Every BetsPlug prediction is classified into one of four quality tiers based on the league and the model's confidence. Each tier has its own historical accuracy — measured on the same public evaluation set used in our track record.
We combine Elo ratings, Logistic Regression, XGBoost and a Poisson goal-model. Weighted average, calibrated probabilities.
Every prediction uses recent form (5/10 matches), venue-specific stats, head-to-head, season averages, Elo differential, rest days, and clean-sheet rate — all computed as of the match kickoff so there is zero future information.
Logistic Regression (40%) and XGBoost (60%) form the main ensemble, with Elo as a sanity anchor. Probabilities are normalised so home + draw + away = 1.00.
A pick is Platinum if the match is in our top-5 leagues AND confidence ≥ 75%. Gold = top-10 + ≥ 70%. Silver = top-14 + ≥ 65%. Free = anything else with confidence ≥ 55%. Picks below 55% are not shown.
The model version that produced each historical pick is the same one live today. We don't backfill old picks with new models to make our trackrecord look better — every pick in our track record is the original one.
Sourced from /api/pricing/comparison. Sample sizes and Wilson 95% lower bounds included — because a point estimate without a confidence interval is advertising, not science.
| Tier | Accuracy | Sample |
|---|---|---|
| … | … | |
| … | … | |
| … | … | |
| … | … |
The accuracy column is the point estimate. The 95% lower bound (Wilson interval) tells you the worst-case accuracy consistent with the observed sample — an honest floor, not the headline.
The most honest backtest we can run: the model never sees data from after the match it's predicting.
A common mistake in football modelling is training on the full 2020–2026 dataset, then testing accuracy on the same years. That leaks information — the model has already "seen" future features like final league standings, season averages, and late-season form. The numbers look great. They're meaningless.
We validate differently. The engine rolls through every match in chronological order. For a match on March 12 2024, the model is only trained on data available up to and including March 11. Form tables are truncated. Elo ratings are reconstructed to their March 11 state. Features are computed point-in-time.
The current v8.1 walk-forward run covered 28,838 out-of-sample predictions across four time-folds. Results directly match what the Per-tier table above shows — we use the same methodology in production.
API-Football Pro tier. 29 competitions, daily ingestion, auditable lineage.
Kickoff times, final scores, in-game events. Ingested hourly during match windows, nightly otherwise.
Match-level shots, possession, corners, cards. Season-level standings, goal differentials. All point-in-time snapshotted.
Pre-match 1X2 odds captured for retrospective analysis. We never place bets and never write to bookmakers — odds are context, not signal.
Educational / simulation only. BetsPlug provides statistical analysis of historical sports data. All predictions are presented as simulations for educational purposes. Nothing on this site constitutes betting advice, financial advice, or a guarantee of future results.
Past performance is not indicative of future results. Tier accuracy figures (85%+, 70%+, etc.) are historical backtest measurements. Sample sizes and confidence intervals are published above so you can verify the scope of each claim.
No gambling licence. No affiliation with bookmakers. BetsPlug operates under an educational / analytical framework (no KSA licence required in the Netherlands). We are not affiliated with any bookmaker and do not facilitate bet placement.
Responsible use. Sports betting involves financial risk. If you or someone you know struggles with gambling, please visit our responsible-gambling page.