Research summary

Our team combined public NFL data on plays, rosters, injuries, draft picks, schedules, and past results to study what relates to a team's total wins across a season. We analyzed 320 team-season observations from 2015 through 2024 and compared seven machine learning models using a chronological train/test split.

The best model, a tuned LSTM, reached an R² of 0.907 and a mean absolute error of 0.887 wins on the held-out seasons. A GRU and Ridge Regression performed nearly as well, underscoring the value of careful feature design.

Several inputs were measured during the season being modeled, so this is a season-level success analysis rather than a forecast made before a season starts. I contributed to data preprocessing, coding and analysis, model testing, results review, and paper development.