Is there a pattern to the chaos? Let's try to find it! Machine learning on the example of a specific sample. - page 6
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
Look, we have entry points - sample lines, and we have a financial result defined by exit points, namely the places where a stop loss or other signal is set. You want to enter the market contrary to the strategy, i.e. buy where you should sell, if the model says so, and for this you need to define new exit points. The question arises, if the exit point has not appeared yet, but the entry point has appeared, what should you do then - close and ask the model about the entry direction, or what?
Do not close. Check all entry points during the markup phase. The teacher must know the price of error. Otherwise, the balance line cannot be built.
Now it turns out that: we lost something. Maybe 10 pts, maybe 100, maybe 500. And so 300 times... what is the validity of such a balance line?
Don't close.
If you do not close, you will miss the signal, and it will be in the sample, i.e. there will be potentially more lines, and then the balance will not be built. I have previously done with forced closing.
As for the current approach with two marks, the balance is very close to the results of the tester, precisely due to this markup. I have already been burned on other approaches, so I believe that this method of markup is the most reliable.
Can you try to prove it?
If you don't close, you will miss the signal, and it will be in the sample, i.e. there will be potentially more lines, and then the balance cannot be built. I have previously done with forced closing.
As for the current approach with two marks, the balance is very close to the results of the tester, precisely due to this markup. I have already been burned on other approaches, so I believe that this method of markup is the most reliable.
Now you are talking about several trades opened at the same time.... Keep even 100 - the main thing is to finish them according to the markup algorithm.
I mean that each line of your dataset should be trained for each class and fill the line of the financial result with the real value. Maybe you have it that way. It's hard to understand without seeing the code.
Tell me better, what are 5000+ features you have invented?
20-30 standard indicators with 100-200 settings variants?
I'm not good at programming( maybe you could try a fiba bias for research?
There are predictors in the sample that use horizontal levels based on Fibonacci ratios.
The sample includes predictors using horizontal levels on Fibonacci ratios.
Are there price deltas in there? 50 to 100 bars. What are the column numbers? Interesting to compare by training on them alone. Suddenly they will be enough and 5000+ features are not needed.
Now you're talking about several deals opened at the same time.... Keep even 100 - the main thing is to finish them according to the markup algorithm.
I am fine-tuning everything for netting accounts - for the Moscow Exchange, and so far I have no virtual position control. Besides, for training I think it is not necessary to overcomplicate the sample - it is better when the average profit and loss per transaction is within reason - this, among other things, allows you to adequately assess the situation and not to select models where accidentally obtained excessive profits. You can already complicate the strategy of position maintenance in the Expert Advisor.
I mean that each line of your dataset should be trained for each class and fill the financial result line with the real value. Maybe you have it that way. It's hard to understand without seeing the code.
I use a very similar version of the Expert Advisor as in the previously published article - the data on the financial result is taken from the results of transactions, not calculated.
Tell me better, what are the 5000+ features you have come up with?
20-30 standard indicators with 100-200 variants of settings?
Indeed, there are different timeframes of the same predictors. The choice turns out to be large and that's why I am inclined to investigate options of selecting useful for better and faster training of models.
Most of the predictors are on the Donchianna channel and ZZ, built on it, some on the regression channel, some on MAs similar, some on parabolic, some on ATR similar, some on returns from different TFs, well, and some part of oscillators, maybe something else small. I have not adjusted parameters for a particular instrument, but perhaps I should - I have planned an experiment on ZZ on this topic.
Are there price deltas? 50 to 100 bars. What are the column numbers? Interesting to compare by training on them alone. Suddenly they will be enough and 5000+ features are not needed.
Probably close in meaning 1041-1489.
You can give me the code for the predictor, and I will make a sample with it.
I can tell you what predictors were used by one of the models - you can check if you successfully train (I have almost no doubt) - do you need to?
There are predictors in the sample that use horizontal levels on Fibonacci ratios.
Not levels, but ranges plus there wave patterns and on candlesticks. Those aren't in the books.
Well, even if they are not in books, how will I know? Describe in detail - if there is uniqueness, I will add these predictors and evaluate their effectiveness on a specific sample.