Discussing the article: "Neuro-symbolic systems in algorithmic trading: Combining symbolic rules and neural networks"

New comment
 

Check out the new article: Neuro-symbolic systems in algorithmic trading: Combining symbolic rules and neural networks.

The article describes the experience of developing a hybrid trading system that combines classical technical analysis with neural networks. The author provides a detailed analysis of the system architecture from basic pattern analysis and neural network structure to the mechanisms behind trading decisions, and shares real code and practical observations.

Imagine you are trying to explain to a computer how to trade on the stock exchange. On the one hand, we have classic rules and patterns — "head and shoulders," "double bottom," and hundreds of other patterns familiar to any trader. Many of us have written EAs in MQL5, trying to encode these patterns. But the market is a living organism, it is constantly changing, and strict rules often fail.

On the other hand, there are neural networks – fashionable, powerful, but sometimes completely opaque in their decisions. Feed historical data to an LSTM network and it will make predictions with decent accuracy. But the reasoning behind these decisions often remains a mystery. In trading, every wrong step can cost real money.

I remember struggling with this dilemma in my trading algorithm a few years ago. Classic patterns produced false positives, and the neural network sometimes produced incredible predictions without any logic. And then it dawned on me: what if we combine both approaches? What if we use clear rules as the system framework, and the neural network as an adaptive mechanism that takes into account the current state of the market?

This is how the idea of a neurosymbolic system for algorithmic trading was born. Imagine it as an experienced trader who knows all the classic patterns and rules, but also knows how to adapt to the market, taking into account subtle nuances and relationships. Such a system has a "skeleton" of clear rules and "muscles" in the form of a neural network, which adds flexibility and adaptability.


Author: Yevgeniy Koshtenko

 
MetaQuotes:

Published article Neurosymbolic Systems in Algorithm Trading: Combining Symbolic Rules and Neural Networks:

Author: Yevgeniy Koshtenko

The main problem is the stability of the calculated frequency of appearance of a white or black candle after the appearance of a pattern. On small samples it is unreliable, and on large samples it is 50/50.

And I don't understand the logic of first feeding the pattern frequency to neuronka as one of the features, and then using the same frequency to filter neuronka signals built on it.


 
Without touching the approach itself, reducing the real ranges of movements to two classes nullifies the useful information that could be extracted by the neural network (for the sake of which we screw it in) - akin to if we started feeding the colour image recognition system with black and white images. IMHO, it is necessary not to adjust the network to the old methods of binary patterns, but to highlight real, fuzzy ones on full data.
New comment