beathespread EA systems
Here I will show you two shots. Both if the shots are out of sample. The training was from
16.01.12 - 23.02.12.
The out of sample testing is from 24.02.12 - today 05.03.12
One of the systems is the Spinal Implant EA. The characteristic of this system is to enter into a trade only when there is larger Hurst exponent. In practice it tries to catch the break - outs and looks very much what an experienced price action specialist would do.
So the characteristics of this system is to know when to trade.
When the decision is taken it is upt to the Neural net to take a directionnal decision.
 |
| От 08 ноември 2011 |
The second EA is using the ASCTrend expert as a basis but it is a little bit more sophisticated because it is using radial basis activatin function in the neurones in order to adapt the indicator settings. This expert is not using the Hurst exponent so it is more vulnerable than the Spian Implant to adverse market conditions. However during the out of sample the market did very good swings and the expert was able to catch them very well. So with this expert you need to monitor the market state carefully. The worse that can happend is ranging and volatile market state, the expert will open a loosing after loosing trade.
 |
| От 08 ноември 2011 |
Comments
Here I would like another shot of out of sample results of the Spinal Implant EA. However in this EA, I am using experimental activation function for the percetrone. And the whole project is not ready for release yet, as it is in the experimental stage.
Here I am adding a continuation of the signals given by this mod of the EA. Needless to say, I was glad to see those results with my experimental neurone activation function.
Here I add the last results with the trading results and the equity curve. Everything is out of sample trading results. The data feed is Alpari (I use alpari demo account under fxcm metatrader).
Here I add the last results with the trading results and the equity curve. Everything is out of sample trading results. The data feed is Alpari (I use alpari demo account under fxcm metatrader).
Is that strat test or live demo? how does it perform on live demo?
This is strat teston ticks with 90 %. There is not a substantial difference with the tests with open prices only. And I use the open prices in order to to the training (Otherwise I will train forever).
In order that my tests results to be relevant with the open charts, the stop loss has to be sufficiently big.
So if the maximum average hour volatility for the last 10 weeks on EUR/USD is 40 pips, having 45 pips as a stop loss on 15 bars is giving some safety margin for the quality of the modeling.
As in theory the volatility is proportional at the square root of time we can quickly can deduct that (we make some quick and dirty assumptions for distribution here):
15m = 40 * sqrt (1/4) // 15 m is 1/4 of 1h.
20=40*0.5
We may need more elaborate tests, but this is a concern that cannot be underestimated. If the result between the tests with open prices and with the tick quality is the same everything is OK. (Anyway we are amateurs, professionals would use confidence intervals everywhere).
It performs well on life demo. However still more tests are required, any supervision of the code is very welcome.
You can go to the site of wolphram alpha and ask to fit gaussian distribution with mean 20 and standard deviation of 9.2.
1 click.
So basically 40 pips volatility is an extreme event (that do happens often especiallt during new releases), but this is enough.
Having a look at the distribution you cannot model safely stop losses with open prices below arround 30 pips on 15 time chart (95 % of the time the volatility will not exceed 29.1 pips in the 15 m bars).
You can easily compute the statistic by using the default atr indicator and the standard deviation indicator.
What you need to use is to apply the standard deviation to the atr readings and not to the close.
With that the recent average true range over 2400 bars is even lower 13.9 pips.
(Maybe this is worth a whole blog post)
40 pips SL is ballpark.
What Take Profit /Trailing Stop/Jumping Stop are you using?
also which versions are above should i be running v3_4 with 2 neurones or 4 neurones?
You need to chech for yourself. I prefer to use the kernel trick with the kernel functions.
The best fermormanve I get with 40 pips stop loss and 35 take profit.
However it also works with 30 pips stop loss and 30 pips take profit (I mean when run on open prices the result is identical as for 90 % of modelling quality).
ok. ive pulled down the indies and have applied them to my chart. investigating with entropy what i will use and lynch some of the code and apply over my base ea template.
The percetrone function is just a few lines of code sitting at the end.
So I consider it as a tool that can be applied everywhere you need it.
The idea is that messing with the mathematical functions and making mathematical simulations (metatrader 5) you gain an insight how they work. And that is what is really precious at the end.
I have explained how the hyperbolic function may have you killed if a trend comes. However it may be more powerfull to model a flat market (imagine that the indicators are lagging all the time, however the hyperbolic tangent function may compensate that).
There is a realationship between the hyperspace formed by the transfer function, the indicators used, the strategy and the market state.
The percetrone function or the kernel requires just the very basic programming skill to understand.
And last but not least I think the use of kernel functions is really the way to go with metatrader allowing us to optimize multiple parameters and still find interesting solutions.
your wrote "There is a realationship between the hyperspace formed by the transfer function, the indicators used, the strategy and the market state." This is for me intriguing.
and we need another kind of indicator, to put inside a perceptron like STO or PFE.. may be AC. or Chaiken Money Flow... tell me more
Yes indeed, as you see from the design in the EA, if your hyper space has negative values that means that if your indicator readings are positive you may end with negative output from the kernel or perceptrone function.
So when we are scalping in the direction of the momentum this is not something we may need to counter the trend and the momentum.
However imagine the situation in EURCHF when the price is antipersistent, that is not a problem at all, increasing the possiblilities of the percetrone function may allow to achieve more profitability.
And yes in the kernel variant for the EURCHF (not released yet), the use of a kernel increases the profitability. What matters is that we do not predict we are looking for a solution that will boost the profitability.
Here I would like to show the last chaos kernel out of sample results with maximum available modelling quality. We do not make assumptions that the solution is an analytic function, we may have an arbitrary function it does not matter. Unfortunately I can't prove that in any way this is highly experimetal.
I get invalid integer number as parameter 2 for indicator call function.... anyone else.
Here is the out of sample performance of another mod using Gauss kernel but I implement a string equation to determine the best period of the pfe, and I use delayed versions of the same indicator.
To tell you the truth I was expecting to nail the break - out from yeterday but it was detected a little bit late, as you know the algorythm is waiting for two consecutive bars to be with fractal dimension below 1.5 (iVAR below 0.5). That is to avoid whipsaws.
The other versions are OK too but this is the best by now.
When I look at the entries it looks like it mimicks the entries of human price action trader. I just got that impression.
I continue to analyse the Spinal Implant EA systems. This shot is from the out of sample analysis. Yes, it was a hard time loosing during the range but this is a part of the system, the idea is not to loose too much. When the market states drifts toward more predictable states the EA system starts to perform really well.