Harwinder
d8b0ba5b47
Trains LogisticRegression and GradientBoostingClassifier on engineered features from trades.csv (engine score/confidence, risk/reward geometry, entry hour/day-of-week, regime one-hot, pair one-hot, grade one-hot). Strict time-based train/test split (first 70% by entry_time = train, last 30% = test) — no random splits, no future leakage. Outputs: - AUC, Brier score, classification report - Per-decile breakdown (1=lowest p_win, 10=highest) - Feature importances - Calibration plot, decile returns plot, equity-compare plot - Summary text with headline lift over engine baseline Result on the 9,659-trade combined pool (7 majors × 1 year M15): - AUC 0.597 (weak but real signal) - Top 20% by p_win on test set: 37.5% WR vs 28.6% baseline - Engine grade features (grade_A, grade_A+, grade_B) have near-zero importance. Raw score and confidence carry signal; the rubric that combines them into A+/A/B does not. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>