How it works

Every pro player has a custom player ELO: five skill ratings (combat, damage, economy, laning, vision), each compared only against players in the same role, plus an outcome ELO that moves with wins and losses. Before a match, the two rosters' ratings are combined with recent team form, head-to-head history, schedule and roster changes, and an XGBoost classifier turns them into the chance of winning a single game, averaged over both sides of the map. That per-game chance is then compounded into the chance of winning the Bo1, Bo3 or Bo5 series.

Training

The model was trained on pro games from Oracle's Elixir since 2017, split chronologically so it never sees the future. Hyperparameters were tuned with Optuna, and the raw probabilities pass through a Platt scaler so the percentages are well calibrated. The numbers above come from a held-out test period (May 2025 to October 2026) that was never used for any modelling choice.

Limits

Predictions are made before the draft, so champion picks, which decide many games, are unknown. That puts a ceiling on any pre-draft model: about 65% per single game is close to it. Most of the signal comes from the outcome ELO alone (picking the higher-rated team is right in 69.8% of series), and many matchups are close to a coin flip.

Explainability

Every prediction is accompanied by SHAP values (SHapley Additive exPlanations) computed from the underlying tree model. The top 5 factors shown on each card represent the features that moved the probability the most for that specific matchup. Green bars support the pick, red bars work against it.