Spring Training Velocity Drops: Do they have Predictive Value?

Velocity drops in spring training are a key area of focus for both fantasy baseball players and sports bettors, and for good reason. Spring training statistics suffer from low sample size and poor predictive value due to low quality of competition and players giving less than maximum effort. But fastball velocity is known to correlate well with getting hitters out while being largely independent of the competition. The key question is whether these velocity drops represent a true decline in performance that will carry over into regular-season play, or simply come as a result of pitchers giving less effort, poorly-calibrated radar guns, or some other unknown factor.

Inevitably, when baseball analysts observe a velocity drop, a reporter will ask the pitcher in question about the decline, at which point, the reporter will be shot down. “My arm feels great, I have never been in better condition, but this is spring, so I am working on a few things” is always the answer provided. As it is a fool’s errand to read into such comments, we turn to a regression study to find the truth behind velocity drops. Sometimes the drop is fake, sometimes it is real, but across hundreds of pitchers who have shown similar velocity drops, we may be able to see a statistical trend.

In this study, we will set up a regression model to predict wOBA allowed using the two best readily-available predictive metrics — strikeout percentage and walk percentage from last season — plus the difference in the pitcher’s fastball velocity between last season and spring. We will focus on regular season games in April, where the impact of spring performance is likely to be most substantial. Along the way, we will also look at the predictive value that both last-season and spring velocity have on several other pitching metrics, both in April and the rest of the season.

We use the MLB Gameday data from 2013–2018 (obtained via the pitchrx and mlbgameday plugins, code used to produce this study is available here), and begin by eliminating pitchers who have a sample size of past statistics too low to be used for a proper comparison. Only pitchers who faced at least 200 batters in the prior season, 40 batters in April of the current season, and had at least 10 fastballs measured by pitchFx in spring training will be used. This limits our sample primarily to established starting pitchers. One challenge with this study is that most spring training stadiums don’t have pitchFx data in the official files, meaning they don’t have spring velocity data. While we would have over 1000 pitcher-Aprils in our sample if all pitchers had met the 10-fastball cutoff, instead we only have 391.

We then must ensure that we are making a proper year-over-year comparison of fastball velocity. Occasionally a “velocity drop” is not due to a true decline in performance, but instead a result of the pitcher in question throwing a different type of fastball. For example, a pitcher moving from a four-seam to two-seam fastball would likely experience a decline in velocity, but this would not represent a true decline in performance. Limiting our study only to pitchers who throw one type of fastball would cut down our sample size too much, so we take a compromise approach. All fastball types are included in the calculation of average velocity, but we also include the pitcher’s average pitchFx z (downward)-movement across all fastballs in both last season and spring games. If the pitcher’s average z-movement has changed by five inches or more between last season and spring, that pitcher is eliminated from the study. This cutoff eliminates about seven percent of pitchers — by far the most common off-season change is a pitcher going from a four-seam to two-seam fastball, a pitch with far less vertical rise:

Z-movement Change

With our sample in hand, we can now run a regression of wOBA allowed in April games versus last season’s K%, BB%, and change in velocity:

woba allowed vs strikeout walk and velocity change

The regression suggests that for every one percent increase in the pitcher’s last season strikeout rate we would expect a wOBA allowed of -0.0027 or 2.7 points lower, for every one percent walk rate we would expect 1.6 points higher, and for every MPH increase in fastball velocity we would expect 5.0 points lower. Despite the low sample of only 395 pitcher-Aprils, the model suggests that velocity changes do have statistically significant value in predicting pitching performance, beyond that of strikeout and walk rate alone. However, the impact is modest: in terms of predicting wOBA allowed, a one MPH difference is worth just a little less than a two percent difference in last season’s strikeout rate.

Because our sample set is small and the size of the impact is not large, it is worth running our model against two metrics that are far more repeatable than wOBA allowed: fastball velocity and strikeout rate. If our model suggests that spring training velocity drops influence these as well, we have further confirmation that the 5 wOBA per MPH estimate is real, and not an artifact of small samples or batted-ball luck. We start with fastball velocity. This time, rather than looking at the year-over-year difference in velocity, we use a regression to predict April fastball velocity based on two variables: last season’s fastball velocity, and spring fastball velocity. If the model suggests that we use a portion of spring velocity to predict April velocity (as opposed to solely relying on last season’s velocity) we can confirm that spring velocity has predictive value. We find that it does:

April FBV

Given the choice of either last season’s velocity or the far more limited sample of spring velocity, the model tells us that the best choice is to use 49.9% of last season’s velocity and 47.1% of spring velocity, suggesting spring training velocity figures do seem to have some predictive value in the early season, over and above last-season’s figures. We next turn to strikeout rate:

April strikeout rate

To build a prediction of April strikeout rate, the model tells us the best choice is to start with 7.9%, then add 64.3% of last year’s strikeout rate plus 0.67% per one-MPH change in velocity. Like all baseball stats, strikeout rate does regress to the mean over time, and changes in fastball velocity are only part of such regression. The fact that the model suggests that over one-thirds of performance in strikeout rate can not be predicted from past strikeout rate alone reflects this.

Overall, the data suggests that spring performance does carry over into April, and we should adjust our performance projections slightly downward on pitchers that have shown velocity drops in spring. Pitchers with increased velocity in spring games can be expected to improve in wOBA allowed and strikeout rate, with this performance likely due to their higher velocity in April games compared to those pitchers who had velocity drops.

To close out our study, we show the results of pitchers grouped by velocity gain or loss. Most pitchers start off the year throwing at lower velocity compared to the year before, then build up as the season goes on. In addition, pitchers naturally fall off each year due to age. The combination of these two factors means that the average pitcher in this sample threw 0.9 MPH lower in spring than they averaged the season before. We therefore group pitchers into three groups — those who lost 2 MPH or more, those between 2 MPH less and the same velocity, and those pitchers who gained velocity:

April and ROY perf

Pitchers with 2+ MPH drops tend to be better overall pitchers (at least by past performance) than those who gained velocity. There are a couple reasons this makes sense: one, since they were better before, regression to the mean means it is “easier” for them to get worse, and two, better pitchers tend to be more established and probably have less reason to perform at max effort in spring training games. Either way, pitchers who had velocity drops of two MPH or more performed worse than the other pitchers in April games, at 21 wOBA points worse than the year before, while velocity gainers slightly over-performed versus last year’s stats.

Moving on from April to the rest of the season, while we have established that velocity droppers perform worse in April, it is possible that velocity drops are not indicative of some type of injury or loss of form, but rather differences in player training patterns, with some players peaking earlier than others. The above table would suggest this the case, as the velocity droppers actually outperformed the other pitchers over the rest of the season while also showing a bounce-back in velocity, and a regression study confirms this. After accounting for the strength of the pitcher in the season before via strikeout and walk rate, we find no relation between spring velocity drops and pitcher performance outside of the month of April, and in fact spring velocity gainers actually performed slightly worse than those with velocity droppers, although the impact was nowhere near statistically significant.

Overall, spring velocity drops are real, and do correlate to early-season performance. A rule of thumb is that the average pitcher drops one MPH from last season to the following spring, and every MPH above or below that is roughly equal to two percentage points of prior strikeout rate in the month of April, in terms of its impact on expected pitcher performance. But the impact of spring velocity tends to be limited to April only and does not carry over into the rest of the season, suggesting that most velocity drops are due to short-term differences in fitness and not long-term loss of form. This means that the season-long fantasy impact of velocity drops is very limited, and may even create a value situation if the news gets to be too widespread. However, daily fantasy and sports bettors may wish to avoid velocity droppers in April, unless the price on offer is extremely favorable.