Our methodology
Every published probability comes from the v4-Pro engine: a LightGBM three-class model over 48 frozen features (Elo, form, rolling tables, attack/defence estimates, rest days, a seamless three-source xG timeline, rolling shot/possession stats, squad strength ratings) plus eight market features including closing prices, blended log-linearly with the market (α = 0.15 model / β = 0.85 market), and a Dixon-Coles score engine whose λ is calibrated to the blended 1X2 and the market Over-2.5 price. High-confidence flags come from a gate threshold chosen only on validation data — every accuracy number we publish is paired with its coverage.
We evaluate on locked splits — train up to 2024, validation 2025, a one-shot 2026 test — and publish the frozen validation numbers next to a live prediction log anyone can check. Top-pick accuracy and RPS are printed on every match page.
What we are not
ScoreLens.net is a data journalism project. We do not accept advertising from gambling operators, we never publish wagering tips, and our outputs are statistical estimates for informational and entertainment purposes only. They are not advice of any kind.