The strike zone is typically called out as knees-to-letters for little leaguers. As players age, the zone tends to shrink and adhere to a rectangle specified by 0 +/- 0.83 feet along the X axis and 2.5 +/- 1 feet along the Z axis. Every player has a slightly different “zone” and how catchers call games also impacts what umpires perceive as strikes.

We included the ability to select different batter/pitcher combinations and see how the umpires “zone” changes in the tool. Typically, umpires call more strikes on the part of the plate closest to the batter as opposed to the part furthest away from them. This is easy to see when analyzing our data.

We also wanted to create a way to visualize how different umpires really are in different areas of the zone. To the right, we can see how much more likely Angel Hernandez is to call a strike when compared to Joe West. As you change grid sizes, you can further refine the differences in 6×6 inch, 3×3 inch, and 1.2×1.2-inch grids. This allows for a fun way to explore how different umpires may be at calling different pitch types, against difference batter/pitcher configurations and along their own careers.

What If I Pitch Here?

This section is a user defined section where user can specify all the game conditions along with user defined choice of pitch location.

Under the hood, the tool filters the data as per given game conditions. A Random Forest model is fit to the given data with pitch types, pitch location and other game conditions as features. Finally, the model takes the defined game conditions as input and calculates the probability of calling a strike at the defined location for different pitch types.

As some filters can significantly reduce the data, we added pitch count in the outcome to show number of samples of each pitch type the model has seen. Generally, the more pitch data fed into the model, the better it performs. Because we wanted to allow for custom exploration, we included a count showing the number of pitches per pitch type used in training the model. The model performs very well across many configurations for fastballs, change-ups, curve balls, and sliders as these are the most common pitches. For less frequent pitches such as knuckle balls, splitters, and others, we can see some varying performance as the pitch counts typically drop faster.

In the image below, we have selected Angel Hernandez and Joe West as our umpires to compare. The parameters selected for this image were an X of 0.35 feet and Z of 2.55 feet and right-handed batters. If you want to look at “Any” batter stance, you will see a wider variance in strike prediction given the effects batter handedness has on outcome. Play around to see for yourselves. For this configuration, we can clearly see that both Angel Hernandez and Joe West are highly likely to call a strike for any pitch type at this location. One spot they differ is if a Splitter were thrown. Joe West has a higher predicted value of calling a strike with these conditions. The coolest thing this does (at least in our eyes) is allow us to playback a previous pitch call from any situation and see if (A) the umpire called it as we would expect and (B) if another umpire might have created a different outcome. The baseball fan in us know that we could leverage this not only to see how consistent an umpire’s performance is relative to a model trained on their past calls, but also just to add fire to water cooler talk among friends. Let us know what you think!