Comments on Statistics Beats the NFL Point Spread: 2009 NFL Point Spread Picks Week 11

It was about time right?

2009-11-23T11:05:04.056-06:00

It was about time right?

Nice week. Probably would've been 4/4 if Warner d...

2009-11-22T22:36:52.675-06:00

Nice week. Probably would've been 4/4 if Warner didn't go down.

Looks like 75% this week. That's got to feel bett...

2009-11-22T18:35:39.256-06:00

Looks like 75% this week. That's got to feel better.

I feel so disconnected! It's already Monday morning here.. and they won't let me switch the TV from Bloomberg to SNF coverage.

Even though I can't (easily) watch the games, I'm probably going to start running my models again this week. I just need to figure out if I can keep up with the news in a timely way.

Good luck going forward!

Jamie, The last few weeks have been a struggle. H...

2009-11-22T10:00:38.207-06:00

Jamie,

The last few weeks have been a struggle. Hopefully we can turn things around. My picks vs yours are:

Den +3
GB -6.5
Stl +9
No -11

I don't think i'll be taking any of these games, but good luck this week.

DDW

Joel, Thanks for your input. I do understand wha...

2009-11-19T17:44:33.718-06:00

Joel,

Thanks for your input. I do understand what you're saying. The trouble is of course with the "who knows." Too many variables in one model quickly becomes untenable from a calculation standpoint. Having many smaller models is much easier and can in fact be computed in parallel.

A compromise may be to use the outputs of the correction factors as inputs into a larger model. This still reduces the number of variables in the master model without assuming additive properties for each of the corrections and without ignoring interactions between the correction factors. This assumes that each correction factor has fewer outputs than inputs.

Thanks again,

Happy

Happy, I would imagine using the number of the fa...

2009-11-19T15:49:14.811-06:00

Happy,

I would imagine using the number of the factors and all factors involved instead of an "if" statement. You may lose alot of information in your data if you apply categories. You very well could be using something other than a regression type model though. Using all the factors, you can always find the relationships between "Heavy defensive lines" and the "Heavy RB1". Of course, you would definitely want to consider the team or skill level. How you would accomplish that is a who knows. The problem is, do you have enough degrees of freedom to do that?

when you are categorizing the factors, you are going to lose a lot of degrees of freedom for other things you may want to implement. If you can spare them, by all means go for it, but I still think you lose too much information by categorizing continuous variables into categorical.

In your model, you can always check to see how well your factors are in determining the point spread with the p-value using whatever test. In a way, it will most likely implement this correction factor in the prediction and detect this behavior.

In terms of prediction, you don't have to worry if the variables are heavily correlated with each other, and if you wanted to check that, check the Variance Inflation Factor or other colinearity measures.

My two cents, don't know if that is what you were trying to imply Happy.

Jaime, Since you are a computer programmer (among...

2009-11-19T15:06:18.094-06:00

Jaime,

Since you are a computer programmer (among other things) and since you have multiple models for predicting NFL outcomes, I thought your site would be an appropriate place to pose a question about object oriented predictive models.

Take for example, a simple base model such as Brian Burke's model based on team offense and defense efficiencies. Then take a tool-box of correction factor models that can be applied based on the situation. For example - I don't have real numbers; this is hypothetical - suppose teams with heavy defensive lines tend to outperform expectations (of the base model) when they play teams with a heavy RB1. An if statement might be if the average DL for team A is greater than (threshold) and the RB1 for Team B is greater than (different threshold) then apply an appropriate correction factor (based on the data that suggested this was a statistically distinct relationship) to the projected result.

The advantage of being able to apply each different relationship as correction factor to the base model has to do with the limits of computational power. Rather than having 3^44 possibilities, it would be more manageable to have 3^8 + 3^10 + 3^6 + 3^12 + 3^8.

The advantage of being able to apply correction factors conditionally has to do with capturing interactions without computing the entire range of possibilities (similar to the advantage of a fractional factorial DOE vs. a full factorial design). Using if statements to apply relationships that are found to be statistically distinct when appropriate and ignoring them when they don't apply holds the best hope, in my estimation, of being able to concurrently harness the benefits of multiple models' strengths without also incorporating their weaknesses.

This approach also allows a model to be constructed in steps, which is nice when starting out (i.e. me), because modules can be added as they are found without nullifying or re-inventing the base model or the previously developed correction factors.

The danger, of course, is that there could be interactions between the correction factors that are developed. Say, for example, teams with inexperienced quarterbacks have a negative correction against teams with strong run defenses and teams with rookie quarterbacks have a positive correction against teams with weak pass rushes. Since a team with a strong run defense may be so in part because they are less aggressive in their blitz packages, these parameters are likely to have an interaction.

I am curious to what degree you have been able to blend your different models and whether you think this object oriented approach has any hope of success.

Best Regards,

Happy

Our models are pretty different this week. Mine d...

2009-11-18T22:57:27.123-06:00

Our models are pretty different this week. Mine doesn't actually bet against yours though, it just doesn't bet with you.

I've got:
PIT -10 (3 units)
CLE +3.5 (2 units)
MIN -10.5 (1 unit)
IND -1 (1 unit)

I've also got:
Over 50.5 in NO vs TB (3 units)
Under 42.5 in MIA vs CAR (2 units)
Under 38 (1 unit) in DET vs CLE

Good luck this week

- Happy

Siggy, Does that mean that we're due to a win in ...

2009-11-18T10:18:31.982-06:00

Siggy,

Does that mean that we're due to a win in that category? The confidence measure takes that into account, i.e., how far the estimate is from the line. Go Cardinals!

These week I don't have any that match your bets. ...

2009-11-18T08:26:40.050-06:00

These week I don't have any that match your bets. Here is what I got:

Vikings -11
Giants -6.5
Patriots -10.5

Vikings: They know how to score, but so do the Seahawks. The difference? Seahawks are mistake prone, any turnovers A.D. has, the seahawks should have double. Why I am wrong: The Brett Farve of late last year starts to return, but if it does, AD will mitigate that effect.

Giants -6.5: Sure they have been losing. Off a bye week, they know they are still in the hunt for the division, they will come out and play hard against an Atlanta Falcons team that was lucky against the 49ers.

Patriots -10.5: They blew last week. Embarrassing. The poor Jets may take the hardest beating of their lives this Sunday, provided Sanchez doesn't produce a miracle.

Jamie--a question on the Arizona line--Do you real...

2009-11-18T06:49:25.230-06:00

Jamie--a question on the Arizona line--Do you realize that in games that your model showed a 10pt. or more variation in favor of the Away Fav. that your record is 0-4.