GameDay

To see where the Pats' offense and defense stand, here are our computer model results.

1- First, explanation of our model , Exhibit 1A shows Scoring Efficiency (ScEf) as a function of starting position of a drive (both offensive and defensive). It shows the mean point expected for a given drive position. The red curve is a simplified model based on NFL historical data; the green bar shows 2007 data of all teams except the Pats. For example, if an average NFL team starts 10 drives at 60 yards away from the EZ, they can expect to score 2x10=20 pts or given up 20 pts.

The most important part of the red curve is the range from -100 yrds to -50 yards, because most drives start in that range. Exh. 1B shows the distribution of drive-start positions. The black curve is the historical model, and the red bars are actual 2007 data.

To evaluate a team offense or defense, one way is to look at their ScEf for the -100 to -50 yd range. In Exh. 1A, the dashed blue curve shows the model for a good offense (high scoring efficiency) and bad defense (high allowed score); Conversely, the dashed red curve shows a bad offense and a good defense.

2- The Pats offense is shown in Exh. 2A and 2B. The ScEf (blue curve) is totally out of this world. They can score almost from any position. The curve was even more amazing before the Eagles and Raven games. The rating has gone down, but still very good. Exhibit 2B shows their actual drive starting positions. In fact, their average score per drive is ~3.4 and given that they have ~ 12 drives/game, they can score ~41 points as expected.

3- The Pats defense is shown in Exh. 3A and 3B. It has been known to us for a while that the Pats D is just a tad better than mediocre, as the purple curve (actual data) is near the red curve (but below). But these data need to be adjusted for opportunities: because the Pats' offense is so good, there were less time and opportunities for the opposite offense. Our opp-adjusted data is shown as the brown dots, which is indeed slightly sub-par (which raised the flag on the Pats D long before the Eagles and Ravens' games).

Even worse, the computer model (black curve) indicates that the Pats' defense is slightly below 2007 NLF average. It is NOT based on actual scoring data, but only from a model based on actual plays (regardless of scores). Before the Eagles and Ravens' games, one can dismiss the black curve as a computer model deficiency (which is fair and probable). But after the last two games, it agrees well with the adjusted scoring data (brown dots).

For reference, Exh. 3B shows the actual Pats' defensive series distribution. We can dismissed their poor performance (the spike in Exh. 3A) in the range from -30 yards and less as statistically insignificant, because there are so few drives there. But the black curve shoots up near there, which is ominous.

Epilogue May be the Pats' D will improve in time.

May be these models are still inaccurate with insufficient statistics.

May be AJ Feeley and K Boller happened to have the performance of their lives. While it is nice to give people deserved credit, it can be self-deceptive to ignore the facts that these guys are pumpkins outside the Pats games. It is equally probable that they are as bad as their records say (Phila lost to Seattle and I'm not holding my breath for the Ravens to beat the Colts) and accept the inference that the Pats' D is simply as bad as the computer model has been suggesting all along.

What I know is that some people bet against the Pats (taking the huge home underdog Ravens) based on this model and win big. I didn't want to bet against the Pats. But I couldn't stop others from laughing their way to the bank.

Caveat: people and team can improve and I surely hope so for the Pats. But it's a downward spiral right now. Computer rating such as Sagarin's (which I disagree with the method) had the Pats ~ 42+ points before the Eagles game and dropped to ~40, and now ~39 after the Ravens. Our model does not rate with a single number like that, but it shows a significant drop also.

The Pats' defense also has a larger variation than average. It means they can be very good in one game and awful in another. This is different from the Pat's offense, which has a narrower performance range.

Attached Images

	Exh 1.jpg (47.0 KB, 230 views)
	Exh 2.jpg (32.8 KB, 225 views)
	Exh 3.jpg (33.9 KB, 224 views)

makewayhomer

could you explain this more?

GameDay

By taking into account all the team's plays, a model can infer their probability for the plays. Then, it can compute the team capability, either scoring or preventing from being scored. The model is validated with a correlation between predicted values and actual. It is validated for historical data and the current season data of all teams.

An analogy, by measuring the performance of an athlete in each event of a decathlon, including the person's initial strength and stamina, etc., one can project the final performance of the whole decathlon events before it happens.

But there is a large variation (or uncertainty) for each team. Some teams can be under performing (unlucky) or over performing (lucky).

Both Pats' offense & defense have been over performing for a while. When this happens, the model eventually adjusts the inferred capability such that it agrees with the data. But before that happens, it can give glimpse into a team capability.

Many of Pats' defensive plays can be dismissed in blown-out games (Miami, Dallas) as being "let up" plays. But if one takes those at face values (without the psychological factor discount), they indicated the Pats' propensity for giving up long drives that eat the clock and lead to scoring. It means less time and opportunities for Pats' offense and it is a double-whammy effect.

Not until the Eagles and now the Ravens that things become more ominous. May be those defensive series given-up in Jets, Cleveland, Miami, and Dallas games are for real. They disappeared in Washington and Buffalo, but reappear now.

5 Rings for Brady!!

Thanks for a interesting post. About the only interesting posts we can count on anymore, so keep up the good work.

I believe that the defense will not stop anybody in the playoffs and this season will rest entirely on our offense. If something changes along the way, then great. But as of today, we have not stopped anybody on defense consistently outside the pathetic Redskins and the benched Losman. I thought that Adalius might actually give the team a lift in Colvin's shoes, but not so far, albeit after one game.

I believe our offense is good enough to do the job. But all our close games have involved one huge X-Factor. The officials have let the other team cheat on virtually every play in the Dolts, Eagles and Ravens games. This is an enormous disadvantage for us on both sides of the ball and really screws with your otherwise excellent stats.

If you eliminated all the plays that where affected by holding or pass interference from your stats, I am afraid you would basically have to cancel out the entire games against the three previously mentioned cheating franchises.

GameDay

Thanks. A detailed analysis may take into account these plays. But that's pretty laborious to do. In spite of these plays. The Pats' offense still delivers! Amazing how productive they are.

They scored 27 pts last night, which is only ~ 0.5 sigma below their capability against an average defense (with 1/2 of TOS). Can't complain about that.

The Ravens D is just slightly above average (regardless of NFL ranking) and hence, Pats' expected mean score should have been 33. Expected special team and turn-over score should have been 1.8 pts. The defense should have given up only 12.4 pts.

The Ravens should have gotten 2.3 pts for home field.

makewayhomer

ok thanks. without completely seeing your model, my guess is that it underestimates situational affects. ie, all 10 yard gains are not created equal: a 10 yards pass on 3rd and 8 is way better than on 3rd and 11. it is possible for defenses to adjust for this and give up "just the right amount" of yards.

ie, "bend but don't break" defenses exist. does your model consider this?

GameDay

Yes. Giving up 1st down is a metric in the model. In fact, a hypothetical team that gives up several short 1st downs can be worse than a team that gives up one occasional big play (not home run), but forces 3 and outs often.

Same with offense. The dink and dunk O, which hopefully Tom Brady will get back to it, can be better than the long bombs. Both AJ Feeley and Boller got greedy, and made play out of their league and paid for them.

Someone once observed that Emmitt Smith was more damaging (to a defense) as a running back than Barry Sanders because he kept on getting 1st down yardage regularly, while Sanders occasionally broke loose a long one but were also often stopped for 3 and outs.

Now, I don't know if this was true, but that's the idea. It's all in the probability.