Who will stop Italy?
#41
Posted 2007-October-05, 16:06
Somehow that doesn't feel quite right.
#42
Posted 2007-October-06, 05:19
Echognome, on Oct 6 2007, 12:03 AM, said:
I fit a standard linear regression model to the per-match IMP differences
{IMP difference in match i} = {strength of hometeam in match i} - {strength of vistorteam in match i} + epsilon
So this model gets 21 coeffiecients as the strength of USA 2 is per definition zero. Plus a standard deviation of epsilon which turns out to be somewhat more than 24 IMPs. For the purpose of modeling an 8 times longer match I scale the coefficients with a factor 8 and stdev(epsilon) with a factor sqrt(8).
Quote
Tx, I will try that. Also I could try
- shrinking the coeffeicients with Mallow's C.
- there will be correlation between the residual variances of matches played with the same boards, maybe that can be exploited
- some time series analysis could be intersting (some teams may be better in the afternoon than in the evening, some may have started bad and then improved etc.)
Quote
The residuals look reasonably normal, but I should probably x-validate it.
Quote
#43
Posted 2007-October-07, 05:33
SA-Ita: 86.1%
USA1-Aus: 99.1%
Nth-Swe: 88.0%
Nor-Chi: 97.0%
#44
Posted 2007-October-07, 05:55
For example, I would expect the volatility of results in later sets to be much higher if the difference is large going into them, as teams start trying to swing. Which means that I think the chance of italy winning is higher than 14%.
#45
Posted 2007-October-07, 09:36
FrancesHinden, on Oct 7 2007, 06:55 AM, said:
I bet lots of people would come to the same conclusion without any fancy math!
(Yes, I am aware of Italy's big deficit as I am posting this.)
#46
Posted 2007-October-07, 09:54
TimG, on Oct 7 2007, 10:36 AM, said:
FrancesHinden, on Oct 7 2007, 06:55 AM, said:
I bet lots of people would come to the same conclusion without any fancy math!
(Yes, I am aware of Italy's big deficit as I am posting this.)
it's all about the fancy math.
the world would have no sense and life would have no meaning, if there was no fancy math.
#47
Posted 2007-October-07, 10:23
- hrothgar
#48
Posted 2007-October-07, 10:25
Hannie, on Oct 7 2007, 10:23 AM, said:
Yeah why wasn't that an option?
#49
Posted 2007-October-07, 13:33
skaeran, on Oct 5 2007, 01:15 PM, said:
helene_t, on Oct 5 2007, 03:31 PM, said:
[1] "NORWAY 0.848930973373835"
84.9% probability that Italy will beat Norway in a head to head match?
That's ridicolous. They've met in matches like this several times before. Those matches has as a rule been very close. Since 1995 the two teams have met three times at the KO stage of the BB. Norway have won twice, Italy once.
Italy probably are favourites if they meet again, but not by much - maybe 55-60%.
And your probabilitiies are likewise way too high for several other teams too.
Let me just point out that this was not Helene's personal opinion. It is just the outcome of a model that has as input only the IMP scores of the round robin matches.
Of course there may be more relevant information, and the model may not be the best model for this purpose, but I expect that the numbers Helene posted are the correct outcomes of her model, and calling them ridiculous doesn't make much sense.
It's like saying that 1+1=2 is ridiculous because we should be computing 1+3.
- hrothgar
#50
Posted 2007-October-07, 14:01
Hannie, on Oct 7 2007, 09:33 PM, said:
skaeran, on Oct 5 2007, 01:15 PM, said:
helene_t, on Oct 5 2007, 03:31 PM, said:
[1] "NORWAY 0.848930973373835"
84.9% probability that Italy will beat Norway in a head to head match?
That's ridicolous. They've met in matches like this several times before. Those matches has as a rule been very close. Since 1995 the two teams have met three times at the KO stage of the BB. Norway have won twice, Italy once.
Italy probably are favourites if they meet again, but not by much - maybe 55-60%.
And your probabilitiies are likewise way too high for several other teams too.
Let me just point out that this was not Helene's personal opinion. It is just the outcome of a model that has as input only the IMP scores of the round robin matches.
