MAYFIELD: Machine Learning Algorithm for Yearly Forecasting Indicators and Estimation of Long-Run Player Development

• Activity Number: 217 - Contributed Poster Presentations: Section on Statistical Computing & Statistics in Sports
• Type: Contributed
• Date/Time: Tuesday, August 4, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Section on Statistics in Sports
• Author(s): Alexander Williams* and Benjamin Clarke and Sethward Brugler
• Companies: The Ohio State University and The Ohio State University and The Ohio State University
• Keywords: Machine learning; KNN regression; Locality-sensitive hashing; National Football League; Time-series forecasting

Abstract

Accurate statistical prediction of American football player development and performance is an important issue in the sports industry. We propose and implement a novel, fast, approximate k-nearest neighbor regression model utilizing locality-sensitive hashing in highly dimensional spaces for prediction of yearly National Football League player statistics. MAYFIELD accepts quantitative and qualitative input data, and can be calibrated according to a variety of parameters. Concurrently, we propose several new computational metrics for empirical player comparison and evaluation in American football, including a weighted inverse-distance similarity score, stadium and league factors, and NCAA-NFL statistical translations. We utilize a training set of comprehensive NFL statistics from 1970-2019, across all player positions and conduct validation on the model with the subset of 2010-19 NFL statistics. Preliminary results indicate the model to significantly improve on current, publicly available predictive methods. Future training with advanced statistical datasets and integration with scouting-based methods could improve MAYFIELD's accuracy even further.