lightElo

Predict a chess player's Elo rating from a single PGN. Engine-derived move-quality and time-usage features feed gradient-boosted tree regressors and benchmarked against a linear baseline.

Try it here

Pipeline

1. Extract - stream-sample the Lichess Jan 2026 database (src/extract_games.py) → 140k games.
2. Evaluate - fill missing positions with Stockfish 17 via Node (app/scripts/evaluate-games.mjs).
3. Features - 27 predictors per player-game (src/create_features.py).
4. Train + evaluate - src/train_model.py, src/evaluate_model.py.

Setup

pip install -r requirements.txt   # needs catboost, xgboost, lightgbm, scikit-learn
cd app && npm install             # for Stockfish evaluation step

Run

python3 run.py --model xgboost    # one model
python3 run.py --all              # all four
python3 run.py --all --tune       # with Optuna tuning

Plots are written to figures/, trained models to models/.

Results

Held-out test set (28k player-games). MAE/RMSE in Elo points.

Model	MAE	RMSE	R²
XGBoost	223.96	279.68	0.4525
CatBoost	224.08	279.57	0.4529
LightGBM	224.66	280.79	0.4481
Linear	241.38	299.79	0.3709

The GBDTs beat the linear baseline by ~17 MAE. Error is lowest mid-distribution (1200–1800) and grows at the tails (2400+ MAE ≈ 550), consistent with engine-feature saturation at high Elo.