Model Methodology
Learn how our Premier League prediction models work.
Machine Learning Models
We train 5 independent machine learning algorithms on every match market (goals and match winner), then combine their predictions into a single ensemble:
K-Nearest Neighbors (KNN)
A simple, interpretable algorithm that classifies matches based on the K nearest historical matches. We optimize K through cross-validation.
XGBoost
A gradient boosting ensemble method that captures complex patterns in match data. We use hyperparameter tuning to maximize predictive accuracy.
TabPFN
A pretrained transformer-based foundation model built specifically for small tabular datasets. It needs no hyperparameter search of its own, just a single fit, making it a fast way to cross-check the tree-based models.
LightGBM
A gradient boosting framework optimized for speed and efficiency on tabular data, using leaf-wise tree growth. Tuned via randomized search over depth and learning rate.
CatBoost
A gradient boosting algorithm designed to handle categorical features natively and resist overfitting on smaller datasets. Also tuned via randomized search.
Accuracy-Weighted Ensemble
Rather than picking a single “best” model, each matchup's final prediction blends all 5 models' probabilities, weighting each model by its own cross-validation accuracy rather than its test-set accuracy — this keeps the ensemble from overfitting to a small held-out test set. The individual model predictions are still shown side by side on the Predictions page for comparison.
Features
Our models use historical and pre-match statistics including:
- Betting odds (home win, away win, draw)
- Points per game (PPG) for home and away teams
- Expected Goals (xG) pre-match estimates
- Average goals per match
- Cumulative team statistics (shots, shots on target, possession)
The feature set is periodically reviewed using permutation importance and correlation analysis, and low-signal or redundant features are trimmed so each market's models train on only the inputs that meaningfully move their predictions.
Predictions
We predict:
- Over/Under 2.5 Goals - Total goals in a match ≥ 2.5
- Over/Under 3.5 Goals - Total goals in a match ≥ 3.5
- Match Winner - Home Win, Draw, or Away Win
Training Methodology
As each new season progresses, models train on a rolling window of roughly one season's worth of matches: the most recently completed weeks of the prior season, plus every completed week of the current season, dropping the prior season's oldest weeks as the new season adds its own. Testing is always done on real, known results from the current season rather than a synthetic hold-out, giving a genuine read on how the models perform going into the next matchweek.
Model Analysis
View detailed performance metrics and visualizations for our models:
View Model Analysis and Charts →Want More Details?
See how each model has actually performed week over week on Past Performance, or dig into confusion matrices and confidence distributions on Current Model Performance.
View Current Predictions →