abhinav314.github.io
NFL Big Data Bowl
As part of the Kaggle NFL Big Data Bowl 2020, I developed a sequential neural network to predict the number of rushing yards an NFL team would gain/lose on each play.
Data Manipulation
The data provided was part of NFL’s NextGen Stats. The training data was broken down play-wise into 22 rows. Each row corresponding to a single player on the field. The data points included fields such as player orientation, positional coordinates on the field, offensive formation, defensive formation, weather conditions, speed, weight, acceleration, etc. I developed around 30 new features based on our understanding of the game. A few features I created were momentum, yards gained vs distance to goal, bucketed jersey numbers, etc. I used an ensemble feature importance algorithm to select a total of 24 features from a total of around 70.
Modeling
A TensorFlow based sequential neural network was used to generate the final predictions. I optimized our neural network using Bayesian Hyperparameter tuning techniques to determine the optimal number of layers, nodes, dropout, etc.
The neural network was then ensembled with a lightGBM model to obtain balanced results
Results
The testing for this project was unique in that live matches were used to test the predictions. My algorithm was run prior to each match from mid-December to the end of the season. My team finished in the top 30% worldwide (out of 5000 teams) on the Kaggle leaderboard, and we were one of the few all-University teams.