Using Machine Learning to Create 2016 QB Projections
perfect fantasy draft

In my last article I identified the five types of quarterbacks and how the career arcs for each tend to play out. In doing so, I also created a machine learning model that gives 2016 QB projections on a per-game basis for each QB.

Make sure to give the quarterback types article a read so you have some background on the nuances behind the projections later in this article.

Model Recap

The model uses the following inputs

  • Draft position (DRAFT)
  • Age (AGE)
  • PPR points in year N (PPR)1
  • Interception Rate (paINTRT)
  • Touchdown Rate (paTDRT)
  • Average Margin (AVGMGN)
  • Rushing attempt market share (ruATTMS)
  • Rushing fantasy points over expectation per attempt (ruFPOEPA)

With these variables I built a machine learning model called an artificial neural network to predict points per game the following year, so long as the QB played eight games in season N and season N+1. I used a technique called k-fold cross-validation (I used 10 folds) to show that the neural network model is useful for making predictions. The R-squared and root-mean-square-error values for the training and cross-validated results are listed below.

Training Cross-Validation
R-Squared 0.634 0.624
RMSE 2.600 2.667

Now that we have a cross-validated model, we can use it to make predictions for the 2016 season. I have included all QBs who played in at least eight games in 2015 (so no Tony Romo) but I did include Andrew Luck because he played in seven games. We’ll have to take his prediction with a grain of salt because of the sample size and the fact that he played hurt in some of them.

2016 QB Projections

Here are the points per game (PPG) predictions for 2016, along with each QB’s 2015 age, 2015 PPG, 2015 games played, and the cluster they ended up in (to see what each cluster is, refer to the article linked at top).

Cam Newton2627.216525.74
Tom Brady3824.816424.74
Russell Wilson2724.216523.87
Andy Dalton2821.613323.75
Aaron Rodgers3221.316323.46
Matthew Stafford2721.116323.36
Andrew Luck2621.47522.93
Drew Brees3623.915322.53
Philip Rivers3420.816321.07
Matt Ryan301816320.36
Marcus Mariota2220.412520.22
Ben Roethlisberger332112320.21
Eli Manning3421.216320.17
Josh McCown3620.28119.76
Tyrod Taylor2621.614519.69
Jameis Winston2119.916518.79
Carson Palmer3622.516318.35
Ryan Fitzpatrick3320.516118.32
Jay Cutler3218.115117.92
Blake Bortles242316117.59
Blaine Gabbert26198217.33
Alex Smith3118.916116.75
Ryan Tannehill271916116.64
Sam Bradford281714216.62
Teddy Bridgewater2314.816116.59
Brock Osweiler2517.38116.29
Kirk Cousins2721.116316.18
Joe Flacco3019.210215.41
Brian Hoyer3017.911414.6
Derek Carr2419.716114.24
Johnny Manziel2312.89113.44
Nick Foles2610.811312.1
Matt Hasselbeck40148111.54
Matt Cassel339.78110.87
Colin Kaepernick2814959.49

I like to use the model to find gaps in value and not to nit-pick the individual placings. As such, here are some observations:

  • Tom Brady doesn’t project to regress any on a per-game basis. He is suspended for four games, but being an elite QB from the Kurt Warner cluster means he’s still near his peak age, so maybe his days of fantasy relevance being over isn’t actually the case.
  • Andy Dalton is in the prime of his career, and the model likes him to finish as a QB1. I personally don’t believe he finishes fourth overall on a per-game basis, but he’s a QB I’m definitely targeting this year at his current ADP of early 13th round, as the 19th QB off the board.
  • Holy Blake Bortles regression, Batman! While Justin Winn’s projections put Bortles at 20.7 points per game, I do think this shows that it is possible Bortles comes down to Earth a bit.
  • The model believes young Daunte Culpepper cluster QBs Marcus Mariota and Jameis Winston have differing outlooks. It gives Mariota a similar per game result as last year, while it views Winston on the low end of this cluster of QBs, where the career arc slowly declines from the start. That certainly gels with the idea that Winston could be overpriced right now.
  • It appears Carson Palmer may have peaked and is in decline. The model doesn’t view him as an elite QB from the Peyton Manning tier, meaning his peak age is likely behind him.
  • Matt Ryan could have a bounce-back year and end up a borderline QB1. He’s currently being drafted as the QB20 in MFL10s.
  • I, like Brian Malone and Neil Dutton, thought that Ryan Tannehill was a value — but according to this model, maybe he isn’t. I personally believe this is one the model is going to miss. Even machines can’t always predict everything perfectly.

What else stands out to you?

  1. Yes, I used PPR for QBs. A few QBs have caught passes, including the occasional TD. Kordell Stewart caught several TD passes in his career.  (back)


Co-Owner and Editor-in-Chief at RotoViz. Mathematics Ph.D. 3x qualifier for the DraftKings NASCAR Main Event.
What we do

Sign-up today for our free Premium Email subscription!

© 2019 RotoViz. All rights Reserved.