Revisiting Stuff+: An Update on Driveline’s Methodology to Quantifying Pitch Design
Introduction
It’s been nearly 3 years since our first post explaining our Stuff+ model. If you haven’t read the post yet, I would highly suggest doing so, as this will largely be building off of much of the content already explained in there. Since the original post was written, we’ve seen the emergence of many public stuff models and have had significant strides from the baseball community in improving these models. Let’s explore what Driveline has done to keep our stuff model among the best in the business.
Why We Update Our Models
Once a model has been created, it is not as simple as just letting it sit and make predictions over time. Even if none of the inputs into the model are updated, it’s important to continue to retrain models on more up to date data, as MLB and baseball in general is constantly changing.
In the case of a stuff model, as soon as strides are made from pitchers to find an edge, hitters begin adapting to it. The best recent example of this is the sweeper. Partially due to the rise in prominence of stuff models, the sweeper began gaining traction a few years ago as an elite pitch shape, specifically against same-handed batters.
The sweeper has been an incredibly effective pitch to generate whiffs. In 2021 sweepers generated the second highest whiff percentage of all pitch types, behind only the splitter. It was after this season that sweeper usage really ticked up, but as of 2024, sweepers are now generating whiffs at rates comparable to changeups and curveballs.
It’s still a quite effective pitch, on the whole, but hitter adaptation is something we need to account for and one of the reasons we must always be on the lookout to update our models with the most up to date data.
Another clear change in recent history that we must account for is velocity. Take fastball velocity, for example.
It’s no secret that pitchers are throwing harder than ever, but there is clear hitter adaptation at play here as well. Despite consistently increasing velocity, hitter whiff percentages have hovered around between 21.5% and 22% without significant variation.
The bar is constantly being moved, and we need to make sure we are on top of the most recent changes across the league from both hitters and pitchers.
What'd We Change?
While updating the model on new data with the same inputs is valuable in itself, we have made some changes to our most recent model. That said, the basis of the model is the same, in that pitch locations are omitted and the primary ball flight metrics considered are as follows:
- Pitch Velocity
- Vertical Break
- Horizontal Break
- Arm Angle
- Extension
In addition, our main pitch types are fed into three different buckets still:
Fastballs
Offspeed
- Changeup
- Splitter
And finally, all pitch type scores can only be compared directly within their respective categories, as the conversion from expected run value to the “plus” scale occurs based on the average run value within each bucket. All this to essentially say a 110 slider is not the same as a 110 sinker.
There were two main changes to the model. The first involved improving our comparisons to a primary pitch. For breaking balls and offspeed pitches we incorporate information about a pitcher’s “primary” pitch to better contextualize any value that could stem from velocity and movement relative to what a hitter may expect. We previously defined the primary as the highest usage percentage out of either a 4-seam fastball, sinker, or cutter. In doing so, we embed a quality of “arsenal effect,” recognizing that difference in velocity or movement from a primary pitch can influence a pitch’s effectiveness.
An interesting trend we’ve noticed since the start of 2021 is an increase in potential primary pitches in a pitcher’s arsenal. This is reflected by a general trend in decreasing 4-seam fastball usage and increase in cutter usage.
It would appear as if we’ve entered a point where pitchers have more pitches that are well suited for specific platoon effects or other situations. Because of this, it is important we better contextualize what the pitcher’s true primary pitch is in a given situation. We have done this by defining a primary pitch’s velocity and movement based on the highest usage out of 4-seam, sinker, and cutter broken down by batter handedness now, rather than just overall.
A good example of this is Aaron Nola this season, who we defined as having a sinker primary vs RHH and 4-seam primary vs LHH. In this case, his breaking balls and offspeed are compared to each respective movement and velocity profile according to what the handedness of the hitter is.
The second major change we made to our model was incorporating adjusted approach angles. As mentioned in the previous blog post, our previous model only considered total movement and velocity, not how that movement occurred.
“The model currently is agnostic towards how the movement is created (spin-induced vs. non-magnus) ; it only cares that the movement and velocity is generated in some manner.”
In this iteration, we included two different approach angle metrics. The first, location adjusted vertical approach angle, gives the vertical approach angle adjusted for the vertical location of the pitch (https://blogs.fangraphs.com/a-visualized-primer-on-vertical-approach-angle-vaa/). The second, adjusted horizontal approach angle, gives the horizontal approach angle adjusted for horizontal location and horizontal release point (https://blogs.fangraphs.com/a-visual-primer-on-horizontal-approach-angle-haa/). In doing so, we improved our grading on pitches that have generally been known to benefit from non-magnus spin along with a more accurate depiction of some pitches that come from deceptive release points; more on these categories will be explained below.
