Game probabilities for week 11 are up at the New York Times' Fifth Down.
This week I take a look at alternate estimates for the BAL-PIT game due to Roethlisberger's injury.
- Home
- predictions
- Weekly Game Probabilities
Weekly Game Probabilities
By
Brian Burke
published on 11/15/2012
in
predictions
Subscribe to:
Post Comments (Atom)
How does one map from 'win probability' to 'predicted point spread' ? Would be interesting to look at each game versus Vegas point spreads.
Once again this season I am tracking the accuracy of Brian's model's prediction compared to the predictions derived from the averages of the six dozen or so computer ratings tracked by ThePredictionTracker (www.thepredictiontracker.com/blog). I posted last year's results in the comments of this link: http://www.advancednflstats.com/2011/12/weekly-game-probabilities_29.html
As was the case last year, Brian's model is slightly worse at predicting game outcomes than an average computer model. Again, one obvious issue seems to be that the model's home field adjustment is too strong (at least relative to the other criteria used). Brian, you acknowledged this problem last year. Any chance it could be fixed?
The games where Brian's model predicted a different outcome than ThePredictionTracker averages are listed below with Brian's model's picks marked with an asterisk.
Week 4: NE at Buf* (L), Min* at Det (W), Cin at Jac* (L)
Week 5: Phi* at Pit (L), Sea at Car* (L), Den* at NE (L)
Week 6: Pit at Ten* (W), Cin at Cle* (W), Den* at SD (W)
Week 7: GB at StL* (L), Dal at Car* (L)
Week 8: SD at Cle* (W), Sea at Det* (W), NE vs StL* at London (L), Mia* at NYJ (W), Atl at Phi* (L), NYG at Dal* (L)
Week 9: Bal at Cle* (L), Car* at Was (W), TB at Oak* (L), Phi* at NO (L)
Week 10: Hou* at Chi (W)
Home: 4-9, Away: 4-3, Neutral: 0-1, Overall 9-13
Correction: Home: 4-9, Away: 5-3, Neutral: 0-1, Overall 9-13
What makes the Steelers good on offense is their ability to convert on 3rd down. Ben is the biggest reason for their high 3rd down conversion rate because of his ability to avoid sacks and keep plays alive. I think the absence of Big Ben is going to hurt the Steelers more than the model is predicting. Unless Mike Wallace is catching 60 yard bombs all day I do not know how this offense is going to put together long enough drives to score points. I see lots of 3 and outs in their future.
You should start calculating probabilities of covering the spreads... not just W-L!!
Anonymous:
I found the following formula works pretty well:
Convert the probability to an odds:
odds = (1-prob)/prob
Take the natural log of the odds and then multiply by 7.
So, for the Atlanta-Arizona matchup, Atlanta's probability is 0.70, which translates to about a 6 point spread in favor of Atlanta. The actual spread, last time I checked, was 9.5 points in favor of Atlanta.
AES your evaluation model is flawed.
For example my system and Brians both had Giants ! 53% @ Cin while yr computers average was ~74%
Using W/L is an inaccurate measure just like using it to rank teams.
We need to measure Brians prediction from actual outcome.
I calculate a true score for each game in this case ~80% Cin when u factor in garbage time turnovers
this means Brians system and mine is ~40% closer to final than computers average. Over a large sample size as long as we are better than comp. average or Vegas line (what ever yr comparison) then model is a winner!
The best way to judge Brian's model is to see is after the full season:
Teams that had Win prob of 50-59% won what percentage of their games?
Teams that had Win prob of 60-69% won what percentage of their games?
Teams that had Win prob of 70-79% won what percentage of their games?
etc.
Anything else is using the model for something it isn't meant to be used for.
That being said. I do clearly recall Brian admitting last season that he was weighing home field advantage too strongly.
When I tried to bring it up earlier this season I didn't get any response. Somewhat curious.
(Higher win probability - 0.5) * -29.65 gives a decent approximation of the point spread.
Buffalo at 52% would be a -0.6 point favorite. So at -2.5 the choice would have been MIA.
By this method since 2007 Brian's system is 618-589 (51.2%) ATS.
But Brian's model is not designed to beat the spread.
He doesn't predict scores. No the best gauge is to measure his error from true score.
anonymous is right.
last weeks average win probability was 62%, indicating that the most likely outcome of favorites vs underdogs was 9 out of 14 (i.e. the fav should win 9 of those games).
Result: 9 of the fav won, 5 underdogs won. Perfectly consistent with the probability model (though obviously with way too few realizations).
This is just like Nate Silver, who predicted probabilities in all 50 states. The difference is, Silver had the favorite win all 50 times which is statistically unlikely (and hence something stinks there).
"This is just like Nate Silver, who predicted probabilities in all 50 states. The difference is, Silver had the favorite win all 50 times which is statistically unlikely (and hence something stinks there)."
Except results in states were strongly linked whereas individual football games are not at all.