for display only
Big Blue Interactive The Corner Forum  
Back to the Corner

Archived Thread

Christian, Gatorade Dunk, others who know NFL analytics...

NoGainDayne : 6/22/2020 2:31 pm
I've begun a more formal exploration process with one of the larger AI companies in the world to build out a fairly all encompassing team win maximization program.

I know you two have mentioned specific knowledge of how some companies like PFF are coming to market with this as well as perhaps some knowledge of the team systems. To this point my research has been mainly on the functions / abilities of those working for teams and articles as they trickle out like.

- The Eagles value process shown here and how they integrate things like nutrition

- The Ravens System which seems to focus more on strategy

- The Pats system which I could link to death but won't

- Teams like the Colts integrated more in game information into their play calls and decisions

The more information I could get on the use cases would be much appreciated, here would be fine. Feel free to email me at bbimock@yahoo.com and i'll respond with my real contact info

To those that would like to turn this into a question of how real i'm being on this, i'm prepared to show proof privately to Eric as I've done recently to solve for these distractions
I presume you are building the model  
kicker : 6/22/2020 2:45 pm : link
on publicly available data or on revealed preferences based on previous actions?

If so, beware that the model you build is an ex post realization that may not necessarily align with ex ante beliefs by the team. Or, in other words, you are witnessing outcomes that are not exogenous. And therefore your model should take that into account.
There are some paid data sources that  
NoGainDayne : 6/22/2020 3:34 pm : link
i'm looking into trials in conjunction with the deal that far eclipse public data.

I think this quote from the article DefenderDog shared yesterday is relevant though

Quote:
You might expect that they would use the historic data on Henry’s runs, based on down, distance, personnel, formation, as well as the tendencies of the defense, including against the opponent they’re facing.

"That wasn’t what Gordeev and Singer did. Instead they completely disregarded the players’ identity and instead focused on the data coming from NextGenStats, namely position and velocity at the time of handoff.

Basically, Gordeev and Singer used the GPS data from NextGenStats to turn the players into 21 (the quarterback was excluded) individual vectors* and use a neural network to compute the probable outcomes of the play. They turned a football play into a (very complicated) physics problem, and had a program tell them the probable results.


*Note: In physics, a vector is an entity with both magnitude and direction."


Before I got into setting up advanced pipelines my experience was building great behavioral models, my coding skills have never been near my theoretical math skills. But you need both. I think trying to model what a person sees and understands vs. what a computer does is extremely difficult and instead of focusing on trying to create equations for that it is always my preference to give ways for people to interact with the results to help show the analysis pipelines what they might be missing. Calculate intermediate values in a pipeline (say a rank order of player values) or interact with game or play grades to help flesh out how to improve what the models may be missing.



RB Analytics Article - ( New Window )
That's a good start  
kicker : 6/22/2020 3:56 pm : link
(though I call paid data just restricted public use; but that's semantics).

However, it doesn't address what I said, which I think you need to seriously consider. If you don't have access to underlying assumptions that teams use (or there is the chance that some data is not there), then the analysis, coding, behavioral, or math, has a flaw.

One that even AI (or machine learning, or simulations, or monte carlo methods, or ...) can't solve to this point; it's an outcomes focused analysis (endogeneity).

It's not necessarily problematic or makes your analysis wrong, it just means you're closer to monitoring an average treatment effect, rather than a causal mechanism.

Just a thought.
That said, NFL analytics is an interesting field.  
kicker : 6/22/2020 4:00 pm : link
One of the key guys in Sports Econ used to work where I now am, and though he's a gigantic douchebag, he's been trying to puzzle through NFL analytics for years now. Baseball, basketball for him are "easy".
All the best  
Bill2 : 6/22/2020 4:16 pm : link
you are trying in the real world

Can't help but learn no matter what happens


Therefore you win in the long best.

Salud
Uggh  
Bill2 : 6/22/2020 4:30 pm : link
in the long run
Yeah so here is where I think the discussion of  
NoGainDayne : 6/22/2020 5:09 pm : link
PFF comes into play. There isn't a data set that incorporates what teams are thinking on plays. You could get one that incorporates just what your coaches are thinking but without the match of the other side, I don't consider that a fruitful data set. My only concern is with data sets that are complete for all teams and players.

I will say areas that i'm focusing on like resource optimization leave more room for error where they can still be high value. IE missing a game score one way or the other could be the difference between being useful or the complete opposite. But the aggregation of forecasting every player and using that to make informed higher level decisions can be useful with lots of smaller misses.
Back to the Corner