Welcome

This project is an attempt to analyze NBA games using neural networks in order to make accurate betting predictions. Currently, this project is able to scrape data for a given NBA season to train a scikit MLP.

Tools for scraping the necessary data as well as generating features and samples are present in util/client.py. For examples on how to use these tools, see the notebook mlp_betting.ipynb.

Currently, new work is being done in bnn_betting.ipynb, which explores the application of Bayesian Neural Networks for this task

Structure

1. Datasets

All collected datasets are contained within the NBA folder:

NBA/games:

Game information for individual seasons, consisting of involved teams and points scored in CSV format:

[Date, Away Team, Away Points, Home Team, Home Points]

No header line included

---

NBA/total:

Old data, uses season totals for generating features. Contains 2 sub folders:

on_off_stats

Pickle files containing on/off statistics for individual players and team totals for a given season

samps_feats

CSV containing feature vectors for each game (16 stats per vector, representing away_stats - home_stats for a given game):

[MOV, SOS, SRS, ORtg, DRtg, Pace, FTr, 3PAr, eFG%, TOV%, ORB%, FT/FGA, eFG%, TOV%, DRB%, FT/FGA]

For information on how this data is collected, see client.py and mlp_betting.ipynb

---

NBA/consecutive:

New data, uses cumulative stats up to a given date for generating features. For details on how this data was constructed, see selen_scraper.py, client.py, and scrape_new.ipynb

conc_feats_samps

Features and samples for NBA games using statistics gathered for each team up to a given day. Feature vectors containe 14 stats, representing away_stats - home_stats:

[OFFRTG, DEFRTG, NETRTG, AST%, AST/TO, AST_RATIO, OREB%, DREB%, REB%, TOV%, EFG%, TS%, PACE, PIE]

Currently, constructed features are based on data from conc_on_off

conc_on_off

Pickle file representing dictionary of the following structure:

stats format: [OFFRTG, DEFRTG, NETRTG, AST%, AST/TO, AST_RATIO, OREB%, DREB%, REB%, TOV%, EFG%, TS%, PACE, PIE]

ABR -> dict(date -> list of len 3):
	0: <list> overall (team) stats up to this date
	
	1: <dict> PLAYER NAME -> on court stats

	2: <dict> PLAYER NAME -> off court stats

conc_stats

Pickle file representing dictionary of the following structure:

stats format: [PTS, W, L, WIN%, FGM, FGA, FG%, 3PM, 3PA, 3P%, FTM, FTA, FT%, OREB, DREB, REB, AST, TOV, STL, BLK, PF, +/-]

ABR -> dict(date -> list of len 2):
	0: <list> overall (team) stats up to this date
	
	1: <list> overall (team) stats over ~ last 10 games

conc_stats_adv (deprecated?)

Pickle file similar to that seen in conc_on_off, however does not contain player on/off data

---

NBA/with_bets:

Game information for individual seaons, including betting lines

These datasets were generated using util/client.py by scraping relevant information from basketball reference and vegas insider

---

2. Utils

• client.py
  • Class for scraping data on an individual season
  • Generates on/off statistics and accounts for injuries
• selen_scraper.py
  • Selenium scraper for getting conseucutive games using process pools
• NeuralNet.py
  • Custom neural network design

3. Notebooks

• Currently, 3 notebooks are included in the notebooks folder
  • mlp_betting.ipynb
    • Scraping NBA data for use on scikit-learn MLPs
    • Makes bets based on MLP predictions
  • bnn_betting.ipynb <- (CURRENT)
    • Creating Bayesian Neural Networks using Pyro
  • scrape_new.ipynb
    • Notebook constructing conseuctive data

Installing

It is recommended to use a virtual environment running Python >= 3.8 when cloning this repository (for instruction on how to do so see here). To install the required dependancies, simply run : python -m pip install -r requirements.txt in the root directory of the project.