v1 vs v2
v2 est un moteur de classement ajusté à l'opposant et pondéré par le temps, construit et contrôlé avec la même discipline de validation progressive que tout le reste ici. Il n'est pas en direct — chaque prévision sur le site aujourd'hui provient de v1. Cette page est le journal des tests de v2, publié avant toute décision de changement.
← Retour à l'historique completWalk-forward backtest (simulated)
v1 (live)
0.633 +2.08%
265,477 settled
v2 (banked)
0.616 +5.34%
50,492 settled
What each model version would have scored on history, forecasting each match using only results that existed before it. This is slice 1's own gate result — v2 is a time-decayed, opponent-adjusted Dixon-Coles rating engine; v1 is the model as it stood when the improvement program began.
Live forecasts
v1 (live)
0.626 +2.62%
21,221 settled
v2 (banked)
0.631 +3.33%
322 settled
Real pre-match snapshots, settled against the result and never edited afterwards. v1 is the live model everyone sees today. v2 accrues its own real forecasts here — banked for research, not shown to users — so its live record can be judged before any decision to ship it.
2026-06-11 to 2026-11-22
v2 is not live: every prediction on the site today comes from v1. This page exists so v2's record — backtest and, as it accrues, live — is visible before any decision to change that, the same glass-box standard applied to everything else here.
Par ligue — backtest
| League | v1 skill↓ | v2 skill↓ | Δ↓ |
|---|---|---|---|
| Serie A | -3.2% | -3.4% | -0.2% |
| Championship | -3.2% | -2.9% | +0.3% |
| La Liga | +2.9% | +2.8% | -0.1% |
| Scottish Premiership | +0.5% | +12.2% | +11.7% |
| Friendlies Clubs | +0.5% | -6.8% | -7.4% |
| Copa Sudamericana | -2.1% | +0.1% | +2.2% |
| Colombia Primera A | -3.6% | -0.6% | +3.0% |
| Austrian Bundesliga | +2.8% | +9.7% | +7.0% |
| Eredivisie | +7.9% | +11.5% | +3.6% |
| B1 | +1.5% | +7.6% | +6.2% |
| D2 | +4.9% | +1.1% | -3.8% |
| Primeira Liga | +12.7% | +14.3% | +1.7% |
| Copa Libertadores | -5.8% | +4.6% | +10.4% |
| T1 | -10.8% | +6.3% | +17.1% |
| I2 | -3.3% | +1.8% | +5.1% |
| Premier League | -10.6% | +9.3% | +19.9% |
| SP2 | -0.9% | -0.1% | +0.9% |
| E2 | -0.4% | +3.4% | +3.9% |
| E3 | -19.8% | +0.7% | +20.5% |
| Argentina Primera Division | -2.2% | -0.4% | +1.7% |
| G1 | +3.4% | +13.0% | +9.6% |
| Bundesliga | +2.8% | +7.4% | +4.7% |
| F2 | -2.1% | +1.3% | +3.5% |
| Ligue 1 | +1.3% | +7.4% | +6.1% |
Compétence de backtest par ligue, validation progressive sur les résultats historiques. Les chiffres de v2 proviennent de sa propre évaluation ; ceux de v1 de l'historique archivé du modèle. Une ligue sans chiffre v1 ou v2 n'a pas encore atteint l'échantillon minimum (n=30).
v2
Time-decayed, opponent-adjusted Dixon-Coles ratings (365d half-life) + per-league fitted rho
Gate: pooled 3-class Brier old=0.6380 -> new=0.6162 (Δ=-0.0217), 22/24 leagues improved. Banked for accrual only (Phase 5) — not the live model. Re-derived edge gate + public comparison page + the live flip itself are separate, later steps.