Re: [amibroker] Monte Carlo Analysis in AMIBROKER?, AmiBroker Email List Archive

On Sun, Jan 24, 2010 at 12:46 PM, Markus Witzler <funnybiz@xxxxxx> wrote:

Hello Howard,

many thanks for your thorough explanation!!! Searching high and low the internet more often than not just adds to already exiting confusion... Thus, I´m really grateful for your summary!

Monte Carlo has been a kind of a black box thing for me and still somewhat is.

I infer from your post that Mersenne Twister is the "best" algorithm to date, so the AB community can be -once again!- glad to have Amibroker at its disposal>G<?

When reading your answer to Joseph (today), I wondered what the difference between intelligent optimizer´s "sensitivity analysis" and a MC simulator´s test might be. Both aim at testing for robustness as far as I understand.

I´m currently in the process of evaluating IO for my needs but I have never yet run a MC to test for robustness. So, I don´t now what it looks like.

I don´t want to take away to much of your time and expect another detailed answer but

1./ could you just shortly elaborate on the difference between a sensitivity analysis, done by IO, and a test for robustness (i.e. sensitivity?) done thru a MC simulator? I understood there is currently no better choice than the combo of AB with Mersenne Twister at this point.

2. Incidentally, I wonder what "modules" one would need to have at hand or write himself (plugins for Amibroker?) to be able to run all the different tests "Know your system software" is able to produce (Amibroker plus Mersenne Twister, ...?). I´m asking since you said to be familiar with IITM stuff. If my questions are to far off-topic and might bore other forum members, just email me privately.

And I am glad for your comments on Van Tharp´s work. Sometimes, he seems to shoot "off the hip", if you know what I mean.. I certainly will look up your posts regarding this subject - thanks for mentioning. I like to verify both sides of the "coin".

I really appreciate your effort to help us newbies out there to broaden our horizins. Forums like this are a great place to be!

Thanks again and all the best for your newest book project!

Peace

Markus

----- Original Message -----

From: Howard B

To: amibroker@xxxxxxxxxxxxxxx

Sent: Sunday, January 24, 2010 7:33 PM

Subject: Re: [amibroker] Monte Carlo Analysis in AMIBROKER?

Hi Markus --

I also own Tharp's books. He does use a proprietary package called "Know Your System" for his analysis.

There are many ways to use Monte Carlo techniques. The different algorithms are more about generating random numbers and defining the distributions from which the random values are drawn and the way they are used.

Tomasz has implemented the Mersenne Twister algorithm in AmiBroker. Mersenne twister is much better (higher in metrics that describe randomness, such as having a long period and not having patterns within the series) than the random number generators that come with C++, Excel, etc. If you use Excel for Monte Carlo analysis, be sure to get a Mersenne Twister Excel addin before you start. Here is one source: http://www.financial-risk-manager.com/risks/analytics/random/rand.html#mt19937. There are others.

The distributions can be very problem specific, but all start out using "uniform" distributions and build from there to "normal", and so forth.

Back to Tharp. I have some criticisms of Tharp's "Definitive Guide to Position Sizing". There are two in particular.

1. He arbitrarily and inappropriately sets an upper limit on the number of data points to be used in calculating his System Quality Number. (His SQN is essentially a t statistic, which he acknowledges in one sentence early in the book, but ignores from then on.)
2. He is completely unrealistic about what level of SQN trading system developers should be able to achieve. As a quick and dirty measure, a t statistic of about 2.0 suggests significance at about the 0.05 level for N of about 20 data points. A trading system that uses expectancy as its metric, computes the t statistic on actual trades or truly out-of-sample results (in-sample results have no value as estimators of future profitability), and has a t statistic of 2.5 to 3.0 will result in extraordinary profitability. Tharp talks about achieving scores of 6 or 7. Give any one of us a system that has a score of 6 and we can buy Manhattan in about a year starting from $10,000.

So, read Tharp. His books do have value. But be aware that there are many aspects of trading system development, testing, and validation that he simply does not understand -- and some of his writing are seriously misleading. I have made many posts on Aussie Stock Forums (http://www.aussiestockforums.com/) on this topic, and also on trading system development. Search using my name.

I have corresponded with Van Tharp about some of these issues and he acknowledges my points. He even gives me credit in DGPS. One of the attendees of my workshops in Australia a few months ago is a personal friend of Vans and agrees with my assessment.

Thanks,
Howard

On Sun, Jan 24, 2010 at 10:42 AM, Markus Witzler <funnybiz@xxxxxx> wrote:

Hello Howard,

just stepping in here since MC analysis may be an issue keeping me busy in near future (though I haven´t yet developed much expertise in it).

Are there different algorithms of MC simulators?

I wonder what to look for when considerung a product, since different "qualities" in algorithms may turn out different levels of "quality" (i.e. what scenarios to expect in the future).

Van Tharp for instance claimes to have Chris Anderson developed a software package called "Know your system" with an MC simulator in it - he uses it for studies for his book on money mangement which I own.

I´m looking forward to buying your book since MC is covered there to as you said below...

Thanks

Markus

__________ Information from ESET Smart Security, version of virus signature database 4668 (20091207) __________

The message was checked by ESET Smart Security.

http://www.eset.com