Research Project | Machine Learning | Behavioral Decision Science
This project investigates the causes of human error in chess and builds machine learning models to predict mistakes before they occur.
Chess provides a unique scientific laboratory for studying human decision-making under complexity. Every move has a measurable correct answer (via Stockfish engine evaluation), making chess one of the few domains where human reasoning can be compared directly to near-perfect play across millions of decisions.
- Can machine learning models predict human chess mistakes?
- Which properties of a chess position increase the probability of error?
- How do time pressure and position complexity interact to cause mistakes?
- Do mistakes cascade — does one error increase the probability of another?
| Finding | Result |
|---|---|
| Amateur blunder rate | 7.81% |
| Grandmaster blunder rate | 1.05% |
| Best model AUC-ROC | 0.7519 (Gradient Boosting) |
| Strongest predictor | Previous move evaluation delta (57.9% importance) |
| Time pressure confirmed | Top 5 feature out of 13 |
| Tilt hypothesis | Confirmed in continuous form, rejected in binary form |
- Source: Lichess.org public game database (February 2026)
- Size: 5,000 games, 333,990 moves analyzed
- Engine: Stockfish at depth 5 for position evaluation
- Blunder definition: Evaluation drop ≥ 200 centipawns
To regenerate the dataset:
python3 src/build_dataset.py| Feature | Description |
|---|---|
n_legal_moves |
Branching factor — number of legal moves available |
eval_before_cp |
Stockfish evaluation before the move |
eval_volatility |
Standard deviation of last 5 evaluations |
time_pressure |
Clock remaining / starting time (0=no time, 1=full time) |
time_spent_seconds |
Seconds spent on this move |
prev_delta_cp |
Evaluation change on previous move |
prev_was_blunder |
Whether previous move was a blunder |
player_rating |
Lichess ELO rating |
material_balance |
Material advantage in centipawns |
| Model | F1 Score | AUC-ROC |
|---|---|---|
| Logistic Regression | 0.4157 | 0.7050 |
| Random Forest | 0.4537 | 0.7461 |
| Gradient Boosting | 0.1866 | 0.7519 |
Chess_Blunder_Prediction/
├── src/
│ ├── read_games.py # Phase 1: PGN parsing
│ ├── engine_analysis.py # Phase 2: Stockfish evaluation
│ ├── feature_extraction.py # Phase 3: Feature engineering
│ ├── build_dataset.py # Phase 4: Full pipeline
│ ├── parse_lichess.py # Clock extraction from Lichess PGN
│ └── model_training.py # Phase 6: ML model training
├── notebooks/
│ └── exploratory_analysis.py # EDA and visualizations
├── data/
│ └── sample_games.pgn # Sample grandmaster games
├── results/ # Generated visualizations
└── requirements.txt # Python dependencies
git clone https://github.com/Coddiwompl3r/Chess_Blunder_Prediction
cd Chess_Blunder_Prediction
pip3 install -r requirements.txt
brew install stockfish
python3 src/read_games.py
Position evaluation momentum (prev_delta_cp) is the dominant predictor
of human chess mistakes, accounting for 57.9% of model predictive power.
Time pressure features rank in the top 5 of 13 features, confirming that
cognitive load under time constraints significantly increases error rates.
The tilt hypothesis — that mistakes cascade — is confirmed in continuous form (evaluation momentum) but rejected in binary form (blunder indicator), suggesting that the magnitude of positional deterioration, not the categorical classification of a move as a blunder, drives subsequent mistake probability.
MIT License — see LICENSE file