Predict Baseball Pitch Selection and Anticipation

Business Problem

In baseball, the pitcher vs. hitter matchup is the central conflict of the game. The most foundational aspect of that conflict is the pitch delivery from the pitcher and swing reaction from the hitter. These seemingly small and often unnoticed decisions happen hundreds of times in every game, and each one has the potential to change the contest’s outcome, so finding even the smallest-of-edges on each of these hundreds of pitches per game is critical.

Before each pitch, the pitcher must make a critical and complex decision of which pitch to throw to maximize his chances for success. This includes both the pitch type (one of a handful in his repertoire) and the precise location across the 2 axes of the strike zone. There are many factors that go into making this selection including his own strengths, the hitter’s strengths, the hitter’s familiarity with the pitcher, the game and at-bat situations, and the most recent pitches seen by the hitter.

Conversely, the hitter must be prepared to react to the pitch thrown. By anticipating the pitch correctly, the hitter can be in a better position to react well and increase his odds of success with a strongly hit ball.

The winner in this pitcher vs. hitter conflict is often determined by talent and chance, but strategy in pitch selection/anticipation can shift the edge from one side to the other; changing the outcomes of pitches, at-bats, games, and seasons.

Intelligent Solution

Note: in this example, we will focus on the decision the defense must make in pitch selection for effectiveness, but every example can be reversed to the offense for batter anticipation and reaction for the pitch coming his way.

The pitcher has hundreds of options with each pitch when they select their target location and pitch type. For most pitches, this decision is driven by the experience and on-the-fly thinking of one player: the catcher, who typically calls pitches and signals them to the pitcher. After running through three basic questions, he makes a fast decision driven mostly by intuition:

1. What does this pitcher throw well?
2. What does this hitter not hit well?
3. Based on the ball-strike count, what is the margin of error for missing the strike zone and throwing a ball?

After a quick consideration of these questions, the catcher must make a strategic decision in just a few seconds that could be the difference between a strikeout with the right pitch or a game-changing home run off the wrong pitch. For such a critical decision, and with such a wealth of information to inform it; this approach can be improved upon with ML-driven insights to lead to better pitch selection.

Using the wealth of data available on historical pitches and their resulting outcomes, ML-driven models can help the pitcher develop optimized strategies for getting hitters out by playing to the pitchers strengths and the hitters weaknesses; keeping the hitter off-balance and guessing; minimizing risk while considering human error in pitch execution and umpire strike zone calls; and minimizing the burden on the pitcher by bringing the at-bat to a resolution in as few pitches as possible. These models could also instantly account for hundreds of variables and thousands of at-bats in the pitcher’s and hitter’s pitch history to build pitch recommendations for every possible situation in the at-bat; developing strategies that are at a level of complexity far beyond anything a human can consider, at a higher degree of quality, and in a fraction of the time.

Implementing these strategies can also be pursued in a variety of ways. Given players can’t take ML-models onto the field of play, these strategies must be generated before the game even starts, creating an informational bottleneck. Even if the pitching team generated strategies for every possible hitter-pitcher matchup and situation, there would be no way for a player to absorb this information and execute it as a script. Therefore, these models must instead be used for insights, heuristics, and building blocks of pitching strategy. For example, the models can identify the best 1-2 pitch combinations for a starting pitcher against each hitter in the opposing team’s batting order, then print those on a wrist band for the catcher to use on-field. Or, pitchers can learn the sequences of pitches that put them at the greatest advantage/disadvantage, either through ball-strike counts or priming the hitter well for the next pitch coming.

Bringing AI-driven intelligence to this pitch selection/anticipation problem will give one side an advantage in this blink-of-an-eye conflict, and potentially swing the balance of power in the contest.