Tuesday, November 1, 2016

Guess the lines - Machine vs Men

I spend a lot of time in my car these days, thanks to a commute that takes me through the worst stretches of LA's 105 and 110 freeways. To fill that idle time, I listen to a variety of podcasts (and a Waze app that likes to talk over them at the worst possible times). A staple of my regular pilgrimage between Pasadena and El Segundo is the "Guess the Lines" podcast from The Ringer, in which Bill Simmons and "Cousin Sal" engage in a friendly prediction contest while previewing the week's upcoming NFL matchups. The goal of the contest is to see who is better at predicting how Vegas will set the line for each game.

I make my own guesses as I listen to the podcast, but it turns out my Vegas intuition is pretty poor. However, while I'm not a natural bookmaker, I do have a ranking system that was explicitly designed for this purpose. My betting market rankings for the NFL (and the NBA, WNBA, MLB, College Football, and College Basketball) are trained and calibrated on the Vegas point spreads, which I use to reverse engineer an implied power ranking.

So, I am going to put my ranking system to the test, and see how it fares against the instincts of Simmons and Cousin Sal, starting with week 10 of the NFL season.

Why Week 10 and not Week 9?

In contrast to standard stat-based rankings, which can only look backwards, my Vegas rankings use data from games that have yet to be played. The rankings as of today have already adjusted for the week 9 lines. Turning around and using those rankings to predict week 9 lines would be a bit of a cheat. I need to make sure my test is properly "out of sample". 

Opening lines usually come out a good week and a half in advance, and my rankings update as soon as they are out. This makes Tuesday the last day to do a proper test of my rankings for week 10. This is somewhat of a handicap for my model, since Bill and Sal can factor in the results of week 9 as well as any injury/personnel news before making their guesses next week.

How the Rankings Work

There are a lot of numbers to be found in my NFL rankings table (and sparklines too, so many sparklines), but the key metric of concern here is Generic Points Favored, or GPF. It is what you would expect a team to be favored by against a league average opponent on a neutral field. To predict the Vegas line for any matchup, you take the difference between the two teams' GPF and then add 2.5 points for home field advantage. 

For example, the Seahawks' GPF is 4.3 points and the Patriots' GPF is 8.8 points. This means my model predicts the Patriots to be a touchdown favorite when Seattle comes to town on week 10: 8.8 - 4.3 + 2.5 = 7 points. The table below runs through the same calculation for all week 10 games. I have summarized the results in the table below (aGPF is the away team's GPF, hGPF is for the home team). I will update the table sometime next week after Bill and Sal have made their guesses and then tally the results.


gameaGPFhGPFmy linebill's linesal's lineactual
CLE @ BAL-7.2-0.3-9.5
CHI @ TB-4.1-2.5-4
SF @ ARZ-6.63.1-12
MIA @ SD-2.30.8-5.5
ATL @ PHI2.71.1-1
HOU @ JAC-1.9-4.0-0.5
LA @ NYJ-2.9-3.0-2.5
KC @ CAR2.72.8-2.5
SEA @ NE4.38.8-7
MIN @ WAS2.4-1.1+1
DEN @ NO3.5-0.5+1.5
DAL @ PIT2.51.6-1.5
GB @ TEN2.3-2.2+2
CIN @ NYG1.50.8-2

1 comment:

  1. This is great. My favorite pod and my favorite rankings come together!

    ReplyDelete