[cda77]

Over the last couple of months I have gotten the feeling that I just wasn't getting good cards. But it has gotten to the point that it makes me laugh. Zero HCP hands happen frequently.

So I looked at my last 100 hands. Here's what I found

Average Hcp's 8.81
Median =8
Mode = 6
Highest HCP was 18
Lowest HCP was 0
34% of my hands have 6 or less HCP
11% 14hcp or more

So what's the deal? pun intended


Not a statistics person but this seems out of whack to me

[1eyedjack]

I'm not a statistics person either. Did a little bit at school some decades ago and am familiar with the basic terminology, but fall down rather where it comes down to interpreting the significance that you can draw from the magnitude of deviation from expectation for a given sample size. It was covered at school but that was too long ago for me.

I am also a little dubious about the meaningfulness of terms such as median and mode in this context, where the distribution will not be "normal", and while the average HCP should be 10 I would be surprised if it ever worked out at that in practice for moderate sample sizes.

Anyway I am having difficulty reproducing your numbers. I am conscious that some hands dealt to you on BBO do not make it to the MyHands database, and also that some hands that do make it to that database are available to view only by you, by BBO admin or by other contestants in the events in which you participated. In a sense that is just as well, because "hidden" results such as BBO "best hand" tournaments have no place in this review. Likewise I do not have access to records of any hands that you might have played under alternate IDs.

I took a look at your hands played in the month of March 2017 to date. "To date" strictly means up to 11 March 2017, my local time in UK. At the time of my measurement you had played 278 hands in the month, of which 264 had been played in the period 01 March to 10 March inclusive. As 11 March has not been concluded at the time of measurement, and as it is not clear precisely when your own 100 hands is measured up to, I choose to exclude the hands played on 11 March and have only looked at the 264 hands played up to and including 10 March.

To attempt to reproduce your figures I ran an analysis that looked at the last 100 hands up to each hand from your 100th hand in the month up to the 264th hand in the month. In no set of 100 hands could I reproduce an average of 8.81 HCP. The lowest that I could get to was 9.32, which was for the final 100 hands up to the close of 10 March. If you take all 264 hands up to that point the average was 9.84.

Likewise the median never fell below 9 for any set of 100 consecutive hands in March ending on or before 10 March, according to my dataset.

HCP.....Frequency.....................Frequency
............over 264 hands .............over 100 hands
............to 10 March 2017...........to 10 March 2017

0............1....................................1
1............2....................................0
2............6....................................3
3............7....................................2
4............10..................................5
5............12..................................4
6............25..................................11
7............22..................................8
8............25..................................9
9............23..................................11
10..........17..................................7
11..........22..................................10
12..........24..................................11
13..........18..................................6
14..........13..................................2
15..........8....................................2
16..........6....................................3
17..........8....................................1
18..........9....................................2
19..........2....................................2
20..........2....................................0
21..........1....................................0
22..........1....................................0
23+........0....................................0

[smerriman]

There's nothing unusual about those stats at all.

For example, the probability of having 18 points or less 100 times in row is about 8% - 8 out of every 100 people will have experienced that in their 100 most recent hands.

In fact, if you consider your last 1000 hands, the probability you've had a string of 100 such 'low' hands in a row is a staggering 91%.

Streaks happen - that's how randomness works.

(Similar stats apply to the other facts you mentioned - an average of 8.8 is much rarer, but it will still be happening to a certain proportion of the population, and if everyone posted about it every time it happened, it would get posted about very regularly).

[nullve]

cda77, on 2017-March-11, 16:51, said:

Not a statistics person but this seems out of whack to me

https://www.youtube....h?v=8_mePjkQW_c

[steve2005]

100 hands isn't a big enough sample size!

[silvr bull]

There are related issues to consider too. Key cards are offsides and critical suits break badly far too often. The BBO "Devil Dealer" deserves its well earned bad reputation. Another frequent problem I have seen in many BBO tournaments is that one side gets all the high cards, while the other side has little more than quacks (and distribution that is unusually good for defense) to work with. That could be expected for a few hands, but not EVERY hand in a tournament. It seems like BBO reuses hands intended for one player against three robots in tournaments with all human players, without changing the orientation of any of the hands, so one side has most of the good stuff throughout the tournaments. BBO is not the place to look for "normal" distribution.

[1eyedjack]

steve2005, on 2017-March-11, 20:31, said:

100 hands isn't a big enough sample size!

