A Rookie's Guide to the NFL
Credit: Cub Studio
Background
In the NFL, being able to accurately predict which play your oppent will call given a set of circumstances would be an invaluable tool, and teams devote many resources towards gaining insights into an opponent’s tendencies on the field. One area of immense interest is assessing a team’s propensity to run or pass the ball in a given situation, which for a defense can inform critical decision-making about strategy, personnel, and positioning on the field – decisions that have a profound impact on the outcome of a game.
In this project, we take a data-driven approach to classifying offensive play types and examine whether tools from machine learning can offer value in characterizing offensive tendencies.
Earlier Work
Much of the work on this topic stems from operations research and economics communities as researchers sought ways to determine optimal play-calling strategies through the frameworks of game theory and decision analysis. This line of inquiry was formally introduced in 1978 by Carter and Machol in their seminal paper outlining optimal fourth-down strategies. These studies were later extended by Boronico and Newbert in 2001 to include goal-line situations, and ultimately all plays by Jordan, et al. in 2008. We sought to further extend on this topic by including an important factor in decision-making in sports, specifically outdoor sports, namely weather.
Weather and Football
Weather can significantly influence how a coach will call a play: Downfield passing and receiving plays are the most affected by the weather conditions. Heavy rains can cause visibility issues for both the quarterback and the receiver, affect grip resulting in players fumbling the ball when attempting to receive a pass in rain. Extreme cold can have a marked effect on the two essential components of any NFL game: the players and the football. Strong winds can alter ball trajectory, resulting in more interceptions. Because running back are the most "waterproof" players, the offense can often continue to run the ball when passing plays are compromised by conditions.
While our work focuses on predicting what play a team will run given a set of conditions, as opposed to what a team ought to do, the foundation established in these early papers in identifying the most critical situational statistics for play-calling decisions proved valuable to us in selecting features for our algorithms.