Sprint Speed, Exit Velocity, and Steamer Projections

The Steamer projections are the best baseball projections available where it is possible to get years of back historical data. It is possible there are better projections out there but they cannot all be tested, whereas one can see that the Steamer projections hold up pretty well over the years and have been getting even better recently. It sounds like PECOTA blew everything up this year and they have some sharp guys over there who sound confident they will crush Steamer, but I will need to see a couple years of results before I go with them. Hence for now, I use the Steamer projections here and in my own MLB-related handicapping and I think most other people do so as well. Of course, since everyone else uses them that means there is no edge to to be had unless you improve on them somehow.

Putting Statcast to the Test

Now that we have five full years of Statcast data I believe we finally have enough of a sample to determine its value in making season-long projections. As I am most interested in handicapping, don't know much about projecting counting stats, and don't play much fantasy baseball anymore I will focus on wOBA here. The two most-useful Statcast stats for hitter wOBA projection are launch speed and sprint speed. You can obtain the sprint speed data here, while I calculated launch speed myself from my baseballr dataset although you can likely find a similar dataset from the official Statcast site that would be about as good for these purposes.

With sprint speed, the best variable is the basic MPH figure from that dataset, while for launch speed I have had the best luck using the average of all non-bunt balls in play. I think there might be more potential with launch speed, perhaps only choosing certain types of hits or adjusting for stadium, but to keep things uncomplicated I will use the simple average here.

To find the potential gains from adding sprint and launch speed over using Steamer alone, we turn to a regression, where the target variable is wOBA in the following season and the independent variables are the Steamer wOBA projection at the start of the season, the hitter's Statcast sprint speed from last season, and the hitter's Statcast launch speed from last season. We are also limiting our sample to hitters with at least 200 PA in the current season and season before. We get the following result:

When adding the Statcast variables to the Steamer projection the regression says that 73% of the Steamer Projection plus 0.0026 wOBA per MPH of sprint speed and 0.003 per MPH of launch speed is best. Without the Statcast variables we would instead use 86% of the the Steamer projection (note that by using the 200 PA cutoff we have biased our measurement towards players not having a terrible season and not getting cut so it makes sense the regression would not suggest 100%). This means that the Statcast metrics make a small but meaningful contribution to the Steamer projection, with some hitters expected to gain or lose over 0.010 wOBA compared to the raw Steamer projection:

In short, for most players the Statcast data won't matter much, but for players with exceptionally low or high exit velo + sprint speed relative to their projection we might make material adjustments.

In-Sample Sprint Speed

A common application of the Statcast numbers has to do with using sprint speed to determine the health of injured players. Last season I was consistently on Cleveland early in the year, especially when Carrasco was on the mound, who eventually wound up having leukemia, which just shows you how sharp the market is; good luck betting baseball for a living in 2020. Eventually I looked into the sprint speed statistics and it was fairly obvious Lindor was a cripple with Jose Ramirez not looking too healthy either. I didn't know what exactly what this was worth at the time, but figured it was worth passing on the team for the time being. It turns out the Indians bounced back in the later part of the year and I would have been better off sticking with them.

Regardless, the idea of sprint speed is that it should converge really quickly to whatever the true talent level of the player is, whereas most baseball hitting statistics don't. This means you can use it as an early indicator of a problem. Unfortunately, we don't have game by game sprint speed statistics, but we can look at the impact of in-sample sprint speed across the entire season, using the same regression we used with the last-season stats:

Last season's speed is already partially baked in to the Steamer projection - whether the Steamer team used it directly or not, players who were faster last year had better stats as a result. So the coefficient of last season's sprint speed will be negative here, but if players maintain the same sprint speed they had, sprint speed will make a net positive contribution, by about the same 0.003 wOBA per MPH as before. On the other hand, if a player slows down, their projection will be worse than what we would expect when compared to an already slow player. It is always dangerous to use in-sample statistics when making out-of-sample adjustments, but I think that a safe application of the above is that for every 1 mph of gain/loss in sprint speed versus last year in early action we can adjust our rest-of-season projection up or down not by the full 0.0067 but perhaps by 0.003 to 0.005 wOBA per MPH or so.

Summary

The models suggest that Steamer projections can be improved in a small but meaningful manner by including Statcast metrics. Of course, this assumes that the Steamer folks do not begin including said metrics in their projections, in which case they would be baked-in and you would end up double counting. For what it's worth, the same regression as before run on a year-by-year basis indicated Statcast data added more to the Steamer projections in 2019 than any year so far. This suggests they were not used by the Steamer folks last year, but I do not know what changes they have made to this year's projections.

As we are all in this together I have included a list of the adjustments I would make to the 2020 Steamer projections based on the 2019 Statcast data, for players who had at least 100 PA last year.