Machine learning in trading: theory, models, practice and algo-trading - page 331

 
Yuriy Asaulenko:

Say, handwriting recognition is an uncomplicated neuron in several layers. There are even copies on the Internet, with training and a demonstration of the results. And one learns in front of an amazed public).

For market patterns, I suppose, if you work on 1m, then 2-3 weeks of history for learning is enough. In general, I'm going to start from the beginning.

In general, I'm going to solve a problem similar to yours (Reshetov's) system, but on honest neuron, and see what comes out of it.


Well, for the text convolutional systems are used, you can quickly train them, yes.

 
Maxim Dmitrievsky:


Ah, well, for the text rollup are used, they can be quickly trained, yes.

How is handwriting recognition so different from pattern recognition? Imho, only at the level of nuances.

I didn't see what was in the handwriting copy, it seems to be ordinary - the code is simple. For starters, let's try to do without convolutional).

 
Yuriy Asaulenko:

How is handwriting recognition so different from pattern recognition? Imho, only at the level of nuances.

For starters, let's try to do without convolutions).


At first approximation, patterns can be very similar, but their continuations will be drastically different, because it is enough for them to have small deviations at the beginning to go in different directions at the end... this is a problem. The letter has a certain image, and the grid distinguishes it by its features, and if these features are minimal but critical, how about patterns? But in any case, you have to try and not theorize :)

I was just recently writing a system on patterns, but it got stuck on the fact that the correlation gives very approximate similarity, in the end even if the pattern is similar in general, but small differences are not taken into account and they often grow and dictate the further outcome (sometimes the opposite)

 
Maxim Dmitrievsky:

... But in any case, you should try instead of theorizing :)

You have to theorize, too. Otherwise you definitely can't do anything).

Yes, of course, patterns can have different extensions, but neuronics are not required to have error-free solutions, just like a system based on logic.

ZZY Patterns for neuronics is a term. It has a slightly different meaning than is customary in the marketplace. Say, the book -Bishop C.M. Neural networks for pattern recognition.

 
Yuriy Asaulenko:

You have to theorize, too. Otherwise you definitely can't do anything).

Yes, of course, patterns can have different continuation, but neuronics is not required to have infallible solutions, as well as the system on the logic.


it is possible to do somehow without correlation... overlay scaled patterns on each other, look at residuals, do something with them... or through matrix operations somehow... but I dunno :)
 
Maxim Dmitrievsky:


Well then the currency snake has long come out of its tunnel and no longer exists

If you have in mind the cluster indicator (a la currency strength), I myself will soon be adding a grid to it :)

Whoa, whoa, whoa, whoa.

You just got a good result in the tester.

Does it match the trading on the demo?

Maybe you really shouldn't elaborate on this topic?

 
Maxim Dmitrievsky:

It is possible to do somehow without correlation... to overlay the scaled patterns on each other, look at the residuals, do something with them... or through matrix operations somehow... but I don't know :)
You can, by feeding not only the price series, but also predictor values to the inputs, thus including them in the image. In your case it is RSI, even without price series.
 
Renat Akhtyamov:

Whoa, whoa, whoa, whoa.

You just got a good result in the tester.

Does it match the trading on the demo?

Maybe it is really not necessary to go further?


This is just the beginning :), I have a more complicated system in my plans, I just picked up the net for my aims, this RNN is enough for now, I'll see.

Ideally - to understand how NS are used for hft trade, gradually moving to smaller time scales

 
Maxim Dmitrievsky:


I have a more sophisticated system in mind, I just picked up the network for my own purposes, this RNN is enough so far, then we will see.

I've got a good idea - to understand how NS are used for hft trading, gradually moving to smaller timeframes

wait. What is there on your demo, at least to understand to begin with?

Then look at the cent real would not hurt.

And only after that we can look forward.

My opinion is...

 
Renat Akhtyamov:

wait. What is there on your demo, at least to understand to begin with?

Then look at the cent real would not hurt.

And only after all this you can look forward.

This is just my opinion.


I think I should stay on the demo. It's a primitive bot, that doesn't take into account many things, but it is able to earn already... it's just a preparation for further, so to say, research... From experience - normal systems can't be created in two days, like in this case :)