Author: Philip Ramirez Date: January 2017

Most work done in sports analytics is related to game scores, outcomes, or player performance forecasting, but modeling game time decisions is a complex - yet potentially fruitful - task. Football is a stop and start game of difficult decisions based on specific game circumstances; It is especially advantageous for coaches and offensive coordinators to be as unpredictable as possible. From the perspective of opposing teams, coaches, and defensive coordinators having a reliable, data driven prediction of the next play would be invaluable. Furthermore, knowing which salient features determine game time decisions can ultimately influence defensive preparation as well as in-game strategy.

There were three major steps involved in acquiring my final dataset.

• The bulk of my data was sourced by NFLSavant.com. It includes full play by play - including playoffs and Super Bowls - of season years 2013 through 1016.
• I then supplemented this data with scraped features from NFLWeather.com and Pro-Football-Reference.com. Some of these features included temperature, humidity, visibility, coaches, defensive alignments, offensive schemes ...
• The test set (data from season year 2017) was scraped from Pro-Football-Reference.com

The cleaned training data set contains a total of around 120,000 offensive plays ranging from season years 2013 - 2016. This data was split into a training and validation sets in order to tune hyper parameters. Finally, the plays run in 2017 were used as the final test set for model evaluation. After feature engineering and the elimination of redundant features, the final feature set includes 31 columns. Categorical variables were then dummied.

The model features fall into 7 major categories: The target is pass. Every pass play attempt is given a value of 1. Every rush attempt is given a value of 0. Three classifiers were attempted and evaluated. Accuracies are supplied with each model:

I've evaluated the Gradient Boosting Classifier's performance against a baseline naive model that always selects pass (the majority class).

One of the advantages of using tree based models are identifiable feature importances. After model evaluation, the fundamentals of football proved to be the most important features.

70% of plays run from the shotgun formation are pass plays. Plays run from under center tend to be more evenly distributed.

The down and the amount of yards until the next first down are both positively correlated to pass attempts. In the last 5 years, plays run on third down -with at least 3 yards to go- are 80% pass attempts.

One of the most important features is how much time is left in each half of a given game. Quarters 1 and 3 tend to be conservative (teams run the ball more), whereas quarters 2 and 4 are pass heavy. The most interesting part of this graphic is the observable dip of pass plays right after the start of overtime. The very conservative play calling in this time period indicates that teams are trying not to lose!

Python

• Pandas and Numpy for data processing and feature engineering
• SciKit Learn Logistic Regression, Random Forest Classifier, Gradient Boosting Classifier, and Grid Search Database
• MongoDB for data storage/querying AWS
• iPython clusters of up to 10 EC2 instances to run grid searches of evaluated models

• Player packages for each play
• Baseball - another stop and start sports with complex in game decisions