Model Performance

AI prediction accuracy and backtest results

Full transparency on how our model performs. No cherry-picked numbers — we show every league, including the ones where accuracy is lower. 140 days of genuine out-of-sample evaluation.

52%
Winner Accuracy (avg)
56%
O/U 2.5 Accuracy
367+
Predictions Tested
140
Days Backtested

Per-league breakdown

Winner Prediction Accuracy

140-day out-of-sample backtest · Aug 2025 — Mar 2026

Bundesliga
59%

Highest accuracy — high-scoring league with clear patterns

Serie A
55%

Strong home advantage signal captured by the model

Premier League
52%

Most competitive league — tighter margins, harder to predict

Ligue 1
49%

PSG dominance helps baseline; mid-table is volatile

La Liga
47%

Tactical and low-scoring — draws are harder to predict

Over/Under 2.5 Goals (all leagues)56%

Putting numbers in context

A random guess on a 3-way market (home/draw/away) would hit about 33% of the time. Bookmaker-implied odds, which represent the market consensus, typically achieve around 48-52% accuracy. Our model operates at 47-59% depending on the league, meaning it provides an edge in most cases — but not all.

The Bundesliga (59%) and Serie A (55%) show the strongest results because these leagues have more predictable patterns — the Bundesliga is high-scoring with clear home advantages, and Serie A has strong defensive tendencies.

La Liga (47%) is our weakest league. It's a tactical, low-scoring competition where draws are common and hard to predict. We include this number because transparency matters more than looking good.

Methodologymethodology

Training

  • LightGBM gradient-boosted models per league
  • Optuna hyperparameter tuning (100-200 GPU trials)
  • Score prediction from LightGBM probability outputs
  • 100+ features from 3 independent data sources

Evaluation

  • Strict temporal split (last 20% of time = test)
  • No random shuffling, no future data leakage
  • 140+ days of out-of-sample prediction
  • Metrics: accuracy, log loss, calibration

What we are honest about

We publish accuracy numbers for every league, including ones where the model underperforms. La Liga at 47% is barely above the random baseline of 33%, and we show that openly.

We don't retroactively fit the model to improve past numbers. We don't exclude bad predictions from the backtest. We don't cherry-pick time periods where the model happened to perform well.

Football is inherently unpredictable. No model can account for injuries announced 30 minutes before kick-off, red cards, or a goalkeeper having the game of their life. Our confidence scores reflect this uncertainty.

Ready to see predictions in action?

Start with a free 3-day trial. No credit card required.

Start Free Trial