![MQL5 - Language of trade strategies built-in the MetaTrader 5 client terminal](https://c.mql5.com/i/registerlandings/logo-2.png)
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
For example, without the filter, optimization for the whole range of basic parameters gives an average profit of 100k and 35% of losses (loss of deposit) with an average of 500 trades. After using the filter, the average profit is 10k and 5% losses with an average of 250 trades - the profit decreased by 10 times, and the percentage of bad outcomes by only 7, while the number of deals has decreased by 2 times - is it good or bad? Which is the higher priority?
Suppose that an experiment is carried out. The subject is to study the effect of some factor. In this case it is the effect of a filter.
Two samples are compared. The first sample is the initial set of results of all the transactions, the second sample is the filtered results.
The null hypothesis is that the filter didn't work, the alternative hypothesis is that the filter did filter something out.
Alex, imho, misses the initial step of estimation - drawing distributions of results of all deals and searching for outliers and then removing them. Otherwise an indicator such as "average profit" is not representative.
Dennis Kirichenko:
Alex, imho, the initial step in estimation is missing - building distributions of the results of all deals and looking for outliers, and then removing them. Otherwise such indicator as "average profit" is not representative.
How do you propose to look for these variations? I haven't yet understood fully what it's about, but I suspect it's about data normalisation? If it is simply about sifting out illogical results, which do not fit into the framework of the theory of strategy, then it is assumed that this step has already been taken, but if it is about excluding the highly prominent results, then it is necessary to have a justification for their exclusion.
While I, on the contrary, try not to exclude unattractive, but acceptable variants, the idea is that the market is volatile (as a rule, the variety of its behavior can be seen on different instruments).
How do you propose to look for these outliers? I don't fully understand what it's about, but I suspect it's about data normalisation? If it's simply about sifting out illogical results that don't fit into the framework of strategy theory, then it's implied that this step has already been taken, but if it's about excluding highly salient results, then you need a justification for their exclusion...
Imho, methodologically it is more correct not to look for some indicators to describe the results of optimization and then play with them but compare 2 samples - before and after :-)
And the samples can be composed of any result criteria. I wrote above about results of deals. But you can take others... By the way, it's an interesting task...
So there is a statistical procedure. Yes, that's right - get rid of population elements, whose values were too big or small. Yes, it is variable. But that is another question. The idea here is simple - we need to assess whether or not the filter affected the performance of the strategy.
So I actually do it by averaging all indicators - this way outliers are leveled, right? I make a convolution of optimization results (with a breakdown into two attributes - but it's for clarity - one of them can be removed - i.e. one of the options will be obviously worse) and compare the convolution (there are maximum and minimum values of some indicators and their structure as percentage of outliers - and you still haven't looked at the file) just to level out the effect of emissions. Besides, in aggregate I do a comparison of 13 currency pairs separately by buying and selling. As the result I compare 13*2*2=52 averaged results of optimization. If the filter is working globally, it should work for all pairs, i.e. there should be a positive trend.
I think it is methodologically more correct not to look for some indicators to describe the optimization results and then play with them, but compare two samples - before and after :-)
And the samples can be composed of any result indicators. I wrote above about results of deals. But you can take others as well... By the way, it's an interesting task...
Samples are not all, but a part - what attribute to take to select this part?
If really interesting, I'm ready to post for detailed analysis the results of optimization - for example one pair without filters and with filters - comparable data.
Then, it is important to see the shape of the sample distribution. If it will be very different from normal, for example having a couple of peaks, it's not good... so that's it, in a nutshell...
Samples are not all, but part of them - what attribute should be used to distinguish this part?
If really interested, I'm ready to post for detailed analysis the results of optimization - for example one pair without filters and with filters - comparable data.
Yes, you can look at it.
Alex, first you need 2 samples(optimization results):
1) these are passes without filter;
2) these are passes with filter (only here filter parameters should not be optimized, but should be fixed).
No. Outliers are taken into account. And they can greatly affect the average. And this is fundamentally wrong.
What if the emissions are not accidental?
Yes, you can see.
Alex, first you need 2 samples(optimization results):
1) these are passes without filter;
2) these are passes with filter (only here filter parameters should not be optimized, but should be fixed).
The file contains optimization results without filter and with filter 7 filter variable positions - split into separate sheets, so there should be no problems with their parsing, and different settings will give an answer to the question how much to filter as well.
Picked the worst option - EURUSD M15 pair for 2013-2015 - buy only.
Without filtering:
With filter - the filter's 7 position switching boundaries are distinguishable on the chart:
Dennis Kirichenko, is there enough data or has your interest faded?
Dennis Kirichenko, is there enough data or has your interest faded?
I'll have a look at it the other day, sorry, I've been very busy...
Good, I think it will be interesting and informative!
Let's first calculate the descriptive statistics. In particular, I want to know if there are outliers.
There are outliers. It concerns the most loss-making trades with the loss exceeding 58,351.76.
Histogram.
Ideally, you should remove the emissions and analyse them further. But there are quite a few outliers here. And removing them is more of a methodological trick. According to the current form, the distribution with outliers does not seem to be normal, which means that there is some factor that influences the result more than others.