Beyond Marcel: Adding Bayesian Regression to NPB Baseball Predictions — A 15-Step Journey

Introduction In a previous project, I built an NPB (Nippon Professional Baseball) player projection system using the Marcel method — a simple "3-year weighted average + regression to the mean" approach. GitHub : npb-prediction Marcel performed surprisingly well: it beat ML models (LightGBM/XGBoost) for pitcher ERA prediction. But it had clear limitations: Limitation Marcel's approach New foreign players Use league average (can't use previous league stats) Point estimates only No uncertainty quantification Uniform age adjustment +0.3%/year for all players Ignores skill metrics Can't leverage K% or BB% To address these, I started a new project: Bayesian regression on top of Marcel . GitHub : npb-bayes-projection Here's what happened over 15 steps. Step 1: Foreign Player Conversion Factors NPB teams sign foreign players every year. If we can convert their previous-league stats to NPB scale, we should beat league-average predictions. What I did Identified 365 foreign players in NPB (2015-2