Starting on April 4th, I ran a neural network each day to predict the player most likely to get a hit that day. It was predicated on that particular hitter starting. If he didn’t start, I used the next one on the list. These picks could by used by fans who played Beat the Streak at MLB.com.
The results are in this spreadsheet. The program was run for 177 days. Based on the top probability of each day’s top hitter, we expected to see 134 days with the batter getting a hit (133.8, to be exact). Those hitter produced 128 games with a hit, so the NN under-performed, but in an expected range. The 95% confidence interval was 122 to 145.
The longest streak produced by the NN was ten games. It produced streak of at least five games 12 times. The longest streak without a hit was three (three times). There were 22 instances when the missed hit lasted just one game.
The best stretch was from 6/15 to 7/16. During that time, the NN predicted 21 hits in 28 games, but the batters produced 25 games with a hit, and no hitless stretch was more than one game.
All in all, it was a successful run. Now to find a way to improve the model. A 75% model won’t do. At around 85%, the predictor would be able to give one a good chance of winning the game. Unfortunately, 75% may be the best the math can do.
from baseballmusings.com http://ift.tt/2keRmRF
No comments:
Post a Comment