This early in the season, when the samples sizes are small and the conclusions fragile, attention tends to focus most especially on pitch velocity. That’s a reasonable thing to do, since velocity stabilizes quickly. Because velocity is the simple result of the acceleration a pitcher can impart to his pitch, it is not as affected by luck as, say, batting average or ERA (or even the advanced metrics like tAV or FIP). Moreover, velocity is of prime importance in predicting a pitcher’s future success.
Velocity is deceptively simple. It is only a single number, and yet it can explain multitudes about a pitcher: the kind of pitches he throws, how he gets hitters out, and in aggregate, the kind of pitcher he is. I set out to explore the early-season changes in velocity, and to what extent they are predictive of the season’s velocity.
I ought to note up front that this analysis is well-trodden ground. However, there is room for exploration and improvement. It’s a new year, and so there’s a new crop of April data to be mined for potential velocity breakouts (and breakdowns).
The Consistency of Velocity
Perhaps the most striking feature of velocity (besides its obvious implications for pitcher performance) is how consistent it is between seasons. Here are 118 pitchers’ median fastball velocities from 2012 and 2013, plotted against each other. For the remainder of the analysis, I’ll limit myself to starters in 2013, attempting to predict their velocities for 2014.
The year to year correlation is R2 = .78, suggesting that most of the variation in 2013 is accounted for by the previous year’s fastball velocity. The big gainers and losers tend to be at the extremes of fastball velocity, and often come from pitchers who are recovering from surgery or who are about to undergo surgery. One other point is worth noting here, and that’s the red line, or regression of 2013 velocity on 2012 velocity. Notice that the slope of that regression is below 1 (the black dashed line), which illustrates that, on average, pitchers tend to lose velocity year to year, at a rate of about .1 ticks per year.
What the above graph suggests most strongly is that there really isn’t that much room for velocity to change between years. Generally, pitchers aren’t going to pick up five mph on their fastball (although they might lose it via injury). The biggest change we ought to expect is on the order of a couple miles per hour.
Predicting the past
Returning to April, we’d like to know whether April fastball velocity has any significant predictive value for forecasting the overall year’s velocity. Unsurprisingly, it does: The linear model’s R2 jumps from .78 to .91 after integrating April velocity. But the better question might be: How much data do you need to reliably guess a starter’s velocity?
I took the first few games of each starter’s season and used that to guess the starter’s velocity for the remainder of the year; the root mean squared error (a measurement of accuracy) is plotted on the y axis. Each line here represents applying that procedure to a different pitcher, and the black dotted line represents the mean performance for all pitchers. Generally, one can guess fastball velocity from only a single start, but the error declines precipitously as you add more data. However, note that there is substantial variation between pitchers in terms of how accurate the guess is, so that some pitchers’ fastball velocities don’t stabilize until much later. What’s more, a single bad start (perhaps due to a hidden injury, or just a bad day) can skew the guess far off toward inaccuracy (note the large spike at game six).
The above graph was generated using only those first few starts, however. That data ought to be regressed heavily to a pitcher’s pre-existing fastball velocity. When we do so, the error of our prediction drops dramatically (as pictured below).
At this point, a pitcher’s velocity for the full year can be predicted within about one mile per hour based on 1) their previous year’s velocity, and 2) the velocity of their first few games. Which is to say that while the April data is not entirely without value, it needs to be taken with a metaphorical grain of salt. More properly, it needs to be considered in the context of a pitcher’s career velocity. When doing so, great accuracy for a given year is, in principle, achievable.
Predicting the future
Now comes the harder job of predicting the future. I took the model from the previous years (2013/2012) and fed it April data from this year as well as fastball velocities from last year. Bearing in mind that the model cannot take into account outside contingencies, such as injury, it ought to be reasonably accurate at forecasting final velocities for this year. Without further ado, let’s proceed to a list of pitchers with velocity gains and losses.
Let’s begin with the good news (what little of it there is). Here are five pitchers predicted to gain the most velocity.
First on the list is Wade Davis. A starter in 2013 who converted to the bullpen this year, we ought to expect his velocity to have increased, and it has, for each of his offerings. Davis tells us that the method is sensitive enough to detect genuine increases in velocity, even when they occur for altogether obvious reasons. (I did re-run the model with and without reliever/starter transitions (2012–>2013), but it was no different in terms of its predictions.)
The other pitchers have much less pronounced increases in velocity, less than a single mile per hour. John Lackey’s velocity sits at the highest level it’s been for years, although improved results have not followed. Joe Saunders has gained a bit on each of his pitches, delaying the inevitable ravages of time for another year. Aaron Harang’s increase is modest, although his secondary pitches are also showing some velocity improvement. Finally, Dillon Gee barely cracks the list with a minor half-mile-per-hour uptick.
Ian Kennedy is an intriguing case. Unlike most of the older folks above, he’s not simply recapturing some velocity he had lost: The four-seam speed he’s showing now is the highest April reading he’s ever had, by about 1 mph. It also coincides with the best FIP (2.52), which outperforms his projections handily. Small sample size caveats abound, but he looks like a good candidate for improvement.
Now for the sad news.
Lucas Harrell has lost speed on all of his pitches, in worrying fashion, but most especially with his cut fastball. This velocity deficit coincides with a dramatic increase in his usage of the pitch, for reasons unknown. Given Harrell’s results thus far, that may be a choice worth reconsidering.
Justin Masterson’s velocity drop is so precipitous and pronounced there may be something more than aging at work. Coming off a great year and a contract extension, Masterson is flagged as something of a low injury risk. However, the profound drop in velocity over three starts looks like something more than a change in strategy.
Bronson Arroyo’s velocity drops once again, although given his relatively junkballing ways, perhaps that drop is not a matter of much concern. It is worth noting that his velocity has been gently declining for several years.
James Shields is next. Shields’ underlying four-seam fastball velocity has remained quite steady for the last several years, but his cutter’s speed has dropped a fair amount. This year, Shields has increased his usage of that cutter a great deal. For these reasons, I’m not convinced that the velocity drop Shields is showing is indicative of a loss of talent or aging so much as a conscious choice to throw a lot more of a slightly slower cutter. This choice manifests in the data as a drop in mean fastball velocity, but it may not lead to any decreased effectiveness. On the other hand, his results have not been up to the usual Shields standard so far.
Finally, Clay Buchholz’s four-seam and sinking fastballs have both dropped a few ticks. This change coincides with increasing usage of a splitter, which interestingly hasn’t lost velocity. Whether this was calculated or not, Buchholz’s velocities resemble the heat we saw from him at the end of last season, instead of his normal, pre-2012 speeds. The decrease may indicate that he has not fully recovered from his last set of injuries, or is hiding a new, cryptic one.
Which fact raises an important point concerning all of the above analysis: namely, that there are external factors, such as injuries, that can affect velocity in ways that the model cannot predict. Whereas injuries are predictable to some limited extent, for the most part, they are still random occurrences. Changes in pitch usage, or the intentional “dialing down” of a pitch, whether for strategic reasons or to prevent injuries, can confound this kind of analysis.
Fastball velocity in a given year is fairly predictable, based primarily on the velocity in the previous year and a few starts’ worth of information in the predicted year. Most of the prediction is relatively straightforward: A pitcher’s velocity in year N is going to be very similar to what it was in year N-1. In rare cases, some of which I’ve highlighted above, April starts can give clues as to the coming year’s velocity changes. Finally, while overall fastball velocity is predictable, that average trend conceals a great deal of variability underneath the surface. Heeding Russell Carleton’s call to investigate the variation between players, I’ll look into how quickly individual players’ velocities stabilizes in a future article.