Monday, April 21, 2008

How will your favorite NFL team do next season? (Part 2)

I'm testing regression to the mean of win-loss records of NFL teams.

I broke down the records of teams from 1993-2007, recorded what each team's record was the following season and how many years it took that team to get improve/regress to .500. In 1993 the current system of free agency was put into place and helped increase parity. Four teams were also added over these years (hypothetically spreading the talent pool more thin).

Here's the breakdown of team records. (Note: I did not include the 6 teams involved in ties in this distribution. I only used those teams records in the year-after recording (ex: The Falcons won 7 games in '01, and 9.5 games the next [9-6-1 record]) .

This analysis is useful for two things:
To see how long, on average, a team with a given record take to regress to .500 (8-8).
To see how some teams take more than 2 standard deviations more than the average to get back to 8-8. Those teams are either run very well or run very poorly. Which teams are those?

Here is the average change in wins for each given record (+/- the standard deviation), and the average number of seasons it took to get to .500 (+/- std. deviation). *NOTE* These averages do not include teams that have not yet regressed to the mean. Therefore, they slightly (in most cases) understate the Years to .500. Example: The Patriots have been above the mean since '01. They did not get a Years to .500 recording in any of their seasons. They're a phenomenon of a few deviations beyond the mean, truly unexplainable by randomness.

*Note 2* Five of the 9 teams that have gone 14-2 returned to .500 or below the following season. The other 4 have yet to return to .500, including two 14-2 Patriot teams.

Notice how every team in the NFL is essentially 2 seasons away from .500 or better/worse. Dynasties are a rarity.

Teams who took longer than 2 standard deviations beyond the mean to regress to .500 from their given record (* = 5%, **=1% significance levels):
49ers from 1993-1999**
Chiefs from 1993-1998*
Packers from 1993-1999*
Dolphins from 1997-2004**
Eagles from 2000-2005*
Patriots from 2001-current**
Colts from 2002-current*

You don't often think about the Chiefs and Dolphins as "dynasties" since they weren't in Super Bowls. But what they did was statistically amazing. How did these teams defy the odds? I can think of a few reasons for the Patriots. 5 MVP QBs represented on the list, but perhaps that's endogenous. Note: If Mike Holmgren has another winning season in Seattle in '08, he will have been a part of 3 dynastic teams.

Now, the worst "dynasties," a monument to historically bad statistical significance. Teams who took a very long time to get to .500 from their given record (*= 5%, **=1%):
Saints 1994-2000*
Rams 1993-1999*
Bengals 1997-2003*
Cardinals 1999-2007** (!)
Texans 2002-2007*
Lions 2001-current**

No surprises here, except the Rams turned things around with a Super Bowl in '99. The Cardinals' management must be really bad. If I were a fan of any of these franchises, I'd be all over the media to publish how the team management is scientifically suspect.

It looks like losing teams regress to .500 slightly slower than winning teams, meaning it's harder to improve than get worse. It's not significant though (if it's even there).

So, how can you use this information? You can't really for your fantasy team. But, remembering regression to the mean is useful in picking players.

The Bengals went 7-9 last year. 7-9 teams have some of the most volatility and are tough to predict. They are slightly likely to improve. There's a 68% chance that they'll go somewhere between 4-12 and 11-5 this year. A 95% chance that they'll go somewhere between 1-15 and 14-2.
If you think they have a good draft and keep Chad Johnson, you might predict an improvement toward the 11-5.

You can compare the chances of a team getting a certain record and compare it to what Vegas predicts (which I'll revisit here at the end of June). Looking at Brian Burke's method is very useful, my data above will just help you quantify the chances a little bit.

No comments: