NBA Prop Betting Value Model

This project builds a machine learning pipeline to identify mispriced NBA player prop lines by comparing public betting lines with independent player performance projections. The goal is to highlight high-confidence over/under picks using real historical performance data, engineered features, and predictive modeling.

---

Key Features

• Web scraping from BettingPros NBA Underdog Props for current lines and projections
• Game log collection using nba_api across multiple seasons
• Feature engineering on player stats (rolling averages, usage rates, home/away splits)
• Modeling using RandomForestClassifier and LogisticRegression with 5-fold stratified CV
• SHAP explainability to surface key features behind predictions
• Output: Top 3 value props for the day with probabilities and suggested over/under direction

---

Folder Structure

nba-prop-model/
├── main.py # Entry point script
├── top3_model_predictions.csv # Output file with daily best picks
├── requirements.txt # Python dependencies
└── README.md

---

Usage

1. Install Dependencies

pip install -r requirements.txt

2. Run the pipeline

python main.py

This will:

1. Scrape props from BettingPros
2. Pull historical player game logs (via nba_api)
3. Engineer features and label outcomes
4. Train and cross-validate a Random Forest model
5. Score and output top 3 model-predicted props

3. View Output

top3_model_predictions.csv 

Player	Stat Type	Line	Proj	RF_Prob	Prediction
J. Embiid	Points	25.5	28.3	0.78	Over
S. Curry	PRA	38.5	41.9	0.74	Over
D. Booker	Rebounds	5.5	4.2	0.25	Under

Model Performance

Classifier: RandomForestClassifier (n=250) with 5-fold stratified CV

Test Accuracy: ~60%+

Log Loss: Reported on console

Feature importance: Visualized via SHAP bar plot for transparency

Dependencies

• undetected_chromedriver
• beautifulsoup4
• nba_api
• sklearn, shap
• pandas, numpy, matplotlib

Acknowledgements

• BettingPros.com for daily prop lines
• FantasyPros for player projections
• nba_api for historical player data

License

This project is licensed under the CC BY-NC-ND 4.0 license.
You may view, use, and share this for educational and research purposes only.
Commercial use and redistribution are strictly prohibited.