Data Preparation

Every transformation is either a deterministic cleaning step (applied once, before splitting, with no learned statistics) or a learned step (fitted inside the pipeline on training folds only). Mixing them up is the most common leak.

Order of work

1. Deterministic cleaning, before the split. Fix types, parse dates, standardize units and category spellings, remove exact duplicates, drop rows with an invalid target, remove post-event columns. Print the row count after each step.
2. Joins with assertions.

   n = len(left)
   out = left.merge(right, on="customer_id", how="left", validate="many_to_one")
   assert len(out) == n, f"join changed rows: {n} -> {len(out)}"

3. Choose the split from the data (references/split-strategies.md):
   • Independent rows, classification: stratified.
   • Repeated entities: group-aware (GroupKFold, StratifiedGroupKFold, GroupShuffleSplit).
   • Time-ordered data: forward-chaining (TimeSeriesSplit or a cutoff date), with a gap equal to the label horizon. Fix and record the seed. Hold out the test set now; it is touched once, at evaluation.
4. Learned steps inside the pipeline. Imputers, scalers, encoders, winsorizers, selectors, resamplers and target encoders go into a ColumnTransformer and Pipeline (or an imbalanced-learn Pipeline when resampling). Fit happens on training folds only, through cross-validation or one fit on the training split.
5. Document every transformation in the feature dictionary: what, why, and when the input is available.

Missing values

Never fill with a statistic computed on the full frame before splitting.

Outliers

Decide error or signal first. Errors are fixed or removed with a written rule and a count. Signal is kept and handled with robust scaling, winsorizing inside the pipeline, or robust losses.

Checks before modeling

• X_train.shape, X_test.shape, positive counts per split, no shared entity ids across splits.
• df.duplicated().sum() before splitting is zero or explained.
• No column recorded after the prediction time.
• The pipeline object prints cleanly and pipeline.fit(X_train, y_train) runs.

Conditions that change preparation (skew, zero inflation, high cardinality, drift): data-profiling/references/scenarios.md.