Of course there may be more relevant information, and the model may not be the best model for this purpose, but I expect that the numbers Helene posted are the correct outcomes of her model, and calling them ridiculous doesn't make much sense.
It's like saying that 1+1=2 is ridiculous because we should be computing 1+3.
If so she should have presented her assuptions and how she computed these probabilities and asked the question what's wrong. Since it should be obvious that none of the figures presented could be anywhere near correct.
Harald
#51
Posted 2007-October-07, 14:12
I suspect there is little to choose between the players in the final stages of the BB.
#52
Posted 2007-October-07, 14:34
skaeran, on Oct 7 2007, 10:01 PM, said:
That's what I did. Matt, Richard and Frances suggested some improvements. I agree that my model is probably too simple, although it is not clear in what direction it would be more fruitful to generalize it. Taking all the problems mentioned into account would obviously lead to an overfit. This data set is very small. Therefor any model will either be very crude, or too complex to fit with these data.
I'm sure someone has made much more serious models for bridge events, based on much larger databases.
#53 Guest_Jlall_*
Posted 2007-October-07, 15:07
Halo, on Oct 7 2007, 03:12 PM, said:
This is not true. If you think it's true though we can definitely bet on the outcome at even money.
#55
Posted 2007-October-07, 19:48
whereagles, on Oct 8 2007, 08:36 AM, said:
Anyone who wants to back South Africa to beat Italy with me, despite the current score (48/96 boards played) please form an orderly queue.
nickf
sydney
#56
Posted 2007-October-07, 23:21
#57
Posted 2007-October-08, 02:14
Maybe I could say something about Frances' hypothesis that there will be more swings now that Italy is trailing, but other than that I'm afraid a statistical analysis is not the kind of analysis you would like. Obviously the performance of SA is far too good to be merely based on statistical fluctuations, but whether this is a time thing or an H2H thing I cannot say.
#58
Posted 2007-October-08, 04:01
nickf, on Oct 8 2007, 11:48 AM, said:
nickf
sydney
alright, all those who placed bets know where to find me. Well done to you all.
nickf
sydney
#59
Posted 2007-October-08, 06:14
The approach that I have been advocating (and that Gerben and Alex implemented) is based on Monte Carlo simulation. To date, the model that we were using made an explicite assumption that the expected value and the variance for board results was constant across the length of a match. A more sophisitcated model should probably build in the notion of deliberately trading off expected value for variance when one is behind by a large amount. I don't think that this would be that much more difficult to program, however, parameterizing the model would become a lot more difficult.
From what I can tell, Helene's approach suffered from a different problem. A number of follks, myself included seemed to feel that Helene's model gave a bot too much credit to the Italians. In particular, the model suggested that the Italians were massive favorites against a number of teams that beat them in head-to-head matches during the Round Robin stage. (In defference to Helene, it seems clear that the Italians also believed that they were clear favorites to beat the SA team despite losing to SA in the RR). I suspect that Helene's model may have fallen victim to one of my favorite theories: Bridge isn't transitive (furthermore, I beleive that Helene's model does have transitivity assumptions built into it). While Italy lost to SA, they scored well against a number of teams that SA didn't fare quite so well against. In turn, this suggested that the Italy v SA was an outlier....
#60
Posted 2007-October-08, 06:39
hrothgar, on Oct 8 2007, 02:14 PM, said:
It is a linear model, i.e. if the coefficients are
Dekistan: 10
Oktistan: 8
Quintistan: 5
so Dekistan is expected to beat Oktistan by 2 IMPs and Oktistan to beat Quintistan by 3 IMPs, then it follows that Dekistan is expected to beat Quintistan by 2+3=5 IMPs.
It could be that the general response function is slightly convex (Dekistan will beat Qintistan by less than 5 IMPs) or concave (more than 5 IMPs). This would not be so difficult to identify and I could probably make up some parametrization of the response function.
Some have implied that the response functions may be different for different teams (e.g. Italy being particulare good at beating weak teams). This may be true but it's not clear how to model it, and in any case it would require a lot more data.
The same goes with independency of the residuals, and constancy of the coefficients throughout the event. Obviously not true either but quite difficult to say in what way dependency between the residuals and variation of the coffeicients should be modeled, and again more data would probably be needed.

Help
