UC San Diego

We test whether large language models (LLMs) add value in commodity-related ETF portfolio construction when information and implementation rules are held fixed across strategies. Hawkish, Dovish, and Debate Agents and a deterministic z-score Rule Agent receive identical macroeconomic inputs and use the same portfolio engine. Across 124 weekly rebalancing dates spanning the 2023 U.S. rate peak and the 2024–2025 soft landing, all three LLM strategies outperform the Rule Agent in return-to-volatility terms. The Hawkish and Debate Agents achieve the largest gains (ΔSharpe=+0.041,p=0.049, 95% CI [−0.007, +0.089]; and +0.037,p=0.074, 95% CI [−0.012, +0.088], respectively; paired stationary block bootstrap, B=5,000) and retain a net-of-cost advantage over passive inverse volatility at one-way trading costs up to 30 basis points. The Debate Agent does not outperform the best single agent (ΔSharpe=−0.004,p=0.802), suggesting that its contribution reflects prior aggregation rather than deliberation-generated return. Results are directionally robust to masked-date prompts and exclusion of the non-commodity proxy; Romano–Wolf correction identifies two masked-date comparisons as jointly significant under the risk-neutral profile and three under risk-averse. However, the advantage is concentrated in the soft-landing sub-period, the sample spans one rate cycle, and primary dated-contract comparisons do not survive familywise correction. Overall, the evidence suggests a modest incremental advantage from LLM-based macro interpretation over a transparent rule layer, with persistence beyond this sample remaining uncertain.