R2 Roy NFL is a sophisticated statistical system for predicting outcomes of NFL games. R2 Roy uses advanced statistics summarizing prior team performance, team-specific home-field advantage, team-specific away performance, climatic differences for away teams, field surface, and matchup-specific information to predict team performance. R2 Roy is the perfect example of a projection model adhering to the R2 Sports Metrics Philosophy.
R2 Roy uses machine learning methods that allow us to test hundreds of different model specifications and select the best amongst them by comparing how they perform on certain diagnostic statistics. Because R2 uses matchup-specific information, it would be inappropriate to use Roy to create an index of relative team strength.
In developing R2 Roy, we tested the model on years 2014 through 2021. In these tests, the results reported are out-of-sample "projections." This means that only information that was known prior to kickoff was used to develop these estimates.
R2 Roy Testing Results 2014 through 2021
In six of the eight seasons tested, Roy has "forecast" winners ATS at or over 53% and even outperformed Vegas in Absolute Error in 2019. There are two years that are exceptions: 2016 and 2017. The full story behind Roy's failure in those years is a little hazy, but some significant share of the blame has to do with the departure of the the Rams from STL and the Chargers from SD and their arrival in Los Angeles. Given that R2 Roy uses team-specific home and away effects, in addition to travel distance and time-zone differences, R2 Roy lost fidelity on those teams for a couple of years -- forecasting them at only a 43.5% rate ATS. But by the 2019 and 2020 seasons, Roy had enough information to forecast those teams at a 51.6% rate ATS.
Important for purposes of making selections in betting markets, R2 Roy performs better the more confident it is. As you can see in the table below, when the difference between the spread and Roy is more than three points, Roy picks ATS winners at over 56%. As outlined in more detail in the philosophy post linked above, R2 selects models that tend to be more complicated, tolerating some reduction in the efficiency of its estimates to capture a bit more "signal." Results such as those presented below are consistent with idea that we can use the bias-variance inflation tradeoff to our advantage.
Roy Accuracy by Degree of Divergence with the Line
As noted in the R2 philosophy, the biased-but-efficient Vegas line will have an edge over the less-efficient/less-biased R2 system in picking winner's "straight-up." I do not recommend R2 Roy for making moneyline picks or for use in your pick-em contest (unless its a ATS pickem - in which case you should use R2).
R2 can furnish full testing results via csv upon request (presuming the volume of the requests is reasonable).