PLONN - NBA Total Points Predictions

This repository contains a script for predicting total points in NBA games using a simple MLP, with a unique feature vector-setup. Feel free to join the free discord PLONN Discord

📄 Script Overview

Filename: nba_total_pts.py
Models: There are 3 models available Raw, Advanced (adv), and 30-day Window (30dw) all predicting a team total score when matched up against another NBA team.

Models

After about 30 days of raw predictions we decided to take a look at the distribtion between the Total Actual scores vs. the Total Predicted scores. Using the distribution leads to a 5% increase in accuracy after 30 days. After the 30-day window we will need to refesh the distribution data to prevent irrelevance.

dist

🚀 Features

Loads and processes NBA game data
Trains MLP model for total team points points prediction
Implements feature engineering and data cleaning
Supports model evaluation and predictions such as the distribution
The raw model takes in 11 basic NBA features: ['PTS', 'FG%', 'FGA', '3P%', '3PA', 'ORB', 'TRB','AST', 'TOV', 'STL', 'PF']
The Adv/30dw models have 15 Features: ['PTS', 'FG%', 'FGA', '3P%', '3PA', 'ORB', 'TRB', 'AST', 'TOV', 'STL', 'PF', 'ORtg', 'DRtg', 'FTA', 'FT%']

The scatter plot shows how close the advance model is predicting previous games (train data). At the end of training we can expect mae to be > 0.5 due to important features such as offensive and defensive rating.

🛠 Requirements

Ensure you have the following Python libraries installed:

pip3 install pandas numpy scikit-learn tensorflow

▶️ Usage

Run the script using the raw model with:

python3 nba_total_pts.py -t -g -m raw

📌 Notes

Modify the script to include additional data sources if needed.
Check model accuracy and tweak parameters for better predictions.

🔗 Author: Jake Giguere
📅 Last Updated: 02/19/2025 Reference: Basketball Reference