Read on Twitter
A little over a year ago, I shared a proof of concept on how weighting EPA could improve predictive power
This summer I made fixes & updates that help the framework reliably outperform EPA, defense adj EPA, and DVOA in predicting future point margin https://www.robbygreer.com/blog/weighted-epa-methodology-amp-performance
This summer I made fixes & updates that help the framework reliably outperform EPA, defense adj EPA, and DVOA in predicting future point margin https://www.robbygreer.com/blog/weighted-epa-methodology-amp-performance
The first iteration had some serious issues with overfitting and didn't do much in the way of testing and training
The new model uses @nflfastR's expanded PBP data (1999-2019) and multiple approaches to validation to minimize overfitting risk
The new model uses @nflfastR's expanded PBP data (1999-2019) and multiple approaches to validation to minimize overfitting risk
The new model explored a much larger group of features to see what additional dynamics could make EPA more predictive
Offensive and defensive weights were determined separately, revealing insights wrt to stability on each side of the ball
For instance, sack fumbles are complete noise for a defense, but not as much for the offense, suggesting pressures and sack fumbles are an offensive stat
For instance, sack fumbles are complete noise for a defense, but not as much for the offense, suggesting pressures and sack fumbles are an offensive stat
The new model outperformed point margin (i.e. net EPA) and DVOA in predicting future point margin, which has proven to be one of the best measures of team performance
As always, all the code is available on github: https://github.com/greerreNFL/wepa_v2
