Not All Signals Are Created Equal: Calibrated Confidence Ranking

The Confidence Gap

L1 XGBoost models output probability scores between 0 and 1. But raw XGBoost probabilities are not well-calibrated. A raw output of 0.70 does not mean a 70% probability of success. S14 applies logistic regression calibration (Platt scaling) to convert raw outputs into calibrated confidence scores.

After calibration, the system ranks signals into quintiles. The top quintile (highest calibrated confidence) achieves 67% win rate. The bottom quintile sits at 52%. That 15-point spread is the exploitable edge that S14 creates.

How Confidence Affects Position Sizing

V7 does not use a binary signal. Every entry carries a confidence weight from S14. Higher confidence signals receive full position size. Lower confidence signals (but still above the entry threshold) receive 75% of standard size. This confidence-weighted sizing means the system bets more on its strongest convictions.

The interaction with cluster-specific thresholds is important. The FOREX cluster threshold is 0.52, meaning calibrated confidence must exceed 52% for entry. CRYPTO is 0.48. These thresholds were optimized per cluster because calibration quality varies across asset classes.

The 67% Win Rate Question

A 67% win rate on top-quintile signals sounds excellent. But you cannot just trade top-quintile signals because there are not enough of them. Approximately 20% of all signals fall in the top quintile by definition. Trading only those would reduce trade count from 4,505 to around 900, concentrating in specific market conditions and losing diversification benefits. S14's value is in sizing, not filtering. You trade all signals above threshold but bet more on the confident ones. This adds approximately +18R over the backtest compared to flat sizing across all signals. Modest but additive, which is exactly what you want from each module in a 43-component system.