If you haven’t had the chance yet to play around with our new Screener app, I highly recommend it. It’s an extremely powerful tool for looking at your assumptions and I now find myself keeping multiple tabs of it open throughout the day.
One of the key features that has recently been added looks at team level stats. The functionality found in the Screener app that you won’t find anywhere else is the ability to run a linear regression, so that you can explore the relationship between Year N variables and Year N+1 variables. I was doing this recently so I thought I would just write a quick post on something I found related to pass attempt numbers in Year N+1 seasons.
The key thing that I saw when going through this exercise is that it appears that passing efficiency has an effect on Year N+1 passing numbers. That makes sense on an intuitive level. Teams that pass a lot in one year, but do it inefficiently, may try to dial back the amount they pass in the following year. Meanwhile, teams that pass very little, but do it efficiently, are essentially leaving points on the table if they don’t try to increase their passing numbers in the following season.
Here’s how I did this work in the Screener app:
- Set the global filters to include seasons 2010-2015. Passing tendencies have changed recently so I don’t want old data to drag down the passing attempt numbers I’m looking at.
- Add paATTS and paFPOE to the variables in the Query 1 selection box. The acronyms are for passing attempts, and then a fantasy efficiency metric that I’ve recently updated for use in this app (Fantasy Points Over Expectation). Think of it as only awarding credit when the fantasy points scored are greater than the average for similar down/distance combos. There are other efficiency metrics you could use in the app, like Adjusted Yards/Attempt, but paFPOE actually worked the best.
- Click “Search the Database” which will generate all of the data needed for this exercise.
- Once that’s done I go to the data output part of the app and select the Team tab.
- Then I go to the Linear Regression tab, where I can create simple linear regression models. In this case I’m just going to experiment with adding two variables to the model, namely paATTS and paFPOE, and observe their relationship with year N+1 paATTS. I do this by putting those variables in the explanatory variables box, and selecting paATTS.Qry1.N1 in the Response Variable box. The N1 appended on the variable name tells you that variable is from the Year N+1 season.
- You can do the steps I did, and then experiment with adding and removing the variables in question from the model to see what happens. You’ll end up seeing that the best fit using those two variables, which is still a very loose fit, includes both paATTS and paFPOE. Basically, efficiency does matter for explaining Year N+1 pass attempts.
When you examine the “prediction” that is made in the table below the model summary, you’ll see that adding that efficiency metric to the model increases SEA’s pass attempt “prediction” by about 30 pass attempts (you can see the table output below). A few caveats to offer is that this is a model that hasn’t been tested out of sample, so the “predictions” aren’t true predictions. Those numbers are just the result of fitting the model to the 2015 data. But that’s ok, because I’m not actually going to use the SEA pass attempt number generated.
Remember that I said this model had a very loose fit? Well there are any number of factors not accounted for by this model that I can add on my own (when I do my projections in the staff projection machine), like the fact that the Cowboys are getting back their starting QB. The model doesn’t know that, and it’s likely one reason that the model fit is loose (r-sqaured of .25 when using both variables). But I can account for things like that when I make my own projections. In the case of SEA, while I think that this exercise yielded some very actionable info, I’m going to combine that with other things that I know about SEA in order to come up with my own projection. I know that teams as efficient as the Seahawks have in the past increased passing, but I also know that the Pete Carroll/Darrell Bevell combo has been very run heavy over the past four years. I’m going to pencil in the team for an increase in attempts, but not to the level projected by the simple model I made. I’m going to pencil them in for more like 514 pass attempts in 2016. That’s an increase from 2015, and it will be a high water mark for the team as long as Russell Wilson has started, but it’s also well below what is likely to be the average among teams in 2016.
All I’m doing by completing this exercise is checking some historical data to see if efficiency has mattered in the past, and then incorporating that information with other things I already know about the team. The linear regression functionality in the app, while it’s limited to creating simple models, can still be used to inform the predictions we make for the upcoming fantasy season.
|OFF||SEAS||paATTS.Qry1||paFPOE.Qry1||Predicted Pass Attempts|