This question is very rare in the trading literature. I found something here [XIV.54.].
The authors, based on tests of various algorithms, came to ambiguous conclusions: if certain conditions are met, neural networks can be used in system trading. Fitting to historical data is a significant problem when using neural networks. Large samples should be used to train neural networks. Different neural networks behave differently in different markets.
I think the main problem here lies in the essence – in the attempt to use neural networks to guess the market. But as we already know, the market cannot be guessed, so using neural networks for these purposes will not give the desired result.
In general, working with a neural network is the following.
1. Choosing the type of neural network.
2. Preparation of data for training a neural network.
3. Training of the neural network by feeding input data to the inputs of the neural network, for trading it will be digital data on price changes or data from several candlesticks displaying the desired figure of those. analysis.
4. Next, for a given input, the output data of the neural network is set – for the case of trading, these are numbers, for example, 0 and 1, according to which a decision is made on a long or short transaction or on the exit of their transaction.
5. The training is repeated until the results of testing the neural network’s “understanding” of patterns satisfy you.
6. Further, in a real trading situation, when fragments of data from the current chart are fed to the inputs of the neural network, the network, based on the available data stock, will look for similar situations that it “knows”. This is done with a certain degree of approximation, which makes it possible to find a solution using a fuzzy pattern, which makes neural networks suitable for such tasks as pattern comparison. This adds flexibility to the network to work with data that is not 100% the same as the templates available in memory. At the same time, we will also get values in the range 0..1 at the output.
7. Next, the neural network decision-making module is added to your trading algorithm, and you start testing the resulting robot on real data.
8. If you consider the tests satisfactory, then you can start trading for money.
According to the book [XIV.54.] neural networks showed excellent results within the sample on which they were trained and poor results outside the sample, which is quite natural.
Given the complexity of understanding the programming of neural networks for practical purposes, it is recommended to look for ready-made packages for creating trading robots on neural networks.
Interesting articles about trading on neural networks are posted here