ML Signal Filter Prototype for Regime‑Aware Strategies

A rule‑based strategy performed well in certain regimes but degraded in others. The goal was to add an adaptive filter without turning the system into a black box.

• • Engineered features from price/volatility/structure indicators.
• • Trained a small set of interpretable models (starting with tree‑based baselines).
• • Used walk‑forward evaluation and sensitivity tests to reduce overfitting risk.
• • Packaged outputs as “gate / no‑gate” decisions with confidence and rationale.

• • Python research repo + reproducible runs
• • Backtest report with assumptions and limitations
• • Model artefact + monitoring plan

Walk‑forward equity comparison — ML filter vs baseline with confusion matrix and feature importance

Model drift monitoring — rolling accuracy, feature PSI heatmap, and fallback mode stats

ML is optional and should be added only when the data, validation, and monitoring plan justify it. Avoid implying guaranteed returns or “AI that always adapts.”