Discussing the article: "Creating a mean-reversion strategy based on machine learning" - page 6

[Deleted]  
Evgeni Gavrilovi #:

Example of market modes. If an order opened on yellow, when changing the mode to red, will the order close?

This is very important, because the trend has already changed, and in your script it may be written in the condition to expect a stop loss

No, the trade continues to hang if there is no opposite signal from the first model. You don't know when the next mode will be switched, they can be frequent.
Forced closing of trades when switching modes worsened statistics, as far as I remember. In MT5 bot it is easy to write such a condition in the section of closing trades and check it.

Plus, price gaps between clusters take part in the markup, i.e. if there was a gap up between two identical clusters, within which there was another cluster, but it was removed, then such a deal is marked to buy, and vice versa. This explains that the TS does not work very well with short stops, because it is allowed to over-sit other unnecessary clusters. Again, this is not a flaw, but a feature of the TS. Because take profit can also be increased and it will remain profitable. But, to increase the number of trades, I make the take profits shorter.
 
Maxim Dmitrievsky #:

Plus, price gaps between clusters are involved in the markup, i.e. if there was a gap up between two identical clusters, within which there was another cluster, but it was removed, then such a deal is marked to buy, and vice versa. This explains that the TS does not work very well with short stops, because it is allowed to over-sit other unnecessary clusters. Again, this is not a flaw, but a feature of the TS. Because take profit can also be increased and it will remain profitable. But, to increase the number of trades, I make the take profits shorter.

For the purity of the experiment it would be necessary to take equal takes and stops - the number of trades would not become less, and over-sitting (and shifted evaluation of correctness/profitability of signals) would be excluded.

[Deleted]  
Stanislav Korotky #:

For the purity of the experiment it would be necessary to take equal take and stops - trades would not become less, and over-sitting (and biased evaluation of correctness/profitability of signals) would be excluded.

The article was written not to get very useful tips, but to show an interesting approach. The code for independent experiments is given.

No matter what code is given, there will always be those who know better "how to do it correctly". Please do not post such things again.
 
Maxim Dmitrievsky #:
The article was written not to get very useful tips, but to show an interesting approach. The code for independent experiments is given.

The approach is interesting and correct - nobody disputes it. I don't understand only why the default settings for the article are chosen to present the results in a disadvantageous light.

PS. While I was writing my reply, there was a note saying that you can't post something like this. I apologise if this is not what you want either.

[Deleted]  
Stanislav Korotky #:

The approach is interesting and correct - no one disputes it. What is not clear is why the default settings for the article are chosen to present the results in a disadvantageous light.

PS. While I was writing my reply, there was a note saying that you can't post something like this. I apologise if this is not what you want either.

Stop settings do not affect the generalisation ability of models, only their selection. What last ones were left in the code, that's what I used. I used them for my bots. There is no hidden message in it. There is no emphasis in the article that these or any other stops should be used. They can be removed at all, it is not crucial in the context of the discussed topic.

There are no prohibitions to use other instruments/currency pairs/TFs/training and testing periods.

For example, you can see for yourself that this approach works worse on trending instruments.
 
Maxim Dmitrievsky #:
(1)Stop settings do not affect the generalisability of models, (2)only their selection.

Why do (2) if (1) is asserted?

If I understand correctly, models without SL and TP are trained, but those that perform better are selected, in this case, when SL = TP * 10. Strange, it would be more logical to give/show Python test results without SL and TP in the article, so that (1) is clearly visible. Ah yes, you can't see equity on python charts anyway, what's the use.....

PS. (1) and (2) in the quote are set by me.
[Deleted]  
Andrey Dik #:

Why do (2) if (1) is asserted?

If I understand correctly, the models without SL and TP are trained, but those that perform better are selected, in this case, when SL = TP * 10. Strange, it would be more logical to give/show Python test results without SL and TP in the article, so that (1) is clearly visible. Ah yes, you can't see equity on python charts anyway, what's the use.....

PS. (1) and (2) in the quote are set by me.

Those who read the article will find a nice bonus at the end - a chart from the terminal, which shows the drawdown.

 
Maxim Dmitrievsky #:
Instead of whining, all you have to do is run it and try it. This is roughly what an answer to such questions should look like. Training in one cycle/export/compile takes a minute. Writing a nonsensical comment like this takes 30 seconds. Some more insights come during, "Oh yeah, oh right, here I am a commenter".

What is it? Where is the irritation coming from? Where are the substantive answers?

I stopped running Python codes from articles as soon as it became clear that what is described in articles is only a small part, and the rest is done and developed by anyone but the author of the article.

You wrote an article? - be kind enough to answer questions instead of writing 3-second meaningless replies.

[Deleted]  
I don't even know how to respond to known "troll commenters". I tried different options.
 
Maxim Dmitrievsky #:
I don't even know how to respond to known "troll commenters". Tried different options.

Do you have any thoughts on how to improve the TC? Share)