The system problem
Market context arrives from different sources, at different times, and in different shapes. A useful workflow must normalize that context before asking a model to reason about it.
The hourly pipeline
The workflow gathers financial headlines, macro sentiment, economic-calendar events, and open MT5 positions. Calendar events receive time-aware labels such as past, imminent, or just_released.
Structured output before execution
The model is asked for a constrained response, not free-form prose. A deterministic layer then validates allowed symbols, actions, stale timestamps, required risk fields, and position conflicts before anything reaches MQL5.
The model can propose. The system must validate.
What I would add next
- Human approval for sensitive actions.
- Risk caps outside the model prompt.
- Replayable input snapshots.
- Full decision and execution logs.
- A kill switch and stale-data protection.