PDA

View Full Version : Standard deviation in Holdem manager - am I interpreting this correctly?



danped
12-27-2008, 11:44 AM
Hi,

My standard deviation in HEM reads 97 bbs. I have a winrate of 5bbs/100. I understand this should mean there is a 68% probability that my true win rate is between -92 (5-97) and 102 (5+97) bbs/100.

The problem is I have 460,000 hands - surely I should be able to more accurately estimate my true win rate from this sample? Maybe the HEM stat is wrong or I am misinterpreting it?

Thanks,
Dan

Paxinor
12-27-2008, 01:59 PM
hi there

you're doing the math wrong

that SD is per 100 hands...

open excel

type CONFIDENCE(0.05;SD(1000);samplsize/1000)

to get SD(1000) you do the following

SD(100) is your Holdemmanager SD

SD(1000) = squareroot(10*SD(100))

the reason for converting your standard deviation from HM to SD(1000) is that No Limit Samples of 100 Hands are not normaly distributed, but samples of 1k are (central limit theorem)

you will get a correct confidence intervall if you do the above steps...

you have a decent sample so can estimate your winrate about +- 2.8 BBs in an 95% confidence intervall

you missinterpreted standard deviation which means the mean of differences of each sample from the samplemean for confidenceintervalls, that tell you that your 95% within a certain range of winrate...

danped
12-27-2008, 02:24 PM
I got it - :D Thanks for the comprehensive answer!

MeekRN
12-28-2008, 11:38 AM
<dl>
<dd><span class="texhtml"><i>A</i><sub>1</sub> = <i>x</i><sub>1</sub></span></dd>
<dd><img class="tex" alt="
A_i=A_{i-1}+\frac{x_i-A_{i-1}}{i}
" src="http://upload.wikimedia.org/math/8/3/2/8329dff4bacb09ce4a81a069d040c086.png" /></dd>
</dl>
<p>where A is the mean value.</p>
<dl>
<dd><span class="texhtml"><i>Q</i><sub>1</sub> = 0</span></dd>
<dd><img class="tex" alt="
Q_i=Q_{i-1}+\frac{(i-1)(x_i-A_{i-1})^2}{i}
" src="http://upload.wikimedia.org/math/f/0/1/f019c960a8a82d4d5ec955d3c11ae540.png" /></dd>
</dl>
<p>sample variance:</p>
<dl>
<dd><img class="tex" alt="
s^2_n=\frac{Q_n}{n-1}
" src="http://upload.wikimedia.org/math/9/b/e/9be1db86f3cc7e91dd114f2af434eabb.png" /></dd>
</dl>
<p>standard variance</p>
<dl>
<dd><img class="tex" alt="
\sigma^2_n=\frac{Q_n}{n}
" src="http://upload.wikimedia.org/math/2/c/c/2cc8b2dc542fa217eb32117a0a73d0c8.png" /></dd>
</dl>
<p>For weighted distribution it is somewhat more complicated: The mean is given by:</p>
<dl>
<dd><span class="texhtml"><i>A</i><sub>1</sub> = <i>x</i><sub>1</sub></span></dd>
<dd><img class="tex" alt="
A_i=A_{i-1}+(x_i-A_{i-1})\frac{w_i}{\sum^{i}_{j=1} w_j}
" src="http://upload.wikimedia.org/math/4/7/5/475f126ffd2894635351239e5af2177e.png" /></dd>
</dl>
<p>where <span class="texhtml"><i>w</i><sub><i>j</i></sub></span> are the weights.</p>
<dl>
<dd><span class="texhtml"><i>Q</i><sub>1</sub> = 0</span></dd>
<dd><img class="tex" alt="
Q_i=Q_{i-1}+w_i\frac{\sum^{i-1}_{j=1} w_j}{\sum^{i}_{j=1} w_j} ( x_i-A_{i-1})^2
" src="http://upload.wikimedia.org/math/2/3/d/23d7d5775a633b27acd95cb818b80269.png" /></dd>
<dd><img class="tex" alt="
s^2_n=Q_n \frac{n'}{(n'-1)\sum^{n}_{j=1} w_j}
" src="http://upload.wikimedia.org/math/9/4/f/94f2b60a9eaa756d60501d7a1bb8ab7b.png" /></dd>
<dd><img class="tex" alt="
\sigma^2_n=Q_n \frac{1}{\sum^{n}_{j=1} w_j}
" src="http://upload.wikimedia.org/math/f/d/d/fddbf9c877c7b9d4d1d3f2ea2f555322.png" /></dd>
</dl>
<p>where n is the total number of elements, and n' is the number of elements with non-zero weights. The above formulas become equal to the more simple formulas given above if we take all weights equal to 1.</p>
<p><a name="See_also" id="See_also"></a></p>
<h2><span class="editsection">[<a href="/w/index.php?title=Standard_deviation&amp;action=edit&amp;sec tion=19" title="Edit section: See also">edit</a>]</span> <span class="mw-headline">See also</span></h2>
<table cellspacing="0" cellpadding="0" class="multicol" style="background:transparent; width:100%;">
<tr>

TierTier
12-28-2008, 12:55 PM
..

