Forecasting running back results is an incredibly difficult business. They’re extremely usage dependent, which means that they are subject to the changes in game plan that a team might have either on the season level or at the individual game level. Running backs take a beating, which means that they’re sensitive to injuries. To make matters even more difficult, the NFL is moving steadily towards become a timeshare running back league. All of this is to say that the the usual difficulties that we might have predicting the future become even more pronounced when trying to project running back results.
Nevertheless, we can still use similarity scores in order to project running back results. Similarity scores will do generally a good job of capturing the way that variables might interact. A variable like total carries could have a different impact on a running back depending on that back’s age or size. Similarity scores recognize the difficulty of forecasting with a bunch of variables that might not be linear and might interact with each other, making the the simple assumption that if you look at a similar group, the results will be similar. If you take the time to backtest this assumption, you’ll find that similarity scores do improve on the predictive ability of a simple linear regression.
The app below makes it possible to create custom similarity scores for individual running backs. Maybe you want to forecast C.J. Spiller by throwing out the games in which he didn’t accumulate 10 carries, or maybe you want to look at Vick Ballard after he became the lead back in week 7. You can do either of these things and a lot more by just checking or unchecking boxes to include those games in the calculation.
I’ve also changed up the plot a little bit and made it more intuitive I think. If you click on the plot tab, it now shows each player/season for the comparables and the % improvement or decline in their fantasy numbers.
One sort of pre-emptive note is that similarity scores are going to have a very difficult time addressing extreme outliers. I’m primarily talking about Adrian Peterson’s season. It’s an outlier on a number of levels. Similarity scores use the past to forecast the future, so when a reasonable approximation can’t be found in the past, it’s going to go to the next closest thing, which might not be that close. I do think it’s reasonable to expect some dropoff from AP next year, but I’m not sure that the similarity scores properly account for what we might expect from him next year.
If you play around with the app a little bit, you’ll probably notice that there is a certain type of back that ends up looking good in this exercise and that is the young pass catching back. Go nuts with the app and have some fun.