What to feed to the input of the neural network? Your ideas... - page 6

New comment
 
Dmytryi Voitukhov #:

Then what's the point of me preparing a detailed answer?

I have the topic in an open tab, in practice, usually when I am doing something else, when an idea comes into my head (what else to export for the grid) I immediately check it. Unfortunately, I don't understand a lot of the suggestions in the thread (I can't figure out what to export, how and where to export it to).

No offence. If there is something formalised (that I can twist), I will be only glad and will definitely check it.

 

- What to feed to the input of the neural network?

- Your ideas...

 
Evaluate the idea (I don't realise such a thing with the means at hand), but I'm interested in your opinion:

In one local article I read that ".... Kohonen networks are usually used in image recognition..." and immediately inputs the price chronology.

Don't you think this is a hint, because traders don't "eat" the chronology, they look at the chart, can mark something in common, see some cluster of prices, mark a level there and evaluate breakdowns/rebounds, figures.

When training a neural network to recognise a cat in images, a set of many images is prepared for the neural network for training. As a result, the neural network can identify a cat in an image that likes to run around the flat at four in the morning without serious problems.
And in the network you can more and more often find articles on how to teach a neural network to identify numbers and signs in an image.


Why not do the same with a price chart? Take screenshots of a price chart before it rises, and vice versa. Since the image has only two colours (black and white), and there will be few details, the images will weigh little, as well as be processed. And feed them all to the neural network, so that it finally simulates/emulates the trader's work, I did not multiply prices.
 
Ivan Butko training a neural network to recognise a cat in images, a set of many images is prepared for the neural network for training. As a result, the neural network can identify a cat in an image that likes to run around the flat at four in the morning without serious problems.
And in the network you can more and more often find articles on how to teach a neural network to identify numbers and signs in an image.


Why not do the same with a price chart? Take screenshots of a price chart before it rises, and vice versa. Since the image has only two colours (black and white), and there will be few details, the images will weigh little, as well as be processed. And feed them all to the neural network, so that it finally simulates/emulates the work of a trader, I did not multiply prices.
Artificial complication, from a 1-dimensional series make a 2-dimensional one. The number of signs will increase by orders of magnitude, training time too. But the result will be the same.

For time series do not make screenshots, but all sorts of transformations like recurrence plot, otherwise the feature matrix will be very sparse.
 
Maxim Dmitrievsky #:
Artificial complication, from a 1-dimensional series you make a 2-dimensional one. The number of features will increase by orders of magnitude, training time too. And the result will be the same.

For time series do not make screenshots, but all sorts of transformations like recurrence plot, otherwise the feature matrix will be very sparse.

Thanks for the advice.

It just seems to me that maybe we should not consider time series, but consider patterns, in which part of the time series will not play a role, and the neural network will look at the "picture as a whole", as a trader does. The task itself is different, i.e.: before a trend (reversal) this part of the chart should be scanned, and "do not tell" the neural network what prices these candlesticks have, the difference of candlestick prices, do not carry out and do not feed it the normalisation of candlestick prices, normalisation of indicator data - all this should be discarded and fed to the actual value during training only "up" or "down", "1" or "0", "bull" or "bear". And, when the neural network "sees" these bulls, at new candlesticks (make a screen of the chart section for the neural network, or somehow automate this process), it will say "well, this is not a bull, there is some bullshit, I don't understand", and on another chart "there is something that looks like a bull, probably it is a bull", on the third chart "there is definitely a bull, the price is about to turn around". By analogy with image recognition (I think I saw an article somewhere on the hacker). There, too, the neural network was fed a million images of cats, and then it: "here is a cat", "this is probably a cat", "this is not a cat")).

 
Another option is to train on each candle. It is labour- and resource-intensive, but maybe it will grow at a distance, since a lot of work is done. Let's say 500 candles or more, train until blue (retraining) and predict only one candle, a new one. Then, as soon as it closed, retrain again and so on. If the learning process will take a lot of time (although, I have not noticed such a thing), then take an hour candle or a 4-hour candle. Maybe it will be possible to bypass this notorious 50/50.

I found an article here by a Brazilian about the reverse error propagation. There is no EA as such, just a script, but it can predict one next value. I will try this approach when I adapt it to an EA.
 
The patterns are where are they located, in nowhere? Or on a time series. Why does everyone have such a strong desire to get rid of it and start training NS on nothing :)

A specific way of training will not work, you need to work out a strategy. Just like in the TS without NS.

Think of the NS as a strategy optimiser, like the one built into the terminal
 
If you train a neural network on everything and try to predict the value, you will get something even worse than ADX
 
Well, you set some condition to predict from now to then, and ignore the rest. It all depends on the strategy, what is expected at the output.

It is possible before training to allocate, it is possible after to put a condition where it should work and where it should not. You can come up with a lot of approaches, as long as they are meaningful.
 

Enter periods into the neural network.

Comparing years.

Then seasons.

Days

Hour candles (considering summer/winter time change).

And then you'll see a pattern in some instrument.

That's it, you're rich. If you make such an analysis, please share it later, I can't get my hands on it myself.

// a mathematician once became a millionaire on the stock market by studying such patterns.


P.S. You can also enter the periods of the planets, the moon and the location of the stock exchange relative to their rotation. Theoretically, you will catch the amplitude of the currency (European exchange - euro, American exchange - dollar, etc.). By comparing amplitudes and combining them, you will see all currency pairs ahead. For those who do not understand, this is humour.

12345678910111213...68
New comment