. Model family (linear, tree ensemble, neural) - it determines which transforms are necessary and which are wasted work.
5. Serving constraints: will these features be computable at inference time with the same code and latency budget?

Step 2: Lock the leakage rules