Mixing systems to reduce risks

Hi, I want to share my thoughts about this topic, maybe others can give me new ideas or point at the errors of my reasoning. With luck the discussion will be of interest and benefit of others.

I'm not a professional trader, I have no experience in the markets, prices, symbols or indicators. I'm a software professional and run only EAs in my accounts.
For my real accounts I NEVER run a single strategy. I do have several (maybe) positive strategies in store and I always thought that using more than one of them in an account is better.

The idea is, to use a biological simile, that when you do have different strategies running at the same time the account is like an ecosystem where the resources are limited (the resources are the free margin and the acceptable drawdown). The different strategies use the resources to produce the output (profit) and you have interest not to put in the same system two strategies requiring margin (operating) just at the same time or that are getting drawdown when the market starts behaving in specific way.

Currently what I do is to maintain an excel sheet where my systems are listed and for each one I have columns with the typical and maximum used margin and drawdowns. For every system I then decide how trend (T) or counter-trend (CT) it is. For instance systems based in MA crosses and going for a high RR ratio are 100% trendy. And, opposite, systems that buy when price is well under an average expecting the price will go back to the average are 100% counter-trendy.

To decide how T/CT the system is I correlate the drawdowns in time of the system with pure T and CT systems.

Once I have all that in the excel it's quite simple to plan the MM for a combined account. Just for math lovers, I use the scalar product of the 2-dimensional system vector of coordinates C & CT.

I'm sure some of you have thought before about this. What do you think? Is it complete non-sense??
t.

Sorry, I forgot to mention that in my case I only worry with T/CT because all my systems operate more or less the same way (multi-symbol, all time, similar operation times, similar TP/SLs).
Of course other 'decouplings' naturally exist just using a system in a symbol and another one or the same in a completely uncorrelated pair.

t.

There is no problem using multiple system at the same time. Since you don't have experience in "currency correlation", you should trade "only 1 pair at all time".

Sure, but my point is that it is better to use some 'special' combinations.
If you have a system that sells when overbought using RSI and another one sells too when price gets far from an average you can expect both systems to couple many times. Those times you are using more margin and risking more DD.
If you combine two systems that dont 'couple' you are optimizing the resources.

t.

In these posts #14 and #15 I propose a solution to modulate the money management to trade N correlated pairs for which you know the pair-wise correlations. Substituing the pairs by the trading systems you should allocate the resource (risk and margin) to each system the same way. You need to know the matrix of pair wise correlations of our systems.

I only found a solution relying on backtest to compute this correlation: you sample the state of each of your systems during a backtest and attribute a value +1, 0, -1 depending on the system is long, flat or short. Then you compute the pair-wise correlation of each of the systems.

Well during the time when both system display "same sell/buy signal" then you should just split the risk in "half".

So if you usualy risk 5% in 1 trade... then when you bet 2 trade because of 2 system signal at the same time. You only risk 2.5% risk each trade...


In forex what matter is your win, so if you want to complex thing out deal with one system say sell, and the other system say buy... then it's up to you.
You might "reduce loss because of the complex signal you to stay out of the market"...

PipMeUp, yours is a much smarter system than mine, Sir.

That's what I'm going to do:
I already have the 'expected-monthly-equity-increase' value for every system I run. That's a positive value, of course! (but it is only expected, unfortunatedly).

That value is equivalent to the slope of a linear regression on the equity evolution curve for each system.

I'm going to create the curve of the actual equity grow MINUS that linear regression historically and that will give me the historical deviations from the average for every system.

Those graphs I can treat as 'price charts' of different symbols and I'll calculate the correlation factor between each pair of them.

And, to optimize the mix I use in real accounts I'll look for systems that are (-1)-correlated. That way I'll be using concurrently systems that not only don't couple, but that help each other.

t.

Sorry victor_ray, I'm not sure I understand (maybe it is the language, I struggle with English and lose many subtleties).

I only trade automatically and when a system says SELL and another one says BUY I just let them both run, ... and smile because it's difficult both systems lose. On average both systems win or one of them wins and the other gets break-even more frequently than I experience a double stop-loss. During the trip I've reduced the margin requirements and that's inherently a good thing.

t.

In my views mixing more than 1 system will totally confuse you and you will suffer from losses.Trade only 1 system in order to make money from the forex market

I have found some excellent discussion somewhere else:

Sharpe sucks, profit factor sucks as well. They don't tell whether you're beating the market.

Everyone can win on a bullish trend - just open and close long positions.

Use Student's two sample t-test: http://en.wikipedia.org/wiki/Student's_t-test#Two-sample_T.C2.A02_test

This will show you how random your trading is, therefore if your idea on what you are doing has anything to do with the market you are dealing with.


Additionally, of all above metrics I use always t-test statistics ,which is a very well-known indicator that is commonly used to determine if a trading system’s observed profitability is due to chance or not. It is applied on the average trades net profit and determines if the average net profit is significantly greater than zero at a specified confidence level. The hypothesis for t-test is:
Ho: μ(profit)=0,
H1: μ(profit)> 0,
The t-test is based on normality and independence. Typically, according to the Central Limit Theorem, if the sample size is at least 30 (30 trades), the distribution of the sample mean approaches normal.

are you really using t-test on the samples of trades? May I ask you what t-values you usually get? Do you reach 6? What markets?

(I mean Student's t test with taking into account the number of trades and the number of rules and/or restrictions of the trading system).

I was forced to involve this measure for trading in forex markets where the leverage is as high as 1:100. In that case you may have 100-300% annual profitability but when you apply t-test you may have t value less than 1. In backtest the best performance was in GBP/JPY which was 6.68 (5.2 on average in different datasets and testing). This was succeeded by involving 6 committees of different neural networks (each committee consists of 15 neural networks trained for different purpose ) using 5 min timeframe.
Of course this a backtest which is not taken into account the slippage - only the spread value. The system is expected to be set on real time account this September. I expect less profitability on real time, but I will see.
Generally, I use this metric as a filter (valid/not valid) - not for selecting the best strategy.