Searching for Under-Used Breaking Pitches with PITCHf/x

You probably didn’t even notice when the Astros claimed Collin McHugh off waivers from the Colorado Rockies last Dec. 18, and I can’t say I blame you. With a career ERA of 8.94 and a fastball that barely broke 90 mph, McHugh looked like your garden-variety, soft-tossing righty. Players like this shuffle among teams on a regular basis, serving as spare parts or emergency call-ups, and the Astros’ claim of McHugh seemed like just another of these unremarkable transactions.

McHugh, however, went on to exceed expectations by a wide margin. After starting the year in Triple-A, the 27-year-old rookie got called up in late April and proceeded to rattle off a 2.73 ERA over 25 starts. McHugh’s 3.3 fWAR campaign may have been the biggest surprise performance of the 2014, and his breakout seems to be largely attributable to a greater reliance on his curve ball and slider.

McHugh threw 47.1 innings with the Mets and Rockies between 2012 and 2013, and his pitch selection looked something like this: 56 percent fastballs, 39 percent breaking balls (curve ball and slider), and five percent change-ups. McHugh relied primarily on his hard stuff, complemented it with some breaking balls, and threw in the occasional change-up, a pretty standard plan of action for a starting pitcher.

The Houston Astros, however, saw something more. Something about those breaking balls caught their eye: They had a ridiculous amount of spin on them. This discovery was brought to light in a recent piece in Business Week:

The Astros’ analysts noticed that McHugh had a world-class curve ball. Most curves spin at about 1,500 times per minute; McHugh’s spins 2,000 times. The more spin, the more the ball moves during the pitch—and the more likely batters are to miss it. Houston snapped him up. “We identified him as someone whose surface statistics might not indicate his true value,” says David Stearns, the team’s 29-year-old assistant general manager.

Upon seeing the data on McHugh’s pitches, the Astros reckoned he was underusing his breaking stuff, and that he had the potential to be a much better pitcher if he were to throw a few more bendy pitches. The Astros were onto something. McHugh upped his curve ball-plus-slider usage to 54 percent last season — the highest for a starter since Jesse Litsch back in 2008 — and he suddenly went from being Triple-A fodder to one of the better starting pitchers in baseball.

Here’s a look at that curve ball in action:

…and the slide piece:

Most teams evaluate — and acquire — players based on how they’ve performed in the past, or in the case of prospects, on their physical tools. You rarely hear of a team bringing on a veteran player and untapping some hidden potential by changing his approach, yet this is exactly what the Astros did. They took a seemingly unremarkable pitcher, and turned him into a borderline star simply by optimizing his pitch selection. McHugh’s unanticipated success poses an obvious question: Are there other McHughs out there? Is it possible that there could be other pitchers who are chronically under-using their breaking pitches?

To identify underused pitches, one would first need to find a way to pick out good pitches. From there, it’s simply a matter of looking for good pitches that aren’t thrown often. The Astros identified McHugh’s curveball by its high spin rate, which is one of the data points tracked by the PITCHf/x system. A curve ball’s spin rate is certainly important, but it’s also not the only factor that makes a curve successful. Other data available in the PITCHf/x database — such as break angle, break length, and velocity — also seem like they’d be useful in identifying effective curves.

PITCHf/x data is useful for identifying pitchers whose curveballs break sharply or have lots of spin, but don’t tell us anything about those pitches’ effectiveness. I think most would agree that more spin and more movement make a curve ball more effective, but it’s harder to say how much these things really matter — or which one is more important. It’s also not clear where pitch velocity comes into play. Is it better to have a slow curve that keeps hitters off balance? Or is a faster pitch more effective since the batter would have less time to react? These characteristics certainly influence a pitch’s effectiveness, but it’s not entirely clear how.

FanGraphs tabulates a metrics called Pitch Values, which sums the change in average run expectancy from pitch to pitch. At first blush, this metric seems like the perfect means for evaluating a pitch’s effectiveness. But here’s the rub: Pitch Values are more descriptive than predictive. In other words, a pitch with a high Pitch Value isn’t overly likely to remain effective going forward. Neil Weinberg’s recent update of the FanGraphs glossary does a good job of explaining why:

Pitching is complicated and interdependent. Sure, you got a batter to swing and miss at a fastball, but that swing and miss didn’t occur simply because you threw a good fastball. That swing and miss occurred in part because of the quality of your other pitches, your location, and the sequencing of those pitches. In other words, when you get a +0.08 on a specific pitch, that single, solitary pitch isn’t the only reason you got a positive outcome.

So PITCHf/x data can tell us what a pitch looks like, but not how effective it is; and Pitch Values can give us some indication how effective a pitch was, but not to the point where it has much predictive value. By themselves, these two data sets can’t get us very far in identifying good curve balls. But by combining them, we might be able to get a sense of the types of pitches that are most often effective in getting batters out.

