Game Probabilities - Conference Championships

Weekly game probabilities are available now at the nytimes.com Fifth Down. This week I also look at how different the Jets are from the other remaining teams.

  • Spread The Love
  • Digg This Post
  • Tweet This Post
  • Stumble This Post
  • Submit This Post To Delicious
  • Submit This Post To Reddit
  • Submit This Post To Mixx

29 Responses to “Game Probabilities - Conference Championships”

  1. Ryan says:

    interesting. i arrived at the same thing for the jets/colts game just using the spread and over/under, which approximates a score of colts 23.3-jets 15.8, giving the jets about a 27% chance to win.

    the saints-vikings spread is lower, however, with a much higher O/U. i'm surprised to see that big a difference in your probabilites to the vegas line, which gives the saints about a 58% chance to win (28.4-24.8), or about the value of home field advantage.

    i do think the saints will win pretty handily, but they both looked damn good last week. should be fun.

  2. Ryan says:

    side note: if you're on board with the vikes-saints O/U at about 50-53, that puts your spread at about -13.

  3. Anonymous says:

    "A quick note on the Vikings-Saints probabilities: Many readers will wonder why my system gives the Vikings such a low chance of winning. Very simply, the Vikings are a fine team, but they’ve faced a very week schedule this year. Despite their convincing win against the Cowboys last week, which is factored in, they face an uphill battle on Sunday. The bottom line is that they’re an above average team that won a lot of games against weaker opponents, and this week they’re on the road against an even better team."


    You do realize that the Saints had an easier schedule than the Vikings this year, right? The Vikings' strength of schedule was .441 and the Saints' was .425. The Vikings also went 4-1 in the regular season against 2009 playoff teams, while the Saints were 3-1.

    I was surprised to see such flawed reasoning to explain the Vikings' supposed 21% chance of winning.

  4. Brian Burke says:

    No they didn't. The expected winning percentage for NO was 0.50, and for MIN it was 0.40. MIN was actually last in the league.

    If you just look at raw win %, you'll get a very distorted picture. For example, teams within the weak NFC West get to play each other twice, which buffets their win%. Had they played a more even schedule, their true win % would have been significantly lower. Teams whose own division get to play weak divisions, like MIN, are going to have inflated SOS numbers.

    So you have to look at 2nd order SOS measures (or 3rd order, or nth order), and you also have to do some other things to get a valid measure of SOS.

    By the way, Sagarin also rates MIN with the 32nd toughest schedule in the league, so I'm very comfortable with my own ratings.

  5. Anonymous says:

    If your numbers show the Vikings having a 20% chance of winning the game in New Orleans, I'm not sure you should be comfortable with them. That's one of those results that really should make you question your methodology, because it's just a laughably bad result.

    Don't get me wrong, I'd probably pick the Saints to win. But the Vikings are just not a 4:1 dog in this game.

  6. Anonymous says:

    Brian,
    I know you factor in home away in your statistical model, but I am curious is there any way to see how much of a difference field type/surface makes for different teams. For instance, the Vikings have looked (seemed) like a much different team on natural grass surfaces than on the carpet. They seem to be built for speed and lose some of that advantage on natural grass, especially grass in poor condition. I know the Vikings this season are likely to small a sample size for any significant correlations to be made, but it would be interesting to see on a larger sample size if there are teams who do better with certain field types than others. Or if any such discrepancies would fall within normal random distributions.

  7. Brian Burke says:

    I wouldn't be so fast to say that. Cardinals fans were all over me last week for the same thing. But one game doesn't prove me right, and this week's Vikings game won't either.

    This model is very well calibrated, although this year it seems to be a few points overconfident. Case in point, in the 42 games where the favorite is predicted 71%-80%, it was right 69% of the time. So 20% might be too low. It might be more like 25 or 26%. Laughably bad result?

    But even 20% is not zero. There is a fair chance we'll see the Vikings win.

  8. James says:

    The 5th down readers just don't get it, do they? "The Vikings and Jets won, so everything you say must be wrong! And yes, we only look up when you were wrong and harp on that, and completely ignore when you were right!"

  9. Brian Burke says:

    Good point. It's true that dome teams playing outdoors in the cold are at a severe disadvantage. A small part of the league-wide HFA comes from that effect alone.

    In this game the Vikings are in a dome, so I don't expect it to be a very big factor, maybe a percentage point or two less than league-average. In other words, HFA is worth +6% points rather than +7 or 8% points.

