This framework exists to stress-test multi-agent trading (MAS) with a simple, rigid REST API based architecture. The goal was to evaluate responses and LLM noise when analyzing financial data. Since free/cheap news APIs provide poor historic coverage, I manually scraped both general and ticker-specific news for the backtest window so the Sentiment Agent’s influence is actually present in the Reasoning Agent’s analysis.
It’s tailored to this specific 20-stock run and not packaged for easy OSS reuse. The current design surfaced considerable noise between the Reasoning Agent and Portfolio Manager; despite some insightful LLM outputs, the end-to-end behavior is suboptimal. I’m now building a new MCP-based system, with tool-calling and database querying with easy integration and it's substantially more configureable (a lot less harcoded) - stay tuned.!
For better understanding read these summary files explaining the framework in two parts.
aws/Mid_Cap/SYSTEM_DESIGN_Part1.md
aws/Mid_Cap/SYSTEM_DESIGN_Part2.md
- Results/updates: X / Twitter — @Carlos_O_B1.
aws/Mid_Cap: Backtest orchestrator, agents, helpers, Design README's.aws/quant_data/: Historical Valuation and Fundamental Data.aws/sentiment_files/: Curated news sentiment dataset.aws/deployment/,aws/scripts/: Deployment and setup utilities.
python3 -m venv venv source venv/bin/activate pip install -r aws/requirements.txt## Run a Backtest (example) cd aws/Mid_cap python process_all_stocks_all_agents.py --data_dir .. --start_date 2025-07-01 --end_date 2025-11-14 //* outputs analysis of the 3 sub agents
python ParallelOrchestrator.py --start_date 2025-07-01 --end_date 2025-11-14 Process all stocks with all three analysis agents: //* Runs the Reasoning Agent on the sub agent analysis, Portfolio Manager takes the Reasoning Agent analysis and allocates available funds accordingly