By Elliot Chin and Nicholas Lopez
Power rankings and game predictions are pillars of sports media and culture. Pundits and fans construct their own lists based on ranging and non standardized criteria to decide who are the best teams. After all, competition and comparison are an essential part of sports. However, the small sample size of an individual game often doesn’t settle debates: even the playoffs don’t always provide consensus as to which teams are the best in the league. Just last week, both No. 1 seeds, the Green Bay Packers and Tennessee Titans, fell to early upsets!
If playoff performance is subject to variance, what is the most accurate way to rank NFL teams? Qualitative power rankings are common, but easily subject to bias and inconsistent criteria. Ranking teams by win-loss record is statistically sound, but ignores strength of schedule. Furthermore, are wins even the most important NFL stat? Statistics like point differential, turnover differential, and total yards are important as well. More advanced stats like Yards/Play or Time of Possession might even be more indicative of team success. Ultimately, there are countless stats that are useful in quantifying team performance; which of these measures actually matter? In this article, we set out to create a ranking system that lets you answer these questions for yourself.
Methodology
Massey matrices are a linear-algebra based method used to rank college football teams. Their use in the sport stems from an ability to account for strength of schedule using matrix algebra (an in-depth mathematical explanation written by the authors is available here). Massey matrices also use point differential to differentiate close games from blowouts. For example, the Cincinnati Bengals’ three point victory over the Kansas City Chiefs (34-31 Week 17) is much different than their three point victory over the Jacksonville Jaguars’ (24-21 Week 4). In a nutshell, Massey matrices are similar to record-based ranking systems except that they account for the quality of wins and losses.
We’ve enhanced the Massey matrix method by replacing point differentials with other football differential stats. For example, teams can be ranked, still using strength of schedule, by interception differential, sack differential, penalty yard differential, or even pure win-loss record. We standardize each of these metrics based on the average standard deviation of the category. A recency metric is used to weight recent game results more than results from earlier in the season. Again, the Bengals win against the Chiefs is more indicative of their current level than their narrow win earlier in the year against the Jaguars. Finally, each of these factors can be combined using custom weights. For example, win-loss record, point differential, turnover differential, and first down differential could each be weighted equally. Most of the statistical differentials we use in the model for the 2022 season can be seen below.
Try It Yourself
Use the model below to construct your own rankings based on what statistics you think are most important. The data is up to date as of the divisional round of the 2022 NFL playoffs. When only wins are used for the model, the Kansas City Chiefs are the best team in the league. On the other hand, if point differentials are used for the model, then the Buffalo Bills come out ahead. If Yards/Play differentials are instead prioritized, then the San Francisco 49ers take the lead. Conversely, the Green Bay Packers are ranked highest when turnover differential is the singular input.
These teams are clearly solid but not necessarily the best; two of them are eliminated from playoffs as of the writing of this article. By incorporating recency—which places higher weight on recent games—and stats like sack differential, penalty differential, and first down differential, any number of analyses can be performed.
Toggle around with the different weights and metrics to create your own custom power rankings based on what you think makes the best team. Comment down below who you think is best!
The underlying code for this model can be found here.
