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

Abstract: We test whether LLMs add value in commodity portfolio construction when the information set and implementation rules are held fixed across strategies. A Hawkish Agent (inflation-tightening prior), a Dovish Agent (growth-easing prior), a Debate Agent, and a deterministic z-score Rule Agent each receive identical FRED macro z-scores and route their tilt signals through 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 Sharpe terms; the Hawkish and Debate Agents record the largest gains (ΔSharpe = +0.044 and +0.040, both p < 0.10 under a block bootstrap) and preserve a net-of-cost advantage over the passive inverse-volatility benchmark at one-way trading costs up to 30 basis points, while the Rule Agent's thin margin over passive disappears at approximately 5 basis points.The Debate Agent does not outperform the best single agent (ΔSharpe = -0.004, p = 0.769); its contribution is bias correction -- averaging out the Dovish Agent's miscalibrated prior -- rather than deliberation-generated return. The performance advantage is concentrated in the soft-landing sub-period, the evaluation window spans a single rate cycle, and the reported $p$-values are unadjusted for multiple comparisons. Within these limits, the results suggest that an LLM acting as a constrained macro-interpretation function can add modest but economically meaningful value over a transparent rule layer, though the margin is small and its persistence beyond this sample is unknown.

• The paper demonstrates that LLM-based agents achieve modest Sharpe improvements (up to +0.044) over deterministic rule-based mappings in commodity ETF construction.
• The study employs an ablation design with four agent strategies using a common 7-dimensional FRED-based macro feature vector to isolate LLM interpretation value.
• Practical implications suggest LLMs serve best as modular, auditable macro interpretation layers during regime shifts despite statistically marginal gains.

LLM-Based Multi-Agent Frameworks for Commodity-Related ETF Portfolio Construction

Context and Motivation

This paper rigorously investigates whether LLM-driven macro-interpretation layers can incrementally improve commodity-related ETF portfolio construction compared to deterministic rule-based mappings, conditional on holding the entire information set and implementation engine constant. Prior research on LLMs in finance has often failed to isolate the value contributed by the LLM itself, given confounding factors such as heterogeneity in data, portfolio construction rules, or access to text corpora. The present study addresses this identification problem by ablation: all strategies observe an identical vector of standardized macroeconomic features and share a common portfolio engine, restricting any systematic outperformance to differential interpretation of macro conditions.

Methodological Design

The agent system comprises four active strategies and a passive benchmark:

• Rule Agent: Deterministic linear mapping from macro z-scores to ETF tilts, aggregating directionally signed exposures from a fixed loading matrix.
• Hawkish Agent: LLM instantiated with a tight/price-stability macro prior emphasizing inflation control, elevated real rates, and dollar strength.
• Dovish Agent: LLM instantiated with a growth-supportive/easing macro prior focusing on labor market recovery, industrial production, and the transience of inflation.
• Debate Agent: Structured aggregation of output from Hawkish/Dovish through a two-round deliberation process, serving as an inter-prior consensus mechanism.
• Inverse Volatility: Passive reference with no macro input, serving as a risk-only control.

Every agent receives a 7-dimensional FRED-based macro feature vector (event-timely, release-lag-corrected, rolling 3-year z-scores) each week, spanning volatility, dollar, funds rate, industrial output, breakeven inflation, real yield, and unemployment. Agents must map these into ETF tilts in $[-2, 2]$, subject to the same base allocation, blend, multiplicative tilt scaling, risk-off caps, single-name limits, and weekly turnover constraints. LLM outputs are deterministic (temperature zero), JSON-formatted, and auditable.

Empirical Evaluation

Performance Metrics and Result Structure

The empirical window covers 124 weekly rebalancing dates (Oct 2023–Feb 2026), spanning a distinct U.S. monetary policy regime shift (rate-peak to soft-landing). Strategies are evaluated on annualized return, volatility, Sharpe ratio, maximum drawdown, and outperformance over both the deterministic rule and passive benchmarks. Pairwise Sharpe differences are assessed via stationary block bootstrap (B=5000, automatic block lengths), with transaction-cost robustness up to 30 bps.

