Experts vs. Equations: A Case Study in the Prediction of NBA Games

• Activity Number: 531
• Type: Topic Contributed
• Date/Time: Wednesday, August 4, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Michele Meisner* and Chamont Wang+ and Danielle Zanghi and James Fitzpatrick
• Companies: The College of New Jersey and The College of New Jersey and The College of New Jersey and The College of New Jersey
• Keywords: Neural Networks ; Regression ; Stochastic Gradient Boosting ; Variable Importance ; Marginal Effect

Abstract

In a recent issue of Journal of Quantitative Analysis in Sports, Loeffelholz et al. (2009) used neural network models to predict the success of teams in the National Basketball Association (NBA). Their results were that the best model predicted the winning team correctly 74.3% of the time, whereas the experts correctly guessed the winning team 68.7% of the time.

In our attempts to re-produce the 74.3% accuracy, we were able to achieve the accuracies of 87% - 90% with n = 30 or 20 games in the training data in cross-validations, whereas the models in Loeffelholz et al. were built with n = 620 games.

Both the results of Loeffelholz et al. and our investigation appear to support a common belief that if we have enough data, then statistical methods would outperform human judgment. Our further investigation indicates that the competition is far from over, at least in NBA prediction.