Composite SOS

October 27, 2019

A discussion this week about comparitive strengths of schedule led me to ponder how to formulate the question from composite ranks. The usual SOS measurement for a particular computer rating is just the average value of opponents' ratings. Average ranks in that case are not really indicative for a specific rating, since rating values are more spread out on the tails of the curve than the difference in ordinal ranks.

We have only the ordinal ranks for the 100+ computers in Massey Ratings College Football Ranking Composite as of Sat Oct 26 06:58:59. A proxy for a composite value can be found in the consensus rank by Borda Count, since we have the values by which the ranks are defined. The distribution is at least similar to what we'd expect from an advanced rating's values. With μ = 664.135 and σ = 3704.711 we see that ℯ-((x-μ)⁄σ)2 approximates a normal curve.

Borda count distribution
So we can use opponents' Borda count to define an SOS based upon a composite ranking.

The next question is why should we care about SOS anyway? All of the advanced ratings apply their version of SOS to derive their ratings to begin with. A technical answer is that the directed games graph is so weakly connected that comparison based upon records vs common opponents isn't possible. Advanced ratings produce their rankings based upon AB→ chains as long as needed to form the comparisons but this results in a necessarily fuzzy measurement.

To demonstrate that we'll use our proxy rating. Tier 1 is the #1 team. Tier 2 consists of the #2 team plus all the teams that are as close to number #2 as #2 is to #1. Tier 3 is topped by the first team not in tier 2 and includes all teams as close to it as it is to the top of tier 2, and so on. With 45 computer rankings for games through 26 October, the process comes up with

CutoffTierRanksTeams
579711{ Ohio State }
56212 2-5{ LSU; Clemson; Penn State; Alabama }
51893 6-14{ Auburn; Oregon; Florida; Oklahoma; Utah; Baylor; Wisconsin; Michigan; Georgia }
48534 15-21{ Cincinnati; Minnesota; Iowa; SMU; Appalachian State; Notre Dame; UCF }
43835 22-33{ Boise State; Washington; Memphis; Navy; Iowa State; Kansas State; Air Force; Southern California; Texas A&M; Texas; Oklahoma State; Wake Forest }
37316 34-43{ Michigan State; Indiana; TCU; Virginia; Louisville; Arizona State; Pittsburgh; UL Lafayette; San Diego State; Florida State }
32117 44-55{ Utah State; North Carolina; Missouri; Washington State; Tulane; Stanford; Wyoming; Kentucky; Mississippi State; Duke; South Carolina; Florida Atlantic }
26608 56-70{ Louisiana Tech; Miami-Florida; California; UCLA; Hawaii; Nebraska; BYU; Temple; Tennessee; Virginia Tech; Arizona; Illinois; Marshall; Boston College; Georgia State }
21139 71-86{ Oregon State; Texas Tech; West Virginia; Western Michigan; Mississippi; North Carolina State; Colorado; UAB; Southern Miss; Houston; Georgia Southern; Western Kentucky; Tulsa; South Florida; Arkansas State; Miami-Ohio }
153010 87-96{ Maryland; Fresno State; Syracuse; Ohio; Purdue; Ball State; Toledo; Kansas; Northwestern; Buffalo }
95611 97-114{ Middle Tenn State; UL Monroe; Eastern Michigan; Georgia Tech; Nevada; San Jose State; Central Michigan; Arkansas; Army; Northern Illinois; Vanderbilt; Liberty; Coastal Carolina; Florida Intl; North Texas; Troy; Colorado State; Kent State }
12612 115-127{ UNC-Charlotte; Texas State; Rutgers; East Carolina; UNLV; Texas-San Antonio; New Mexico; Connecticut; Bowling Green; Rice; South Alabama; Old Dominion; New Mexico State }
-57713 128-130{ UTEP; Massachusetts; Akron }
(The breakdown will be different as more of the 100+ computers contribute to this proxy rating.)

So the notion that SOS can provide context for comparing teams in the same tier makes sense. The next question is how to present SOS data in a useful way. I decided to produce a report with a one-line "resume" for each team, listing the opponents' ranks separately for the teams' wins and losses ordered by the average openents rating for wins-only (a team could easily have the highest SOS and finish 0-12.)

For our tier-3 teams, the report looks like this:

Strength of schedule based upon the average computer rankings of teams' opponents. Ordered by average opponent rank in teams' wins.
RankBordaTeamSOSRankRankSOS(W)    Wins    Losses
65449Auburn377367 3183 7 30 48 52 104 114 8 2
75400Oregon30303617 2685 23 47 49 58 77 101 131 6
135240Michigan33732019 2679 17 20 67 97 105 117 12 4
125254Wisconsin30393524 2621 13 34 84 95 103 114 67 1
85377Florida29804325 2590 6 51 54 57 64 131 131 2
105288Utah27426127 2484 39 47 58 62 71 106 131 29
115264Baylor243978322439 26 27 32 72 120 124 131
145189Georgia25377033 2421 20 51 64 85 107 131 54
95352Oklahoma26806535 2403 31 59 72 73 80 94 131 27
The opponents' ranks for wins are ordered best-to-worst and losses worst-to-best. The temporal and location contexts are provided by links from the team names to what I call the resume view of their schedules.

I will add the report to the Ranking Analysis section of the home page, and its current version will always be at http://football.kislanko.com/2019/borda_sosW.html.

© Copyright 2019, Paul Kislanko
Football Home