Discussing the article: "Neural Networks in Trading: An Ensemble of Agents with Attention Mechanisms (Final Part)"

 

Check out the new article: Neural Networks in Trading: An Ensemble of Agents with Attention Mechanisms (Final Part).

In the previous article, we introduced the multi-agent adaptive framework MASAAT, which uses an ensemble of agents to perform cross-analysis of multimodal time series at different data scales. Today we will continue implementing the approaches of this framework in MQL5 and bring this work to a logical conclusion.

To capture significant price shifts, the agents employ directional movement filters with varying threshold values. This allows the extraction of key trend characteristics from analyzed price time series, improving the interpretation of market transitions of different intensities. The proposed method introduces a novel sequence token generation technique, enabling cross-sectional attention (CSA) and temporal analysis (TA) modules to effectively identify diverse correlations. Specifically, when reconstructing feature maps, sequence tokens in the CSA module are generated based on individual asset indicators, optimized through attention mechanisms. In parallel, tokens in the TA module are constructed from temporal characteristics, making it possible to identify meaningful relationships across different time points.

The correlation assessments of assets and time points, derived from the CSA and TA modules, are then combined by MASAAT agents using an attention mechanism, with the goal of detecting dependencies for each asset in relation to every time point over the observation period.

The original visualization of the MASAAT framework is provided below.

The MASAAT framework exhibits a clearly defined modular architecture. This makes it possible to implement each module as an independent class and then integrate the resulting objects into a unified structure. In the previous article, we introduced the implementation algorithms for the multi-agent object CNeuronPLRMultiAgentsOCL, which transforms the analyzed multimodal time series into multi-scale piecewise-linear representations. We also reviewed the algorithm of the CSACNeuronCrossSectionalAnalysis module. In this article, we continue this line of work.


Author: Dmitriy Gizlyk

 
Why do they print such disgraceful articles? Profit for the month is 0.27%, my cat, accidentally tapping his paw, will make 50 times more.