Features selection in NBA outcome prediction through Deep Learning

TL;DR

This paper explores feature selection for NBA game outcome prediction, demonstrating that models based on Elo ratings or victory frequency outperform traditional box-score predictors using deep learning and cross-validation.

Contribution

It introduces a novel feature selection approach emphasizing Elo and victory frequency, improving prediction accuracy over traditional box-score features.

Findings

Elo rating and victory frequency features outperform box-score predictors.

Deep learning models achieve better fit with selected features.

Home court factor significantly influences prediction accuracy.

Abstract

This manuscript is focused on features' definition for the outcome prediction of matches of NBA basketball championship. It is shown how models based on one a single feature (Elo rating or the relative victory frequency) have a quality of fit better than models using box-score predictors (e.g. the Four Factors). Features have been ex ante calculated for a dataset containing data of 16 NBA regular seasons, paying particular attention to home court factor. Models have been produced via Deep Learning, using cross validation.