WinPIQ Lab · Model Room
How WinPIQ thinks
Transparency without giving away the formulas: which model families vote, how they are weighted, what the backtest says, and what changed between versions. No proprietary code, no invented numbers.
ens_v0.1
current model
7
signals used
5
competitions covered
141
matches analyzed (live)
0
matches in backtest
—
Dixon-Coles ρ
Model families and weights
| family | role | logit weight |
|---|---|---|
| Elo | team strength from results | — |
| Poisson | scoreline distribution from attack/defence rates | — |
| Dixon-Coles | low-score dependency correction (ρ fitted) | — |
| Gradient boosting | non-linear features, regularised | — |
| Market model | no-vig prices as context only — never in the published probability | — |
| Isotonic calibration | maps ensemble outputs to observed frequencies | — |
Weights come from services/ml/artifacts/weights.json, fitted walk-forward per season. Market weight applies to context_adjusted_1x2 only; the published probability is model-only.
Model changelog
- ens_v0.11X2_ensemble · 28/08/2026Ensemble of Elo, Poisson, Dixon-Coles (rho fitted per league) and gradient boosting (HistGradientBoosting, median-imputed, regularised). Logit-average with walk-forward weights per season; isotonic calibration on held-out seasons. Market blend reported only as context_adjusted_1x2 — published probability is model-only (spec §31). Signals: adjusted form, league home advantage from history, xG attack/defence ratings where the CSV carries xG (2026/27), draw-risk radar. Known limit: does not yet beat the closing line (backtest log-loss ≈0.99 vs market 0.96–0.98).