Calibration Dashboard

AI accuracy

Brier score, Platt scaling parameters, and confidence calibration metrics for BWC-generated signals.

Brier Score

0.640

Poor

Platt Scaling

a=-0.08 b=-1.19

n=158

Graded Signals

500

98W / 402L / 0B · 19.6% WR

Avg Confidence

0.868

vs 19.6% actual win rate

Calibration table

Confidence Bin	Count	Predicted	Actual Frequency	Bias
0–20%	0	10%	—	—
20–40%	0	30%	—	—
40–60%	21	50%	19%	-31.0%
60–80%	32	70%	28%	-41.9%
80–100%	105	90%	19%	-71.0%

Outcome distribution

Failure Mode	Count	% Total	Avg R	Avg Brier
tp1 hit	96	19.2%	2.82R	0.020
tp2 hit	2	0.4%	2.16R	0.000
stop loss	306	61.2%	-1.00R	0.240
expired	58	11.6%	0.56R	0.375

Recent graded signals

Symbol	Side	Confidence	Calibrated	Outcome	Failure	R	Brier
ONDOUSDT	SHORT	—	—	LOSS	stop loss	-1.00R	—
SOLUSDT	SHORT	—	—	LOSS	stop loss	-1.00R	—
UNIUSDT	SHORT

Methodology

Brier score measures the mean squared difference between predicted probability and actual outcome (1 = win, 0 = loss). Scores range from 0 (perfect) to 1 (worst). Platt scaling fits a logistic regression to adjust model confidence toward empirical frequencies. The calibration table compares predicted vs actual win rates within each confidence bucket — a perfectly calibrated model would show actual frequency matching the predicted bin midpoint.