How accurate is this, really?

Walk-forward backtest: for every past match, the model is scored using only data that existed before that match was played, never with hindsight. Lower is better for RPS, Brier score, and log-loss; higher is better for accuracy. What's RPS?

Evaluated 18058 matches from 2020-08-01 to 2026-09-13.

✅ The model beats the naive home-advantage baseline on RPS.

Accuracy above only counts the single top pick, so a draw almost never wins it — even bookmakers make the draw the favourite in only 0.21% of matches, and draws are still 24.9% of real results. That puts the practical ceiling on top-pick accuracy at 75%, not 100% — on that basis the model is right 70% of the time vs 72% for the market. Every prediction still carries a real draw probability — RPS (above) is the metric that actually uses it; top-pick accuracy just can't reward it.

By competition, best to worst RPS

RPS shown as model vs market (lower is better either side) — the gap is how far we are from what the sharpest bookmaker line already knows.

Live tracking

Genuinely out-of-sample: every prediction the site actually showed before a match, scored once the real result came in, not a historical simulation like the backtest above.

Tracked 252 predictions · 133 right · 119 wrong · 52.8% top-pick accuracy (95% range 47–59%) · RPS 0.196