Machine learning in trading: theory, models, practice and algo-trading - page 3177
![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
Yes
I.e. I need to generate a target binary for a sample, let's say, and see how often quantum segments will be found by my method for different predictors, and so 10 times?
If the number of quantum segments will be found about the same number on average as it is now for all predictors, then the method does not work, do I understand the thought correctly?
Tail and other parameters of the series will remain the same. I think that's a plus.
There is a risk of getting bogged down in pointless arguments again. What is the difference between a randomly found set that works on oos and one that was invented through the hardest mental suffering, but also without fundamental justification? When the method of validation is the same. Rhetorical question.
What is the difference between a random search and a search with an element of randomness of choice? ))The goal is to detect rule breaking, and to do that you need to know what they look like in the correct form. In other words, what statistical descriptive features can indicate that it is indeed a pattern. If you know them, then deviation from these rules would be a signal to stop the model.
And, assembling a model from such rules still seems to be a more stable solution than assembling pure randomness.
A couple of rules that give the main profit may get into the random, but then the market may forget about them for a while. With random it is more likely to panic and switch off the model at the beginning of a drawdown, but with the approach I suggest - waiting for the order of rules to change, i.e. a drawdown can be recognised as a normal phenomenon.
And the effect described above can be clearly seen on the gifs that I made, where in one case the sample test in plus, and in another case - although the rules are all taken from the sample train, but their (rules) appearance is not uniform in time.
The bonus of the approach, ideally, is to avoid exam sampling altogether.
An alternative could be deep model estimation - there is some work in this direction as well.
As a result, using only a random model:
1. We don't know why it works.
2. We don't know why it stopped working.
3. We don't know how to "fix" it.
4. We don't know if it's time to stop trading.
5. The lifetime of a random model will be shorter because it contains erroneous rules.
At randomisation, I personally get about the following - out of 100 models 30 of 30 are selected within half a year in the profitable zone about 10. I want at least 50% to 50%, then we can make partfiles.
Wouldn't it be enough just to mix the column with the target column?
Tail and other parameters of the series will remain the same. I think that's a plus.
Can you elaborate - I don't get it.
The question was rhetorical
You have no need to understand what I am explaining, and for that reason I have no desire to waste energy on re-explanation. Stay with your opinion about the identity of the result with different approaches.
You have no need to understand what I am explaining, and for this reason I have no desire to waste energy on a repeated explanation. Stay with your opinion about the identity of the result in different approaches.
I.e. I need to generate a target binary for a sample, let's say, and see how often quantum segments will be found by my technique for different predictors, and so 10 times?
If the number of quantum segments will be found about the same number on average as it is now for all predictors, then the method does not work, do I understand the thought correctly?
Well yes, the point is to repeat your procedure many times on a large number of obviously meaningless problems. Then see how the application on specific real data looks relative to them - if it doesn't stand out much, then the method is bad. Usually formalise by counting some number on each application of the method and draw a sample, and then see if the number calculated on real data falls far into its tail - if yes, then ok.
Behold, for if you are not given a pattern in the original series, the Hilbertian path will not lead you to your cherished goal. Your endeavours will turn into devilishness, and you will find an ignominious slaughter instead of paradise.
The problem of sectarians is the fear of testing their religious dogmas.
There are always many patterns - the question is the right choice.
Wouldn't it be enough just to mix the column with the target column?
Tail and other parameters of the series will remain the same. I think that's a plus.
It's not bad either. It's probably better to try both methods. If there are a lot of outliers in the traits (heavy tails), the results may be different, which may give additional information.