Welcome to Football Season 2017!
After another offseason that has seemed to fly by, college football has already begun, and the NFL kicks off tonight. We’ve been a little haphazard in getting everything up on the site for the season (I just started a yearlong trip abroad), but our numbers should be as solid as ever.
Every offseason, we look into improving our numbers, trying to find workarounds to tough modeling problems (separating QB value from team offense in-season, for example), and this season was no different. The biggest “new” thing I looked at was quantifying the value of coaches–both head coaches and coordinators–in a Bayesian way. In the NFL, I worried that our QB priors, by virtue of using many seasons of past data, ended up conflating QB skill with organizational elements such as coaching and good personnel management. After all, some organizations have been consistently better than others over the years, and that (obviously) doesn’t just reflect QB play. Just to be clear, our priors do factor in more than QB play, but we don’t look back further than a year for our team offense and team defense priors. Spoiler alert: while there are some coaching effects, it is very difficult (and costly in terms of degrees of freedom!) to segregate head coach, coordinators, and upper management effects. We found coaching effects to be (predictably) more meaningful in the college game, since the coach plays a much larger role organizationally. Additionally, I’ve personally done extensive work on home field advantage, which will not impact our power ratings but will effect our game predictions.
Last season, our NFL picks struggled. It was our first losing year since we’ve been doing this. Was it variance, the market catching up, just a strange year? I don’t think anyone can tell you with certainty (but I’m totally blaming ESPN). Looking at the predictive value of our NFL numbers relative to the closing market in all games, our model showed little to no value last year, and was similarly weak in 2015. But building narratives around small samples goes against everything we believe in. In 2014, the closing line showed no predictive power all relative to our number. There is a lot of volatility looking at individual years, and we’re not looking at a truly independent sample. tl;dr: I still think our model (especially with improvements) has value relative to betting markets, but you can decide that for yourselves.
Last year wasn’t ALL bad though. Our college football numbers were excellent. We stopped releasing picks a few years ago after we starting moving the market too much for my liking, but when you compare our numbers to other quantitative systems out there, using data from thepredictiontracker.com, we would have been the 2nd best system out of 64 (full disclosure: we didn’t submit numbers to any third-party trackers so you’ll have to take us at our word).
As always, all our content is free. We are not trying to sell you anything. We both are perfectly happy with our current gigs and are not looking to make a quick buck selling picks. The downside to this is that I’m going to take care of my own interests first, so to speak, before I release picks, so you may get the same value that I did.
Here’s what we’re planning to provide on this site this year:
- NFL (all) and college football (top 35) power ratings, released midweek
- Futures numbers from our season/playoff simulation results for both NFL & CFB
- Weekly game grades for both NFL & CFB
- NFL picks against the spread, likely posted on Thursday mornings (European time)
Here’s where you can find us elsewhere:
- Jeff Ma and I are launching a
blogpodcast called Bet The Process, where we’ll discuss topics in the world of sports betting, make fun of touts, dispel false narratives, and talk about where we find value each week…and have fun doing it.
- Cade co-hosts the Wharton Moneyball radio show each Wednesday morning on SiriusXM’s Wharton Business Radio.
- The Wall Street Journal will be hosting our college football simulation results.