Machine learning in trading: theory, models, practice and algo-trading - page 888

 
Aleksey Vyazmikin:

So I will take only the ones that have great support and reliability. This is how I see it working - I have indicators that generate predictors in real time and on history, put them together into a string, and then this string is looked for in the array; if it is found, then we mark the bar as favorable for entering, and if not, we do nothing. Accordingly, the doubled lines will only increase the array. Of course, it is possible to make gradation by color, where will be information about reliability and support (by multiplying one by the other we will get the coefficient, which will change color depending on the value), but for this it is easier to just make a separate array of int type with an index. Or maybe I don't understand something....

Theoretically if NS feed all without sifting out invalid strings, itself should find them and increase their importance. But you can of course do it yourself, as you described. At the same time it is interesting and compare - what is better.

Although in the articles there are examples of screening out noisy strings.
 
Alyosha:

I agree, python is the most topical language among HFTs, it is just those who have negative error, I mentioned above, they also study quantum entanglement there to transmit signal faster than light, there are also libraries on python, in darknet.

you don't have to talk nonsense again, it shows your level

 
Alyosha:

Yes, I'm just kidding with some 90% ahtungs, do not take it seriously, python is a student language, to quickly try a bunch of bibles and draw graphs, for example to consolidate the course Vorontsov, and "research" well maybe for graduate students or something distant from reality, In serious companies all the tools are written in C++, more than half of the algorithms are exclusive or modified well-known, the problems are very specific and highly specialized, and do not require running parameters for 100500 libraries, so Python is present in trace amounts.

Aleshenka, go on mastering MO, about C++ api and Cython and migration of models and individual modules and deployment on servers after research/training haven't you heard?

And without a negative error in trying to predict returns don't come back again. Otherwise you get very upset - a lot of rhetoric, but negative amount of useful information.

 
elibrarius:

Theoretically, if NS fetches all strings without sifting out invalid ones, it should find them and increase their importance. But of course you can do it yourself, as you described. At the same time it is interesting and compare - what is better.

Although there are examples in the articles with screening out noisy strings.

In general, I do not fully understand how to organize the work of NS, if it takes data from the history, theoretically from any point... But with a tree/forest, everything is easier, they do not analyze history, so they need to input a limited number of variables, and the output results are already known, and I wrote them into an array; after training, the tree cannot create something new, but the neural network can react to new data in different ways, Since it inputs weights (coefficients) and they are constant for each neuron, but if there are 10 or more neurons, and each feature has its own coefficient, then it turns out that there are a lot of choices - the network obviously does not train by all variants, so it may have different surprises, or I don't understand something.

And I have made an indicator, but the array enumeration is not fast, and the code itself is very slow - I need optimization.

 
Aleksey Vyazmikin:

In general, I do not fully understand how to organize the work of NS, if it takes data from the history, theoretically from any point... But with a tree/forest, everything is easier, they do not analyze history, so they need to input a limited number of variables, and the output results are already known, and I have written them into an array, something new after training the tree cannot create, but the neural network can react to new data in different ways, Since it inputs weights (coefficients) and they are constant for each neuron, but if there are 10 or more neurons, and each feature has its own coefficient, then it turns out that there are a lot of choices - the network obviously does not train by all variants, so it may have different surprises, or I don't understand something.

I have made an indicator, but the array enumeration is not fast, and the code itself is very slow - I need optimization.

The NS feeds the same table with data - no difference with the forest. Just algorithms are different inside.

 
elibrarius:

The same table with data is fed to the NS - no difference with the forest. Just the algorithms are different inside.

Then why does the NS need to look at the history when training, if it does not use it when working?

 
Alyosha:

Now I'll know that the coolest thing is not python api to c bits but vice versa, c++ api to python, with deployment on servers. keep on burning)))

The mercury in the thermometer of your sarcasm has shrunk into a bubble and just exploded, or expanded so much that it exploded from the scale of stupidity and drooling, whatever

api to frameworks in different languages, including python and ++

Or maybe you didn't, but it's just sad?

 
Alyosha:

No way! Myrandom wandering is just as awesome predicted as forex! 90% acurasi if without overtraining, and 100% with overtraining.

Whoo! It seems to be starting to get to the natives that overtraining is not only not fatal, but, in many cases, even the best).

 
Maxim Dmitrievsky:

Well here is a terrible overfit, but still tradable in the first half oos :) I could make it even worse, but there's no point. The money is somewhere in the middle between over and underfits.

There's money everywhere except underfits. It's only a question of application.

 
Aleksey Vyazmikin:

Then why would the NS look at the history in training, if it does not use it when working?

During training it got weights and offsets for neurons, and according to them it calculates output on new data.