  10. Brian Burke says:

    James-They're largely Jets fans, so they have emotional attachments. They are essentially searching for validation for their hopes, and a post like mine is not what they want to read. (I confess I do the same thing for my own team.)

    I was really nervous this year about the predictions because they're featured in the Times. I thought the model would finally have an off year and I'd be humiliated. But luckily it had its best year so far in terms of accuracy. I'm not sure what else would convince skeptics.

    I have to say I do respect their skepticism. They're basically saying, "What makes you so smart, mister?" That's a good thing. Also, keep in mind that there is a bias in comments on sites like mine. People are far more compelled to speak out when they disagree than when they agree.

    I try not to respond in the comments at the Times. I'll answer questions, but only rarely defend against criticism. I figure I've had my say. (On my own site I will engage the morons who make misleading or ad hominem attacks. It's kind of a sport to me.)

  11. Anonymous says:

    Thanks Brina (re:Dome Teams Playing outdoors). I guess my question is that Leaguewide it may be a pretty small effect, but would there be a way to determine if some teams are more affected by it than others. i.e Team a--playing on grass away makes them 15-% more likely to lose than on turf away, whereas Team B is equally unlikely to lose away whether on grass or on turf. I know these numbers are likely way to high, just wondering how evenly distributed effects like this are.

    Thanks,

    Erik

  12. Anonymous says:

    Sorry Brian,

    Didn't mean to typo your name. Really enjoy the site.

    Erik

  13. Anonymous says:

    If I'm not mistaken the model is about 70% accurate (excluding week 17) this year. That is better than probably any media expert has been this year. Its even better than most models at thepredictiontracker.com

  14. Brian Burke says:

    About 71 or 72%. If you count week 17 it's closer to 70%. It's better than all the other models, in fact, including the Vegas lines. It has been now for 4 years running, although my first year these were unpublished, so I have to really say 3. Also, to be fair, it only predicts games starting in week 4, because it needs some data to crunch. I'm also 1 game behind the consensus in the playoffs so far.

    I realize it's unpleasant to toot my own horn like this, but considering some of the criticism lately, my hand is forced!

  15. Anonymous says:

    I would include the accuracy in you NYTimes posting for now on and include it on the main page.

  16. Anonymous says:

    Brian,

    I appreciate all that you do.
    Look forward to reading your
    insights each week. How these
    fans can get so emotional over
    probabilities is beyond me.

    Zeke Tomber

  17. Anonymous says:

    Brian,
    I just looked over all your previous HFA Posts. You suggest that the HFA for dome at cold is .88. My question is as follows, does this impact the input for the Effiecency stats. Should the stats from such games be somehow normlized to take that extra 38% back out?

    i.e. 2 fictional exactly equal dome teams, Team A has 3 dome at cold games, Team B does not. Team A will have presumable lower efficiency stats for having the Dome at Cold games (The extreme HFA against would likely result in worse production), but it would have no effect on a Team A / Team B. Is this acounted for in the model, or is it a miniscule enough occurance not to matter.

    Bottom line does the number of Dome at cold games a team has affect their WP in the model for non Dome at cold games. Should it? Does it matter?

    Thanks,

    Erik

  18. conquistador says:

    Brian, as a math major myself, I enjoy your posts (probabilities), but do think you have made a mistake on Minnesota a little bit. I think the 80% chance given to the Saints should call into question something regarding your model, though I know none will be perfect. I think you acknowledged that your ratings will be high. I suspect (perhaps you have the data) but what about the teams you give a 80-90% chance of winning? it is probably in the 70's. I see our friends at footballoutsiders (sorry for the mention) have the saints with just a 51% chance of winning Of course, we will never know what is the THEORETICAL probability, but it almost certainly has to be between (though I lean towards your number a little more) perhaps 70%.

    I suppose another way to look at things is that an 80% favorite translate to about a 10 point favorite, which means the oddsmakers are off by 55 points. I have followed spots and betting for the better part of 40 years and while that is very possible during the regular season, I have found the playoff lines quite tight (again, nobody can ever verify the 'true' probability)

  19. Jim Glass says:

    At Pro Football Reference.com they have a thread running on why the #1 seeds get to the Super Bowl so infrequently.

    One of my favorite articles that you did here was the one showing that more than 50% of NFL game outcomes are determined by random chance.

