Weekly game probabilities are available now at the nytimes.com Fifth Down.
This week I regurgitate what I wrote earlier on the importance of playoff seeding.
Weekly Game Probabilities
published on 12/29/2011
Subscribe to:
Post Comments (Atom)
I have been tracking the accuracy of Brian's model's predictions this season compared to the predictions derived from the averages of the computer ratings tracked by ThePredictionTracker (see www.thepredictiontracker.com/blog for its game predictions). In the games where Brian's model has predicted an outcome that is different than ThePredictionTracker average, Brian's model is 15-19. So at least this year, Brian's model is slightly less predictive than an average computer model (many of which use only final scores as inputs). The game-by-game details are below. Brian's model's picks are marked with an asterisk. Of these games, the model seems to do worse when it picks the home team, suggesting that perhaps the model's home field adjustment is too strong.
Week 4: Min at *KC (W), NYG at *Ari (L), Atl at *Sea (L), NE at *Oak (L) [1-3]
Week 5: Cin at *Jac (L), SD at *Den (L) [1-5]
Week 6: *Car at Atl (L), Buf at *NYG (W) [2-6]
Week 7: Chi vs *TB at London (L), Was at *Car (W) [3-7]
Week 8: Dal at *Phi (W), NE at *Pit (W) [5-7]
Week 9: Atl at *Ind (L), SF at *Was (L), Cin at *Ten (L) [5-10]
Week 10: *Pit at Cin (W), *NO at Atl (W) [7-10]
Week 11: Jac at *Cle (W), *Oak at Min (W), Sea at *StL (L) [9-11]
Week 12: Chi at *Oak (W), Ari at *StL (L), GB at *Det (L), NE at *Phi (L) [10-14]
Week 13: Den at *Min (L), Ten at *Buf (L) [10-16]
Week 14: *Phi at Mia (W), *Chi at Den (L) [11-17]
Week 15: Mia at *Buf (L), NYJ at *Phi (W), *Pit at SF (L) [12-19]
Week 16: *Oak at KC (W), Den at *Buf (W), *NYG at NYJ (W) [15-19]
"In the games where Brian's model has predicted an outcome that is different than ThePredictionTracker average"
Different by what measure?
I've come to the same conclusion. But it's not that the home field is too strong, it's that everything else is too weak. I'm regressing the team stats too much, but not sure how or exactly where I need to turn the dials just yet.
At least we had a very good week last week. +4 games.
"But it's not that the home field is too strong, it's that everything else is too weak. I'm regressing the team stats too much"
Can you explain this a bit more Brian for those of us with a shallow-at-best understanding of the nuts and bolts of statistical analysis
Hi Brian,
I think Houston has to play to win at home vs Tennessee. They can't afford to go into the playoffs with 3 straight losses. Play the starters and get your momentum back.
Minnesota played well last week and Chicago is done. I like Minnesota at home even with Peterson out.
I expect the other top teams to play hard in an effort to obtain a first round bye. And, of course, I'm holding out hope that my Raiders can come through and make the playoffs. They will need some help, though.
I'm tied for 8th out of 200 in my football pool going into the last weekend. Wish me luck!
Keith Rosenkranz
Brian, may I suggest an idea for improving your regression-to-the-mean methodology?
If you were to create each team, prior to the first game of the year, as an idealised average team, average in every respect that the model considers, and then take the exponential average over time with these initial ratings, then you may find your model can be both tuned to avoid over-regressing, and tuned to pick up on changes in team strength due to injury or otherwise.
It would then be a simple case of minimising the error that this would incur over your historical data by varying the exponential weighting to maximise predictivity. Whilst the way you consider strength of schedule would need to be updated to account for this, it is possible to neatly entwine the two concepts.