UCLA

The purpose of this study is to evaluate the performance of offensive linemen in the National Football League by using play-by-play tracking data to predict a player's annual salary. We ran an algorithm on a frame-by-frame data set of plays during the first 8 weeks of the 2021 NFL season that created new variables based on existing data for every offensive lineman on every play. With this information, we tried to use both an unweighted stepwise regression model and a weighted version that emphasized certain play types, but both did not have appropriate residual distributions. As a result, we used box-cox transformations to log transform the response variable, salary, and this change led to evenly distributed residuals, so we tried this model on the test set and got an RMSE of 4.7. We improved upon this number with a series of random forest models, where we iterated over 5 different numbers of trees, ranging from 50 to 150, and discovered that the lowest RMSE of 4.36 was achieved by the model with 75 trees.