Swinging, Fast and Slow: Interpreting variation in baseball swing tracking metrics

TL;DR

This paper develops a Bayesian framework to interpret baseball swing metrics, accounting for timing and recognition confounds, and estimates how bat speed and swing length causally affect contact and power outcomes.

Contribution

It introduces a novel Bayesian hierarchical model and instrumental variables approach to disentangle and estimate causal effects of swing metrics on batting performance.

Findings

Reducing bat speed can lower strikeout rates.

Tradeoff between contact and power due to bat speed is approximately balanced.

Framework accounts for timing and recognition confounds in swing data.

Abstract

In 2024, Major League Baseball released new bat tracking data, reporting swing-by-swing bat speed and swing length measured at the point of contact. While exciting, the data present challenges for their interpretation. The timing of the batter's swing relative to the pitch determines the point of measurement relative to the full swing path. The relationship between swing metrics and swing outcomes is confounded by the batter's pitch recognition. We introduce a framework for interpreting bat tracking data in which we first estimate the batter's intention conditional on ball-strike count and pitch location using a Bayesian hierarchical skew-normal model with random intercept and random slopes for batter. This yields batter-specific effects of count on swing metrics, which we leverage via instrumental variables regression to estimate causal effects of bat speed and swing length on contact and power outcomes. Finally, we valuate the tradeoff between contact and power due to bat speed by modeling a plate appearance as a Markov chain. We conclude that batters can reduce their strikeout rate by reducing bat speed as strikes increase, but the tradeoff in reduced power approximately counteracts the benefit to the average batter.