This is the second follow up to the post on predicting wins. In the original, a neural network (NN) fed the winning percentage of the road team* and the home team was compared against the Log5 method. In the second post, team OBP was added to the mix, for the batters and the pitchers of both teams. While the NN in the original underestimated the actual wins of road teams, the NN in second over-estimated the road team winning percentage, but not as much as the Log5 method.
*Since this is binary, it’s 1- the probability of the home team winning.
The third try added five parameters. For each starting pitcher, the one and three-year OBP allowed up to that date, (four parameters), plus the league OBP of the season up to that date were included. This did make an improvement. This new NN now predicts 2975 wins versus 2851 actual wins in the test set. With 6149 games in the test set. That’s down from 3054 without the starting pitcher information. In terms of winning percentages, that’s .463 for the test set, .484 for this run, and .497 for the previous run.
The NN is also giving a wider range of values, with the low coming in at 0.30 and the high at 0.66.
The numbers continue to move in the right direction. Power number get added next.
from baseballmusings.com http://bit.ly/2EUtH2l
No comments:
Post a Comment