Dual-Tower Transformer Agents for Trading Scenarios

Summary

The article describes a trading adaptation of the Hidformer multivariate time-series architecture. Its dual towers analyze temporal patterns and frequency-domain information separately, using recursive attention for sequence dependencies and linear attention for frequency features. The author modifies the encoders into independent trading agents that generate trade scenarios, adds role-based processing and learnable segment shuffling, and includes recurrent analysis of earlier decisions. The output is framed as trade volume, stop-loss, and take-profit parameters.

The implementation is presented in MQL5, with training and out-of-sample testing on historical data. The article reports positive test observations, including a majority of profitable trades and larger average winners than losers, but the supplied excerpt does not give enough context to establish robust performance or costs. The conclusion itself says a more representative dataset and comprehensive testing are needed before live deployment. The reported results should therefore be treated as preliminary evidence for a complex model design, not as proof of a reliable trading edge.

Key ideas

• The Hidformer design separates temporal sequence analysis from frequency-domain analysis in two encoder towers.
• The trading adaptation treats the towers as agents that propose trade scenarios and review their prior decisions.
• Recursive attention and linear attention are used for different forms of input analysis.
• The model's output includes trade volume, stop-loss, and take-profit settings.
• The article reports out-of-sample testing but calls for broader data and further evaluation before live deployment.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.