An Exploratory Data Analysis of BBO Tournament results

[hrothgar]

Barry and Uday were gracious enough to share some of their board results with me.

 

The following document describes the results of an Exploratory Data Analysis of a large corpus of online bridge tournaments that were played on Bridge Base Online between April 2018 and April 2019. Three different data sets were examined:

 

1. BBO Daylong (MP) tournaments (8 boards in length)

2. ACBL Daylong (MP) tournaments (12 boards in length)

3. The Spring ACBL Online National tournament (24 boards in length)

 

The primary motivation behind this project was to examine the variance in board results for the different types of tournaments.

 

http://www.filedropper.com/exploratorydataanalysisofbbotournamentresults

[thepossum]

Thankyou Hrothgar

 

Very interesting analysis!

 

I'm sure I will have many questions once I've read it.

 

I guess my first question is how much you can conclude about individual behaviour by looking at distributions across all instances and players in a tournament. I imagine you would need much more data on individuals and more complex analysis of the variances. It seems that those torunament/instance level average distributions are relatively stable distributions/profiles

 

Just by way of example, in my own hands I havent noticed any real significant difference in variance between different types of tournaments, my means vary but not my overall variance (althogh there is some difference). However you would need thousands of hands for many individuals to see if there was a difference across player behaviour on average. Note my personal hand sigma is around 30-31% :) EDIT (not that it matters to anyone but me), but my sigmas on different tourneys are closer to 22-28 (small samples and time periods). I overstated my variance by making calulcations on mixed up tournament types. I would hate to think that anyone who knows me read these threads :) However I am seriously interested if anyone actually thinks its possible to separate individual variances from all the others in such complex tournament and game as bridge

[hrothgar]

Please note: Chart 2 should be labeled "Distribution of Sigma Across Board IDs"

[hrothgar]

FWIW, part of the reason that I did this analysis was what I perceived to be a very high variance in my own results in some of these events...

[dbl118]

FWIW, part of the reason that I did this analysis was what I perceived to be a very high variance in my own results in some of these events...

 

I feel you are not controlling for the most important variable in all of this... the user pool. I play the daylong MP tournaments multiple times per week. They are free, and easy to do. I have never played an ACBL tournament, and despite winning free entry in the spring ACBL tournament, I didn't play that either. I believe my competition is made up of a moderate size of relatively elite players, and a lot of people that are fairly subpar players, or not trying very hard. Indeed if I take my time and possibly even take notes during the hands about cards played, it is very rare that I will score under 65% or so, largely based on my understanding of how GIB works, as well as how to fool it. If unfocused, my results can vary a lot.

 

There are different types of players in each tournament. I play exclusively online, and almost exclusively with robots. I imagine the population that played the spring ACBL were people far more likely to play in clubs or other tournaments. They likely had less inside knowledge about how GIB bids. I am not interested in ACBL tournaments and I do not buy into the point system (I won free entry to the spring ACBL but chose not to play since I'd need to become a member).

[TylerE]

I find your Chart 1 a bit confusing. Some sort of funky interpolation going on? It shows values <0 and >100.

[hrothgar]

I find your Chart 1 a bit confusing. Some sort of funky interpolation going on? It shows values <0 and >100.

 

I used a kernel smoother

 

Standard way to visualize this sort of data, but it does spread stuff out some

[hrothgar]

I feel you are not controlling for the most important variable in all of this... the user pool.

 

Thanks for the comments

 

I think that there is merit in what you are suggesting. The set of players that you get matched against should certainly impact the scores that you receive. However, given the relatively large number of comparisons that are used for each board I would expect that this would wash out a lot of the time.

[dbl118]

Thanks for the comments

 

I think that there is merit in what you are suggesting. The set of players that you get matched against should certainly impact the scores that you receive. However, given the relatively large number of comparisons that are used for each board I would expect that this would wash out a lot of the time.

 

I don't see why it would wash out at all if what I'm saying is true, that the quality of players differs between types of tournaments. Let's say in each tournament I'm compared to 100 players. If in one type of tournament, 30 of those players on average are "bad", I'm going to do a lot better than a tournament that requires you to pay a decent amount, where on average maybe 10 players are bad. Doesn't matter how many comparisons there are, you are comparing apples to oranges anytime you compare two totally different tournaments. In fact, the more comparisons there are, the more likely what I'm saying will be true.

[hrothgar]

I don't see why it would wash out at all if what I'm saying is true, that the quality of players differs between types of tournaments. Let's say in each tournament I'm compared to 100 players. If in one type of tournament, 30 of those players on average are "bad", I'm going to do a lot better than a tournament that requires you to pay a decent amount, where on average maybe 10 players are bad. Doesn't matter how many comparisons there are, you are comparing apples to oranges anytime you compare two totally different tournaments. In fact, the more comparisons there are, the more likely what I'm saying will be true.

 

Sorry, I misunderstood your comment.

 

I thought that you were suggesting that IF you get matched with a lot of bad players in tournament A and a lot of bad players in tournament B then ...

 

Intuitively, I agree with you that I would expect that tournaments that cost more $$$ might have a stronger set of players, however, I would expect to see this manifest itself in the variance of the board results that doesn't seem evident to me...

