The Virtue of Transparency in Sports Analytics: An Application to NFL Quarterbacks
- CSA Authors
- 3 days ago
- 2 min read
Megan Czasonis is a Founding Partner of Cambridge Sports Analytics in Cambridge, MA.
Miles Kee is a data scientist at Cambridge Sports Analytics in Cambridge, MA.
Mark Kritzman is a Founding Partner of Cambridge Sports Analytics in Cambridge, MA, and a senior lecturer at the MIT Sloan School of Management in Cambridge, MA.
David Turkington is a Founding Partner of Cambridge Sports Analytics in Cambridge, MA.
Date Written: February 17, 2026
Abstract
A model-based approach to prediction, such as a regression model or a machine learning model, estimates parameters from a sample of observations and applies those parameters uniformly across all prediction tasks. These approaches fail to consider the unique circumstances of individual prediction tasks, and they conceal the influence of observations and predictive variables on individual predictions. Additionally, model-based approaches to prediction offer no guidance about the reliability of individual predictions. We describe a novel model-free prediction system called relevance-based prediction (RBP) that overcomes these limitations, and we show how it enables prediction-specific transparency that is beyond the reach of model-based approaches. RBP reveals the specific reliability of each prediction before the prediction is made, the importance of each prior player to each prediction, the contribution of each predictive variable to each prediction's value, and the contribution of each predictive variable to each prediction's reliability. We illustrate this new prediction system by applying it to predict passing yards per game for NFL quarterbacks. The prediction-specific information given by RBP stands in contrast to R-squared, t-statistics, and beta which only give information about average effects, as we illustrate with specific player examples.
This research was authored by Megan Czasonis, Miles Kee, Mark Kritzman, and David Turkington.
Published as an MIT Sloan Research Paper (No. 7358-26) and available via SSRN
Comments