Pats' rating and the Eagles game

[GameDay]

Just want to share a different perspective on the Patriots' power. We do some computer modeling and it shows some aspects of the Pats' that may or may not be correct, but intriguing for thought.

A. Exhibit 1A and 1B show a model for Pats' scoring capability (both for and against) vs. a hypothetical NFL opponent with an average O, D, and ST. The bar charts are actual data to date. The curves are not from the data. They are computerized model based on the Pats' plays. Notice that the data of the "for-score" (green bars in Exh. 1A) is to the right of the green curve, which may indicate that the Pats have been scoring higher than expected, probably from playing with sub-par defense.

The red bars and the red curves in Exh. 1B seem to be a bit more consistent, which may indicate that the Pats have been playing more even offense.

B. But more interestingly, the Eagles game result is not about "blue print" or as shocking as what one may think. Exhibit 2 shows the Pats' win margin, and while one can expect the Pats to win by ~19 pts against an average team, the chance for such a team to pull an offset is ~6.5%, which is not that small, and to pull within 3 points is 10%.

At the same time, winning by 40 pts or more is also not probable. But one can be selective to think that the Pats should do that all the time while thinking that winning by 3 is rare. The Eagles are better than average, and after the game with update play capability, they can pull within 3 pts with the Pats at 18% chance. So, the Eagles game is not any more surprising than the Redskins and Bills' games.

C. Finally, what is most intriguing is seen in Exh. 3 that compares the Pats O and D. Here, the scoring is calculated only for net O and net D, not including special team. The Pats' O did not score as high as we think. For example, they scored only net 35 pts in the Miami game (7 pts from special team and minus 7 pts INT returned). The D is not as bad as it seems, since they gave up 28 pts to the Eagles, but scored 7 pts to their credit for a net 21. The model is consistent with both cases.

The most intriguing suggestion is shown in Exh. 3C, when the Pats' O and D are normalized so that they can be compared with each other. Here, the Pats' O curve is narrow, which means that it is much more consistent than the Pats' D, which has a broad curve. In other words, in games like the Eagles, the Pats' O did its job as expected (might be a bit on the low side), but the Pats' D can be more "Jekyll-and-Hyde". They can shut out or give up huge # of points. Their performance curve is much worse vs. better defenses. If there had been a TBrady clone driving the Eagles on the last drive, I think the Eagles would have won. In other words, TBrady + Moss+Welker+Stallworth et al would have lost to TBrady +Eagles offense.

Consistency is really the power of team. The long right tail of the red D curve in Exh. 3B and 3C means that there is a very significant risk they'll give up 30 or 40 pts a game, a la 2006 AFCCG.

I believe that the Pats will go 19-0, but if they lose one game (regular or play off), this computer model indicates that it will be most likely on the D performance, not the O. Thus, the Ravens should not worry us, but Green Bay and Favre should be cause for concern.

 

[ClevTrev]

I believe that the Pats will go 19-0, but if they lose one game (regular or play off), this computer model indicates that it will be most likely on the D performance, not the O. Thus, the Ravens should not worry us, but Green Bay and Favre should be cause for concern.

Or for that matter, Indy and Manning or Dallas and Romo.

 

[Tony Cahncuhd]

Nice Work! It would be interesting to see what Graph 3C for other teams looks like. The Pats offense is such an anomaly this year.

 

[bbabin]

Great job. This is some interesting stuff. Based on this, it looks like we have a pretty good chance for our margin of victory to be positive for each of the remaining games. Here's to that! :rocker:

 

[NEGoldenAge]

Pretty interesting stuff, I wonder how accurate this model would be in applying to future games. Just out of curiosity, this looks very much like a school project... is it?

 

[GameDay]

Or for that matter, Indy and Manning or Dallas and Romo.

Yes. They are threats.

Nice Work! It would be interesting to see what Graph 3C for other teams looks like. The Pats offense is such an anomaly this year.

Thanks. Believe it or not, but Pitts' D and Indy's D are both more consistent than Pats' D. I can't figure out why. But seems that a lot of scores against Pitts' D are from their poor ST?

But injuries can change all. Freeney and Colvin etc.

Great job. This is some interesting stuff. Based on this, it looks like we have a pretty good chance for our margin of victory to be positive for each of the remaining games. Here's to that! :rocker:

Pretty interesting stuff, I wonder how accurate this model would be in applying to future games. Just out of curiosity, this looks very much like a school project... is it?

Nah.. it's just for betting (small amounts, not high roller). A most satisfying thing of this model is that it correctly produces the shape of NFL score distribution for the last 12 seasons (Also both total pts and win differentials).

Here is the Pats' remaining game (as of today rating, not including or knowing the impact of Colvin's loss)

Ravens: 95.4%
Steelers: 82.3%
Dolphs: 96.7%
Jets: 95.8%
Giants: 89.4%

Chance for Pats to go 16-0 (from today): 65%

All these numbers are ridiculous. Most NFL game splits are 70%-30% to 50%-50%.
Any game higher than 70%-30% should be considered as non-competitive.

 

[Pujo]

Pretty interesting stuff, I wonder how accurate this model would be in applying to future games.

It's tough to apply statistics to football because the sample size is so small (very few games in a season). You can show some trends, but exceptions tend to happen rather often.

