Finding Skill Changes in Major League Baseball Player Development: A Bayesian Approach with LSTM Networks and Hidden Markov Models

Abstract

Major League Baseball (MLB) organizations invest heavily in player evaluation and development, often relying on end-of-season statistics and traditional regression-to-the-mean models to assess talent. However, regression-to-the-mean assumes fixed skill levels and fails to account for the dynamic nature of player performance over a season. My thesis presents a novel approach to evaluating and forecasting MLB player development using Bayesian inference and changepoint detection models, including CUSUM, Bayesian Online Changepoint Detection (BOCPD), Hidden Markov Models (HMMs), and Long Short-Term Memory (LSTM) networks.

I use a Bayesian framework to iteratively update beliefs about a player's true skill level across various performance metrics such as batting average, slugging percentage, and weighted on-base-average. This approach incorporates uncertainty and provides richer comparisons between players than single-point estimates. I tested my models on both synthetic and real MLB play-by-play data, with synthetic data used to benchmark changepoint detection accuracy across controlled scenarios.

My analysis shows that while Bayesian inference effectively captures player skill trends and variation, the changepoint detection models struggle to identify subtle but significant shifts in skill due to the high noise inherent in binary baseball outcomes. The LSTM model initially showed promise but ultimately failed to outperform simpler methods in accuracy or consistency. Nevertheless, this work provides a foundation for future efforts to disentangle random fluctuations from true skill changes in athlete performance.

By offering a probabilistic framework for evaluating player development, this thesis contributes a more nuanced perspective to player scouting and performance forecasting, with implications for team decision-making, player strategy, and contract valuation.