    I'd expect that this percentage is higher in the playoffs than during the regular season, as a during the regular season there are a lot of "good team v bad team" matchups, while in the playoffs the teams are all withing a much narrower quality band.

    Would you have a back-of-the-envelope way of estimating that?

    I mean, commentators never mention this concept -- if anything, they "hero-ize" the top teams to make it seem like they should be more likely to win in the playoffs. Championship character revealing itself through "clutch play" and all.

    If you have another article coming up in the Times, this could be a good angle: "If your team wins the Super Bowl what does it mean? That you were lucky or good? Mostly lucky!"

    That would get some conversation going in the comments over there, I bet.

  20. Unknown says:

    Brian, you wrote that you have beaten Vegas predictions. Can you give a summary on why you think you have?

  21. Tim Brady says:

    So I just spent an hour or so analyzing your model and I think it is pretty clear that the model was overconfident this year. I analyzed the 192 games you predicted between week 4 and week 16.

    My first pass graph looked pretty good for the model: http://web.mit.edu/tfbrady/football/modelCheck.png plots the model predictions (win%) and the actual win% of the teams, binned into 5% units. No obvious trend for the actual win% to be shallower than predicted by the model, although certainly a hint of that.

    But if you look at it more precisely, using the log likelihood of the data under the model, then it is clear that making the model predictions less extreme improves the model. Taking the sum of the log likelihoods for each game (sum(log(winProbabilityOfWinningTeam))), closer to 0 is better (if you predicted all games with certainty 1 and were always right you'd have a 0), and closer to -Inf is worse (if you predicted all games with a certainty of 1 and were always wrong you'd have -Inf). If you predicted 0.5 for each game, you'd have a score of -133.1. The model as is has a log likelihood of -116.1. So the model is doing far better than just saying all games are toss ups.

    But you can also look at what the log probability of the model is if you make it uniformly more or less conservative. The way I did this is to take the model's predictions and move them a percentage closer to or further from 0.5. The results are plotted here: http://web.mit.edu/tfbrady/football/modelTooExtreme.png As you can see, if you make all of the model predictions 1.2 times less extreme (e.g., make 75% into 70.1%, 90% into 83%, etc), the model achieves its best log probability (closest to 0). So I think that it is pretty clear that in general the model is overconfident in its predictions (they should be 20% closer to 0.50).

    I'm curious what you make of this.

  22. Brian Burke says:

    Yup. I agree. It's about an average of 5 % points overconfident this season. I'm a little surprised because this year I increased the strength of the regression of the early season team stats. If anything I thought it would be underconfident.

  23. Brian Burke says:

    Each week I only really pay attention to the games where I disagree with the spread. If my favorite is different than the opening line's (I usually look at USA Today's lines), I'll note if the model was correct or not. I think I'm up at least 2, maybe 3 games this year--not enough to make money if you're a gambler. I only count weeks 4 through 16 each season.

  24. Tim Brady says:

    The overconfidence interacts with home field advantage pretty severely. When you predict the home team will win, you do very well (in fact, the model is quite underconfident). However, when you predict the away team will win you are pretty much randomly guessing. Probably means you are underweighting home field in some situations.

    E.g., http://web.mit.edu/tfbrady/winAndHomefield.png

  25. Tim Brady says:

    By the way I love the site -- I realize I just jumped in and starting critiquing the model, but that is just because I think it is really interesting and fun to play with (and I don't feel like doing real work right now!). I hope you keep it up for a long time! :-)

  26. Brian Burke says:

    Thanks for doing this by the way.

    I'm really surprised by that. I thought it would be the opposite. My HFA coefficient is the highest it's been this year. It gets up over 9% (59 vs 41) when teams are evenly matched.

    Maybe I need to interact HFA with the other variables. Perhaps increase it just for home underdogs?

    Home field can vary pretty wildly from year to year. About three years ago home teams only won 53 or 54%. Then another year it was up well over 60%. Maybe 2009 was just a very high year. I can't see increasing the HFA coefficient any higher than it already is.

  27. Anonymous says:

    Does your HFA coefficient adjust for dome vs. Cold games. Say for example if the Colts were going to the Jets, would the Jets HFA be higher in your model because it's a cold vs. Dome game?

  28. Brian Burke says:

    No. I factor that in inside my own head, but the model doesn't automatically consider it.

  29. Anonymous says:

    2-0 for this week not too bad.

Leave a Reply

Note: Only a member of this blog may post a comment.