This criticism seems to be supported by my review of this dataset.
The average HCP over the 100 consecutive hands to the end of 10 March was, as noted above, 9.32. Over the 264 consecutive hands then ended it rose to 9.84. If I then include the first 14 hands of 11 March it rises again to 9.93:
1 @ 6
1 @ 7
1 @ 8
1 @ 9
1 @ 10
6 @ 12
1 @ 13
1 @ 17
1 @ 22

Each time we increase the sample size we get a closer approximation of the expectation of 10

[1eyedjack]

silvr bull, on 2017-March-11, 23:56, said:

There are related issues to consider too. Key cards are offsides and critical suits break badly far too often.

This myth is trotted out occasionally in forum threads. Any serious attempt to back it up to date has always foundered.
It has actually been a while since resurrection in these forums and was about due. Had it any basis in fact it would have been rather more prevalent a topic, I think, and some hard evidence in support provided by those who understand the numbers would have been forthcoming by now.

BBO could possibly help itself by publishing statistical data periodically about deals such as suit breaks, for comparison against expectation. Separating out the "best hand South" deals perhaps into a separate population. "Offside key cards" is more difficult to look at globally because what is offside for one side is onside for the other.

silvr bull, on 2017-March-11, 23:56, said:

BBO is not the place to look for "normal" distribution.

So I should hope. If you found it, that would be evidence of departure from theoretical expectation.

[fromageGB]

Yes, this happens. The best cure I know is to walk three times round your chair anticlockwise when viewed from above. Please note that if you are playing using a mobile phone, that while relativity speaking it would be the same thing to rotate your mobile three times in front on you clockwise while you sat still, THIS DOES NOT WORK.

[barmar]

1eyedjack, on 2017-March-12, 04:48, said:

"Offside key cards" is more difficult to look at globally because what is offside for one side is onside for the other.

The usual complaint is that the human's finesses fail significantly more than the expected 50% of the time when playing with robots, or the complaining user always loses his finesses. But as you said, we've never been able to confirm this claim. Other than the "Best Hand" games, our dealing software isn't biased towards any particular seat, and definitely not to or from a particular player.

We don't do routine statistical analysis of our deals, and we know that we don't use the best dealing algorithm. But we did provide a large set of deals to the author of BigDeal, who is generally considered the foremost expert on computer dealing and analysis of it. He analyzed the deals and said that they were within statistical expectations.

[1eyedjack]

barmar, on 2017-March-12, 13:58, said:

The usual complaint is that the human's finesses fail significantly more than the expected 50% of the time when playing with robots, or the complaining user always loses his finesses.

Doubtless mentioned before, but for the benefit of this thread there is a reason for this phenomenon that does not require "fixing" the deals.
Robots have a particular aversion to leading away from unsupported honours.
So, say you have in hand
S:AQx
H:Axx
(with nothing opposite)
and you receive a Heart lead.
You might think that the Spade finesse is 50%. It is not. It is significantly worse. Because with nothing in either major (and all other factors being equal) the robot would with equal holdings regard a Spade and Heart lead as equally favourable and would (without the Spade King) be about as likely to choose a Spade lead as a Heart lead.

Sometimes in similar situations the robot lacks the Spade King and, on a proportion of those occasions, leads a Spade to his partner's King and Ace. The complainers conveniently overlook that event in their "tally" of the working v. failing finesses because they were not instrumental in the decision to take the finesse.

Instead, they take the Spade finesse and complain that it fails more than 50% of the time when in reality a failure rate of more than 50% actually matches the statistical expectation having taken into account all available information. Not to say that the Spade finesse is the wrong line - just that its success rate should not be regarded as 50%.

barmar, on 2017-March-12, 13:58, said:

we know that we don't use the best dealing algorithm.

That is disappointing. Having identified the fact I would expect it to be a high priority for correction, even if the existing algorithm meets statistical expectation. It is rather a core component of the product.

[barmar]

1eyedjack, on 2017-March-12, 14:18, said:

That is disappointing. Having identified the fact I would expect it to be a high priority for correction, even if the existing algorithm meets statistical expectation.

We looked into it using BigDeal, but it would have taken considerable work to incorporate it into the system. We felt that the flaws in our dealing algorithm weren't serious enough to warrant it.

ACBL recently switched to BigDeal, because people discovered how to predict later deals in a set from the first few deals. I'm not so worried about this because we're constantly producing deals for so many different activities, I don't think it's likely that someone could tell how deals fit into a sequence.