Who this is for — Anyone building ML models for markets: feature quality often beats algorithm choice. Raw data rarely suffices.
Feature engineering is the process of creating, transforming and selecting input variables from prices, volume, fundamentals or alternative data — normalization, lags, rolling stats, interactions, regime encoding.
In plain terms — Prep ingredients before cooking the model: make explicit what the market «hides» in raw ticks.
Examples in finance
| Family | Typical features |
|---|---|
| Price/volume | Return, rolling vol, RSI-like, order-flow proxy |
| Structure | Distance from MA, breakout flag, range compression |
| Macro | Surprise vs consensus, yield curve slope |
| Text |
Good features are interpretable, stable over time and computable without future information — otherwise data leakage kicks in.
Healthy pipeline
- Define horizon and target (direction class, return regression…)
- Generate features only with data available at time t
- Validate with temporal split / walk-forward
- Monitor drift and overfitting
Common mistake — Hundreds of features without selection — model memorises the past (overfitting) and fails live.
Example — Predict 5-bar forward return: features = return lags 1–5, ATR(14)/close, volume z-score 20 — all computed with `.shift()` to avoid look-ahead.
Card
- Principle: fewer sensible features > many random ones.
- Test: OOS + ablation (drop one feature at a time).
- Hub: AI & markets.
Sources
- scikit-learn, Feature extraction - converting raw arrays and text into machine-learning features.
- scikit-learn, Common pitfalls and recommended practices - fit preprocessing only on training data.