Practical FMZ Workflow Guidance for Data, AI Nodes, and Backtesting
Summary
This FAQ explains implementation details for quantitative strategies built with FMZ workflows. It covers supported runtime requirements, JavaScript code nodes, sequential node execution, trigger timing, passing outputs between connected nodes, and using persistent global storage to share state. It also describes the need to return data from code nodes and to aggregate multiple records before passing them to downstream processing.
The AI sections cover model credentials, API configuration, usage costs, output variability, and risk controls such as position limits and stop logic. The article says workflow backtests limit actual AI calls and reuse cached responses, so results do not represent repeated live model decisions; using current news against historical prices also creates a time mismatch. It recommends cautious live observation with small allocations. These are platform-specific operational notes rather than evidence for a trading strategy, and API behavior may change over time.
Key ideas
- FMZ workflow nodes execute sequentially, and time-based triggers wait for their specified bar to close.
- Connected node outputs can be read directly, while persistent global storage can share state across workflows.
- Code nodes must return output, and multiple records may need aggregation before downstream processing.
- AI calls incur token costs and model decisions can vary, so strict position and loss controls are advised.
- AI backtests use limited live calls with cached results, making their output an unreliable measure of live decision quality.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.