Macro Economists in the Machine: A Multi-Agent LLM Framework for Commodity-Related ETF Portfolio Construction

Battle Count: 78

Highly relevant to quantitative trading as it provides a controlled framework for evaluating LLM-based macro interpretation in portfolio construction. The paper demonstrates a clean ablation study isolating the interpretation layer from the portfolio engine. Results show modest but economically meaningful improvements (Delta Sharpe +0.04) over rule-based baselines, with transaction cost robustness up to 30 bps. The multi-agent debate mechanism and Hawkish/Dovish prior framework offer practical design patterns for LLM-augmented trading systems. However, the small sample size, single rate cycle, and statistical fragility limit immediate practical deployment.

Moderate complexity. The framework requires: (1) FRED data pipeline with rolling z-score computation and release-lag adjustments; (2) LLM API integration with structured JSON prompt engineering and output validation; (3) Multi-agent orchestration (Hawkish, Dovish, Debate with disagreement threshold and deliberation rounds); (4) Portfolio construction engine with inverse-volatility weighting, macro-tilt scaling, risk-off caps, single-name constraints, and turnover limits; (5) Block bootstrap inference implementation. The LLM API cost is minimal (~$0.002/week), but maintaining prompts, validating JSON outputs, and monitoring model behavior adds operational burden. The deterministic Rule Agent provides a simpler baseline.

All parameters are explicitly specified (Table B2), prompt templates provided in Appendix D, loading matrix in Appendix A. Data sourced from public FRED and Yahoo Finance. However, processed weekly panel, generated LLM contracts, and replication materials are only available from the corresponding author upon request. The LLM outputs depend on GPT-4o-mini API which may change over time. Cached contracts mitigate this but are not publicly distributed. No GitHub repository is mentioned.