In the last year, I've come to realize that potential readers of this blog would much prefer to know what FredRank is, rather than trying to figure it out by reading through older posts of mine.
The goal of this post is to explain why I've created FredRank, how it is calculated, and which factors go into its various iterations for different leagues and divisions.
What is FredRank? Why was it created?
If you've ever taken a look at more than one football computer rating, you've probably noticed that the scales are different. For the National Football League, an Elo rating will have numbers in the thousands; Jeff Sagarin's ratings average out to 20; Ken Massey's ratings and Football Outsiders' DVOA average out to 0. And perhaps just as importantly, different systems don't always rank teams in the same order and the gaps between those teams can scale differently.
FredRank was created to put different rating systems on a single scale, in order to create an average rating for easy comparison of teams.
How is FredRank calculated?
All ratings are put on a scale of 0 to 1. While some measurements, such as win percentage, are left alone, because they naturally fall in that range, most ratings need to be scaled down to that range.
Take, for example, Sagarin ratings. As of today, following week 1 of the 2019 NFL season:
The goal of this post is to explain why I've created FredRank, how it is calculated, and which factors go into its various iterations for different leagues and divisions.
What is FredRank? Why was it created?
If you've ever taken a look at more than one football computer rating, you've probably noticed that the scales are different. For the National Football League, an Elo rating will have numbers in the thousands; Jeff Sagarin's ratings average out to 20; Ken Massey's ratings and Football Outsiders' DVOA average out to 0. And perhaps just as importantly, different systems don't always rank teams in the same order and the gaps between those teams can scale differently.
FredRank was created to put different rating systems on a single scale, in order to create an average rating for easy comparison of teams.
How is FredRank calculated?
All ratings are put on a scale of 0 to 1. While some measurements, such as win percentage, are left alone, because they naturally fall in that range, most ratings need to be scaled down to that range.
Take, for example, Sagarin ratings. As of today, following week 1 of the 2019 NFL season:
- The New England Patriots, the highest team in his ratings, have a rating of 27.19.
- The Miami Dolphins, the lowest team, are rated at 8.92.
- The Seattle Seahawks are rated at 20.15. This is the closest any team comes to the average rating of 20.00.
No matter what the highest or lowest rating nominally is, the highest rating will be converted to be equal to 1, and the lowest will be converted to 0.
Ratings in the middle will be assigned a value based on which rating they are closer to, and how close they are to that rating.
Another way to think of it: each rating becomes a percentage based on the difference in the highest and lowest ratings.
The current difference in the ratings for the Patriots and Dolphins is 18.27.
- Patriots = (27.19 - 8.92) / 18.27 = 1.0000
- Dolphins = (8.92 - 8.92) / 18.27 = 0.0000
- Seahawks = (20.15 - 8.92) / 18.27 = 0.6147
So even though the Seahawks are the closest to the nominal average of Sagarin's ratings, the difference in their rating from the Dolphins' rating is about 61.5% percent of the difference between that of the Patriots and Dolphins.
What factors go into FredRank for different leagues?
As of 2019, I publish three versions of FredRank: one for the NFL, one for the Football Bowl Subdivision (FBS) of NCAA Division I, and one for the Football Championship Subdivision (FCS) of NCAA Division I.
While I publish ratings for the Canadian Football League (CFL) as well, these ratings are not properly referred to as FredRank. There are not enough computer ratings out there to create a composite for the CFL. More properly, my ratings are an application of Jeff Sagarin's method of rating teams, as applied to the CFL.
Factors in NFL FredRank:
- My Sagarin-style ratings, which are not weighted toward recent results. All games have equal weight in my ratings.
- Jeff Sagarin's actual ratings, which I believe are weighted. Recent games have more weight in these than early games.
- The AP Pro32, a human poll. Instead of being indexed, teams are scored by how many points they received in the latest poll, as a percentage of the maximum points possible, which is 32 * 12 = 384.
- ESPN's Football Power Index (FPI)
- Football Outsiders' DVOA.
- Adjusted DVOA, also from Football Outsiders. The adjustment depends on what point of the season we're in: early adjustments factor in preseason expectations, while later adjustments weight recent games more heavily.
- Consistency, a proprietary factor of mine which measures how a team's actual game results deviate from expected results, based on my Sagarin-style ratings.
- Ken Massey's computer ratings
- Elo ratings from FiveThirtyEight
- Strength of schedule, as measured by Ken Massey. This factor includes only games that have already been played.
- Strength of future schedule, also from Massey. This includes both past and future games. I use Massey for both strength of schedule factors specifically because he calculates both, so the two can be easily compared.
- Steve Pugh's "Compughter" ratings
- Games above or below expected record. This penalizes teams for losing games which, according to my Sagarin-style ratings, they "should have" won, while rewarding them for winning games they "should have" lost.
- Current win percentage
- Expected final win percentage. Does not include any postseason games until the postseason is scheduled/played.
- Expected final conference win percentage
- Expected final division win percentage.
Factors in Division I football FredRank (some factors FBS only):
- Sagarin ratings. I don't generate my own Sagarin-style ratings for Division I football because there's too many teams for Excel to handle that.
- Football Outsiders' F/+ ratings (FBS only)
- Massey ratings
- Pugh ratings
- FPI (FBS only)
- Future strength of schedule, from Massey. Unlike NFL FR, I do not use current strength of schedule for college.
- Comparative Performance Index ratings (CPI)
- Elo ratings from Warren Nolan
- Pythagorean expectation, also from Nolan
- Congrove computer ratings. I use Power rather than Rank, as the latter appears to be affected by win-loss record. (FBS only)
- Anderson & Hester ratings (FBS only)
- Billingsley ratings (FBS only)
- Wolfe ratings
- Games above or below expected record, as explained above
- Consistency, as explained above
- Expected final win percentage. Unlike NFL FR, this will factor in any predicted or actual postseason games, every week.
- Expected final conference win percentage. For independent teams, overall win percentage is instead used twice.
I create two versions of FredRank for college football. Traditional FredRank also incorporates human polls (3 for FBS, 2 for FCS), while the NoPoll index omits them. I find that it's helpful to illustrate when a team may be overrated or underrated, but human polls also help to account for a team's qualities which are not or possibly cannot be quantified by a computer.
The reason why I don't calculate a NoPoll index for the NFL ties into the section below.
What is FredRank used for? How is it used?
My primary purpose for calculating FredRank is to make predictions regarding the postseason for any league.
In the NFL, the postseason is determined by overall win-loss record. Teams qualify for the postseason either by winning their division or by being one of the two best non-division-winners in their conference. While some metrics such as strength of schedule can be used to break ties, it largely never comes to that. This can often result in teams who would have a better FredRank being left out of the postseason in favor of worse teams.
Thus, NFL FredRank is better used for predicting games outcomes.
However, the postseason in college football is largely invitational. While there are conditions which must be met for a team to be eligible, the games and matchups are frequently determined by measurements which FredRank takes into account.
As of 2019, the FBS bowl system has 40 games and requires 78 teams to fill every spot, as one of those games is played between the winners of two of those bowls. The filling of each spot is flexible enough to be, in many cases, hard to predict.
The FCS can be even harder to project. A committee will choose, as of 2019, 24 teams to play a single-elimination tournament. While the FBS bowls will often exhaust the list of eligible teams, or at least choose all but maybe five, at most, of the eligible teams, choosing 24 teams, when twice as many teams have a winning record, will frequently come down to analyzing who is better from a perspective well-suited to a method such as FredRank.
Questions? Comments? Complaints?
Feel free to leave any of those in the comments below! I can also be reached on the Fans of FredRank Discord server.
Comments
Post a Comment