Separating Intent from Execution: A Probabilistic Approach to Pitch Location Accuracy

TL;DR

This paper introduces xCTRL, a personalized metric for measuring pitcher control by estimating individual pitch intentions, improving accuracy and predictive power over traditional uniformity-based metrics.

Contribution

The study proposes a novel probabilistic method to infer individual pitcher's intended locations, moving beyond traditional uniformity assumptions for control measurement.

Findings

xCTRL shows strong stability across data

xCTRL has greater predictive power than existing metrics

Personalized inference improves understanding of pitching control

Abstract

Control has long been recognized as a critical component of pitcher performance, reflecting a pitcher's ability to execute pitches in alignment with his intended targets. However, accurately inferring a pitcher's intentions presents a persistent challenge. Traditional metrics typically rely on uniformity assumptions, inferring intent based on the behavior of a ``typical'' pitcher across similar situations. In this study, we propose an alternative, individualized approach to measuring control, one that eschews such assumptions in favor of personalized inference. We estimate a pitcher's intended location on a pitch-by-pitch basis, conditioning on both individual tendencies and specific game contexts. This allows us to assess control by comparing the actual pitch location to the inferred intended target, thereby aligning measurement more closely with the unique strategies of each pitcher. We introduce xCTRL, a novel metric that quantifies control as the distance between a pitch's actual location and its estimated intended location. We find that xCTRL exhibits strong stability and greater predictive power than existing control metrics. By capturing pitcher-specific intent, xCTRL enhances our understanding of control and offers a more intuitive and accurate representation of pitching performance.