v8.1 ensemble
An XGBoost classifier paired with a calibrated logistic model. The two outputs are blended so the published probability matches the empirical hit rate — no overconfident 70% calls that hit at 55%.
Engine
The model, in plain English.
No black-box mystery. The same engine drives Free picks and Platinum calls — only the strictness of the tier filter changes. Numbers below are pulled live from the public evaluation table.
IV Platinum
85.9%
291 graded picks · conf ≥ 0.75
Wilson lower bound: 81.4%
III Gold
76.5%
618 graded picks · conf ≥ 0.70
Wilson lower bound: 73.0%
II Silver
70.4%
858 graded picks · conf ≥ 0.65
Wilson lower bound: 67.3%
I Bronze
59.4%
1,698 graded picks · conf ≥ 0.55
Wilson lower bound: 57.1%
What's under the hood
Point-in-time features
Elo ratings, last-N form, head-to-head history, league standings, and bookmaker odds. Every feature is materialised at kickoff time — no post-hoc data leaks into the prediction.
30+ competitions
Premier League, Champions League, La Liga, Serie A, all the way down to MLS, J1 League and Saudi Pro League. Each league shares the same engine; the tier filter decides which confidence bands surface to which subscriber tier.
Pre-match lock + grading
Predictions are timestamped and stored before kickoff. After the match, an automated scheduler grades them against the real result and writes to the public evaluation table — the same table this page reads from.
Walk-forward validation
The engine was validated using strict walk-forward — for every test fixture, only data observable strictly before kickoff is used as input. No model knows the future. Out-of-sample accuracy on 28,838 historical fixtures matches the live numbers above within tolerance, which is the test that matters: a model whose backtest beats reality is a model that's leaking data.
Educational simulation only. 18+. Past model performance is not indicative of future results. No financial or betting advice.