Can prior season WL and ATS records predict future season performance?

Theory Two is that there are always biases in how people respond to the events of the prior season (note for instance the poor historical record of the Super Bowl losers in the next year), and capitalizing on this should be something possible with enough data and the right tools.

We built a database with the straight-up and spread records for each team during the regular season from 1990 on. Then we ran some correlations and regressions on the data to see how well you can predict the next year by the previous ones.

So on the surface it looks like there is a fair amount of correlation. Ah, but the problem, as Bob Dylan has sung, is times have changed. If we change the data set to only reflect the 1999 to 2003 timeframe (using 1998 history for '99 projections) we get some notably different results:

All the correlations drop, and indeed two of them switch signs! A good prior season WL record or ATS record has been a negative indicator for the ensuing season's ATS performance. Some would say the year the Rams came out of nowhere to win it all in 1999 was the line in the sand between the old NFL and the new game...

All right, you may say, last year is not going to tell us very much in itself about this season, but what about using some kind of rolling number of years?

Well, we're glad you asked and we'll move next to regressions:

Ugh. Even with a fair amount of data points (n=305), the R squared is very small, and the standard errors are bigger than the coefficients in some cases! This suggests that you're not going to fare very well trying to predict the current season Wins using the last four WL records.

The next track then is to up the ante, let's use both the past four seasons Win records and Spread records to see if we might not find some tidbits of value -- for instance a team that posted a nice number of wins but a lousy spread record might be deemed an underachiever which could have a trend one way or the other in the following year.

Even with eight factors in the regression, the R squared values are terrible. Basically we are being told that we should look elsewhere if we're intent on trying to predict next year performance.

Given this, the following projected win tallies should be taken very lightly as they will likely have little predictive value. Not to worry though, we have better regression tools in store!

So according to this admittedly dubious formula we've concocted from the regression outputs, the top three spread value teams for the NFL 2004 season are projected to be the New York Giants, Oakland Raiders, and Detroit Lions, whereas by far the worst spread value for the coming season is predicted to be the New England Patriots. Hmmm, there may be some hint of truth to that after all...