Machine learning in trading: theory, models, practice and algo-trading - page 72

 
Mihail Marchukajtes:
This is why you guys are failing. I'll explain for the last time and I won't come back to this question. The classifier defines the current state of the system. The regression determines the future state of the system. That's it... the subject is closed.

This needs to go in the anatomy of the forum.

A new word in machine learning. I wonder, does the world community know about this great discovery?

The horror

 
Yury Reshetov:
Now I finished the code generator to unload ternary classifier models in Java. I'll test it to see if it works. Then I'll make it for mql - there are some differences, for example, mql doesn't have Math.signum() function. Then I think I will have to make a model code generation for R as well, because local figures don't understand anything else.
Please, throw me a link where to download it, I can't find it on the forum :-(
 
Vladimir Perervenko:

This needs to go in the analogue of the forum.

New word in machine learning. I wonder if the world community knows about this great discovery.

Terrible

It seems that everyone knows except for the inhabitants of this forum :-) To know is one thing, but to understand the essence, is quite another.....
 
Mihail Marchukajtes:
It seems that everyone knows except for the inhabitants of this forum :-)

If you troll, your fate will be sealed. Smile )

It is explained to you by knowledgeable people (who do not train models on 10 predictors and 20 observations) that the state of the system, to which the classifier is trained, can mean anything, from the current phase of the moon, to the forecast that tomorrow it will rain. You are closed on the current state and cannot understand that a state can include a forecast for the future.

 
Alexey Burnakov:

If you troll, your fate will be sealed. Smile )

It is explained to you by knowledgeable people (who do not train models on 10 predictors and 20 observations) that the state of the system, to which the classifier is trained, can mean anything, from the current phase of the moon, to the forecast that tomorrow it will rain. You are locked into the current state and can't understand that the state can include a forecast for the future.

What? You're scaring me. I don't have much to be afraid of... I'm doing just fine. And I understand perfectly what is going on, we are talking about the same thing, only with different terminology.... Anyway....

And about the highlighted. By determining the state of the system we make a conclusion about the future. But there is no way state includes prediction. Although there is truth in these words. But the classifier doesn't make the forecast itself, the forecast is made by the system it classifies, but not by itself.....

 
Dr.Trader:

I will try, train the model within a week. Then you can even check on fronttest, on yet unknown continuation of that file. But there is a nuance that the algorithm I work with is built on two classes, on three will be a problem with my fitness function to evaluate the model.

Right now I have two possible classes in the model, with a classification result of 0 or 1 for sell/buy. I'll apply the same code for your three classes, but with the output scaled to (0;0.5;1), but that's not the best approach. It would be better to make 3 outputs in neuronka corresponding to three classes, and take the result of classification as the output with the largest value. I have no idea which of these ways would work best on your data, I'll make both of these models, I wonder which one will give the best result.

Validation data:https://drive.google.com/file/d/0B_Au3ANgcG7COGpJb24wbkxoaTg/view?usp=sharing

It has 5.5 years for 5 pairs. The dates are strictly after the training set. In the far right column is the celestial: price increment in 181 minutes (what was coded with categories in the tray).

Try to get an MO on validation greater than 0.0001. I got about 0.00013, which corresponds to the average spread. That is, zero.

At the same time I do not trade every observation, but only there where the signal is strong (about 5-10% of observations).

Thank you,

Alexey

 
Mihail Marchukajtes:

And about the highlighted. By determining the state of the system we infer the future. But there is no way the state includes the prediction. Although there is truth in these words. But the classifier doesn't make a prediction itself, the prediction is made by the system it classifies, but not by itself.....

Can someone translate this into Russian? It looks like Chernomyrdin's speech, but it makes no sense.
 
Vladimir Perervenko:
Can anyone translate this into Russian? It looks like Chernomyrdin's speech, but makes no sense.
I am shocked myself :-)
 
Mihail Marchukajtes:
Shocked myself :-)
It happens.
 
Mihail Marchukajtes:
Please drop me a link where to download it, because I can't find it on the forum :-(
https://bitbucket.org/jprediction/jprediction/downloads