Neural Sabermetrics with World Model: Play-by-play Predictive Modeling with Large Language Model
A new AI can forecast the next pitch in a baseball game with surprising precision.
Deep Dive
Researchers have trained a large language model on over ten years of MLB data, comprising seven million pitch sequences. This 'world model' predicts the complex flow of a baseball game pitch-by-pitch. It correctly forecasts the next pitch type 64% of the time and predicts a batter's swing decision 78% of the time, outperforming previous neural network approaches. This demonstrates LLMs can be effective simulators for dynamic sports environments.
Why It Matters
This shows AI's potential to model complex real-world dynamics, with applications beyond sports into logistics and strategy.