[barmar]

There are different types of players in each tournament. I play exclusively online, and almost exclusively with robots. I imagine the population that played the spring ACBL were people far more likely to play in clubs or other tournaments. They likely had less inside knowledge about how GIB bids. I am not interested in ACBL tournaments and I do not buy into the point system (I won free entry to the spring ACBL but chose not to play since I'd need to become a member).

Wasn't that the point of graphing the statistics for each different type of tournament? And what he found was that the shapes of the graphs were pretty similar.

[0 carbon]

To control for number of players in a tourney, you could winnow out the bigger & smaller tourneys so you can compare better between ACBL & BBO tourneys.

 

Are the BBO tourneys barometer & ACBL not? This could account for wilder bidding/play in later rounds. Did you do any comparison between, say, the first 2 boards of a tourney and the last 2 while using barometer scoring? This might indicate if high-variance bidding occurs.

 

 

To eliminate the "user pool" bias, you could divide the results into 3 categories & analyse separately to see if there are differences:

• Players playing only BBO tourneys
• Players playing in both ACBL & BBO
• Players playing only ACBL

-- tom

 

[barmar]

To control for number of players in a tourney, you could winnow out the bigger & smaller tourneys so you can compare better between ACBL & BBO tourneys.

The daily ACBL and BBO tourneys both have on the order of 1,000 players. The NABC tourneys had around 1800 players. None of them are small.

Are the BBO tourneys barometer & ACBL not? This could account for wilder bidding/play in later rounds. Did you do any comparison between, say, the first 2 boards of a tourney and the last 2 while using barometer scoring? This might indicate if high-variance bidding occurs.

No, they're not barometer.

To eliminate the "user pool" bias, you could divide the results into 3 categories & analyse separately to see if there are differences:

• Players playing only BBO tourneys
• Players playing in both ACBL & BBO
• Players playing only ACBL

-- tom

The information we gave Richard doesn't identify specific players, so he can't correlate players between tourneys.

[hrothgar]

The information we gave Richard doesn't identify specific players, so he can't correlate players between tourneys.

 

I will also note in passing that trying to build system that was able accurately account for individual players is roughly akin to creating a rating system and I so don't want to go there...

[leftfoot]

I think there are two driving forces in the comparison:

 

1. There is always going to be a larger variation of scores in shorter tournaments vs. longer tournaments.

 

2. The master point scale of the BBO daylong highly rewards swinging. A 70% game will get you close to 7 BBO points while a 60% game might get you half of one BBO point. That is 14 times the amount of points.

 

The same comparison in ACBL ... 70% will get you .9 ACBL points while

60% will probably get you .43. The 70% game only gets you roughly twice the amount vs. a 60% game. The incentive for swinging is much less.

 

3. You should also exclude the "Just Declare" tournaments in your analysis because there will be more flat boards since the opportunity for deviation in the bidding is eliminated.

[mugsmate]

I don't understand what I'm looking at.

If you have all the players, how can the median not be 50%?

Any result, R%, a player gets, their opponents will get (100-R)% so the distribution has to be symmetric unless you haven't got all the data.

[hrothgar]

I don't understand what I'm looking at.

If you have all the players, how can the median not be 50%?

Any result, R%, a player gets, their opponents will get (100-R)% so the distribution has to be symmetric unless you haven't got all the data.

 

Robots don't place...

[mugsmate]

"2. To what extent is tournament success “predestined” (determined more by the type of hand that people were dealt rather than how the board was played)"

 

You could probably look to do this from the data set you hold.

 

You calculated the "Board variance" for each hand played.

 

Define a player's: Average_Board_Variance as the average of the board variances for all the boards they happened to play in the tournament.

 

Then for every given tournament, compute the distribution of "Average_Board_Variance" for all participants and the Average_Board_Variance of the winner.

 

For every tournament played, return the percentile position of the winner's Average_Board_Variance on the distribution for all the players.

 

Now, you can plot that, and see how often players win tournaments as a function of how flat their boards were relative to he boards in play. You can plot that for tournaments of different lengths.

[mugsmate]

Robots don't place...

 

OK, understood, so Robots take over from pairs that drop out?

In which case, your asymmetry has to be due to how people perform against Robots versus against people.

 

So "when playing robots" 0-50% is flatter than 50-100%

 

Well, that makes sense. The robots make very few zero inducing errors relative to the human performers, but are calibrated to play at the average level of the field.

[hrothgar]

These are tournaments in which one human plays at a table with three robots

[mugsmate]

These are tournaments in which one human plays at a table with three robots

OK - but the skewness of your MP/board distribution is still more a function of how the robots play than anything else.

[hrothgar]

OK - but the skewness of your MP/board distribution is still more a function of how the robots play than anything else.

 

Opinionated and ignorant...

[mugsmate]

If you aren't interested in other people's ideas, don't bother posting in future. I thought there were some interesting ideas here, and would have happily helped. Absolutely no cause for your rudeness. I note no reply to my suggestion for how to tackle proving your hypothesis.

 

Best regards.

[mugsmate]

Remember, I read your article and I spent some time thinking about how to prove your hypotheses. I also attempted to address your comment regarding your uncertainty about why the board distribution was asymmetric.

 

I was very keen to get your thoughts on what I'd written. Instead, a dismissive, rude remark that makes me feel I wasted all that time.

 

That is not a good way to treat people.