How does Monte Carlo work exaclty in NT7 ?

[TakisTakis]

Hello NT friends,

I was looking through http://www.ninjatrader.com/support/h...simulation.htm Monte Carlo manual and there are some important questions to be asked:

It says:

Monte Carlo Simulation will randomize your trade results over and over again in multiple simulations to provide you with a normal distribution of simulation performance.

How does it do that exactly, if I may ask ?

a) Does it add a normal distribution produced variable on top of the actual data ?
b) Does it use only a normal distributed variable and ignoring the present data ?
c) Does it mix the sequence of the data ? ( candlestick 4 - candlestick10 - candlestick22 etc )

... or something else ?

If the case a and b are true, is there any way for me to use another distribution ?


I would be very happy if someone could answer the above questions.

Thanks in advance

[NinjaTrader_Tim]

Hi TakisTakis,

The Monte Carlo does a randomized mix of the orders and creates a normalized curve as described at the link you provided. So, there are no other distribution types or options available.

[TakisTakis]

Could you please elaborate on that ?

First of all thank you for attempt to help me.

However, the answer does not really clarify the situation and attention to detail is really key here.

a randomized mix of the orders and creates a normalized curve as described at the link you provided.

The link doesnt really describe with details how this happens. It just has the cryptic quote as mentioned below.

What exactly do you mean with saying randomized mix ? It mixes the sequence of the orders ? It mixes the quantity ? It mixes the sequence of the bars ? It mixes the prices ?

If we cannot understand the exact function of NT's monte carlo application, then it is not much of a use as it can lead us to wrong decisions that will cost us money.

Sorry, for pushing for details and I will understand if you do not have the knowledge on this very new feature. In that case please just pass a demand the developers to add more information in some future version of the manual as, as it is it is not much of a use.

Thanks for you quick answer and sorry if I ask too much

[NinjaTrader_Tim]

Hi TakisTakis,

Thank you for your deeper interest in the subject, the Monte Carlo just randomizes the orders, and the sequence in which they are stringed together.

This is described in detail here, as well...

http://en.wikipedia.org/wiki/Monte_Carlo_method

[TakisTakis]

Thank you for your answer, however MC is known of randomizing the prices above all.

[NinjaTrader_Tim]

Hi TakisTakis,

I see, and understood. In this case, the prices are not manipulated away from the historical prices.

[fury661]

@TakisTakis

Actually, it's a commonly accepted industry practice to expect a Monte Carlo simulator to randomize the sequence of orders. No serious trader would expect the MC sim to "randomize" prices. That would make your system useless.

For instance, growth stocks have price patterns that are different than currencies. Oil futures behave differently than tech stocks. Indexes behave differently than commercial paper yields.

If you randomize price... you lose the unique characteristics of the instrument you are testing.

I think you need to check your assumptions of what an MC simulator is used for as it relates to trading.

You should read Martin Pring's "Trading Systems Explained". I've found that book to be quite helpful.

[bunyip]

Hi Takis Takis, I've only just started using NT but have decades of experience with Monte Carlo and randomisation methods. I'm stunned by the lack of understanding in every post and blog I've read about the NT MC test - including those by the developers. First, the MC test is designed to test whether or not your system performs better than if you were to simply place trades at random. Second, it makes no difference whether you randomise the order of price data or trades - the net effect will be the same. Third, and most importantly, the MC graph that NT provides is of little value unless the result of your original system is included.

An example: Lets say your trading system gave you a cumulative profit of $1000.00. You then run the MC simulation and plot the graph for the cumulative profit. If the curve of the graph intersects the point for $1000.00 at 50% on the X-axis, then you can say with confidence that your system performed as well as a system that placed trades completely at random. On the other hand, if the intersection was at 95% or higher, then you have evidence to suggest that your system performed significantly better (in a statistical sense) than if you had placed trades at random. You do not want to trade any system where the intersection occurs lower than about 95%. If you did, you may as well place trades by flipping a coin. At least that way you'd save yourself the time you would otherwise spend develop trading systems. The only exception would be if the intersection was <= 5%. In that case, you could consider a system where your trades were actually the opposite to what the system said, i.e., you buy when the system says sell, etc.

The way the results are presented in NT are, in my mind, not helpful. Instead of a backwards cumulative frequency distribution (i.e., the axes in NT are reversed - the statistic should be on the X-axis and the probability (or %) value on the Y-axis), the developers should change it to a standard histogram and they should include in that histogram a vertical line showing the value of the statistic from your original strategy. At least that way, anyone who has studied basic statistics and who has seen the normal distribution will be able to interpret the results much more easily. As it is, it's quite confusing and the documentation for the MC test is simply misleading.

For what it's worth, I do all of my system development in the statistical programming language R. I then bring those results into NT as a custom indicator and use the NT order handling features as a final test of profitability. I always run MC tests on my systems in R. To date, every system I've developed that passed the MC test gets wiped out by transaction fees and slippage. Still, it's a hobby...