Looking over SI.com’s and ESPN NFL “Power Rankings,” I wondered what drives them. They are part beauty contest trying to attract readers but contain some kind of informational content. Particularly, do the power rankings suffer from “path dependence” — punishing teams for an early loss or two, or do they weight a bad loss or two a lot. In the NFL, performance, even for good teams, varies considerably from game to game in a long regular season based on injuries, game-specific strategy, officiating, weather, and bounces of the ball (See Media and Probability Distributions.) Also, did they account very well for the problem of discreteness of W-L records versus more continuous performance measures like point differentials. How do they stack up versus market-based, betting rankings.
First, I collected data on game-to-game point differentials for Super Bowl winners over the past 5 regular seasons to get a sense of the level and variability of performance. (Similar numbers appear if using just the first 7 games of the season.)
Average Differential = 7.0
Median Differential = 7.0
Standard Deviation = 13.3
The outcomes form a near perfect normal (bell) distribution based on all the typical measures. Cross-game performance varied more than I had expected. These very good teams can easily range from two touchdown winners to touchdown losers and with some regular frequency lose by two touchdowns or more. An NFL season is a lot like running a marathon — it’s a long, long race and everybody has at least a bad mile or two.
Second, what would a “Power Ranking” based on point differentials at this point of the season would look like? I used median differentials rather than averages so that a blow-out in seven or eight games would not carry too much weight. The Table below presents these results along side the SI and ESPN rankings as well as a ranking based on current odds to win the Super Bowl and VegasInsider.
| Team | W-L | Median Pt. Dif | Rank by Pt. Dif | SI | ESPN | Vegas |
| Colts | 7-0 | 21 | 1 | 2 | 2 | 2 |
| Eagles | 5-2 | 19 | 2 | 8 | 9 | 6 |
| Saints | 7-0 | 18 | 3 | 1 | 1 | 1 |
| Cowboys | 5-2 | 13 | 4 | 9 | 8 | 8 |
| Vikings | 7-1 | 12 | 5 | 3 | 3 | 5 |
| Steelers | 5-2 | 8 | 6 | 7 | 7 | 4 |
| Broncos | 6-1 | 7 | 7 | 4 | 4 | 10 |
| Cardinals | 4-3 | 7 | 8 | 13 | 13 | 11 |
| Falcons | 4-3 | 7 | 9 | 11 | 11 | 16 |
| Patriots | 5-2 | 6 | 10 | 6 | 5 | 3 |
| Ravens | 4-3 | 5 | 11 | 10 | 10 | 9 |
| Packers | 4-3 | 5 | 12 | 16 | 12 | 14 |
| Giants | 5-3 | 4 | 13 | 14 | 15 | 7 |
| Chargers | 4-3 | 4 | 14 | 15 | 16 | 12 |
| Bengals | 5-2 | 3 | 15 | 6 | 6 | 13 |
Next, for comparisons of these rankings, I estimated correlations between each pair:
| Rankings | Correlation |
| PtDif-SI | 0.64 |
| PtDif-ESPN | 0.66 |
| PtDif-Vegas | 0.71 |
| SI-ESPN | 0.96 |
| SI-Vegas | 0.68 |
| ESPN-Vegas | 0.67 |
The SI and ESPN rankings track each other nearly perfectly. The Point Differential ranking more closely correlates to Vegas odds but only slightly more than the SI or ESPN rankings. The Vegas market for predicting the Super Bowl winner seems to place more weight on “reputation” from recent season success such as New England which jumps up to third in Vegas and Pittsburgh which jumps to fourth.
