Weekly Game Probabilities

Weekly game probabilities are available now at the nytimes.com Fifth Down. This week's (triple guarantee lockdown put-it-in-the-bank) upset special is the Panthers-Falcons matchup.

8 Responses to “Weekly Game Probabilities”

1. Xandu says:

Might I ask what factors aside from GWP factor into game probabilities? For example, looking at something like the Dallas-New England matchup, it would seem from team rankings that Dallas (0.69GWP, #2 offense, #5 defense) would actually be the statistical favorite over New England (0.59GWP, #1 offense, #29 defense). Does home field advantage really hold that much sway in the weekly game probabilities? Injuries? Number of days off? Type of stadium? Average player salary?

2. Anonymous says:

Homefield advantage was strengthened in the model this year according to Brian.

Brian another question here. You said in a previous comment that you don't factor in the boost teams get after coming off a bye week, but that there is definitely a statistically significant boost there. My question is: why not include it? Seems like it would be easy enough.

3. Brian Burke says:

The honest reason is I haven't gotten around to it! It's very small, I think it might be 2% advantage.

http://community.advancednflstats.com/2009/10/bye-weeks.html

Dennis O'Regan did some good work at the Community site, and according to his analysis, it's more complicated than a simple +2% advantage. The advantage appears to primarily go to road teams playing after a bye.

4. Brian Burke says:

HFA is only fractionally higher this season. The difference is very slight. The thing to remember is that HFA is variable based on relative team strength. The closer the two opponents are in ability, the more determinative the effect. The effect remains the same, it's just more decisive.

(Say Notre Dame is playing at Caltech, and they'd have a 99% chance of winning. If they played at ND instead, they couldn't have a 106% chance of winning.)

Logit regression captures this effect very well.

5. Anonymous says:

Correction Brian, Notre Dame doesn't have a chance at a 99% win probability vs. anyone.

6. Brian Burke says:

That's true! They came pretty close against my alma matter for quite a few years though. (By the way, I picked Caltech because it doesn't have a football team.)

7. Jonathan says:

Why would you expect an undefeated juggernaut to have a 99% chance of losing?

8. Anonymous says:

I have been tracking the game probability model by making (hypothetical) bets against the moneylines. For example, this week the model predicts that Houston has a 43% chance of winning. The moneyline is 340, so if Houston has a greater than 22.73% (100/100+340) of winning, you will expect to make money betting on Houston.

I am testing the game probability model under three systems. Each assumes I start with \$100. Under System 1, I bet \$1 on each game as described above. Under Systems 2 and 3, I vary the bet between \$1 and \$6 based on the amount by which the model indicates the moneyline is off. That is, if the moneyline is 100 (i.e., 50/50) and the model says a team is a 54% favorite, the bet is small, but if the model says that team is a 75% favorite, the bet is larger. If the model's probabilities are accurate, you would be better off using something like System 2 or 3 to maximize your advantage.

Results through two weeks:

System 1 (all \$1 bets):

Expected Result (based on EV of wagers) - \$110.87 (gain of \$10.87)
Actual Result - \$91.52 (loss of \$8.48)

System 2:

Expected Result - \$134.95 (gain of \$34.95)
Actual Result - \$82.78 (loss of \$17.22)

System 3:

Expected Result - \$133.91 (gain of \$33.91)
Actual Result - \$86.74 (loss of \$13.26)

Not a great start, but too early to draw any conclusions.