The name "Monte Carlo Simulation" is quite prevalent, but the point of running the model through many randomly generated possibilities is to see what are the possible outcomes/metrics and what their likelihood is. Significant improvements in the performance/convergence of these methods can be achieved by being more systematic when sampling. http://cscs.umich.edu/~crshalizi/notabene/monte-carlo.html http://en.wikipedia.org/wiki/Markov_chain_Monte_Carlo http://www.math.nyu.edu/faculty/avellane/ConqueringTheGreeks.pdf
That is not what he does with his software. He takes the output performance paramaters of a system and randomizes them, probably assuming a normal distribution. Many sets of those parameters violate the fundamental asusmption of MCS of equal probability for a specific system. It is not MCS simulation. It is some ad-hoc simulation based on many unjustifiable assumptions. It is quite intriguing that he provides no explanation how he does it other than how-to guides of using the software. In legitimate MCS, one would pair system returns, including neutral positions, to market returns through resampling. Then one would test the hypothesis that the system is not intelligent in matching the returns. I think someone named Masters had a book once that fully explained the legitimate use of MCS for trading system design. I read it long ago at my library.
@Intradaybill, your arguments are ridiculous. You critizise and have even not got the point. The MCS method is a "number cruncher" model, independent from normal distributions etc. There are many methods to combine your trading assumptions with this general MCS methodology, but there's not a proven no.1. I'm sure (and all my clients are sure) that my implemented methods are a good instrument to analyze chances and risks of trading systems concerning a so-called system simulation stress test and that the additional implemented data simulation methods are helpful for alternative system tests! And it's 100% pure MCS! bye, zentrader
How is that? Provide a reference of how you "crunch" the numbers without assuming any distribution at all. I said already if you provide the details of a sound methodology I will accept it. You just come back with aphorisms and statements like "this is ridiculous", "you do not get it", etc. but you resist providing details of your number crunching. How to you crunch numbers? I visited the links you provided and the only thing I found was a how-to guide of your software.
IMO Bill is correct. I also find your attitude very irritating. I think he asked you several times a specific question, namely the procedure you follow to get results. Each and every reply of yours is about not understanding it, but what exactly doesnât he understand? Do you want him to take it for granted that your alleged MCS works fine? Can you reply with a link to the mathematics you use in that specific product of yours? People have found problems and omissions in very popular math packages, like BOUNDS for example, that produced significant error for years without anyone noticing. We know that many backtesting platforms generated wrong results under certain conditions for many years. Why should we discount this possibility in your case?
Dave, i've the feeling discussing with intradaybill is talking against a wall. I've clearly said and shown (links) how a typical MCS stress test based on the known trade ratios and pay-off-ratios can be done. Again Wikipedia concerning the underlying method: "..."...Monte Carlo methods (or Monte Carlo experiments) are a class of computational algorithms that rely on repeated random sampling to compute their results..." Sampling a base model using pseudo random numbers in huge numbers is the basic principle of MCS and that's what i do - it's not difficult to understand or? (i don't hope that you expect the exact programming code here?) Apparently Intradaybill has a problem to accept this basic MCS principle (that I don't have invented - i'm using that like all other users of MCS methods do). But that's his problem not mine. Only my two cents, zentrader
It is of course your problem because in generating pseudo random numbers for trade parameters you probably end up with outcomes of parameters that do not correspond to actual system performance. Averaging these outcomes may overestimate or underestimate distributions. You still refuse to explain what you do but instead you seek refuge behind general wikipedia statement about MCS. I have asked you specific questions. I am not asking you to give out code. Just tell us what you do with the parameters and how you get your results. You start with a set of trading system performance parameters. How do you proceed, in general terms, just outline the steps: 1. Parameters 2... 3... 4... If you don't respond, I will assume you do not know and that you basically make up the results. I am asking you straightforward questions here . It is an opportunity for you to demonstrate that you know what you are doing.
@intradaybill, see my contribution of 03-11-11. I thought, it was clear enough...? I've already said, that I use trade off ratio and pay off ratio of a historical system test as input parameters. If you put this input model of the historical episode into a MCS stress test with multiple simulation runs you get other time series of wins and losses as in your one and only historical test. That also leads to different possible results belonging to profits and risks (spec. DDs!), which are system-inherent even under same market/system conditions (profit factor unchanged etc.). A second step using MCS in this context is the possibility to create also new test data (based on the historical data) to test the system against other market conditions / time series. Both functions are using heavily common MCS methodology. bye, zentrader
Not clear enough, let us see about this one I understand that by "trade off ratio" you mean win ratio (winners over total trades) and by pay off ratio you mean the ratio of average win to average loss. Is that right? Let's take it one step at a time...