With sports betting becoming more widespread, I thought it would be fun to try to develop a Neural Net (NN) predictor for estimating the winning percentages of two teams involved in a game. The first model kept it simple. It uses the year-to-date road team winning percentage and the year-to-date home team winning percentage. Those are adjusted as follows:
- If the team has played less than ten games, the winning percentage is regressed toward .500. A team with no wins would be .500. A 9-0 team would initially be .860.
- Any team above .750 is adjusted down to .750 (including after the above adjustment). Any team below .250 is adjusted up to .250.
I selected 30 dates from each even year from 1980 to 2010, and used the games on those dates as training samples. That should have resulted in around 6000 training samples, 75% of those were used for training and the other 25% for validation. I ran 40 iterations of training, although the model converged quickly.
The test set consisted of 30 dates from the years 1979 through 2009, 6149 test samples. In addition to calculating the probability of the road team winning, I calculated the log5 probability of the road team winning. The results are in this spreadsheet.
Log5 predicted 3084 road wins, the NN predicted 2782 wins. The games produced 2851 road wins. So the NN model was much closer to the actual result than the Log5 method. The mean squared error for Log5 was .502, for the NN, .452.
The NN did not do a good job, however, of predicting winners. The NN basically did a regression that decided the best predictor was something very close to a .45 probability of the road team winning. Log5 had a maximum prediction of 0.9 and a minimum prediction of 0.1. The NN had a maximum prediction of 0.48 and a minimum prediction of 0.42. It would appear that being on the road conveys much more information than the YTD winning percentage of either team.
What this provides is a baseline. The idea will be to add more information to the NN, and see if it improves. The upshot of this experiment, however, is that betting against road favorites looks like a good strategy.