Machine learning in trading: theory, models, practice and algo-trading - page 1204

New comment
 
Maxim Dmitrievsky:

is solved by selecting optimal weights... for example how to vary the posterior... from uniform to exponential

I'm not a proponent of "black box" models. Better when everything is transparent, with a simple "physical" meaning.

For example - we consider the a priori probability of a correction to become a reversal and then recalculate it for each certain correction in the a posteriori one depending on the time of day or the trend characteristics.

 
Aleksey Nikolayev:

I am not a supporter of "black box" models. It is better when everything is transparent, with a simple "physical" meaning.

For example - we consider the a priori probability of a correction to become a reversal, and then we recalculate it for each specific correction to the a posteriori probability, depending on the time of day or the trend characteristics.

Unfortunately, we do not know the simple physical meaning of the Forex laws.

 
Maxim Dmitrievsky:

The simple physical meaning of the laws of forex is not known to us, unfortunately

It's about the interpretability of the model.

 
Aleksey Nikolayev:

It's about the interpretability of the model.

then you are not in the subject of MO :) although metamodels are easily interpreted through their metrics

 
Maxim Dmitrievsky:

then you are not in the MO theme :) although metamodels are easily interpreted through their metrics

why? imho, the problem is the same as to distinguish a cat from a dog for MoD

 
Igor Makanu:

Why? imho, the task is the same as distinguishing a cat from a dog for the MoD

Because it's like speaking different languages, in my opinion...

distinguish by signs, of course... they're chips
 
Maxim Dmitrievsky:

then you are out of the MO theme :) although metamodels are easily interpreted through their metrics

I'm not sure it's always easy, but somehow it can be done. I believe that it is not the neural network itself, but its simplified approximation that should deal directly with trading.

Without methods of MO (you can call it "clever exploratory analysis") in our case, you can't do without it.)

 
Aleksey Nikolayev:

I'm not sure it's always easy, but somehow it can be done. I believe that it is not the neural network itself, but its simplified approximation that should deal directly with trading.

Without methods of MO (we can call it "smart exploratory analysis") in our case we can't do without)

Now I want to add the dependence of signals on distributions to the parameters to be optimized, I have done it for the beginning, to see

     double arr[];
     CopyClose(_Symbol,0,0,100,arr);
     double kurt = MathKurtosis(arr);
     double skew = MathSkewness(arr); 
     if(kurt > 2.0) if(rand()/32767.0<0.5) res = 0; else res = 1;
     else {
      if(skew >0) if(rand()/32767.0>prob_shift) res = 0; else res = 1;
      if(skew <0) if(rand()/32767.0<prob_shift) res = 0; else res = 1;

If the kurtosis is higher than a certain value (you can opt for it), then a flat situation is observed, and it is possible to buy/sell with the same probability (and then fix all the wrong ones)

further on asymmetry, if there is a certain side, then the probability of the signal to buy or to sell is displaced

This is a primitive one, but this is the way we can select the targets in the optimizer

All you need to get from metrics is classification error on the test sample (to be trained on the training sample). The optimizer goes through the hyperparameters and chooses the model with the lowest error. What is non-interpretable here? It is enough to know, looking at the errors on the test data, whether such a model is capable of generalizing or not.

I just made an example of working with such bullshit


 
Maxim Dmitrievsky:

Now I want to add the dependence of signals on distributions to the optimized parameters.

And it seems that there is a dependence...

I trained "SMM" (hidden Markovian model) on returnees, divided it into 10 states and trained it without a teacher, so it divided different distributions by itself


state distributions.


And here I grouped returns by states, i.e. each row is a separate market state

Some states (1,4,6,8,9) have too few observations, so you can not take them at all

And now I will try to restore the series, that is to make a cumulative sum, if some tendency is found in some of the states - the regularity in the direction

I did a cumulative summation.

states 5 and 7 have a consistent structure, 5 is for the bay and 7 is for the village

 
Maxim Dmitrievsky:

Now I want to add to the optimized parameters the dependence of signals on distributions, I did it for the beginning, to see

If the kurtosis is higher than a certain value (you can opt for it), then a flat situation is observed, and it is possible to buy/sell with the same probability (and then to fix all the wrong ones)

further on asymmetry, if there is a certain side, then the probability of the signal to buy or to sell is displaced

This is a primitive one, but it's approximately the way the optimizer can select the targets.

Why prices and not their increments?

1...119711981199120012011202120312041205120612071208120912101211...3526
New comment
Reason: