Here's how it breaks down: each team plays 16 games and collects anywhere from zero to 16 wins. That win total translates to a percentage of the possible wins, which we all recognize as "winning percentage." Tampa, Philadelphia and Green Bay shared the wins crown with 12, in an unusually dense pack of results at the top--19 of 32 teams finished .500 or better, but none with more than 12 wins. Because this is the only set of outcomes that the teams are able to control themselves (ie, it's within their power to increase their win total if they play better), this percentage should be heavily weighted in the power ratings. After all, the idea is to win your games, right?

But 16 games isn't much of a sample, and you can't really say too much about the difference between the Packers' 12-4, and say, the 49ers' 10-6. Is 125 percentage points a big gap? It's hard to rely on the significance of that difference by itself, but any discussion of power starts with your ability to win, so we assign a weight of 35.5% to the base win percentage.

We also need to go deeper and see how their opponents fared in their own games, to see whether one team played a relatively tougher slate of games against better opponents. For each team, we compile the win total of all 16 teams played during the season, to come up with an opponents' win total, out of a possible 256 (16 opponents times 16 games). The Eagles' opponents won 120 games between them, for a .469 average. By contrast, the Raiders--a team with one less win--faced opponents who posted 135 wins, or .527. That difference definitely elevates Oakland's power rating, and for Philadelphia's, a 30th-best .469 winning percentage among opponents takes some of the shine off the achievement...not enough to drop them out of the top three, however.

The opponents' percentage can be skewed by the presence of very good or very bad teams within one's own division. Even more shocking than how poorly the Eagles' opponents fared is that Green Bay's drop a full 20 points lower. But when six of your games are against teams with three, four and six wins all year, your number will suffer. One serendipitous result of expansion to 32 teams is that the weight of divisional games is reduced from half the schedule (eight) to between a fourth and a third (six). So some of that divisional skew goes away this year. Still, these numbers, while less volatile in their range than basic win totals, are subject to small sample sizes (only 256 games). So we assign a weight of 30% to this percentage.

Finally we go even deeper into the numbers, and record the winning percentage of the opponents of the opponents played in a season. This yields a sample of 4096 games for each team, which is plenty to minimize statistical error. These numbers are, like opponent win percentage, out of a team's control, but because there are so many games to review, noticeable differences between teams are more likely to be significant. Thus, we will weight these final figures 34.5%--not as high as the base win percentage, but more than opponent win pct.

To compute the power rating, first all the wins must be equalized. Because there are so many fewer games in a 16-game sample, than the 4096-game sample of OO Wins, a difference of just one win in a 16 game season can disproportionally skew the results. Therefore, each win must be multiplied by its proportion of all possible wins in the sample. For base wins, the number is .0625, or 1/16th. For opponent wins it's 1/256th (.00782), and for OO wins it's 1/4096 (about .002445). We then take these adjusted wins, wins of their opponents, and wins of _their_ opponents, weight them (multplying them by .355, .3 or .345), and divide that number by the weighted result that would equal an average, .500 season. The same process is conducted as for the actual wins, and then that is divided by each team's result. The raw power number is then multiplied by 1000. What you end up with is the power rating you see in the table. The perfectly average score would be 1000, so the higher you are above 1000, the better. Conversely, teams under 1000 are less than average.

Having the benefit of a week's worth of hindsight, we can suggest a possible reason why the Packers were caught flatfooted by an apparently inferior Falcons team: they hadn't really been tested, and didn't fare all that well when they were. On the other hand, you couldn't tell by looking at the Falcons' power numbers, that they were ready to break the 35-game Curse of Curly. Despite division opponents who won no fewer than seven games apiece, their opponents were by and large subpar. And their OOPCT is even worse, reinforcing their suspect achievement at backing into the playoffs. You can't suspect the way they dismantled the Packers, though.

If we were to predict the NCAA-like "Final Four" from these rankings, we should expect to see an Oakland/Tennessee -- Tampa/Philadelphia conference finals matchup. From here, those seem like the safest, most rational choices as well. Oakland in particular really faced a high caliber of opponent overall, and put together two 5+ win streaks. All they need is a three-game roll now. Tampa will give the Niners trouble, although clearly counting anyone out after this past weekend is a bad idea. Philly's bad games against suspect teams during the season--chinks in their armor of dominance--and the possibility of a rusty McNabb, plus the 2nd marquee QB matchup of the playoffs, should make for a great storyline.

However, before we get carried away reading predictive tea leaves on who's in and out in the playoffs, a single game situation is entirely unpredictable--Michael Vick and Brett Favre trump any statistical theory. So what goeth before does not presage what cometh after. They do not take into account margin of victory, scoring differentials, injuries or the unusual motivations of homo sapiens. This rating concentrates on the sole outcome that matters...winning.

One last note--the Steelers and Falcons got no credit for their tie game, since neither received a win. The game counts in their totals as a "game not won."

Enjoy reading over the results!

:: comments to mark bunster