Arm angle and Pitcher type

Danny Young

New York Mets pitcher Danny Young’s Baseball Savant movement profile is as follows.

He throws a sinker and a sweeper with significant horizontal movement from a low arm angle. He’s a pitcher with an interesting pitch mix. It’s clear that he confuses hitters more with differences in horizontal movement between pitches rather than vertical movement.

Horizontal vs. Vertical Movement Differences

This gave me an idea. I looked into which pitchers have the greatest differences in horizontal movement and which in vertical movement when pairing their pitches two by two. As previously seen, Danny Young falls into the former category.

The image above shows the movement profile of Cubs pitcher Caleb Thielbar in 2020. Unlike Danny Young mentioned earlier, he features a four-seam fastball and curveball combination from a high arm angle. He falls into the latter category, characterized by a large difference in vertical movement.

Based on data from 2020 to 2025 for pitchers who faced at least 25 batters, I calculated the horizontal and vertical movement differences for each pitch combination per pitcher. To isolate movement solely induced by the pitcher, I used IVB (Induced Vertical Break) for vertical movement.

1. For each pitcher, I created all possible pitch combinations using two pitches at a time. For example, a pitcher who throws a fastball, curveball, slider, and changeup would have the following six combinations:

[Fastball, Curveball], [Fastball, Slider], [Fastball, Changeup], [Curveball, Slider], [Curveball, Changeup], [Slider, Changeup].

For each combination, I calculated the differences in horizontal and vertical movement, using the absolute values for comparison. Then, I identified the pitchers with the largest absolute movement differences in any pitch combination.

Below are the pitchers with the largest horizontal movement differences.

Here are the pitchers with the largest vertical movement differences.

x_type and z_type

We compared the maximum possible horizontal movement difference and vertical movement difference among each pitcher’s pitch combinations. If the horizontal movement difference was greater, the pitcher was classified as x_type; if the vertical movement difference was greater, the pitcher was classified as z_type.

As mentioned earlier, Danny Young falls under x_type, while Caleb Thielbar is categorized as z_type.

Arm Angle and Pitcher Type

After dividing pitchers into two types, we added the arm angle as a variable. Using this, we built a model to predict a pitcher’s type based on their arm angle. Since there are two types of pitchers, we used a logistic regression model. x_typewas assigned as class 0, and z_type as class 1.

Accuracy is the proportion of correct predictions out of the total data. An accuracy of 0.755 means the model correctly predicted approximately 75.5% of the 1,174 total samples.3

A precision of 0.858 for class 0 means that about 85.8% of the predictions made as class 0 were actually class 0.

A precision of 0.639 for class 1 means that about 63.9% of the predictions made as class 1 were actually class 1.

A recall of 0.727 for class 0 means that the model correctly identified 72.7% of the actual class 0 samples. In other words, it missed about 27.3% of the true class 0 samples.

A recall of 0.801 for class 1 means that the model correctly identified 80.1% of the actual class 1 samples. That is, about 19.9% of the true class 1 samples were missed.

F1-score is the harmonic mean of precision and recall, both of which range from 0 to 1. The model’s F1-score is 0.787 for class 0 and 0.711 for class 1.

The arm angle threshold that separates class 0 and class 1 is approximately 40 degrees. In other words, pitchers with an arm angle above 40 degrees are classified as z_type, and those with an arm angle below or equal to 40 degrees are classified as x_type.

Conclusion

It is true that a pitcher’s arm angle influences their type, but it is not the whole story. The classification method used in this analysis was simple, and it considered only the arm angle as a variable, which is a limitation. It would be interesting to perform a classification using multiple variables by selecting a few appropriate ones.