Quantitative Research Partnership
Strategic validation seats · full model transparency · direct quant liaison.
Strategic Partnership · Program Brief
We are inviting a limited number of institutional partners and quant desks to validate the v1.7 CORE engine. Access is provided on a strategic partnership basis — full model transparency, walk-forward validated backtests, Dixon-Coles parameter exports, and direct liaison with the engineering team.
Partners receive a replayable harness covering 14,483 matches across 5 seasons, 6 leagues and 7 markets, together with feature-level SHAP attributions and fold-by-fold validation telemetry sufficient for independent risk and compliance review.
FAQ · Technical & Program
What exactly is Calibre v1.7 CORE?+
A walk-forward validated sports prediction engine — not daily tips. Dixon-Coles bivariate Poisson fundamentals fused with a LightGBM + XGBoost dual-engine ensemble. Every signal is verified against Pinnacle's closing line before it reaches production.
How is CLV+ measured?+
Every signal is benchmarked against Pinnacle's closing line for the identical market selection. CLV+ means the model-implied probability exceeded the market's final assessment at settlement. It is the quantitative gold standard for edge verification and the only KPI that cannot be faked with cherry-picked samples.
What leagues and markets are supported?+
v1.7 CORE covers Premier League, La Liga, Serie A, Bundesliga, Ligue 1, and Super Lig. Markets: 1X2 (match result), Over/Under 2.5, Asian Handicap (full + quarter). Deep alpha concentration sits in high-odds Under 2.5 (3.00–5.00+) where defensive-solidity mispricing is most persistent.
What does the validation partnership include?+
Full model documentation, walk-forward validation datasets, feature importance rankings, Dixon-Coles parameter exports, and a direct quantitative liaison. Partners receive everything needed for independent risk, compliance, and quantitative review — including replayable backtest harnesses.
What is the anti-overfit protocol?+
Walk-forward validation with purged splits and embargo windows. Zero look-ahead by construction. Aggressive L1/L2 regularization, capped tree depth, constrained feature interaction, monotonic constraints where the sign of the effect is theoretically fixed. The model is intentionally slightly underpowered in-sample to guarantee out-of-sample robustness.
What is the underlying technology?+
A hybrid ML engine: Dixon-Coles bivariate Poisson fundamentals (with τ correction and exponential time-decay) anchor the prior; a LightGBM + XGBoost stacked ensemble learns the residual. Pinnacle closing lines anchor the Bayesian prior for drift detection. Full methodology is available to validation partners.