Testing Neural Networks

I’ve been thinking about applying neural networks (NN) to baseball research, and have been starting to build code. My first project was to build an age classifier for batters. I picked eight parameters, Plate Appearances, AB rate, hit rate, doubles rate, triples rate, HR rate, walk rate and strikeout rate. All rates were per plate appearance. I then put the player into four classes bases on seasonal age:

• <25 = Young
• 25-29 = Prime
• 30-34 = Past Prime
• >35 = Old

I tried a number of different NN models, but all produced pretty much the same results, the age distribution of the group of players. In other words, if 40% of the players in the training group were in their prime, the NN told me the probability of the player being in his prime was around 40%. Just looking at a season of these parameters gives almost no information about the player’s age. That was disappointing.

I was hoping if the age classifier worked, I could go back and look at players from the PED era and see how their careers looked. One might expect the PED users to have Old years that classify as Prime or Young.

Now, at least, I have some code in place that could be modified for other projects, maybe a projection system.

from baseballmusings.com http://ift.tt/1KujWWh