Creating NFL model

**Indecent** · 01-11-11, 01:26 PM

Originally posted by simplydusty

So I have my database now filled with around 20-25 stats for each team, each game from 2002-current. I want to create a predictive model using regression analysis by the start of next season.

My plan is to use all 9 seasons when running my regression analysis. Then use 6 of the 9 seasons to create my model, and the other 3 to test the model.

Does that seem like the best plan of attack? Granted more data is better, but 9 full seasons worth of data should work, right?

Thanks for any help

Only run the regression analysis using the first 6 of the 9 seasons.

**simplydusty** · 01-11-11, 01:34 PM

So basically keep 3 season out of anything involved with creating the model?

Do you think 3 is about right for the number of 'testing' seasons?

**Justin7** · 01-11-11, 02:24 PM

6 seasons is a good base. Adding 3 might give you a little more power, but not a lot.

It's not a huge datamining issue to use the last 3 seasons for the regression though... as long as none of your regressions look at a game spread, total, or any other market you want to wager.

**MadTiger** · 01-11-11, 02:29 PM

Originally posted by simplydusty

So basically keep 3 season out of anything involved with creating the model?

Yes. I was trying to look for a link that talks about the statistical analysis principle behind this warning.

But yeah, you don't want to use data that helped create the model as a test data sets.

**Peregrine Stoop** · 01-11-11, 03:18 PM

lol @ 25 stats. You're going to find a lot of auto correlation in your model.

**greg66** · 01-11-11, 03:51 PM

Originally posted by Peregrine Stoop

lol @ 25 stats. You're going to find a lot of auto correlation in your model.

Hetereoscedasticity as well.

**rfr3sh** · 01-11-11, 03:51 PM

I would say start with something besides NFL
since the limits are like 70k on pinny