The Problem

What actually drives MLB game attendance, and how should a team allocate a fixed seasonal marketing budget? The drivers are correlated — opening day, a pennant race, market size, and promotions all move attendance at once — and marketing has carry-over and saturation effects that naive regression cannot capture.

What I Built

• Fit a Bayesian Marketing Mix Model in PyMC on 11,972 real MLB home games across 5 seasons and all 30 teams using NUTS (Hamiltonian Monte Carlo) — decomposing log-attendance into a hierarchical team intercept, game-context and performance covariates, and 4 marketing channels passed through geometric adstock and Hill saturation transforms
• Built a fully Bayesian variant that places priors on and learns each channel's adstock decay, Hill α, and Hill γ from data using a reset-aware pytensor.scan — full posteriors over every parameter, uncertainty included
• Six-page Streamlit dashboard for effect decomposition, posterior predictive checks, channel diagnostics, and interactive budget scenario planning with return-on-spend curves

Impact

A production-grade Bayesian causal model on real sports data — demonstrating the same technique used by major advertisers to attribute and optimize spend, applied end-to-end from data to interactive planning tool.