Topic: Satoshi Dice -- Statistical Analysis - page 31. (Read 192896 times)

I checked, and there really were 65 losing "lessthan 32000" bets in a row in this block: http://blockexplorer.com/b/184391.  Search for 25088, which is the winning bet before the losing streak starts, then pair up the txids in the first column with the bettx column in this (large, and so not inlined) collection of screenshots: https://i.imgur.com/GhH7b2H.png

Updated the stats.

I DO think there is another whale out there with at least 10k distributed across other addresses.  That guy uses a similar betting pattern to the first whale but much smaller bet sizes.

https://bitcointalksearch.org/topic/m.1515541

THIS ^

https://bitcointalksearch.org/topic/m.1555528

But also see this ^

Thoughts on why profits are so low this month?

I'm regenerating the list such that I can sort it into the order that the bets were placed rather than payout order.  That should give better numbers I think.  It will take a few hours though.

Edit: the list of bets I'm generating has the block number and transaction number within the block, so I can sort it once it's done.  I just checked a few early entries to make sure the bets were being listed right, and it showed me:


ie. the 8th transaction (counting from 0) in block 178122 contains 3 bets of 0.04 BTC.  I checked this against blockchain.info (http://blockchain.info/block-index/215305/0000000000000583451e3795919b1afbc6c6b96a0da71773015999a70da6aeb7) but found that these 3 bets were in the 9th transaction, and that there were bets in the 3rd transaction which show up in my list as being the 37th transaction.  So I was worried my code is listing the transactions in the wrong order.

It turns out though that my code lists the transactions in the same order as blockexplorer (http://blockexplorer.com/block/0000000000000583451e3795919b1afbc6c6b96a0da71773015999a70da6aeb7) so it's blockchain.info that's listing the transactions in an apparently random order.

It's only got as far as last October so far, but already found sequences of 30 and 26 wins in a row at lessthan 32768, and 23 losses in a row too:


Edit2: It's up to December 28th now, and I've improved the code that finds the longest sequences so it tells you which block and transaction number each sequence starts and ends at:


That looks like a genuine run of 30 wins on lessthan 32768.  Isn't that incredibly unlikely?  Like a 1-in-a-billion chance?

Edit3: I see a problem.  It used to be that all the bets in a single transaction would get the same lucky number.  It's also possible to bet on the same address multiple times in a single transaction.  That means you can win the same bet twice in a single transaction, and I'm counting that as 2 in a row.  See the 'lessthan 256' above - it says "2 wins (208181-71   to 208181-71)".  That's because he bet on lessthan 256 twice in the same transaction, and both bets got the same lucky number:
  http://blockchain.info/tx-index/33164014/9
  http://blockchain.info/tx-index/33164014/19
or see lines 3 and 8 here.

OK, so it's finished making the list, and I've changed the script to ignore multiple bets in a single transaction in blocks earlier than block 213879 (when the rule changed).  Here's a list of the current losest winning and losing streaks per bet:


So the 30 wins in a row on lessthan 32768 was real.  Even more incredible is the run of 34 wins on the slightly harder lessthan 32000 game.

Edit76: I just noticed "65 losses (184391-665  to 184391-739" on the "lessthan 32000" game.  That's a huge number of losses in a row.  They're all in the same block, too.

Here's the list of them:

Can I ask what shows up if you type the following?

$ awk '{print $3, $1}' allbets.txt | grep '^32000 ' | uniq -c | sort -n | tail
$ awk '{print $3, $1}' allbets.txt | grep '^32768 ' | uniq -c | sort -n | tail
$ awk '{print $3, $1}' allbets.txt | grep '^48000 ' | uniq -c | sort -n | tail
$ awk '{print $3, $1}' allbets.txt | grep '^52000 ' | uniq -c | sort -n | tail
$ awk '{print $3, $1}' allbets.txt | grep '^56000 ' | uniq -c | sort -n | tail
$ awk '{print $3, $1}' allbets.txt | grep '^60000 ' | uniq -c | sort -n | tail
$ awk '{print $3, $1}' allbets.txt | grep '^64000 ' | uniq -c | sort -n | tail

I started producing a list of all the bets to find the longest winning and losing chains and found a problem:

SatoshiDice has at times in the past run low on confirmed funds in their hot wallet.  When this happens, they can easily pay out losing bets (since all they need to do that is the incoming bet itself), but they have to wait before they can pay out winning bets.  This results in large groups of consecutive bets with the same result.

