Inside the Huddle: Reproducing ESPN Player Projections Using Public Data

Abstract

This thesis asks how well a purely statistical model, trained only on public data, can project weekly NFL fantasy football performance for skill-position players and how such a model compares with ESPN’s commercial projections. I construct three per-position XGBoost regression models for running backs, wide receivers, and tight ends using a public-data panel of 32,900 player-weeks from the 2022–2025 seasons. The system is extended in two ways: first, through a tier-adaptive ensemble that blends the statistical model with ESPN’s own projections; second, through quantile regression heads that produce 80% prediction intervals. On the 2025 NFL regular season, the standalone per-position model achieves a mean absolute error of 2.909 fantasy points against ESPN’s 3.074, a 5.4% improvement, and beats ESPN in 17 of 18 individual weeks. The tier-adaptive ensemble improves slightly further to 2.907 MAE and beats ESPN in all 18 weeks. The largest relative gain is at tight end (11.0%), while the most practically consequential gains occur in the starter and flex tiers where lineup decisions matter most. Feature-importance analysis shows that Vegas-derived variables dominate the models at every position, with betting-market features accounting for roughly 40–52% of total feature gain depending on position. A played-only ablation further shows that much of the model’s advantage comes from implicitly learning the will-they-play component of the task, not merely from better conditional projections for active players. The thesis contributes a reproducible public-data benchmark for weekly fantasy projection and shows that a disciplined statistical pipeline can match or exceed a major commercial baseline on aggregate error while also producing useful measures of uncertainty.