by bpleone
LEARNING

🧠 Live Learning

The model updates itself from settled outcomes. Every few hours, fresh results retrain the probability calibration, re-weight which pick-sources to trust, and re-fit per-sport bias — no human in the loop. This is what it's doing right now, straight from the live artifacts.

1 · Calibration — the displayed probabilities get truer

An isotonic transform is re-fit from every settled outcome so a posted "70%" actually hits ~70%. Lower calibration error (ECE) = more honest numbers. This is the single most-retrained piece.

Loading…

2 · Source trust — learned from realized hit rate

Each pick-source starts at a prior weight and moves toward what it has actually done. Sources that keep hitting earn trust; chronic losers get downweighted (Wilson-guarded so a hot streak can't overfit). The biggest moves:

Pick sourcePriorLearnedMoveSettledHit%
Loading…

3 · Per-sport bias correction

Each desk's model is checked against real results and nudged for systematic bias once enough games settle (Brier = calibration quality, lower is better). "Applied" means the correction is live in the boards.

SportGamesHit%BrierBiasApplied
Loading…

4 · Training readiness — is it safe to size real money yet?

A composite gauge of whether the model is calibrated and sampled enough to trust for sizing, vs research-only. Built from calibration quality, sample sufficiency, residual stability, and realized ROI — plus the loop's own most recent notes.

Loading…