I've only looked through a fraction of the bets and already found a chain of 319 losing "lessthan 32000" bets in a row, and also a chain of 287 winning "lessthan 32000" bets in a row.

This makes me think it's not worth continuing with this analysis, since it doesn't tell us anything other than that SatoshiDice sometimes bunches winning bets together.

Here's the progress so far.  In less than 400k bets I see all these chains:

It looks fine, as long as you have the win/lose probabilities right.

You can also calculate the expected number of rolls before n losses in a row:


I've seen a few methods of proving this. The one I have to hand uses Markov chains and is a bit dense, and I can't find the other which is more inuitive to follow. If I find it I'll post it.

Hi,

I got the latest numbers from dooglus post and tried the theoretical formula for consecutive losses in a row and came up with this.

bets     p-win   q-lose  loss-row
604294   0.5     0.5     18.20489587
1118230  0.48828 0.51172 19.71766168
244037   0.36621 0.63379 24.99911792
243386   0.24414 0.75586 39.27274907
94075    0.18311 0.81689 48.22841691
182495   0.12207 0.87793 76.89902593

So, those playing the 50% game can expect to lose 18 or 19 times in a row if they played that game since the beginning. Those playing the other games can expect to lose much more.


Did I do this correctly?

Those colourful violin plots are very nice.   

Meanwhile, back to the regular ugly graphs...

ggplot doesn't do multiple axes. I read (I think) that the devs planned on support for multiple axes when the axes scale linearly to each other, but they don't plan on the sort of multiple axes you're thinking of. I ccould annotate the chart with the actual numbers, but that's not very pretty.

So, number of bets as well as volume of betting? I have a think about it. I hadn't considered that, although it's might be interesting data to have.

If the 2-3 month cycles are valid and hold, May-June ought to be the start of a very active period.

Amazing charts OoC, does GGplot let you do multiple axes? It would be good to overlay a line showing total number of bets per month, or perhaps median number per day as well as max/day and min/day for each month.

I played around and got some quite beautiful tile plots and contour plots, but they didn't really help show bet size of volume trends. Then I realised that all I had to do was use the proportional violin plots and use colour as a guide to the volume. I hope this works for you. Please don't tell me you're colour blind - I didn't use the colour blind palette.

The two below just use different colour schema.

I plotted your request below - normalising to the volume of bets under 50btc (as an arbitrary cut off). I also plotted all bets before December, to make those trends a bit more obvious, too. Tbh, I should have written "the size and volume of bets in December 2012, January 2013 and February 2013 have increased compared to the previous months". Even so, it's not obvious from the plots.

I'm working on something similar to the "percentile" plots I did that will hopefully show this sort of information in a more intuitive and cntinuous way than either the violin - er scary hypodermic needle - plots or the percentile plots. Hopefully I'll be able to post it tonight.

Hi, Sorry for adding to the stress. I guess you just pick one way of ordering it, and use that for consistency. Actually, a real world example of a 50% game had 26 losses in a row, in a casino.

If you're bored ... you can do this. When you got more bored, you can also do the different ways of ordering.

I don't know if it will tell us anything useful, but I think it will be interesting. And the minute you post your analysis, all the gamblers who are reading this thread will suddenly start betting a little bit more, thinking that they can attempt 21 consecutive attempts to martinegale from 0.01 to 500. (I just might do it, hehe.)

If you're being grouchy, let me know what I need to download and I will try to do this. (Or at least someone else will, so you have less stress.)

Unfortunately, I'm on a crappy OS machine (windows), so I don't know if it's even possible.

It seems to be as expected since the analysis is done by block - a chart of betting would've levelled off for ~6 hours or so, but payouts would follow the consistent climb. Transactions still had to be processed into the blockchain block by block.

I think my graph only shows payouts though.  It ignores bets until they have been paid out.