This repo contains code from an independent project where I predict NBA player career using LSTM recurrent neural networks.

Code was revisited and cleaned in January 2023.

The basic idea is that a career is a time series of vectors $X_1, X_2, ...., X_N$ where $X_j$ is a $k$ dimensional vector of points, rebounds, and other statistics that measure performance over the course of a season. My goal is sequence prediction - given observations $X_1, ..., X_p$ I want to predict $X_{p+1}, ..., X_N$ to determine how a career changes over time. The vector prediction nature of the problem lends itself to recurrent neural networks.

Modeling was done using python 3.7.15 and virtual environments were managed using conda.

Requirements are stored in requirements.txt

I found it easiest to install deps with both pip (out of habit) and conda (to circumvent long tensforflow compilation / builds from scratch) .

To install deps and build the environment, run source install.shin a terminal shell.

The repo is set up as a python package, to make imports between modules easier.

To make forecasts (after installing), run

conda activate nba_forecasting && streamlit run app.py

The application should look something like this:

For information on validation of the model, including performance comparison against a baseline model, look here.

Future work here should:

1. Better tune hyperparameters
2. Calibrate confidence intervals