About the model

ChessPredict estimates win, draw and loss probabilities from a chronological Elo model. The first real-data build uses completed rated games from the public Lichess PGN archive (January 2013) normalized into stable white/black player IDs and chess time controls.

Win, draw or loss

Chess games can be drawn, so the site uses the Davidson ties model: the rating gap sets the win/loss odds while a draw-propensity term assigns probability to draws — heaviest between evenly matched players. A colour bonus captures first-move advantage for White.

Backtest protocol

Ratings are updated walk-forward after each completed game. K-factor, White colour bonus and draw propensity are selected by a simple chronological holdout grid search, then the site publishes historical probabilities, rankings and calibration metrics generated from that configuration.

Current limitations

This is not yet a live pro-event forecast. FIDE and chess-results.com adapters remain planned, and forward pairing ingest is not wired. The current site is an honest real-data ratings and backtest slice from Lichess rather than betting-style predictions for upcoming games.

These are model probabilities for entertainment and analysis, not betting advice.