The methodology behind the NFL kicker latent ability project.
Field goal over expected (FGOE) isolates a kicker's true performance from the situations their head coach puts them in. Using historical play-by-play data from 1999 to 2025, our custom PyTorch model calculates the exact probability of an average NFL kicker making a kick based on factors like distance, score differential, and game clock.
If a kicker makes a 55-yarder that the average kicker only makes 40% of the time, they gain +0.60 FGOE.
Standard kicking percentages are flawed because a kicker who goes 10/10 on 30-yard kicks looks better on paper than a kicker who goes 9/10 on 50-yard kicks. To fix this, we track a continuous latent ability index.
This acts similarly to a chess Elo rating. Every time a kicker attempts a field goal, their hidden ability rating updates. Because this is modeled as an amortized state-space model, it also factors in temporal decay—meaning a kicker's perceived talent will slowly regress toward the league average during the offseason or if they miss significant time due to injury.
This dashboard and the underlying Bayesian model were developed by Kevin Kim, a MSCS student at UC San Diego, passionate about sports analytics, machine learning, and building interactive data tools during his time at Emory University under the advising of Dr. Kevin McAlister.
Questions or collaboration? Reach out at kek018@ucsd.edu