wOBA Using Exit Velocity and Launch Angle

After reading this post on predicting batted-ball type based on exit velocity and launch angle, I thought it might be neat to see how it could be applied to reducing the effects of luck and defense on wOBA.

The idea is that wOBA correctly values outcomes; however, wOBA implicitly assumes the batter has total control over the outcome of a batted-ball event based on his input to the system. I tried awarding the batter value based on the expected value of balls hit with similar launch angles and exit velocities. Instead of treating the outcomes as events, I assigned them a value based on the wOBA weights used for the season of interest.

I used the random forest classifier from the referenced post but looked at outcomes relevant to the wOBA formula. The probabilities of each outcome based on the exit velocity and launch angle of the batted-ball events are multiplied by the wOBA weights to give an expected value for the batted-ball event.

The classifier was trained on all batted-ball events from the 2015 season. The model accurately classified only 70% of the 2016 batted-ball events, so there may be a problem with over-fitting. The use of probabilities rather than plain classification should help to reduce error.

The plot below is a graphical representation of the classifier. This shows what the classifier believes to be the most likely outcome for different levels of EV and LA. When using the individual probabilities, the model is more smooth.

I compared the number of at-bats for a player to regress to their season-long mean for this value metric with the number of at-bats for the player to regress to their season-long mean for the corresponding section of their wOBA. It should be noted I only counted at-bats where the batter put the ball in play.

It looks like it is capturing some information about the batter that is lost by considering only the true outcome of the batted-ball event. It takes fewer at-bats for the error to stabilize.