Is there a pattern to the chaos? Let's try to find it! Machine learning on the example of a specific sample. - page 9

 
elibrarius #:

At 448 bars, the best is 0.03000.


By the way, you can take the volatility predictors in the code from the article, so I see - they are really often used by the models.

elibrarius #:
I calculate profit and chart after training.

Well, it is logical, but then it turns out that it is not realistic - you can't know from the beginning what the profit will be.

 
Aleksey Vyazmikin #:

Well, logical, but then it turns out that it's not realistic - because you can't know from the start what the profit will be.

As the model predicts, so will the price. It is possible to calculate on the exam sample and look at the chart, but in real trading we will hope that the model has been trained on long-acting signs.

 
Anyway, thank you.
I haven't done the MO for six months. I'm kind of into it now, thanks to your enthusiasm and interesting experiment. I'll continue my experiments now.
 
elibrarius #:
Anyway, thank you.
I haven't done the MO for six months. I'm kind of into it now, thanks to your enthusiasm and interesting experiment. I'll continue my experiments from now on.

You're welcome!

By the way, I can remove predictors for your target - for this I need an array of input dates - add them to yours and see if all these predictors make sense. If you give me a column with markup, I will also try to find informative predictors for your target.

 
Aleksey Vyazmikin #:

You're welcome!

By the way, I can remove predictors for your target - for this I need an array of input dates - add them to yours and see if all these predictors make sense. If you give me a column with markup, I will also try to find informative predictors for your target.

I don't have a target selected. Everything I tried was not interesting in conjunction with the used chips. That's why I almost abandoned MO.
Thanks for the offer to test your chips.
Let's try EURUSD H1 - all bars. I will use either Alpari or MQL demo server (If you have Alpari, it is better to upload the features on it). If you have another server, then add Open price, to check in case there will be jumps with time zones and winter/summer time.

Target no, will try different ones. The same combinations of TP/SL can be dozens and each must be tested. So there is too much work to do with selecting informative features. I will see what training on all of them will give. And then maybe I will search for them myself.

 
elibrarius #:

I don't have a selected target. All that I have tried did not interest me in conjunction with the used features. That's why I almost abandoned MO.
Thanks for the offer to test your chips.
Let's try EURUSD H1 - all bars. I will use either Alpari or MQL demo server (If you have Alpari, it is better to upload the features on it). If you have another server, then add the Open price to check in case there are time zone and winter/summer time jumps.

Target no, will try different ones. The same TP/SL combinations can be dozens and each one should be tested. So there is too much work to do with selecting informative features. I will see what training on all of them will give. And then maybe I'll search for them myself.

Here is a sample of hour openings from the MQ server - I haven't used Alpari for a long time - their moves from host to host are scary.

One is a sell deal - respectively zero is a potential buy deal, columns can also be used modularly.

 
Aleksey Vyazmikin #:

Here's a sampling of the opening hour from the MQ server - haven't used alpari in a while - their host-to-host moves are scary.

One is a sell deal - accordingly - zero is potentially a buy deal, the columns can also be used modularly.

Thanks!

 
elibrarius #:

Thank you!

You're welcome!

Let me know the results :)

 

Hmm, with the last sample, the accuracy turns out to be 54%.

The balance is very strange, I won't post it yet - very much like a mistake in terms of interpreting the results...

 

I ran the tests in the terminal

Here is the report without using the model

It turned out that the accuracy of my calculations is not quite correct - I will look for the reason - probably the financial result differs a bit. Maybe spread - I don't know.

Below is the chart with the model applied. Only sales are opened by the model, this is due to the fact that the model has a small Recall. Maybe it makes sense to train the model for buying separately - I will try.

Below I made a pass with closing on the signal (entering a possible reversal zone) - the percentage of profitable trades has slightly increased.

I can see that the chart is not bad until the end of November 2021, and then there is a breakdown and fluctuations - apparently, there are violations of patterns - the market began to change. Whether it will stabilise or the pattern will stop working is an interesting question.

What did you get? What was the previous best result for this target?

I think it is necessary to split the sample to identify similar entry points, it can improve learning.