Calibre v1.7 Full
Dixon-Coles bivariate Poisson fused with a LightGBM+XGBoost dual-engine · walk-forward validated.
Input
Raw Market Data
Unfiltered probability signals across all evaluated fixtures.
Mathematical Filter
Dixon-Coles + Steam + CLV+
Bivariate Poisson edge estimate > threshold, confirmed by closing-line movement.
Output
Institutional Alpha
Top ~1.5% by edge. 100% CLV+. Live: +$118.87 PnL, n=490.
Dixon-Coles Bivariate Poisson Core
Team-level attack/defense rating vector estimated by maximum likelihood on bivariate Poisson match scores, with the Dixon-Coles τ correction rebalancing mass in the low-scoring cells (0-0, 1-0, 0-1, 1-1) where naive Poisson structurally under-fits. Exponential time-decay weights (ξ) bias fit toward recent form without discarding sample size. Outputs feed the full scoreline matrix used by both 1X2 and Under 2.5 pricing.
LightGBM + XGBoost Dual-Engine
Gradient-boosted residual learners stacked on top of the Dixon-Coles prior. LightGBM (histogram, leaf-wise) captures dense low-latency features; XGBoost (level-wise, depth-capped) handles sparse interaction terms. Blended by inverse-variance weighting on a held-out fold. L1/L2 regularization, feature-subsampling, and monotonic constraints hold in-sample performance deliberately close to OOS.
Pinnacle Steam Detection
Open-to-close divergence gate on the sharpest book in the market. Signal fires only when institutional flow confirms the Dixon-Coles edge — i.e., Pinnacle moves the line toward our prior between open and close. Rejects reflexive moves driven by retail volume. No signal survives without independent market confirmation.
Walk-Forward · Purged + Embargo
Chronological train → test rolling across 5 seasons and 14,483 matches. Each fold's test window is purged of fixtures temporally adjacent to training matches (to kill overlapping-label leakage) and embargoed by a buffer of 6+ match-days on each side of the split. Zero look-ahead by construction. OOS window: Aug 2024 – Jan 2026 · 490 live signals · 100% CLV+.
6 Leagues · 7 Markets
EPL, La Liga, Serie A, Bundesliga, Ligue 1, Super Lig. Markets: 1X2 + Asian Handicap (full and quarter) + Over/Under 2.5. Deep alpha concentration in high-odds Under 2.5 (3.00–5.00+) where defensive-solidity mispricing is most persistent.
Variant G filtration simulation
Institutional alpha · CORE +118.87u PnL · 56.70% win · n=490 · Live store
Figures below are from the v1.7 walk-forward validated backtest. 14,483 matches across 6 leagues and 5 seasons. 42-dimensional feature space. Every metric is reproducible from the whitepaper.
- Signals (n)
- 490
- Net ROI
- +24.26%
- Win Rate
- 56.7%
- Sharpe
- 3.09
- Max DD
- -12.5%
- CLV+
- 100%
Feature Importance · Production Ensemble
Model architecture and feature weights reflect the current CORE production version. Walk-forward validation results are derived from historical data with zero look-ahead bias; live performance may differ. All CLV metrics are calculated against Pinnacle closing lines at settlement. Oracle V1.7 CORE covers EPL, La Liga, Serie A, Bundesliga, and Super Lig.