 

[ClevTrev]

Thanks for this information, as it adds an element of analysis that hasn't been discussed here before. Hope you plan on providing updates prior to all remaining games. While there is definitely a margin of error in this and knowing it isn't intended to be a lock, it still appears to provide a meaningful breakdown. This is good stuff!

 

[GameDay]

It's tough to apply statistics to football because the sample size is so small (very few games in a season). You can show some trends, but exceptions tend to happen rather often.

No, this is not based on descriptive statistics - as you said, because of small sample size. Model is based on performance of plays which obey fairly solid statistical law and which do have large sample sizes to obtain the parameters.

That said. Game is a random event. The higher randomness (or entropy of the outcomes), the more interesting it is. The model does not do prediction. It can only say the confidence level for each outcome of our interest. So, we can say that the model is 95% confidence that the Pats will score more points than the Ravens. That's all it does. We human use the info to predict events. (All in the spirit of fun... or profit if you bet on games).

 

[makewayhomer]

No, this is not based on descriptive statistics - as you said, because of small sample size. Model is based on performance of plays which obey fairly solid statistical law and which do have large sample sizes to obtain the parameters.

That said. Game is a random event. The higher randomness (or entropy of the outcomes), the more interesting it is. The model does not do prediction. It can only say the confidence level for each outcome of our interest. So, we can say that the model is 95% confidence that the Pats will score more points than the Ravens. That's all it does. We human use the info to predict events. (All in the spirit of fun... or profit if you bet on games).

gameday, great stuff and post. very interesting.

are you familar with football outsiders?

 

[makewayhomer]

how have your results been, wrt to betting? care to share any results?

from looking at your win probabilities, it matches pretty closely to the (predicted) moneylines of future regular season games. Vegas is basically saying we also have a 65% ish chance to go 16-0.

another comment to explain offensive vs defensive variance: this has a lot to do with opponents faced.

according to Football Outsiders DVOA ratings, we have faced the following offenses

2,3,5,6,7,11,20,21,22,23 (twice)

and the following defenses:

3,6,11,14,16,19 (twice), 26,27,30,32

so, we have been playing better offenses (average ranking: 13) than defenses (average ranking: 18.5)

 

[ctpatsfan77]

Here's a question worth examining, GameDay: how has the Pats' performance on D (and O) been relative to what you would predict for the opposing teams?

 

[braveht]

I don't understand yuor abscissa in Ex 3C. The spread at 1% points(0.01 pdf) is ~2 for D and 0.8 for O a ratio of 2.5(and thus your conclusion). But the point spreads between 1% pts is 27pts on O (48-21) vs 30pts(32-2) for D, a much lower variance. Dividing by the mean(average) to normalize produces a distortion usually reserved for politics and economics. Otherwise good interesting data.

OBTW The distribution of the difference between two normally distributed variables can be derived from the individual means and variances; but I haven't any texts with me and it's been since 1963 since I last did this.

Kudos

 

[GameDay]

how have your results been, wrt to betting? care to share any results?

from looking at your win probabilities, it matches pretty closely to the (predicted) moneylines of future regular season games. Vegas is basically saying we also have a 65% ish chance to go 16-0.

betting result is one thing I'm bound not to say (other people are in it too).
But I can say we win modest amount. Other partners have used "home dogs" strategy and successful for years. We began this ~3 yrs ago and the model was not as sophisticated as today, yet we were winning more than now. The spread is just too good to beat.

Dogs at home however, seem to be harder to win anyway. Home teams win an average of only ~2.2 pts in recent years.

BTW, Accuscore is based on a similar concept, but we don't know the details of their models. About DVOA: we don't understand enough to make good use of their results. Perhaps it's worth it.

 

[makewayhomer]

betting result is one thing I'm bound not to say (other people are in it too).
But I can say we win modest amount. Other partners have used "home dogs" strategy and successful for years. We began this ~3 yrs ago and the model was not as sophisticated as today, yet we were winning more than now. The spread is just too good to beat.

Dogs at home however, seem to be harder to win anyway. Home teams win an average of only ~2.2 pts in recent years.

BTW, Accuscore is based on a similar concept, but we don't know the details of their models. About DVOA: we don't understand enough to make good use of their results. Perhaps it's worth it.

yeah, it is complicated. I have been using a DVOA based betting system (very small sample size of 43 games) with good results (28 correct). my variance is very, very positive - there is now way I can continue to be 65% correct. my guess is that it's good enough to beat the vig, maybe 55% or so

 

[GameDay]

I don't understand yuor abscissa in Ex 3C. The spread at 1% points(0.01 pdf) is ~2 for D and 0.8 for O a ratio of 2.5(and thus your conclusion). But the point spreads between 1% pts is 27pts on O (48-21) vs 30pts(32-2) for D, a much lower variance. Dividing by the mean(average) to normalize produces a distortion usually reserved for politics and economics. Otherwise good interesting data.

OBTW The distribution of the difference between two normally distributed variables can be derived from the individual means and variances; but I haven't any texts with me and it's been since 1963 since I last did this.

Kudos

Look at it this way: Pats' offense always scores more points than D. Scaling it allows us to see the variation of Pats' O scoring vs. D's allowed scoring relative to mean of each. This way, we can see that Pats' O is more consistent than D.

BTW. These distributions are not normal, they are too complex for some analytical pdf. But can be approximated. (The model involves random variables with Lognormal, binomial, Poissons, Chi, Beta, Gamma, Weinbull...) - Don't want to turn this thread into a nerd thread.