Let’s start with curve balls. Considering all pitcher seasons since 2008, I regressed pitchers’ curveball PITCHf/x characteristics onto their wCB/C to see how these characteristics influenced a pitch’s effectiveness. I did some out-of-sample testing in order to determine which particular variables to include in my final model. In other words, I built a few potential models using four years of data, and tested them on the remaining two years to ensure I wasn’t overfitting the data. Here are the regression coefficients I came away with:

Regression Coefficients

Variable	Coefficient
Intercept	55.62
Velocity	-1.496
Spin Rate	-1.86E-03
Velocity^2	1.02E-02
Spin Rate^2	9.48E-07

In everyday English, this means that a higher velocity and a higher spin rate generally make a curve ball more effective, but the effect isn’t linear. The marginal benefit of each mph or rpm is greater on the higher end of the spectrum.

Applying this model to pitchers’ 2014 performances, the following pitchers have the best curve balls (minimum 200 pitches).

Best Curveballs, 2014

Name	Avg. Spin	Avg. Velocity	xCB/C
Yoervis Medina	1,746	85	1.6
Corey Kluber	1,945	82	1.3
Felix Hernandez	2,095	80	1.3
Carlos Martinez	839	86	1.1
Sonny Gray	1,913	81	1.1
Charlie Morton	2,075	78	1.0
Marcus Stroman	1,660	83	0.9
Brett Cecil	707	84	0.9
Jarred Cosart	1,999	79	0.9
Jose Fernandez	1,626	82	0.8

In addition to being statistically significant, this model also seems to pass the smell test based on this list. By this analysis, Corey Kluber and Felix Hernandez have a couple of the best curve balls in the game, which probably isn’t much of a stretch, and most of the other pitchers near the top also have the reputation of having excellent curves.

Now, let’s repeat this exercise for sliders. Regressing pitchers’ PITCHf/x characteristics onto their wSL/C results gives us this model:

Slider Model

Variable	Cofficient
Intercept	-13.4
Velocity	0.109
Spin rate	-2.62E-03
Break angle	-3.02E-02
Break length	1.231
HandR	-0.3545
Spin Rate^2	1.47E-06
Break length^2	-6.04E-02

There’s a little more going on here. Spin rate matters in much the same way that it does for curve balls. Velocity matters again, but the relationship is linear this time. A slider’s break angle (adjusted for handedness) shows a negative relationship, which suggests that sliders that break further away from same-side hitters are more effective. A higher break length is a good thing, but the added benefit really starts to diminish once you get past six or seven inches. And lastly, with all else being equal, a slider is more effective when it’s thrown by a lefty.

Applying this model to pitchers from last season, the following sliders grade out best:

Best Sliders, 2014

Name	Hand	Avg. Velocity	Avg. Break Angle	Avg. Break Length	Avg. Spin Rate	xSL/C
Dana Eveland	L	84	-18	10	1,302	1.6
Bryan Shaw	R	81	-22	11	1,653	1.5
Yu Darvish	R	79	-19	12	1,809	1.4
Charlie Furbush	L	81	-17	11	1,355	1.3
Adam Ottavino	R	82	-19	10	1,472	1.3
Tyler Matzek	L	84	-3	9	457	1.3
Glen Perkins	L	83	0	9	337	1.3
Andrew Miller	L	84	-12	9	910	1.3
Aaron Barrett	R	85	-17	9	1,256	1.2
Mike Dunn	L	87	-7	7	606	1.2

I know what you’re probably thinking. All of this fancy math is cool and all, but can these formulas actually pinpoint guys who could improve by throwing a few more breaking pitches? It seems so. Looking at all starting pitchers (minimum five games started) who increased their percentage of breaking balls thrown by at least 10 percent in consecutive seasons, those who graded out well by my model saw larger drops in FIP relative to their Marcel projections. That regression line has a slope of -.344. Note that I calculated the weighted average (which is on the X-axis) with the following formula, using values from the initial year: (CB*xCB/C + SL*xSL/C) / (CB + SL).

Now for the fun part: identifying pitchers who possess good breaking balls, but for whatever reason, chose not throw them very often in 2014. The graphic below plots the percentage of breaking balls a pitcher threw in 2014 versus the weighted average based on my models. In addition to sliders and curve balls, I included knuckle curves in my count of breaking pitches. Knuckle curves are different enough from standard curve balls that I chose not to lump the two together, but I also didn’t want to ignore them completely.

Unsurprisingly, the regression line trends upwards, which suggests that pitchers with better breaking pitches throw them more often. Basically, the dots above the line represent the players who threw more breaking balls than you’d expect based on their stuff, while the dots below the line signify the guys who threw fewer. I took the liberty of labeling a few outliers in the latter category who aren’t projected to pitch particularly well in 2015. Lets take a closer look at these pitchers.

Kevin Gausman – 3.91 ERA in 2015 per Steamer

Gausman doesn’t quite fit the McHugh mold in that he’s already a pretty good pitcher. Gausman posted a 3.57 ERA and 3.41 FIP over 20 starts in his rookie campaign, but the data suggest that his slider might be under-utilized. The 23-year-old relied primarily on two excellent pitches to get batters out last season: a mid-90s fastball, and a change-up that moves so much that PITCHf/x calls it a forkball. Gausman’s only other pitch was his slider, which he spun just 7.6 percent of the time, or about 11 times per nine innings.

