Screen Machine:
Pre-Snap Play Prediction in the NFL

External Advisor: Dr. Karim Kassam, Teamworks

Krish Bothra, Ethan Larimore, Jenna Lau

Motivating Example: LA vs. CAR

Data Overview

Sources: NFL Big Data Bowl 2025, nflfastR & nflreadr packages

Model Observations: 2022 Passing Plays from Weeks 1-9

Data Subsets:

• Game
• Play by Play
• Player play
• Tracking

No Significant Difference Between Share of Screen Targets by Position & Direction

Modeling Approach

Methodology:

• Compare LASSO and XGBoost models
• 3-Fold Cross Validation
• Evaluation Metrics: Baseline, Kappa, ROC AUC

Model 1, Screen Probability:

Feature Set Overview:

• Team-Level Historic Tendencies
• Play Context
• Spatial Positioning

Conditional Probabilities:

Model 2, Lateral Direction:

Model 3, Receiver Group:

Case Study Application

Model Predictability by Team

Lateral Direction:

Target Receiver Group:

Discussion

XGBoost/LASSO hierarchical modeling framework to predict screen pass tendencies for coach-usable insights

Limitation:

• Sparse number of single-season screen plays per team

Future Work:

• Sequential models for continuous pre-snap probability

• Find individual teams’ pre-snap “tells”

Acknowledgements

Special thanks to Karim Kassam, Erin Franke, Sara Colando, Dr. Ron Yurko, Quang Nguyen, and the CMSACamp TAs for their guidance throughout our study.

Appendix

Lateral Distance Model VIP:

Appendix Continued

Receiver Group Model VIP:

Screen Model VIP: