Machine learning in trading: theory, models, practice and algo-trading - page 784
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
If there are no questions to the target and most agree, let someone know and continue, if there are corrections, we listen to the adequacy, correct and move on.
It is clear that the parameters of the target can be anything, but we chose these and will work on them in the absence of objections.
Next, you need to decide and choose one of the suggested approaches or propose your own.
1. We make a forecast on the whole chart without missing bars.
2. Making a forecast at certain moments of the market, such as: The opening of Europe at 10:00 MSK. That is, at this point, we try to determine the change in the slope in 10 bars (hours) or at any other point.
That is, you must select either every bar or at a certain point in time. The choice is up to you. As we determine the above questions we will continue. Looking forward to hearing from you!!!
If there are no questions to the target and most agree, let someone know and continue, if there are corrections, we listen to the adequacy, correct and move on.
It is clear that the parameters of the target can be anything, but we have chosen these and will work with them if there are no objections.
Doesn't such a small window result in just following the trend? I.e. the trend is 100 bars - 10 windows, and 3 of them have approximately an error (internal correction), while the flat has 1 guess out of 3. As a result, if a trend was high during the learning period, then we will select the parameters for the trend, and for the flat, for the flat. I.e. the trading system will be simply focused on the market situation during the training period.
A real TS should determine itself, which situation to expect and, depending on expectations, apply this or that trading algorithm. So why not teach the neural network to predict the future market phase?
Doesn't such a small window lead to simple trend following? I.e. the trend is 100 bars - 10 windows, and 3 of them contain approximately an error (internal correction), while in the flat, 1 guess out of three. As a result, if a trend was high during the learning period, then we will select the parameters for the trend, and for the flat, for the flat. I.e. the trading system will be simply focused on the market situation during the training period.
A real TS should determine itself, which situation to expect and, depending on expectations, apply this or that trading algorithm. So why not teach a neural network to predict the future phase of the market?
As for preparing the data for training, this is a separate story, and I will tell you what options you can try. As for the result of NS in the form of predicting the change of klos in 10 bars, it is not a TS yet. The task is to prepare such a model or group of models in which at the REF segment the error in divergence from the real will be minimum and only then we will make TS with entry and exit from these results. The forecast itself is not a TS yet. For example: if the current bar the ST says that in 10 bars the change will be more than 100 pips, then we go in, if not, we do not go in. The task of the AI is not to learn this or that section better or worse, but to be able to generalize. That is, if in the training there were more trend vectors than a flat, then at the onset of a flat, the NS should still recognize it..... Let's take it one step at a time. There are two questions on the agenda. Approval of the target and the choice of the whole market or at a certain point. As you choose let me know. We will fix them as initial data and go on...
About data preparation for training is a separate story and I'll tell you what options you can try. As for the result of working of NS in the form of predicting changes in klos in 10 bars, it is not a TS yet. The task is to prepare such a model or group of models in which at the REF segment the error in divergence from the real will be minimum and only then we will make TS with entry and exit from these results. The forecast itself is not a TS yet. For example: if the current bar the ST says that in 10 bars the change will be more than 100 pips, then we go in, if not, we do not go in. The task of the AI is not to learn this or that section better or worse, but to be able to generalize. That is, if in the training there were more trend vectors than a flat, then at the onset of a flat, the NS should still recognize it..... Let's take it one step at a time. There are two questions on the agenda. Approval of the target and the choice of the whole market or at a certain point. As you choose let me know. We will fix them as the initial data and will go on...
Micha, you want a regression target, but you're saying that the simpler the better. As for me, the classification target is much easier to use. That is, for the same for example ten bars three variants of events 1 class profit happened, 2 class profit happened, 3 class no profit or loss happened....
and you no longer calculate the target itself with the number of pips (which itself introduces a distortion because of the error), but only the probability
Now get more air in your chest and exhale... Inhale and exhale again. And now reread my last posts, especially in the beginning where I said that the classification I have no questions and the project can be considered as completed. Working completed project. I got bored and decided to twist the topic of regression, so let's put classification aside and try to apply the principles to regression models. About classification not a word, except during the reference to the organization of works and any general principles.... Now we make a prediction. Are there weighty arguments against my proposed target function? There is an alternative to the target function. What can you offer. If you can not, then agree with what is and continue...
Now get more air in your chest and breathe out... Inhale and exhale again. And now reread my last posts, especially in the beginning where I say that the classification I have no questions and the project can be considered completed. Working completed project. I got bored and decided to twist the topic of regression, so let's put classification aside and try to apply the principles to regression models. About classification not a word, except during the reference to the organization of works and any general principles.... Now we make a prediction. Are there weighty arguments against my proposed target function? There is an alternative to the target function. What can you offer. If you can't, then go with what you have and continue...
I'm sure you won't get anything out of the regression version of the target function.
About data preparation for training is a separate story and I'll tell you what options you can try. As for the result of working of NS in the form of predicting changes in klos in 10 bars, it is not a TS yet. The task is to prepare such a model or group of models in which at the PP section the error in divergence from the real will be minimum and only then we will make TS with entry and exit from these results. The forecast itself is not a TS yet. For example: if the current bar the ST says that in 10 bars the change will be more than 100 pips, then we go in, if not, we do not go in. The task of the AI is not to learn this or that section better or worse, but to be able to generalize. That is, if in the training there were more trend vectors than a flat, then at the onset of a flat, the NS should still recognize it..... Let's take it one step at a time. There are two questions on the agenda. Approval of the target and the choice of the whole market or at a certain point. As you choose let me know. Let's record them as initial data and move on...
Okay, you don't want to talk about that, let's talk about something else.
Why do you take the delta of 10 bars and not the delta between opening and maximum/minimum for those 10 bars?
The prognostic probability varies around 50%, i.e. it is good if 55%, and taking into account the error that almost any NS produces is like a finger in the air. I am sure that you can hardly get anything out of the regression variant of the target.
How do you get such percentages? There are tests that have been done. We are still unable to decide on the initial data, and you are already talking about the result ....