Machine learning in trading: theory, models, practice and algo-trading - page 881
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
I have a question about the target variable.
If our target variable is a financial result of a trade, then it is reasonable to normalize this result, as I thought. But here I'm looking for information on the site, and everywhere it says that the target variable must have two values - buying or selling. And if I will have a loss in any case - buy or sell (and it turned out so!), then why should I cut all negative variables? And if it is the presence of negative options that affects the statistics?
In general, I would like to know what networks work (and where to get them?) in the extreme case with a trigger - buy/sell/never do anything, and in the best case with a function (earlier I asked for a function here since I was looking for a theoretical solution, but now I made a script that summarizes predictors) that does the ranking.
> If our target variable is financial result of a trade, then it's reasonable to normalize this result, as I thought
I don't normalize the targetets, I use them as they are (price increase). If you use neuronics, then it's better to normalize predictors (another name - inputs, chips). For the forest somehow I don't really need to bother with normalization, it will work fine in any case.
Note that neuronics very often have an output that also goes through an activation function, and can only be within (0;1) for a sigmoid. Then the target also needs to be normalized if it does not fall within this interval. Or you can remove the activation for the exit so it may take any value.
> But here I'm looking for information on the site, and everywhere else I'm talking about the fact that the target variable must have two values - buy or sell.
This is called classification. When instead of price, it's just some set (0 and 1; -1 and 1; "buy" and "sell").
When you predict the price itself or its increment, it is not called classification, but regression.
> If I will have a loss in any case - buy or sell (it turned out so!), then why should I cut all of the negative choices? And what if it's the presence of negative options that affects the statistics?
It all depends on your predictors, you can only find out the answer experimentally, by trying both options. For example I tried to create my own fitness function for forest - I built a trade diagram (taking into account the spread) using forest forecasts and used the diagram to define the sharpe ratio, that was the value I tried to increase as a result.
> In general, I would like to know what networks work.
I am currently taking open prices and using them with indicators (homemade) to create new features. I train neuronics which predicts price increase per bar. I need a lot of time to create new indicators, otherwise the model will not beat the spread.
I've looked through your files and see that you already have a lot of predictors. If your targeting is just a set of -1,0,1 - use the forest. If you want to predict the price - use neuronka instead.
I misunderstood... yes, the points where the maximal profits are possible of course.
for the purpose of classification, a multilayer perseptron with a softmax output layer (outputs probabilities of belonging to a class)
Have you read this? https://www.mql5.com/ru/articles/497 on the example of the 1st neuron. And then imagine that there are many of them, that's the whole neural network
There just described the threshold functions you were asking about.
Well, it looks like a good result, yeah.
Missed that article - thanks for pointing it out, something cleared up! But not everything at once - you have to read these things a couple of times... I understand about spreading out the coefficients and listening to their sum with a function.
Tried the 1st file, split it into 3 parts:
Training
Predicted
Actual 0 1
0 28107 1244
1 3045 4119
Test 1
Predicted
Actual 0 1
0 5950 356
1 742 776
Test 2
Predicted
Actual 0 1
0 5945 333
1 779 769
Calculated on nnet with 10 neurons in the hidden layer (NS of Rattle package from R)
Worse than your scaffold, but not bad either. Second file will probably have the same results.
Thanks! I think these results can be used as a filter, i.e. a ban on trading - since guessing zeros is more stable.
I missed this article - thank you for pointing it out, something cleared up! But not all at once - you have to read such things a couple of times... I understand about spreading out the coefficients and listening to their sum with a function.
Thanks! I think that these results can be used as a filter, i.e. a ban on trade operations - since guessing zeros is more stable.
Well there are more of them in reality - that's why it's easy to guess)
> If our target variable is the financial result of the trade, then it is reasonable to normalize this result, as I thought
I don't normalize targets, I use them as they are (price increment). If you use neuronics, then it's better to normalize predictors (another name - inputs, chips). For the forest somehow do not really need to bother with normalization, it will work fine in any case.
Hm, and on the contrary, I thought, that for the forest is more important to make the predictors logical, like yes / no, which I did and that's why there are a lot of them, and so it would be much less. I'll try to do otherwise - to give one predictor 8 values, and assess whether the result will change.
