Here are the top picks my programs produced for use in Beat the Streak. This post mostly explains the ideas behind the calculations. In addition, this post shows tests on the Neural Network (NN).
First, the Log5 Method picks:
0.374 — Martin Prado batting against Luis Perdomo
0.354 — Marcell Ozuna batting against Luis Perdomo
0.350 — Jacob Realmuto batting against Luis Perdomo
0.349 — Christian Yelich batting against Luis Perdomo
0.343 — Ichiro Suzuki batting against Luis Perdomo
0.328 — Yunel Escobar batting against Tyler Duffey
0.321 — Jose Altuve batting against Adam Wainwright
0.318 — Derek Dietrich batting against Luis Perdomo
0.314 — Melky Cabrera batting against Mike Pelfrey
0.310 — Adeiny Hechavarria batting against Luis Perdomo
Luis Perdomo is a rookie pitcher for the Padres who allowed 63 hits in 36 innings. That works out to a .382 BA allowed. That’s a BABIP of .431, and the opposition has seven home runs. He does walk and strikeout quite a few batters however. It’s a small sample size, but his BABIP in the minors was pretty high as well.
The NN looks at it like this:
0.321, 0.757 — Jose Altuve batting against Adam Wainwright.
0.309, 0.742 — Xander Bogaerts batting against Kevin Gausman.
0.374, 0.738 — Martin Prado batting against Luis Perdomo.
0.303, 0.734 — Jose Abreu batting against Mike Pelfrey.
0.328, 0.734 — Yunel Escobar batting against Tyler Duffey.
0.314, 0.730 — Melky Cabrera batting against Mike Pelfrey.
0.279, 0.729 — Josh Harrison batting against Noah Syndergaard.
0.261, 0.728 — Ender Inciarte batting against Anthony DeSclafani.
0.297, 0.728 — Danny Valencia batting against Derek Holland.
0.266, 0.728 — Francisco Lindor batting against Ian Kennedy.
The NN seems to value long-term pitching trends more than short term ones, so it sees Perdomo in a slightly better light. Martin Prado appears to be the consensus pick, however, with Yunel Escobar and Jose Altuve getting some consideration as well.
As always, these picks are not sure things, and will be wrong about 25% of the time. Good luck!
You need to include park factors for these programs to be valid. More robust models almost always yield at least a few Colorado players in the top 10 (provided they are playing at home).
(Would be very interested to see your results if you did at the park factors, though. Have never used NN.)
Hueyduey » It’s a good idea, although a lot of parks never really get a chance to play long term before they are modified. I’ll see what I can do. One of the nice things about NN is that once the program structure is in place, it’s fairly easy to try new things.