i-Poker has now been tested for pre-flop all-ins - these were similar to other bad beat tests that have been performed.

In total, over 18 million hands were tested. The reason the testing was so extensive was that in the first two tests a significant bias was found in favour of dominated hands and it was deemed necessary to carry out widespread testing.



Here is a brief run through the results and some discussion (for full details see this web-page). Each hand sample is 1.1 million hands apart from sample F which is 11.5 million hands.

Samples A and B were from 5c/10c games (FR and 6-max respectively) and were from May/June/early July 2011

Sample A, Dominated hands: +3.38 standard deviations from expectancy (this should occur once in approx. every 1380 tests)

Sample B, Dominated hands: +2.78 standard deviations from expectancy (this should occur once in approx. every 230 tests)

Samples A and B were merged to see how likely both results occur consecutively was. The result for samples A and B combined was +4.28 standard deviations from expectancy (this should occur once in approx. every 53,500 tests).


Widespread testing was then carried out to see if the bias was widespread. 6 more hand samples were used and more than 16 million hands tested. Hand samples were taken from different stakes during a similar time periods to samples A and B and also from the same stakes as samples A and B (5c/10c) but during a different time period in 2012.

Results from these tests showed no significant deviations from expectancy.




There are many possible explanations for this isolated bias, although it should be noted that these explanations are speculation and the aim of these tests is to test for a bias and not to explain why it occurs. However, here is some speculation...

1) The bias could be due to variance. However, this is unlikely as the result should only occur once in every 53,500 tests (about 50 tests have been run). This does not however mean that it is impossible.

2) The bias could be due to a bias in the deal. However, this also seems unlikely as it would be expected that a bias in the deal would be consistently shown throughout all tables and time periods.

3) The bias could be due to widespread collusion at 5c10c 6max and full ring tables during May and June 2011. It has been suggested that players colluding would introduce a bias into the results of the bad beat tests since if a players holds AK, AQ etc. and was aware of that another player that they are colluding with also holds an ace or a king they would be unlikely to call an all-in from a third player. If such collusion was widespread enough to influence the results of tests which are run over millions of hands it is likley that it is 'bots' that are colluding together rather than players that are colluding together.



Some points worth discussing here:

Margin for Error:

The margin for error due to the Monte Carlo part of the Poker Tracker report was significantly reduced for the dominated hands test of samples A and B. This was achieved by running each test 10 times and then finding the mean result.

Card Removal Effects:

The results in samples A and B cannot be attributed to card removal effects as card removal effects should be reasonably consistent across all bad beat tests and the results of tests on samples A and B have not been witnessed in any other tests that have been run by Online Poker Watchdog.

Interesting but i think reason 3) alone gives us enough reason to stay extremely optimistic that's it's not rigged for the time being.

Statistics geeks can chime in, but combining the samples to get the 4.28 standard deviations seems like bad methodology to me. You're introducing a high degree of selection bias.

This is a point that was wrestled with but here are some reasons to combine the samples:

Samples A and B were taken from games at one poker site, at identical stakes during the same time period. i.e. 5c10c tables during a time period in May/June/Start of July.

The only differences between the tables is that sample A is for full ring and sample B is for 6 max tables (the time frame for the FR tables is slightly longer than the time frame for the 6max tables due to less hands being played at FR).

Therefore the circumstances are as similar as is possible to achieve.

The individual tests on these tables both came up with a considerable bias of approximately 3 standard deviations in each case (when errors were reduced the results were both very similar and close to 3 standard deviations: +3.098 for sample A and +2.988 for sample B).

Each of these results on their own is about 1000 to 1 (note that in none of the tests from all other sites has a result close to this far from expectancy occurred) and here it happens twice in a row under very similar circumstances.

Speculation again... if someone ran bots that colluded at 5c10c it is quite likely that they could run them at FR and 6max tables simultaneously - the games are very similar compared to... say HU at the same stakes.

Speculation again... this makes sense - however if players are playing against bots that collude then they have very little chance of winning (if it is the case it is still definitely not a fair game).

Interesting post LA. I agree with the other poster that the interpretation of the results of the combined A + B samples is not quite appropriate given that you already knew A and B were off individually before you combined them. That is, the p-value is not accurate.

But who cares about the p-value. There are two issues here: 1) Clearly these results would indicate a deviation from expectation that is unlikely due to chance alone. 2) Therefore, one also wonders how large this effect is? Or put another way, how far from expectation are the results in these samples?

Also have you contacted i-Poker about this?

I haven't contacted i-Poker about this yet - I am going to let it have some time on the forum beforehand in case there are any kinks that need ironing out.

When I ran the tests samples A was the first one that I did - I figured that one result on it's own could easily be variance and didn't expect to get another similar result. Sample B was the second test that I ran and I deliberately chose the sample that resembled sample A the most in terms of game type, stakes . When the results came through I nearly lept out of my seat!

Then it was a matter of getting other similar hands to see if the pattern was widespread. Most were a little different in terms of stakes or date-range. The most similar was the 11.5 million hand sample F that is at the same stakes but covers a longer period that actually included all hands from sample B.

When these came out to have no significant bias it was pretty obvious that it was an isolated case and it wasn't quite as worrying as I'd imagined at first.

If taking a larger sample results in a smaller offset from the mean as measured in standard deviations, rather than a larger one, that seems to me to point to either normal variance or a narrowly isolated effect (collusion?) in the smaller samples. If the effect were consistent throughout then of course the offset would be larger in a larger sample (as measured in SD units). That's what we would expect to find if the effect were due to a non-random deal.

That said, I'm not sure I agree with the collusion theory as a possible cause either. It seems like additional skew in the card removal effects would be somewhat self-cancelling and not all one way.

You didn't list as a cause a possible bias in the hand collection itself for those isolated smaller samples. We know the hand collection sites don't have every hand, and a systemic bias for a temporary period could happen, where for some reason certain hands aren't in the data.

This is unlikely as if there was a systematic bias in hand collection it would show up in all the other tests too. Also, if there was a systematic bias in hand collection then it should not be temporary.

In practice, the missing hands from data-mining sites are likely due to server down-time etc. and therefore would be random.

wykh has run a 10k all-in simulation with 3 colluders and this shows a slight swing in favour of the underdog: see previous thread.

If you want to think of it in terms of theory instead of practice: if the collusion was effective it would obviously skew the tests...

the colluders would be folding hands that to players that could see only their own hole cards would seem like 'ahead' hands but in reality with information about other player's hole cards the colluders know these hands are way behind.

This means that they don't play as many hands that appear 'ahead' but are actually behind... skewed result in favour of the underdog.

There's no reason for that to be true, and cases of temporary issues where certain hands were not collected have been documented in these forums. But I agree it's unlikely to explain your results.

would it be possible for you to post a link

thanks

Another thought - if there was a temporary systematic bias with the datamining sites then all hand samples during that period would be affected - not just sites at 5c/10c games.

A few weeks ago I talked to the head of security and fraud prevention of pokerstrategy. He confirmed that bots on Ipoker work together. All you have to do is compare stats, tracked from date, screenames etc. Its obvious. Higirl82 - Lowride72, Perchesko, Sancheskos, Gulleks/Silleks. On the first screen you can see Higirl82 and Juged666. Look at the stats + betsizing. Most of them played on Offsidebet + Dafa Poker. Ipoker = bot infested

Wow!!!!

But how do the bots collude? Has someone built a bot communication network, i.e. botnet for poker? Serious question.

I'm no botter. I cant answer this question. I played them alot and it feels like playing vs 1 player on different tables. I think they share one database.