Garbage Time Points Matter, Just Not as Much
In this post I’ll do some quick analysis to both show the functionality of our new Screener app, and also answer a commonly asked question in fantasy circles: is garbage time scoring reliable from year to year?
First let me define how I intend to look at that question. I’m just going to call garbage time fantasy scoring any points scored while a team is trailing. That’s a really broad definition of garbage time fantasy scoring and someone else might take a more narrow view of the issue. But that’s how I’m looking at it. Then I’ll call fantasy points scored while winning the game the non-garbage time fantasy scoring.
Now let’s get started with the app. The app will let you set up to three separate queries, or splits. For each query you can display multiple variables. So you could look at receiving yards, touchdowns, etc.
For the first query I’ll get PPR points scored when the scoring margin is zero to 59. Those are the points scored while winning. Then I’ll set the second query to be scoring margin from -59 to -1. Those are the garbage time points. Here’s a screenshot so that you can see what I’m doing in the app.
You might notice that I’ve used all three queries here.1 For the third query I haven’t set any slider values because I want to catch all fantasy points using that query.
I also limit the results to just the WR position using the Display Positions input. Then I click submit and wait literally a second while the app plows through a Play by Play database of over 500,000 rows.
When the data crunching is done the app gives me all of the display variables I’ve asked for, subject to the splits I specified, and it also gives me all of those values in the Year N+1 season. This is probably the killer functionality of the app. You might be used to trying to do analysis in Excel and relying on a convoluted series of VLOOKUPs in order to match data. This app does it all for you. It then labels that data in columns based on which Query the data is coming from.
The other killer functionality of the app is that it lets you do a little linear regression as well. We’re not proposing that this linear regression take the place of serious analysis. But because of the way the app controls work, you could look at the relationship between lots of variables in the time you’re waiting for Excel to load.
To illustrate how the linear regression functionality works you can click over to that tab, which I have a screenshot of below:
You can set your desired explanatory variables and the app will give you an adjusted r-squared, along with the coefficients of each variable. The app also returns the common tests of significance that are often looked for, in the form of t-tests and p-values. In the example above you can see that I’m using the PPR points from Query 1 and Query 2 to explain the N+1 PPR points for Query 3. So basically PPR points while winning and losing are used to explain all PPR points in year N+1. Both explanatory variables pass the common tests of significance, and you can also see that the PPR points scored while winning have a larger coefficient. They’re roughly worth about 1.5 times the points scored while losing in terms of explaining Year N+1 fantasy points.
The model is then applied to 2015 data so that you can see how that would impact predicted values based on the 2015 season. Note that while the last column in the table is called “predicted” that just means that it’s predicted based on this model. Because the model hasn’t been validated on out of sample data, it’s not a true prediction. But this is a good place to start. Based on this exercise we can say in the past it’s been the case that points scored while winning have been about 1.5X as valuable as points scored while losing, when it comes to explaining year N+1 fantasy points. This app is not meant to be a one stop shop for analysis, so if you wanted to take it further you could work on validating the model with out of sample observations. I’m also going to think about ways that I could introduce the out of sample validation to the app so that the value in the predicted column is a true prediction.
Some other notes on the app:
- There’s also a graphs tab that let’s you create scatterplots and do some labeling on the plots.
- Both the screen results tab and the regression tab have separate data download buttons.
- You can also use global filters to only look at rookies, or only players above a certain weight, or only within a certain draft pick range.
- the third query only displays if you’re already using the 2nd query (back)