Neural Networks for Basketball Substitution Prediction
In this project, I utilised Long Short Term Memory (LSTM) networks to research the problem of predicting the occurrence of substitutions 10 game events in the future using play-by-play data from the National Basketball Association (NBA). I created a dataset, derived features, performed ablation studies, and tested the model on six teams. The final model achieved an accuracy of 60.2%. I concluded that the temporal features, namely time into the match quarter and quarter number, had the greatest benefit on the model.
The full research paper can be found below. For a more informative overview, please read the abstract of my paper.