Feature engineering and preprocessing · insight-0132

Take clipping limits from training-fold quantiles and record them as model parameters

Situation
You winsorize heavy-tailed features or targets to limit the influence of extremes.
Action
Fit clip limits on the training fold only, store them with the pipeline, apply them unchanged at serving, and log how often serving values hit the limits.
Reason
Clip limits are learned parameters like a scaler's mean; treating them as constants keeps evaluation honest and serving stable.
Boundary
Linear and distance models, squared-error losses and any feature with rare extreme values.

Symptom or risk

Limits computed on the full dataset leak test information; limits recomputed at serving time drift with each batch.

Physical bounds (load factor between 0 and 1) can be hard-coded because they come from definitions, not data.

Clipping reduces variance but biases predictions for genuinely extreme cases; check errors on the clipped rows.

Clipping the target in evaluation as well as training, which hides exactly the errors the business may care about.

Synthetic raw airline data contains six load_factor values above 1: those are integrity errors to fix, not values to clip.