That stats chick on the NFL Network has us at 9-7.
NOTE: This response is a bit mathy. I tried to keep this as brief as possible. This would have been multiple pages had I put in examples, but I'll be glad to respond to any questions/responses.
One of the reasons I left econ after my prof called me a prodigy was that I had tremendous disdain for how poor the modeling was. I mean, if economic modeling was worth anything, even on a macro scale, then it would be increasingly difficult for mass speculation using instruments like derivatives, for example. I believed this in 1988, but 2008 proved what I was saying better than any paper I could ever have published.
In baseball, the analytics play out to a greater degree due to the much larger number of iterations across 10x more games than in the NFL. Also, the context is easier to manage since the game revolves around a one on one matchup, hitter v pitcher, and then expands from there.
In the NFL, it's 11 v 11, and since each player affects the other 21 players on each play to some degree which can be measured even if negligible (and yes, much like one can't pee in the ocean in Los Angeles and affect the water temperature in Tokyo, there would be instances of limits which any algorithm would have to take into account), we're talking about 22ˆ21. That's a massive difference. Moreover, the NFL variables are much more sensitive to context. Thus the algorithms can go way off by being only slightly incorrect and we routinely see predictions that are wildly off. Were it not for limits built within some of the algorithms, I feel very confident some games for example would have teams scoring thousands of points in a game. The limits aren't the context, they act like bumpers in the gutters of a bowling lane when kids bowl.
Also, the trend in the NFL with analytics don't take into account the effects of changes to installed systems, new installers of systems...that being new offense/defense and/or new OC/DC. Moreover, there isn't generally a proper accounting for how a player fits within a scheme. Lastly, players are often devalued not for their play, but for moves made as a result of the salary cap or how a FO/HC view a player's fit into a scheme.
One example would be Brandin Cooks. Why hasn't he stayed with one team? Well, Brees is at that place where Payton's scheme has to rely more on timing than speed and he's struggled to hit speed receivers.
Cooks went to NE where it takes time to learn that EP offense, still had success, but Belichick has been reticent to pay people coming up on big contracts. Thus, coming up on his last contract year, they ship him out for draft picks which is NE's SOP. Is Cooks a receiver who bounces from team to team because he's not a fit? Not at all.
Unfortunately, that's the lazy narrative. Cooks is a PERFECT skill and personnel fit for this offense.
And Cooks is but one of many players that are misunderstood from a numbers perspective.
Analytics only works when the inputs are accounted for properly. You could say that the context of the variables is as important as the variables themselves. And routinely, with NFL analytics, they get the context wrong.
Let me make it more simple. Target predicted that a 16 year old girl was pregnant because they understood the context of a person purchasing multiple items at the same time. There was a correlation which could be measured. Those purchasing X number of items from a list of things most pregnant women buy within the first trimester meant a statistical likelihood that the buyer was in their first trimester of pregnancy. The more items purchased, the greater the statistical likelihood.
However, that only works if Target picks the products which actually directly correlate to the purchases of women in their first trimester. If they misunderstand the needs of women in their first trimester or misweight certain variables, they could end up thinking camping gear denoted pregnancy as opposed to saltine crackers and cocoa butter.
Could the Rams go 9-7? Not with a healthy starting lineup. Not even with moderate injuries. Which leads me to believe that without modeling for massive injuries that her algorithm, the only way to articulate context in a math based system, is simply way off.
And based on statements I've listened to and read from her, I'm not convinced that she truly understands player context to a degree that she can create profiles accurate enough to mean anything in a model (remember the 22ˆ21 is 1.55 x 10ˆ28. Thus, with so many permutations PER PLAY, even the slightest discrepancy in a model can render the outputs moot.
Now, I don't think she's modeling a season on a per play basis, but that means she's using even greater approximations for player or unit performance over a season. That only works if the unit modeling is extraordinarily accurate. Clearly, they are not.
TL;dr Modeling is using known variables and known equations that describe behaviors or phenomena and extrapolating future outcomes. Analytics is essentially contextual modeling or using modeling within a specific framework with variables and correlations which only have value within that system.
NFL analytics is still in its infancy and is held back by the knowledge of true and accurate player evaluation and how the variables of scheme, salary cap, FO and/or HC disposition, etc play out. A prime example would be Jeff Fisher or Jon Gruden. One cannot predict unit performance without accounting for their predisposition for older players and thus the muted contributions of first and second year players. Without the ability to properly weight the variables, the algorithms used to determine efficacy, overall performance or success simply output garbage.
And as we see year in and year out, the algorithms aren't remotely close enough to be of value. They'll get there, but they're not there, yet.
