The Vulture: Non-Closers Preying on the Win

A vulture, according to a quick Google search is:

a large bird of prey with the head and neck more or less bare of feathers, feeding chiefly on carrion (the decaying flesh of dead animals) and reputed to gather with others in anticipation of the death of a sick or injured animal or person.

Mike Baumann, of the Baltimore Orioles, not FanGraphs.com, has been given the nickname, “The Vulture” due to his 2023 fondness for swooping in and picking up the win once the starter leaves the game. It’s a pretty badass nickname and though by literal definition may not be flattering, it’s still pretty cool. I think Bauman and the rest of the O’s bullpen leaning into it would be fun. Just imagine the reliever taking a huge bite out of a cherry snowcone right before running out of the pen. He begins his warm-up pitches from the mound, red dripping down his chin. The vulture has entered the game.

Bauman is tied with Colin Poche for most wins among non-closer relief pitchers with nine (as of this writing). Does that mean that they will continue to vulture wins? This time last year the non-closer relievers who led the MLB in wins were Adam Cimber with eight and Diego Castillo with seven. Cimber finished the season with 10 and Castillo did not record another win, finishing with seven, though he dealt with a few injuries. However, from mid-August to the end of the season in 2022, there were a handful of non-closer relievers who recorded at least three more wins.

In fact, Alexis Díaz, Adrian Morejon, and Chris Stratton each recorded four more wins, though you could say that Díaz had taken over the closer role by that point in the season. Regardless, rest of season (ROS) projection systems are not going to attempt to call out those relievers. Take Bauman for example. He has nine wins in 54 games good for a 16% win rate. He’s projected (by Steamer) to appear in 16 more games. If we multiply his win rate (.16) by his expected ROS game appearances (16), he should be good for 2.6 more wins. Yet, Steamer only has him for one. Projection systems are more conservative in this instance because, well, they should be. Predicting Bauman for three more wins would probably be irresponsible. But, it could happen!

The question is, can we attempt to identify the pitchers on the outer edges of the norm, the pitchers who are likely to gain three and maybe even four more wins? Jeff Zimmerman got wind of my intent to answer this question and did his best to dissuade me, referencing his last attempt and the uncertainty that he came to conclude himself. I often find, when the idea for my weekly RotoGraphs topic finally elbows its way into my cramped and clustered brain, that Jeff has already written a great piece on that very topic. My work here is slightly different from his, yet, we both try to find strategies to identify relievers who are easily available on the waiver wire and good for a few more wins this season.

For this experiment, I took data from 2021 and 2022 from the start of the season through August 14th of each year. I spliced some data together to look at relievers with at least 30 innings pitched in that time period and brought in additional information such as the amount of high-leverage innings they pitched (HL_IP) and how long (IP/GS by team starters) their starters typically threw into the game. I also added in each team’s starting pitcher ERA (TEAM_SP_ERA). Lastly, I brought in the number of wins each reliever recorded for the rest of the season. That gave me the following features in my dataset:

Features: W, L, SV, IP, K/9, BB/9, HR/9, BABIP, LOB%, GB%, HR/FB, vFA (pi), ERA, FIP, IP/GS, TEAM_SP_ERA, HL_IP

Target: ROS_W

The model I built takes on a lot of serious assumptions, mainly, that in-season performance is predictive of rest of season wins. It would take me way more than a few after-work hours of coding to really determine if that assumption is correct, but that’s what we’re working with here. With these predictive features, I built a decision tree regressor model to make predictions on how many wins each reliever could be expected to earn for the rest of the season.

I’d like to point out that tuning this model was challenging, and that predicting wins is hard, and while I am somewhat offering rest-of-season predictions, I’m not claiming to have better predictions than anyone else. ATC, Steamer, THE BAT, are all sound projection systems that you can utilize here. My goal is to try and identify a few relievers who could potentially get one or two more wins than what those projection systems have identified.

In the end, the best model I could muster had a mean squared error of 1.5, which means on average, my predictions are off by 1.5 wins. Considering most relievers will earn between zero and two more wins this season, the model ends up being just slightly helpful. The tree builds decisions like so:

I’ll be explaining the leaf all the way to the right that is predicting 2.455 more wins. From the starting split at the top of the tree, the decisions move “Yes” candidates to the left and “No” candidates to the right. To get to our 2.455 more wins leaf, the tree breaks out relievers who have thrown in more than 8.15 high-leverage innings. That might not seem like a lot considering Félix Bautista has thrown in 30.2 IP so far this season, but remember that closers have been removed from this data and they are the ones who typically throw in high leverage.

This model is telling us that if a middle-reliever has thrown in more than 8.15 high-leverage innings, he’s that much more likely to pick up a win. Moving to the next node we see a split on fastball velocity, greater than 94.3 MPH. Move to the right, and high-win-probability relievers need to also have thrown in at least 46 innings this season.

So there we have it. It’s that simple. Rest of season high-win-probability relievers, according to this simple model, must have pitched in high leverage, have a good fastball, and pitched in a high number of innings in total so far. This is basically telling us that relievers who have established confidence and have a good fastball end up in line for the win most often.

You can do that same exercise with the tree to see that there are other reasons that include pitching less often in high leverage, but recording very little losses and maintaining a high LOB%. Finally, here are the relievers who end up in the leaf all the way to the right, predicted to record another 2.455 wins this season:

I’ve highlighted both Brazoban and Thompson as they are injured and might not get enough chances to record wins moving forward. Note that I’m not suggesting you should add and roster these pitchers the rest of the way. I have made no mention of what they could potentially do to your ratios.

However, keeping an eye on these relievers using the Roster Resource Closer Depth Chart will allow you to keep track of their usage and keeping in mind who is starting and the opponent will be key. With any experiment, there come suggestions for future work and I have the following:

• dig in further to the predictors, it’s odd that team starter IP/GS and starter ERA did not pop in the model
• play with time periods
• add more predictive features
• conduct backtesting

We all know there are vultures out there, waiting on the trim of the bullpen wall, sitting and staring as the starter gets more and more tired and less and less likely to fall in line for the win. They can smell it, they can smell the thick stench of a win filling the stadium air and entering their featherless nostrils. Bite the cherry snowcone, enter the game and take the W…and try to do it three more times this season.