Methodology
Oracle V3 is a production-grade probability engine designed for operators who require closing line value discipline, transparent constraints, and auditability. The following sections summarize the architecture at a level suitable for quantitative and risk review.
Objective function
Oracle V3 optimizes for calibrated probability scores that outperform the closing line on the 1X2 market, with emphasis on away-side structural inefficiencies observed across Serie A, EPL, and La Liga.
The model does not chase headline accuracy in isolation — closing line value (CLV) is the primary quality gate for production signals.
Pinnacle anchoring
Final model probability is a convex combination of Pinnacle’s closing implied probability and the LightGBM output: 60% market + 40% model. This anchors decisions to liquid reference prices while allowing the learner to identify residual edge.
LightGBM architecture
Depth and capacity are intentionally constrained (depth 3, leaves 8, L1=5, L2=10) to limit variance and reduce overfit to idiosyncratic season noise.
Feature selection retains the top 35 inputs by mutual information against the target, evaluated on rolling temporal splits.
Validation protocol
Training uses purged, walk-forward splits with embargo windows to prevent leakage across adjacent fixtures.
Hold-out AUC on the primary ranking task has remained stable near 0.77 across refits; production monitoring tracks CLV drift by league bucket.
Feature importance (ranking stability)
Relative contribution after mutual-information screening and constrained tree fitting. Values are illustrative aggregates from the latest refit.