That slider, however, appears to be better than most, due in large part to its break length of 10 inches. Gausman’s success makes it a little hard to justify a change in pitch selection, and perhaps he shouldn’t try to fix what isn’t broken. Nonetheless, I think there’s decent a chance Gausman could be even better if he were to work in his slider a little more often.

Clay Buchholz – 4.20 ERA

Buchholz pitched excellently in his injury-shortened 2013 campaign, but saw the wheels fall off in 2014. He struggled all year long, and ended the year with a dismal 5.34 ERA over 28 games started. Buchholz’s plan of attack revolved around four different fastballs — a four-seamer, a two-seamer, a cutter and a splitter — that he threw a combined 72 percent of the time. The balance of Buchholz’s pitches were change-ups (12 percent) and curve balls (16 percent).

With an average velocity of 77 mph, Buchholz’s bender doesn’t come in particularly hard, but what it does do is spin. His curve averaged 1,950 rpms last season, which was the 11th highest of the 187 pitchers with at least 100 curve balls thrown. Buchholz’s fastball velocity has been on the decline for a few years now, so a shift away from the hard stuff might serve him well.

Rubby de la Rosa – 4.64 ERA

At 82 mph, De la Rosa’s slider comes in fairly hard, but it’s the pitch’s break that makes it tick. De la Rosa’s slider rates above-average in both break angle (-9.2 degrees) and break length (nine inches). Despite the pitch’s break, De la Rosa threw his slider just 13 percent of the time last year, relying instead on his mid-90s fastball. This has been De la Rosa’s plan of attack since he broke in with the Dodgers back in 2011, but things have yet to really click for soon-to-be 26-year-old. Despite having excellent stuff, he holds a mediocre 4.34 ERA over 174 big league innings.

Joe Kelly – 4.66 ERA

Joe Kelly makes three mediocre Red Sox pitchers who fall below the regression line. Kelly’s an interesting case in that he possesses a slider and a curve ball that grade out well by my analysis, yet he threw them just a combined 23 percent of the time.

At 86 mph, Kelly’s slide piece gives opposing hitters little time to react, and also has a decent amount of break. The pitch is a more of an afterthought for Kelly, however; he threw just 97 sliders all of last season. Kelly’s curve ball has been a much bigger weapon. The pitch comes in at 79 mph, which is much slower than his slider, but with an average rpm of over 1,800, the pitch possesses an awful lot of spin. Both of Kelly’s breaking pitches have their redeeming qualities, but have always taken a back seat to his fastball.

Allen Webster – 4.90 ERA

Yet another Red Sox hurler, Webster throws a hard 84 mph slider that also possesses a lot of movement and spin. The pitch spins at a better-than-average 938 rpms, and has an average break length of 8.5 inches. Yet like each of the players mentioned above, Webster has historically been more of a fast ball-change-up pitcher, turning to his slider just 17 percent of the time.

Webster was a fairly-well regarded prospect with the Dodgers a few years ago, but he hasn’t had much success at the big league level. A 25-year-old with a career 6.25 ERA is exactly the type of pitcher who should be tinkering with his approach to stay relevant, and for Webster, a few more sliders might help do the trick.

Honorable mentions

Jeremy Hellickson – 4.15 ERA

Hellickson’s curve ball had the third highest spin among pitchers with at least 200 curve balls thrown last year, but he threw it just 19 percent of the time.

Jared Cosart – 4.51 ERA

Cosart’s curveball ranked fifthh in terms of spin, and 10th overall. He turned to his bender 26 percent of the time.

Trevor May – 4.81 ERA

May puts an above-average amount of spin on both his curve ball and slider, but threw them only a combined 24 percent of the time.

…

In Buchholz, De la Rosa, Kelly and Webster, the Red Sox have four pitchers who seem to have good breaking stuff, but don’t use it very often. While it’s interesting that so many on this list play for the same team, I’m not fully convinced that it has anything to do with the types of pitchers the Red Sox acquire. My methodology for choosing these pitchers was based on two criteria: guys who threw fewer breaking balls than you’d expect based on their stuff, and those who also were projected to pitch relatively poorly in 2015. Frankly, the Red Sox don’t have a particularly good collection of starting pitchers on their roster, which likely plays a role in their having an outsized share of players on my list.

This analysis shows that factors like movement and velocity are decent predictors of a breaking pitch’s effectiveness, but it’s worth keeping in mind that there’s plenty more that comes into play. The R^2 on my models only came out to around .05, meaning about 95 percent of the variation among Pitch Values isn’t explained by PITCHf/x data. At least part of this is likely due to the volatile nature of the Pitch Values metrics, but I think it also speaks to the importance of attributes that aren’t tied to a pitcher’s raw stuff, like deception or command.

Unfortunately, deception isn’t easy to quantify, and the same goes for command — at least as long as COMMANDf/x isn’t publicly available. Nonetheless, the factors that can be captured by PITCHf/x make up a big piece of the puzzle, and can provide us with insight into which pitches are most likely to be effective.