PrizePicks ML — NBA & WNBA Player Prop Prediction

Preprint: Predicting NBA and WNBA Player Prop Markets: A Machine Learning Approach to Sports Betting Efficiency — see paper/main.tex

A machine learning pipeline that predicts whether NBA and WNBA players will go OVER or UNDER their PrizePicks prop lines, using only publicly available ESPN gamelog data.

---

Key Results

	NBA (2026 test)	WNBA (2025 test)
Overall accuracy	61.2%	57.6%
≥60% confidence	68.4%	62.6%
≥70% confidence	76.5%	68.3%
≥80% confidence	85.8%	78.4%
Permutation test	z = 87.4, p < 0.0001	z = 23.7, p < 0.0001

• 499,770 NBA training samples · 55,695 WNBA samples
• Walk-forward backtest: models frozen on 2023–2025, tested on full 2026 season (144,846 predictions)
• Model confidence is well-calibrated — higher confidence = higher accuracy, monotonically

---

How It Works

ESPN Gamelogs → Feature Engineering → Walk-Forward CV → GradBoost Model
                                                              ↓
                                          PrizePicks Lines → Daily Email Picks

Features: Rolling averages (L1/L3/L5/L10/L20), season avg, momentum, acceleration, coefficient of variation, hit rate, cross-stat correlations, home/away, game number

Models: Gradient Boosting (best), Voting Ensemble, Random Forest, SVM, Logistic Regression

Validation: Walk-forward cross-validation — never trains on future data

---

Repo Structure

├── paper/                  # LaTeX preprint
│   ├── main.tex
│   └── refs.bib
├── data/
│   ├── nba_props_dataset.csv       # 499,770 rows
│   ├── wnba_props_dataset.csv      # 55,695 rows
│   ├── models/                     # Saved GradBoost pickles
│   ├── plots/                      # Backtest figures
│   └── live_results.csv            # Daily pick outcomes
├── ml_pipeline.py          # NBA model training
├── wnba_ml_pipeline.py     # WNBA model training
├── dataset_builder.py      # ESPN gamelog → features
├── backtest.py             # Walk-forward backtest + plots
├── email_picks.py          # Daily automated picks email
├── main.py                 # Core prediction logic
├── prizepicks.py           # PrizePicks line fetcher
├── vegas_lines.py          # Odds API comparison
├── moneyline.py            # Moneyline value finder
├── flex_pnl.py             # Flex Play P&L tracker
├── check_results.py        # 1am results checker
└── significance_tests.py   # Permutation tests

---

Automated Pipeline

Two GitHub Actions run daily:

• 5pm CDT — fetches live PrizePicks lines, runs model, emails top picks
• 1am CDT — checks ESPN box scores, logs outcomes, sends results email

---

Setup

pip install -r requirements.txt

Set environment variables:

APIFY_TOKEN       # PrizePicks line scraper
ODDS_API_KEY      # The Odds API (moneyline/vegas comparison)
KALSHI_API_KEY    # Kalshi prediction market
GMAIL_USER        # Sending email
GMAIL_PASS        # Gmail app password

Run backtest:

python backtest.py

Train models:

python ml_pipeline.py       # NBA
python wnba_ml_pipeline.py  # WNBA

---

Paper

See paper/main.tex for the full preprint. Key findings:

• NBA prop markets are approximately 94% efficient by Brier Skill Score
• Primary inefficiency: short-horizon momentum — lines lag hot/cold streaks, especially for role players
• Secondary: class imbalance in rare props (blocks, steals) — market systematically overestimates these, creating UNDER value
• WNBA markets show consistent but slightly lower inefficiency, consistent with thinner market action

---

University of North Texas · 2026