[ Calibration & Quality ]

Calibration & Quality Metrics

How well do our probabilities align with reality? Three public metrics — Brier, CRPS and DSR — continuously recalculated from our resolved signals. No editing, no cherry-picking.

← Track RecordMethodology →

Brier Score

Fair

Direction forecast calibration (bullish/bearish/neutral). Lower is better.

0.268

290 resolved forecasts · target < 0.20

BS = (1 − p)² when correct

CRPS

Poor

Quantile forecast quality (P10 / P50 / P90). Lower is better.

7.204

258 quantile forecasts · target < 0.15

∫ (F(x) − 1{x ≥ y})² dx

Deflated Sharpe

Insufficient data

Probability that our edge survives multiple-testing bias correction. Higher is better.

Insufficient data

Insufficient data

DSR = Φ((SR − SR₀) · √(N − 1) / σ̂)

Last updated: 18 Jul 2026, 11:38 · Window: 90 days

How to read these metrics

  • Brier measures how well our probabilities match the binary outcomes of forecasts (bullish/bearish/neutral). Correct trinomial variant: correct prediction at 70% confidence → ~0.09; opposite prediction → ~0.49. Lower = better. Target: < 0.20.
  • CRPS (Continuous Ranked Probability Score) evaluates the quality of quantile forecasts — not just direction, but the entire distribution. Lower = better. Target: < 0.15.
  • DSR (Deflated Sharpe Ratio) would be the probability that our edge survives multiple-testing bias correction (Bailey & Lopez de Prado 2014). Higher = better; a meaningful reading needs ≥ 0.95. We do not have a value yet — the metric requires at least 30 resolved trades in the published cohort and we have not reached that. The card below reports “insufficient data” and will keep doing so until the sample exists. We show it empty rather than quietly omitting it.

See live signals from the AI Council.

Get Started →