NFL Pre-Snap Play-Calling Prediction
Predicts run vs. pass using only what a defender can see before the snap.
- out-of-sample accuracy, 2018–2025
- 76.5%
- above majority-class baseline
- +15.5 pts
- plays evaluated out of sample
- 262,953
- offensive snaps curated, 2016–2025
- 343k+
- Weekly expanding-window backtest: retrains on the prior season plus completed weeks, then holds each target week fully out of sample.
- Chronology and perturbation audits check for target leakage.
- Chose a regularized depth-7 global XGBoost model and rejected a more complex expert-blend ensemble that gained only 172 net correct calls.
More detail
- 76.4% accuracy on the 2025 season alone.
- The ensemble’s gain worked out to about 1 extra correct call per 1,529 plays. Not enough to justify the complexity.
- Defensive alignment is deliberately excluded: box count reflects the defense reacting to its own read.
- Logistic regression and a two-hidden-layer neural network written from scratch in NumPy: backpropagation, mini-batch Adam, early stopping.