Feature Engineering

Two different ideas

• Feature extraction compresses raw or high-dimensional data into a compact representation (PCA, truncated SVD, NMF, TF-IDF, pretrained embeddings, FFT and rolling statistics). Always state what information is lost and how you inspected it: explained variance, reconstruction error, loadings.
• Feature engineering builds informative inputs from domain knowledge (ratios, lags, cyclical time, group aggregates, flags). Every feature needs a rationale.

Rules

1. Each engineered feature has a one-line rationale and passes the leakage question: would this value exist at prediction time?
2. Features are computed inside the pipeline (ColumnTransformer, FunctionTransformer with a named importable function, or a custom transformer in src/<package>/transformers.py) so training and serving share code.
3. Anything that learns from data (encoders, binning edges, selectors, PCA, target encoding) is fitted on training folds only.
4. Lags and rolling windows use only past data: shift(k) with k above zero, trailing windows, grouped by entity, computed per snapshot.
5. Target encoding is out-of-fold: sklearn.preprocessing.TargetEncoder (cross-fitted in fit_transform) inside the pipeline.
6. Record a feature dictionary: name, source, transform, type, rationale, availability time. mlkit.features.feature_dictionary() writes it.

Selection

Selection lives inside the pipeline and is evaluated with cross-validation. Remove leaky and near-duplicate features first.

Ablation

Report the CV metric with and without each feature group (raw, ratios, lags, encodings) with fold spread. Keep groups whose gain exceeds the fold-to-fold noise.

Output

• Pipeline code in src/<package>/features.py, imported by the notebook and the API.
• artifacts/feature_dictionary.csv.
• Ablation table under a findings heading.

Recipes with code: references/feature-recipes.md.