Models and forecast quality

Variants compared on the same held-out period. Lower Brier and log loss mean better forecast quality on this sample.

This evaluation covers the Value Detector model. Daily picks, runs, cards and corners use separate model versions and require their own evaluation.

Validation period: 2026-10-01 – 2026-10-19

Variant

Variant	Matches	Brier ↓	Log-loss ↓
Poisson: goals, xG and time decay Current variant	402	0.6087	1.0146
ELO Research candidate	402	0.6129	1.0190
Dixon–Coles Research candidate	402	0.6132	1.0219
Poisson: goals only Research candidate	402	0.6155	1.0240
League average Baseline	402	0.6415	1.0584

Three-outcome Brier is the sum of three squared errors. We compare identical matches. Top-outcome accuracy does not replace calibration.

Rho was fitted only before the validation period. ρ=-0.1.

Probability	Matches	Mean forecast	Observed share
20%–30%	18	26%	28% [12%–51%]
30%–40%	113	36%	35% [26%–44%]
40%–50%	164	44%	46% [39%–54%]
50%–60%	79	55%	61% [50%–71%]
60%–70%	25	64%	60% [41%–77%]
70%–80%	3	75%	100% [44%–100%]

One historical validation period. Results do not establish profitability. ROI and CLV require historical executable prices.