TierTier
12-28-2008, 12:58 PM
Uhm....standard deviation doesn't decrease with sample size. Standard error does. Not much to do with central limit theorem there. Multiplying out the standard deviation is completely wrong.

Paxinor
12-28-2008, 06:51 PM
it is the standard deviation OF samples...

standard deviation increases with bigger samples, that should be intuitive...

central limit theorem is used here for confidence intervalls!

the "CONFIDENCE" function of excel assumes normal distribution... you cannot use this function on not normaly distributed variables... because it would be biased

the standard deviation of Holdem Manager is the standard deviation of samples of 100...

samples of 100 No Limit pokerhands are not normaly distributed (see http://forumserver.twoplustwo.com/15/poker-theory/depth-analysis-variance-poker-23281/#post378062 for proof)

you have to use 1000 pokerhand samples...

the law of the sums of variance says that VAR(SUM(X)) = SUM(VAR(X)) (see http://en.wikipedia.org/wiki/Variance for proof)

therefore the variance of samples of 1000 pokerhands is ten times the variance of samples of 100 pokerhands

again if you want to use unbiased confidence intervall you need mean and standard deviation of normaly distributed variables, so you cannot use the HM sd and mean...

oh and i'm not multipling the standard deviation, i'm multiplying variance....

thats why you do the conversion...

john2
01-15-2011, 01:48 AM
Sorry to bump this old thread, but I have been through every post yielded by searching for "Standard Deviation" and cannot understand what HEM standard deviation is based upon.

Can I plug the Standard Deviation shown by HEM directly into a variance calculator or do I need to doctor it as explained in the above posts?

I cannot understand the above posts or the formula given by Roy in at least two separate posts. I took statistics in college, but that was during the last century.

The above posts suggest that the Standard Deviation shown in HEM cannot be used directly because NL HE hands are not normally distributed, but I have little idea exactly what this means mathematically.

Thank You.

john2
01-18-2011, 01:01 PM
Bump

morny
01-20-2011, 02:24 PM
Ive asked the developer but he is quite busy at the moment but he will respond as soon as he has some free time. Unfortunately i have no idea of the correct answer :)

john2
02-05-2011, 11:41 PM
REBump

morny
02-07-2011, 02:47 PM
Hi, the developer said the formulas posted above should be an accurate representation of it, if you think there is an error with this please post an example and ill get it checked out

john2
02-10-2011, 01:58 AM
Sorry to be a burden, but as stated in my previous post, I have no idea what the above formulas mean.

I tried entering the initial formula in Excel, but it didn't recognize the given function.

___: ___:_
02-10-2011, 06:23 AM
Have you guys even looked at what a 1000 hands looks like in a histogram? It's far from normal. If you graph 80 100-hand samples (so 8k hands), then only do we see something "slightly" normal.

Use this calculator for confidence intervals:
uDevil's Poker Results Calculator (http://www.castrovalva.com/~la/win.htm)

Again, make sure your sample is >8k hands (periods > 80) for the results to be meaningful.

john2
02-10-2011, 11:56 AM
This is the variance calculator I am attempting to use, but I still need to know what to enter as "Standard Deviation".

From what I gather from the above posts, HEM Standard Deviation stat cannot be plugged directly into this calculator, so I'm trying to find out how I need to doctor the stat.

The issue here isn't finding a big enough sample to produce lower variance; rather, it is how many hands are needed to produce an accurate Standard Deviation to plug into the calculator, and how to manipulate HEM's Std Dev stat to work with it.

___: ___:_
02-10-2011, 05:17 PM
This is the variance calculator I am attempting to use, but I still need to know what to enter as "Standard Deviation".

You enter exactly what HEM gives you.



From what I gather from the above posts, HEM Standard Deviation stat cannot be plugged directly into this calculator, so I'm trying to find out how I need to doctor the stat.

Yes you can. Export the hands into excel and use excel to calculate SD. You will get the same #.



The issue here isn't finding a big enough sample to produce lower variance; rather, it is how many hands are needed to produce an accurate Standard Deviation to plug into the calculator, and how to manipulate HEM's Std Dev stat to work with it.

This was the entire point of my last post. You need at the very least 8k hands (80 100-hand samples) for the calculator I posted above to be useful.

You cannot use (or manipulate SD etc) the calculator under any circumstances for samples less than 8k hands. Confidence intervals are constructed on the basis of the Central Limit Thm, that states a sums (or averages) of a random variable converge in distribution to normal as samples increase to infinity. I've seen that 8k hands (80 samples) is where the histogram of winrates starts to look somewhat normal.

john2
02-10-2011, 07:44 PM
Aha, thank you for your reply...

I just use the calculator for longer term (70k) possiblities.