Machine learning in trading: theory, models, practice and algo-trading - page 1016

 
SanSanych Fomenko:

I brought the link because of the table: a fresh look at the predictors and as a development of Alexander's thought about the ACF.

Fresh, are you serious? Old as ... the mammoth. How is it better than Fourier or other decomposition or autoregressive coefficients and their values or filters?

None of them work.

 
Maxim Dmitrievsky:

Fresh, are you serious? Old as ... the mammoth. How is it better than Fourier or other decomposition or autoregressive coefficients and their values or filters

None of these work.

Fresh for this thread.

So, what have you tried?

What specifically? For what target?

 
SanSanych Fomenko:

Fresh for this thread.

It turns out that you tried?

What specifically? For which target?

I tried different decompositions, autoregression (no difference from acf) together with coefficients too.

Works on well correlated series (take the 2nd series as a feature, build a VAR for example between instruments, take the predictor and train the model). Well, how it works - as long as the correlation is not broken.

On the 1st instrument it works the same way as many other predictors, with overfit. It works in the same way as many other predictors with overfit.

 
Ivan Negreshniy:

Thanks, I am interested not so much in membership, which, as I understand, involves significant difficulties, but to look at the level, which is probably not less significant.

"Membership" is right))) any membership is usually the result of chance acquaintance or coincidence, and there are several advanced guys improvising their "kuklusklan" with its culture and rituals, I've been there only a couple of months, until the super-valuable information has been poured into my head, But the fact that the ecosystem is at least relatively closed is a big plus for such discussions and exchange of spare parts for their algotrading infrastructures, but I don't think it will last long with such hazing, you need to look for similar groups that don't humiliate geks that badly

You wrote that this group is considered a unified representation of the models of the MO, that's these models and I would like to see.

Yes such thoughts were made out loud and not only there I heard about it by the way, on elit-tradera they were talking, may be here I heard something similar, I propose to discuss in private if you want how to make the format of the model exchange settled, I also have thoughts on this subject, in fact every coder knows I have some ideas abouthow to do it, it's just a question of standards, something like fullstack C++ dll-grall pattern that takes raw data and outputs forecasts and a couple lines of code to connect from Sharp's python and so on. Who cares where.

For comparison as ready to show their, modest developments, I trained models serialize in binary or text format and in the form of source code.

In private message all well or create a separate topic here, in this do not need here "deep" garbage

 

Zhenya:

You wrote that in this group the unified representation of MO models is considered, here these models and I would like to see.

Yes such thoughts were voiced out loud and not only there I heard about it by the way, on elit-trader it was said, maybe here I heard something like that too, I propose to discuss in private if you want how we can make the format of the model exchange clear, I have thoughts about it too, in fact every coder knows I have some ideashow to do it, it's just a question of standards, something like fullstack C++ dll-grall pattern, which would receive raw data and output forecasts and in a couple lines of code can be connected from Sharp's peyton and so on. Who cares where.

There is PMML for example.

https://ru.wikipedia.org/wiki/Язык_разметки_прогнозного_моделирования

There's a library for R (also called pmml) which can convert most popular models to this format.

you can do it this way:
1) train the model in R as usual
2) convert the model into pmml
3) save the pmml model to an xml file and share it

Some time ago I wanted to make a script for mql which could read xml files with pmml models, in this case I could embed into EA a prediction based on gbm from R with couple lines of code (including the script + pmml as a resource). But I haven't got round to doing it.


This is not suitable if you want to keep the model itself a secret, because in the xml file will be prescribed weights of neuronka or forest branches.

For full secrecy of the model arrangement, you can use an idea from one datascientist contest - they require a file with hundreds of thousands of predictions. They can then use that file to interpolate the predictions to get a prediction next to the existing one.

 
Dr. Trader:

There is PMML, for example.

https://ru.wikipedia.org/wiki/Язык_разметки_прогнозного_моделирования

There is a library for R (also called pmml) which can convert most popular models to this format.

you could do it this way:
1) train the model in R as usual
2) convert the model to pmml
3) save the pmml model into an xml file and share it

Some time ago I wanted to make a script for mql which could read xml files with pmml models. Then I could embed into EA the prediction of trained gbm from r with couple lines of code (including the script + pmml as a resource). But I haven't got round to doing it.

This is not suitable if you want to keep the model itself a secret, because in the xml file will be prescribed weights of neuronka or forest branches.

Unfortunately, the secrecy is the main requirement here))) It is about the exchange protocol obfuscated C++ models that take raw data from the exchange and produce forecasts, so you can take it with description of its inputs and outputs, use it for example a month or for how long it is designed without modification (additional training, etc.) and draw conclusions (buy, rent, etc.)

It is desirable that it was just a folder with the files, different binaries in which to understand the details is not cost-effective.

For total secrecy of the model device, you can use an idea from a datascientist contest - they require a file with hundreds of thousands of predictions. Then using this file they can interpolate the predictions to get a prediction next to the existing one.

If you mean numerai, their approach will not work, in this case, by "model" we mean the full set of transformations of raw data into a prediction, first of all it's signs, numerai has a student contest for marketing their coin (NMR), it can not be somehow related to the real markets, the last thing a hedge fund needs is to outsource the pure classification. If you mean the other contest then please tell me.

 
Can anyone say anything about this use of MO for non-stationary series: combined models? We introduce some additional parameter corresponding to the state of the process and on which the model parameters depend. This parameter is also predicted. The approach is used in the article to predict temperature. Surely such a complication of the model is fraught with overfitting, but at least we can somehow deal with non-stationarity.
 

Numerai, yes.

Their way makes sense. I tried my models to predict hundreds of thousands of random instances. Then for "black box" prediction, I looked for the closest point by coordinates, and used its result as the prediction itself. This prototype worked, but we could improve it for real - find 3 closest points and triangulate the average result. But this is computationally expensive, even with the opencl view can take a couple of seconds to predict.

 
Aleksey Nikolayev:
Some additional parameter corresponding to the state of the process is introduced and on which the model parameters depend. This parameter is also predicted.

It's very similar to memory in RNN (recurrent neural networks)

Predictors and one more value (memory) are input into the model. The model outputs two numbers in the prediction - the target itself, and a new memory value that will be used together with the predictors in the next prediction. That's why the recurrence network, its output will be used as an input for the next time, and so on in a circle each time.

RNNs in forex are very overfeeding, it's bad, you can't do everything by textbook and put trade.
But for a model with only a couple of parameters it has quite good accuracy, comparable to a regular neuron with a big hidden layer, it still amazes me.

 
Dr. Trader:

This is very similar to memory in RNN (recurrent neural networks)

The model inputs predictors and another value (memory). The model outputs two numbers in the prediction - the target itself, and a new memory value that will be used together with the predictors in the next prediction. That's why it's a recursive network, because its output will be used as input the next time, and so on in a circle each time.

RNNs in forex are very overfeeding, it's bad, you can't do everything by textbook and put trade.
But for a model with only a couple of parameters it has quite good accuracy, comparable to a regular neuron with a big hidden layer, it still amazes me.

Thank you, it really does look like that.