Written by: Kai Franke
Follow him on Twitter: @KaiFranke3
Follow Prospects Worldwide on Twitter: @ProspectsWorldW
To see the last post from the Pitch Grades series (Fastballs and Curveballs), go to this link here.
For Slider, Part 3: https://prospectsworldwide.com/2020/08/02/part-3-sliders-predicting-pitcher-performance-based-on-stuff-alone/
With my fastball and curveball grades write-up posted a couple of weeks ago, I wanted to continue with this series and present to you guys my changeup pitch grades. I will be utilizing these grades to evaluate and rank pitchers using Rapsodo data from Minnesota Blizzard pitchers. As a reminder, these grades are based on data that the Rapsodo 1.0 unit uses, so there will only be metrics from that particular device.
To start here, I decided to use more metrics than I had used in my last article, where I only used pitch velocity and true spin of the pitches. I continued to use those two metrics, but I added horizontal break, vertical break, the difference between the fastball and changeup velocities, and the true spin Bauer units.
Movement should be a big factor in determining how often a hitter gets out and spin rate doesn’t completely paint the picture of the movement profile. Changeups are off-speed pitches, so it would make sense that if a changeup had more of a difference in velocity compared to the fastball, it would be more effective. That’s why I decided to make that a factor in my metrics. Finally, true spin Bauer units is the true spin rate of the changeup divided by the velocity of it.
With these metrics now added I went into R-Studio to create my equations for pxwOBA and pWhiff% using a linear regression. The p’s in both metrics mean predicted. Here are the formulas below:
Changeup pxwOBA = 0.8596 – (Velo. Difference+ * 0.0003189) – (Changeup True Spin+ * 0.0023589) – (Avg. Velo+ * 0.0036778) – (True Bauer Units+ * 0.0011765) – (Z Break+ (Vertical Break) * 0.0013178) + (X Break+ (Horizontal Break) * 0.0008533)
Changeup pWhiff% = -103 + (Velo. Difference+ * 0.1125) – (Changeup True Spin+ * 0.09129) + (Avg. Velo+ * 1.019) + (True Bauer Units+ * 0.1122) + (Z Break+ (Vertical Break) * 0.1707) + (X Break+ (Horizontal Break) * 0.003217)
As I had determined with my last article, the average velocity has the highest effect on both statistics as they are given the highest coefficients. Horizontal break has the lowest value towards the metrics, which makes sense as the vertical drop compared to fastballs would end up creating weak contact and swings and misses.
With pWhiff%, velocity difference ends up being the third most important metric in estimating whiff rates, while vertical break is the second most important variable, but are still very far behind velocity itself.
Let’s take a look at the best and worst pitchers based on pWhiff%:
*pWhiff% average is 29.6, pxwOBA average is .295
**Whiff% average is 29.7, xwOBA average is .295
Something new I added to the chart since last time was the difference between the actual stat and predicted stat, this way, I could see which types of players my model was missing.
Luis Castillo is known for his excellent changeup and he comes in at third on this list, as expected. Josh James, who was just named a starter in the Astros’ rotation, has gone a little bit under the radar, especially in terms of pure stuff, as he is ranked here fifth. His 50.7% whiff percentage was tops for changeups in the bigs last year, he was also third in xwOBA with a .186. My model predicted that he would’ve had a whiff% of 36.5 and an xwOBA of .266, these numbers are pretty far off and I will get to why that is later on.
A guy that I predicted very well to not be very good was Jon Lester. He was about average in 2019 as he had a 4.26 FIP and struck out 21.6% of the hitters he faced. However, I predicted his changeup to have a .309 xwOBA and a 25.3 whiff rate. Lester had about that with a .319 xwOBA and a 23.3 whiff rate. The changeup wasn’t torched a ton though, as it was Lester’s least used pitch, being thrown just 12% of the time. Even though he was average last year, x-stats (or expected statistics) don’t like him and he very well could be on his way out. He had a 22nd percentile xwOBA, a 26th percentile xERA, and a 10th percentile xBA, his changeup, along with himself, are heading in the wrong direction.
Three pitchers that beat out their projections and surprised me in the bottom twenty were Wade Miley, Joe Biagini, and Jacob Waguespack.
Miley, now with the Reds, has had a solid past couple of seasons and his changeup has done the same as well. He ended up with a 37.8 whiff% on it, 27th in the bigs, and had an xwOBA of .231, which was 17th.
Biagini and Waguespack on the other hand, weren’t very good in 2019, with .356 and .343 xwOBAs respectively. Their changeups were the opposite of that, though, as Biagini had a .227 xwOBA and 38.1 whiff% on it and Waguespack had a .228 xwOBA and a 39.5 whiff%. My predictions for them were a lot worse than that, however; Biagini had a pxwOBA of .315 and a pWhiff% of 25.5%, Waguespack had a .300 and 25.2%.
With all of these guys mentioned, I believe the reasons why I was off on them could be due to a multitude of factors. That could be pitch sequencing, pitch location, or tunneling, all were mentioned in the last article. One thing I noticed with the charts was that it seemed as though the top and bottom extremes were not predicted well. This goes back to Josh James, he had a 50.7 whiff%, but my model’s prediction was way off by 14.2%, which was one of the worst predictions in the set.
This made me want to explore more and see if this theory was, in fact, true. So, I made graphs of the stats on the x-axis with the differences of the stat on the y-axis to see if extremes in the stats lead to a worse prediction. Here are the results:
And my suspicion was proven true! As the stat goes from the low to high extremes, the difference between the stats lower as we get closer to the median. This trend is shown the best for xwOBA as the R^2 is .769, that’s pretty good. Also, the points make a true ‘V’ looking shape, while in the Whiff% graph, it isn’t as correlated and doesn’t have as sharp of a ‘V’ as the xwOBA one. This shows that my model doesn’t predict extreme numbers very well, which is to be expected as they are outliers.
Finally, I wanted to see how well my model estimated a pitcher’s performance simply based on the metrics I used in the equation. Again, I made a graph and measured the R^2 value using the summary() function in RStudio. Here’s the code I used to create the graph and find the correlation:
And here’s the graphs for both pWhiff% and pxwOBA:
Similar to the fastball graphs in the last article, pWhiff% correlates better with the changeup than pxwOBA, albeit at low .138. This again means that this metric can’t predict the statistics very well. However, I still feel this is a solid model, especially for off-speed pitches, since I was only able to get the curveball to about a .050 R^2 for both stats. This is an improvement and adding the metrics I used in this model would help the curveball one.
The changeup is a pitch that must be sequenced well to be effective, the pitch characteristics can help decide what you want it paired with. Such as a 3:00 changeup that is tunneled with a 9:00 slider, a changeup and a fastball coming out of the same window (say a 12:30 fastball is best paired with a 12:30 changeup), having almost the same spin rate as the fastball or way less spin for the “tumbling” action, being able to have it 18.6 mph slower than the fastball like Aaron Sanchez does, or a combination of some or all of it. If a pitcher can find a way to find that perfect pairing and locate the pitch well, they will be extremely effective.
To see the last post from the Pitch Grades series (Fastballs and Curveballs), go to this link here.
Follow us on Twitter! @ProspectsWorldW
One thought on “Part 2: Changeups – Predicting Pitcher Performance Based on ‘Stuff’ Alone”