What'd We Find Out?
At the highest level, cutters were the biggest gainers from the old model to the new model, with 4-seam fastballs and sweepers getting hit the hardest. It’s worth noting that, as discussed previously, the sweeper still grades out well above average, it just isn’t rated as highly as it was when the last model was trained and it was less common across the league. On the opposite end, the cutter jump is the most significant we saw, but only brought the pitch up to average.
Let’s dig into some more specific insights broken down by pitch type buckets.
4-Seams and Sinkers
The most obvious difference we found between the previous and newest stuff model was the relationship between velocity and Stuff+.
Previously this relationship was essentially linear, with every MPH in velocity being worth slightly over 6 points of Stuff+. We now recognize a more exponential relationship between the two, indicating a potentially widening gap between elite fastballs and everything else. Based on the hitter ‘s ability to hit fastballs staying relatively constant despite league wide velocity increasing, there is lots of reason to believe this is the case.
This relationship is further supported by the consistency of the “dead zone” until the ~96 MPH point. It isn’t until this point that the dead zone begins to finally shrink significantly and velocity overpowers shape.
Extending this trend to 4-seams and sinkers separately shows another interesting relationship. Our newest model generally views sinkers as more valuable than 4-seam fastballs until velocity reaches ~97 MPH.
Again, this can likely be tied back to the fact that the purpose of these two pitches is different. 4-seams are generally better at generating whiffs, while sinkers are better at generating ground balls. With whiffs being harder to get as batters have adjusted to higher velocity around the league, the sinker profile becomes more valuable than 4-seam at lower velocities. More on this from Director of Pitching Chris Langin here:
Examples of reclassifying Four-Seamers.
Instead of bucketing all four-seamers into the same classification, we would separate Pete Fairbanks' relative cutting 4-seamer from Jimmy Hergert's 4-seamer, which has significantly different action on it. pic.twitter.com/KOuWVDVQV2
— Chris Langin (@LanginTots13) April 1, 2021
Beyond velocity trends, the addition of location adjusted approach angles helped to better contextualize vertical break in fastballs. Not all vertical breaks are created equal, and low release points can create effective carry while putting up induced vertical break (IVB) numbers that aren’t conventionally impressive. A prime example of this is Alexis Diaz, who falls into the traditional dead zone with a 9.5 inch HB and 13 inch IVB profile.
Diaz’s extremely long extension (7.7 feet!) and low release point allow him to consistently reach a flatter than average vertical approach angle, giving him an above average Stuff+ on his 4-seam fastball. This relationship can be generalized to the broader population when looking at the progression of Stuff+ by shape by different location adjusted VAAs.
Sliders, Sweepers, and Curveballs
Our model’s view on traditional breaking balls remains relatively unchanged. Regardless of movement profile, the simplest (not always easiest) way to improve Stuff+ is to increase velocity. Take the difference between successful movement profiles from 81-85 vs 86-90 MPH.
With that being said, there were more general trends in the value of different pitch types, specifically with curveballs getting a slight bump in average value, and sweeper receiving a slightly larger hit. One particular trend is that our newest model now values low spin efficiency pitches higher than our old model previously did.
Cutters and Gyro Sliders
Looking specifically at pitches that can generally have lower spin efficiencies, like cutters and sliders, we can see that total movement is largely tied to spin efficiency.
This correlation matches intuition and the laws of physics, as efficiency spin creates Magnus force which contributes directly to movement. However, not all movement is created the same. Gyro spin, for instance, can create unique profiles that don’t necessarily have the largest shape but are still effective.
Including location adjusted approach angles accounts for some nuance of these effects. In doing so generally applies a higher value to low spin efficient pitches, with the two models beginning to agree as spin efficiency increases.
So, What's Next?
Stuff models will continue to remain crucial in our pitch design and on-field analysis. The changes discussed here are not all inclusive, but an example of how we can continuously improve and account for changes across MLB and baseball in general. With that being said, “stuff” is not the only component of pitching. Context from the perspective of building an effective arsenal and commanding pitches remains crucial, and Stuff+ is only one of the models we use to evaluate pitchers.
Comment section