It should be taken into account that very often the output of a neuronkey also passes the activation function, and can only be within (0;1) for a sigmoid. Then the target should also be normalized if it does not fall within this interval. Or you can remove the activation for the output so that it can take any value.
How do you take away the activation for the input? I don't know what it is at all - activation...
> But here I'm looking for information on the site, and everywhere it says that the target variable must have two values - buy or sell.
It's called classification. When instead of price - just some set (0 and 1; -1 and 1; "buy" and "sell")
If you predict the price itself or its increment, it's called regression, not classification.
Yes, but if I need not Buy or Sell, but just a signal to skip, then what should I do? So I decided to make two separate NS, to take into account the moment - not to trade.
> If I will make a loss in any case - buy or sell (it may be so!), then why should I cut all negative variants? And what if it's the presence of negative options that affects the statistics?
It all depends on your predictors, you can only find out the answer experimentally, by trying both options. For example I tried to create my own fitness function for forest - I built a trade diagram (taking into account spread) using forest forecasts and used the diagram to determine sharpe ratio, that was the value I tried to increase as a result.
I.e. for some predictors the history is not very important, and for others it is critical, and since there may be both, it turns out that cutting out is not desirable...
I'm currently taking open prices and using them with indicators (homemade) to create new features. I train neuronics which predicts price increase per bar using these chips. It takes a lot of time to create new indicators, otherwise the model will not beat spread.
Partly most of my features works with prediction, not of the price but of its level - I'm using my ATR, but the standard one should work the same way.
Well there are more of them in reality - that's why it's easy to guess)
So it is not important - the main thing is to know in what areas there is an increased risk for trade, which they kind of showed.
The other thing is that I have no idea how to turn all this into the same indicator - is it really necessary to rewrite all the rules that have been formed or what?
Hmm, and on the contrary, I thought that for the forest is more important to make the predictors logical, like yes / no, which I did and that's why there are a lot of them, and so it would be much less. I'll try to do otherwise - I'll give one predictor 8 values, and see if the result changes.
The forest is just worse with sparse features, there will be few splits
and if there are a lot of sparse features and one of them is not sparse, then the forest will overfit to it and it will have the biggest imports, and the others will have very little effect on the result
Forest just works worse with sparse traits, there will be few splits
And if there are many sparse features and one among them is not sparse, then the forest will overfit to it and its imports will be the biggest, and the others will have very little effect on the result
Normal forest or random forest, or both?
I put a Rattle and R (well, and glitches all this stuff ...), and now I can not understand how to make a comparable setting, as in the screenshot below? Because standard Rattle settings gave worse results than the program, which I used before.
Forest just works worse with sparse traits, there will be few splits
And if there are a lot of sparse signs and among them is one unsparse, the forest will overfit to it and its imports will be the biggest, and the others will have very little effect on the result
Same signs, same settings, but the signs are collapsed, not expanded as before.
Old variant
New variant
I found a little more zeros, but considerably fewer ones - almost 2 times less! I didn't think that collapsed and expanded variables can have such an effect...
Normal woods or random woods, or both?
I put Rattle and R (well, the whole thing glitches...), and now I can not understand how to make comparable settings, as in the screenshot below? Because the standard Rattle settings gave worse results than the program I was using before.
normal forest and random forest and tree forests are the same thing :) Forest is an ensemble of Trees
are the traits collapsed, meaning there are less of them or what? by collapsed traits are rarely changing and/or categorical ones like ones and zeros (well, this is a high level of understanding)
I do not use R because I was inculcated with an aversion to it by the local gurus of their golden spectacles
so long you'll bother, study the theory of what is trees and what is a forest of trees
https://habr.com/post/171759/
https://habr.com/post/116385/
If you want to change the settings of Rattle, you should take it to Sanych, but I don't know what for, to spoil yourself, or to bother with the integration of the EA on mt5 and R
https://www.mql5.com/ru/articles/1165
On your data, file Pred_004_Buy divided in half, in head-on you can get 0.85.
The data is rubbish and better thrown out. The rest is catching up on your own. In silence...
Why does the data suck? It's a very good result.