Skip to main content

Feature Engineering Cookbook

Use this checklist when adding or changing ML features.

Design Rules​

  • Prefer causality-safe features only.
  • Keep naming conventions consistent (TF*_ALT_*, TF*_BTC_*).
  • Avoid leaking future information through timestamp alignment.

Train/Infer Parity​

  • Ensure feature trimming and windows match inference path.
  • Keep schema changes backward-compatible only when needed.
  • Add tests for transformed output shape.

Practical Workflow​

  1. Add feature in transform path.
  2. Regenerate dataset export.
  3. Train model and inspect holdout/walk-forward quality.
  4. Verify inference still loads and predicts with new feature set.
  5. Document feature intent and impact.

Anti-Patterns​

  • Adding highly unstable features without drift monitoring.
  • Mixing experiment-only features into production path silently.