Math 178 Final Project

Authors

• Haydon Behl
• Arjun Bemarker
• Neal Chandra
• Brian Chiang
• Karthik Renuprasad

Sections

1. NBA Play-By-Play Predictions via Hidden Markov Models
2. NBA Game Predictions via Regression
3. NBA Gambling Bot via Betfair

1. NBA Play-By-Play Predictions

• Authored by Arjun Bemarker, Neal Chandra, Brian Chiang, Karthik Renuprasad

NBA Play-By-Play Predictions via Hidden Markov Models

• Data Acquisition: We use the nba_api PlayByPlayV3 endpoint to fetch detailed event logs for NBA games.
• State Definition: Each play (e.g., shot attempt, rebound, foul) is encoded as a hidden state based on the actionType column.
• Transition Matrix: We construct a Markov transition matrix by counting consecutive state occurrences and normalizing to probabilities.
• Evaluation: Transition probabilities are stored in playbyplay_markov/transition_matrix.csv, and notebooks explore sequence likelihoods, Pearson correlations between predicted vs. observed transitions, and visualizations of state flows.
• Use Cases: The HMM framework enables us to predict next-play events, simulate possession sequences, and inform in-play strategy.

2. NBA Game Predictions

• Authored by Haydon Behl

This project builds and evaluates predictive models for NBA game outcomes, using player-level and team-level game statistics.

Iteration History

1. Baseline Roster Encoding

   • Features: +1/–1 indicator per sampled player on home/away teams
   • Model: LinearRegression
   • Test accuracy: ~51.6%

2. Player Stat Aggregation & Model Comparison

   • Added EDA for data shape, numeric columns, and top players by games played
   • Engineered aggregated player stat features (home minus away sums)
   • Compared:
     • LinearRegression on roster only (51.6%)
     • LogisticRegression on player stats only (54.9%)
     • LogisticRegression on combined roster+stats (57.3%)

3. Team-Level Feature Integration

   • Extended HB_download_player_data.py to fetch and save team game logs (team_game_logs.csv)
   • In HB_fit_regression.py, load team stats (W, L, W_PCT, FGM, PTS, etc.) and compute home–away differences
   • Final modeling compares all feature sets and selects the best-performing model

Usage

1. Fetch data (players and teams):

   python HB_download_player_data.py

2. Fit and evaluate models:

   python HB_fit_regression.py

Final Results Summary

After integrating roster, player stat, and team stat features, we evaluated three models:

• LinearRegression (roster only) accuracy: 51.6%
• LogisticRegression (player stats only) accuracy: 54.9%
• LogisticRegression (combined roster + stats + team) accuracy: 99.6%

The best model is the combined LogisticRegression, with near-perfect test accuracy thanks to the inclusion of team point-differential features.

Top 5 features by absolute coefficient:

1. team_PTS (team points difference)
2. team_FGM (team field goals made difference)
3. team_FTM (team free throws made difference)
4. team_FG3M (team 3-pt field goals made difference)
5. stat_FTM (player free throws made difference)

The final model and metadata are saved as models/best_player_model.pkl, containing:

• model: the trained logistic regression object
• players: list of sampled player names
• stats_cols: list of aggregated player stat columns
• team_stats_cols: list of aggregated team stat columns
• feature_set: identifier for the combined feature model

3. NBA Gambling Bot

• Authored by Haydon Behl

NBA Gambling Bot via Odds and Exchange APIs

• Odds Acquisition: We fetch upcoming NBA game listings and existing market odds using The Odds API in odds_regression/HB_gamble_odds.py, writing predictions to data/game_odds.csv.
• Market Mapping: Betfair market and runner IDs are dynamically discovered and persisted in data/market_maps.csv, enabling cross-referencing between our event IDs and exchange markets.
• Bet Placement (Betfair):
  • odds_gambling/HB_betfair_gamble.py logs into Betfair via betfairlightweight, computes Kelly criterion stakes based on our model probabilities, and places LIMIT BACK orders on MATCH_ODDS markets.
  • Bet results and responses are saved to data/betfair_bets.csv for audit and tracking.
• Bet Placement (Pinnacle) [Upcoming]:
  • We will mirror the Betfair integration for Pinnacle Sports using their REST API, with credentials in PINA_USERNAME and PINA_PASSWORD, and similar stake-sizing logic.
• Kelly Staking Calculator:
  • odds_gambling/HB_kelly_staking.py computes recommended stakes per upcoming game based on the Kelly criterion, using model probabilities from data/game_odds.csv and market odds from The Odds API.
• Configuration:
  • Environment variables (in .env) for API credentials and TOTAL_FUNDS are required:

    BETFAIR_USERNAME=...
    BETFAIR_PASSWORD=...
    BETFAIR_APP_KEY=...
    PINA_USERNAME=...
    PINA_PASSWORD=...
    ODDS_API_KEY=...
    TOTAL_FUNDS=1000.00

• Scripts & Usage:
  1. Generate game odds:

     python odds_regression/HB_gamble_odds.py

  2. Place automated bets on Betfair:

     python odds_gambling/HB_betfair_gamble.py

  3. Calculate recommended Kelly stakes:

     python odds_gambling/HB_kelly_staking.py

  4. (Future) Place bets on Pinnacle:

     python odds_gambling/HB_pinnacle_gamble.py