What the FPS Model says about Week 12 in the NFL
For the past 6 years we have been fine tuning a week-to-week model that accurately predicts NFL outcomes at a profitable level. Although we call this FPS 2.0, it may as well be called version 1,294. Our hope is do do a deep dive into the what, how and why of the model next week, but for now we wanted to release this week’s output before tomorrow. Here is how Week 12 shaped up:
*One important note: As you probably noticed the model has two outputs for every game. The first, “FPS” is simply using the raw DAVE as a base (see below for more explanation). The second “FPS Var” uses the same DAVE base team metric, but adjusts for variance throughout the season (almost as if using standard deviation to help hone in on the most likely outcome of THIS ONE game). It’s fair to interpret the first number as the absolute outcome (both team’s playing their best ball), while the second number is the tighter, more likely outcome (using how a team CONSISTENTLY performs).
The Model’s Key Elements are net values between two team’s…
- Base line level of (team) productivity, using Football Outsiders flagship data point: DVOA, or DAVE (which is used to combine preseason projections with actual season results in hopes to not skew the data due to low-volume, early season outcomes).
- An adjustment for each coach’s impact on the game. This is a work in progress metric, but much of the early thanks goes to EDJ Sports and their coach rankings based on “Game Winning Chance” a coach has added/taken away based on decisions they have made in aggregate during the season. (i.e. empirical proof Matt Patricia is the worst coach in the league…possibly in the history of the NFL).
- Opponent Turnover Likelihood as explained in an earlier post.
- Injury Adjustment, using a very complex process along with SIS On/Off report that depicts expected points added when a player is on/off the field.
Again, we will deep dive on this next week (particularly how we weight each variable), but we have seen a season long correlation between our model and NFL outcomes at between 35%-77% (with 63% correct selections). Very proud of that number.