Earlier today I brought up the Friday contest at MLB.com, trying to pick 57 hitters who will get hits in a single day. I suggested two strategies for picking players from the Beat the Streak Neural Network list. The first was to take the top 57 players starting that day. The second was to take a random sampling by rolling two dice as you move down the list, and if the roll was a 2 through 7 or an 11, take the next player.
This afternoon I wrote a simulator to see how the two systems compared. The results are in this Google spreadsheet. I used 80 days with 15 games, 20 each from 2003, 2005, 2007, and 2009. None of those games were included in the data that built the NN.
For picking the top 57 players, the best runs yielded 50 players with hits. That happened twice. The expected number of players with hits on those two days were 40.7 and 40.8. So the NN got lucky both days. The worst runs were 33, also twice, with expected hits those days of 40.9 and 40.6 The average expected number of hits was 40.7 with a minimum of 39 and a max of 41.7. So the predictions were very consistent.
For picking down the list based on a random value, the best runs yielded 50 hits, with an expected value of 40.6. The worst run yield 32 hits twice, with expected values of 39.7 and 40.6. The average expected number of hits was 40.2 with a minimum of 38.7 and a maximum of 41.
Finally, the top 57 beat the random selection 43 times, the random selection won 23 times, and there were 14 ties. The top 57 averaged 41.9 batters with hits, the random selection 41.0. Note that both out performed the NN prediction, but it looks like choosing the top 57 is the better system.
Good luck!
from baseballmusings.com http://ift.tt/2uXL8qA
No comments:
Post a Comment