A year ago, I decided to experiment a bit with lineup optimization. If you have toiled around baseball data science a bit, you know that lineup optimization tools come down to a few realizations:
1. No one uses what a tool would consider an optimized lineup.I recognized that state of the research and decided to take on a challenge that most tools do not consider: the linear relationship of a lineup.
2. Everyone is not far off that optimal lineup.
3. The difference between the best and worst conceivable lineup is about 30 runs usually.
4. These tools rarely include enough of the right information to make an informed lineup.
What I mean when I talk about lineup linearity is that each member of a batting lineup exists in context of those who bat around him. While much of baseball data science is about isolation, isolation, isolation; I tried to consider context, context, context. Yes, true talent level is best measured in a vacuum, but talent effect might well be best measured by recognizing how a player's talent is impacted by the talent of others.
Let us consider an extreme example. Let us say we have a singles hitter. Let's say that this singles hitter is very fast. Let's say that he walks a lot, too. How about we given him a line of 350/450/400 and 80 steals out of 85 attempts. The individual we created is a super Juan Pierre. A player like that would be worth about 5-6 WAR. Now, what if I told you that all of his teammates struck out. They struck out every single time without exception. While super Pierre has 5-6 WAR "talent," his "talent effect" is below replacement level.
Why? The way WAR works is to assign a run value to every event. That run value is determined by league averages. What WAR considers is this, what is super Pierre's talent in the average lineup, in the average position in that lineup, in the average base-out condition, with as many other considerations averaged out. You can see how that is a great way to determine Pierre's true talent, but not his effect. Because his effect is linked into his context.
With that in mind, I created (part 1, part 2) a lineup optimization tool that considered how a player does in a particular position in a lineup in relationship to those who bat before him. The model I put together worked well and correlated to actual run production. The model weighs heavily on doubles, home runs, walks, and strikeouts. Those were the primarily determinants in run scoring. It should be noted that one is limited by who actually plays in each position in the lineup. A big bruising hitter batting leadoff is highly uncommon, so the model may well be extrapolating beyond its data capabilities. Weird things may well happen outside of the data set. But what was remarkable about that work was that it suggested that perhaps it was a bad idea to group power hitters. That maybe your best home run-centric power hitter who gets on base should bat lead off.
The model declared that Chris Davis was the best leadoff hitter for the Orioles.
Fast forward to this Spring Training and a major point of discussion was that Chris Davis was in fact leading off games. It was noted as being done to get him more plate appearances, but also noted as testing out the idea that maybe he should well be batting leadoff. It is a scenario that Davis tends to do well with. Last year, when he was confronted with a situation where there were no outs and the bases empty (130 PA), he hit 259/338/534 (129 wRC+) and fared more poorly in other situations with a 184/308/308 (64 wRC+). Now, all that is just gravy. The model does not know those situational stats. What it recognizes is Davis' overall statistics and what it means based on how leadoff hitters have hit in the past.
This post will only look at two different lineups. Yes, the season will offer a myriad of sequences, but we will just play around with this iteration.
This is one of those stunning model results. Optimizing the lineup to the model results in a prediction that major gains in run scoring would happen at the leadoff position (+12 runs), sixth position (+14 runs), and seventh position (+26 runs). The leadoff difference can pretty much be explained by Davis' increase in power over Beckham. Sixth has more to do with run opportunities than differences in hitter makeup. Seventh has nothing to do with the hitter and all about the opportunities he now sees. Still, I really want to reiterate, that it is astounding that the model predicts a difference of 52 runs between these lineups. That would be worth five wins and would greatly improve upon the runs scored by last year's team (743).
Traditional Optimized Tim Beckham 3B Chris Davis 1B Trey Mancini LF Trey Mancini LF Manny Machado SS M. Machado SS Jonathan Schoop 2B J. Schoop 2B Chris Davis 1B Adam Jones CF Adam Jones CF Tim Beckham 3B Anthony Santander DH A. Santander DH Caleb Joseph CF Caleb Joseph CF Colby Rasmus RF Colby Rasmus RF 738 runs 790 runs
These results, however, are not astounding to us because we came to this conclusion last year and that surprise wears off. We also saw about a month or two after publishing our results that several teams began experimenting with our approach (i.e., Kyle Schwarber batting leadoff). With a club like the Orioles, a club without an obvious leadoff hitter and a need to find value in something that few others are doing, this might well be a kind of advantage they can exploit if the model is actually correct.
Maybe the Orioles will venture and give this idea a chance. Or, maybe they will do what everyone else is doing and hope to beat them by playing the same game.
4 comments:
Chris Tillman should be the Opening Day starter. He is the key to being competitive. If his velocity is back, they are OK, if not, the trade deadline will be the only interesting part of the season. Note that if he had his typical season in '17, they would have been tied for the 2nd wild card one week before the season's end.
This is very interesting, I really appreciate it.
I was with you until your model recommended Caleb playing CF
Either autofill or we really stumbled onto something.
Post a Comment