Calibration

Real numbers measuring how well our predicted probabilities track actual market outcomes.

Latest model

typepolymarket_logistic
trained_at2026-05-06 07:14:11
training_set_size2295
Brier score0.1996
log loss0.5879
resolved markets in DB2413
feature snapshots in DB2689

Honest caveat: v1 trained on post-hoc snapshots; Brier optimistic. Retrain when forward-looking snapshots accumulate.

Coefficients (raw-feature space)

entity_count -0.0943
sum_gravity -0.5976
mean_momentum -0.8735
events_30d 0.0000
edge_weight_30d_internal 0.0003
(intercept) -0.8520

Sign of each coefficient shows the feature's effect on YES probability. Magnitudes are in the raw feature scale, not standardized.

Calibration curve

Each circle is a bin of predictions; the x-axis is the mean predicted probability in the bin, the y-axis is the mean actual outcome (YES = 1). Circle size scales with sample count. The dashed diagonal is perfect calibration: points above it are underconfident, points below are overconfident.

Bins

Range	Mean predicted	Mean observed	n samples
0.10 – 0.20	0.194	0.000	3
0.20 – 0.30	0.272	0.270	2131
0.30 – 0.40	0.336	0.424	66
0.40 – 0.50	0.450	0.442	86
0.70 – 0.80	0.736	0.556	9