Adapting HiSSD Skill Hierarchies for Multi-Agent Forex Trading
Summary
The article describes an implementation of HiSSD, a hierarchical skill-discovery framework, adapted for algorithmic trading. Its central design separates common skills that can apply across situations from task-specific skills derived from local observations. A Planner processes shared information and predicts future states and value, while a Controller combines each agent’s observations with both skill types to produce actions. The text explains a MQL5 controller architecture intended to support parallel agents with separate action-decoding parameters.
The authors report training on historical EURUSD quotes from 2024 and testing on unseen data from early 2025, and state that each of the three testing months ended profitably. Those results are presented without enough detail here to assess robustness, risk-adjusted returns, costs, or comparison with baselines. The framework is an implementation interpretation of research developed for offline multi-agent cooperation, and the article explicitly cautions that broader datasets and testing across varied market conditions are needed before live use.
Key ideas
- HiSSD divides learned behavior into common skills and task-specific skills.
- A Planner predicts future states and expected value from shared information.
- A Controller combines local observations with common and task-specific skills to generate agent actions.
- The MQL5 implementation is evaluated on historical EURUSD data, with the stated test period ending profitably in each month.
- The limited reported evaluation does not establish robustness across market regimes or live trading conditions.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.