NFL Game Prediction and Calibration System

A reusable NFL win-probability project built around a 2026 Philadelphia Eagles forecast. The system emphasizes leak-free feature engineering, season-forward validation, model comparison, and a local Streamlit dashboard for portfolio review.

Problem

A single predicted winner is not enough for sports forecasting. The useful object is a calibrated probability that can be compared across games, seasons, and models. This project frames NFL game prediction as a probability-quality problem rather than a binary pick sheet.

Method

The pipeline builds pregame features only: Elo difference, rolling win rate, rolling point margin, rest differential, home field, market spread when available, and weather fields when present. Models are evaluated with season-forward validation from 2018 through 2025 using Brier score, log loss, accuracy, and calibration buckets.

• Elo baseline for transparent probability estimation.
• Logistic regression as the primary calibrated model.
• Histogram gradient boosting as the nonlinear comparison model.

Result

Logistic regression selected as the primary model by average Brier score. The dashboard surfaces the 2026 Eagles schedule, win probabilities, model comparison metrics, calibration diagnostics, and caveats around preseason uncertainty.