Temporal Context in Deep Nets for Pitch Modeling: A Study of Lag Length and Architecture

MIT 6.7960 (Deep Learning) · 2024

This post uses an MLB pitch-modeling dataset (3.8M+ pitches, 2021–2025) to examine how inductive biases in LSTM versus MLP architectures affect autoregressive prediction of delta run expectancy. We find that LSTMs filter noisy contextual signals more effectively than MLPs, suggesting that simply adding more lag features to an MLP is not an optimal approach for modeling pitch sequences.