Main Results

1. All LLM agents outperform the deterministic Rule Agent in full-sample Sharpe ratio, with the largest differentials for Hawkish ($\Delta$Sharpe=+0.044) and Debate ($\Delta$Sharpe=+0.040) agents. These gains are statistically weak (p<0.10, unadjusted), and confidence intervals include zero, with no outcomes surviving multiple-testing adjustment.
2. Outperformance is regime-conditional: LLM edges are concentrated entirely in the subsequent soft-landing regime (2024–2025), during which macro signals diverged and deterministic rules suffered from sign/rank rigidity. During the rates-peak, all signal-based methods are dominated by the passive inverse-volatility benchmark.
3. Debate Agent is not incrementally superior to the best single prior: The Debate Agent's Sharpe is statistically indistinguishable from the Hawkish Agent (p=0.769), indicating its core role is bias-averaging rather than synergistic deliberation.
4. Transaction cost robustness: LLM-based strategies maintain net Sharpe advantages over passive for cost assumptions up to 30 bps one-way; the Rule Agent’s performance degrades irrecoverably at lower cost levels (~5 bps).

Detailed Interpretations

• The observed excess performance of LLM-based agents is economically modest (~40–44 basis point Sharpe premium relative to rule-based mapping) and is entirely driven by flexible aggregation of familiar macro predictors.
• Attribution analysis indicates that improvements are realized via superior cross-sectional allocation, especially in precious metals (SLV, GLD, PALL), rather than universal commodity overweighting.
• Within-family comparison reveals a lack of deliberation premium: the Debate process primarily stabilizes against prior miscalibration rather than extracting novel or consensus-driven value.

Theoretical and Practical Implications

Theoretical Insights

• The results support the hypothesis that interpretation flexibility in the mapping from structured macro states to asset allocations provides incremental value in non-stationary, regime-shifting environments, a setting where directionally fixed loading matrices are suboptimal.
• The absence of deliberation alpha in the Debate Agent challenges the notion that inter-agent LLM debate (as in open-ended question answering) translates into enhanced signal extraction in strictly bounded quantitative regimes.
• Methodologically, the study establishes a template for attribution-controlled evaluation of financial AI, isolating marginal value by rigorous ablation of all non-interpretation layers.

Practical Implications

• In a production environment, the utility of LLMs for regulated, audit-sensitive asset allocation is linked to their role as constrained macro interpretation layers, rather than as unrestricted, end-to-end trading engines.
• Given modest gains, operational deployment should focus on reducing prior-selection risk and providing regime-aware robustness via interpretable aggregation rather than on speculative pursuit of deliberation alpha.
• Real-world persistence of Sharpe differential is unproven; performance is highly sample-dependent and could vanish in other cycles.

Limitations and Directions for Future Research

• Sample constraint: Only one cycle (rates-peak to soft landing) is covered; result generalizability is unknown.
• Non-vintage macro: Macroeconomic data is not fully vintage (ALFRED recommended), exposing potential lookahead biases.
• Asset universe impurity: Inclusion of equity cash-flow proxies (e.g., COWZ) dilutes pure commodity exposure, warranting robustness over a strictly commodity-only universe.
• Prompting/stability risk: Absence of masked-date robustness tests leaves open the possibility that LLMs could leverage residual pretrained calendar knowledge.
• No formal multiple-testing adjustment: Reported p-values should be interpreted as exploratory, not confirmatory.

Future work should prioritize longer out-of-sample evaluations, full-vintage macro handling, masked-date prompt protocols, stricter asset universes, and evaluation of dynamic prior-weighting or meta-learning mechanisms for prior selection.

Conclusion

Within a regime-controlled, ablation architecture, LLM-based macro interpretation exhibits small but consistent Sharpe improvements over deterministic rule-based portfolio mappings for commodity-related ETF allocation. The practical advantage lies in interpreting conflicting or ambiguous macroeconomic signals where classical sign-mapping fails, with contributions concentrated in periods of cross-signal ambiguity. The multi-agent "Debate" mechanism serves primarily to hedge prior misalignment, not to generate substantive independent value. Real-world adoption should be cautious, as sustained alpha and operational cost justification remain unproven. As such, LLMs are best deployed as modular, auditable interpretation layers, not all-in-one trading systems, until more comprehensive evidence emerges regarding their regime-invariant efficacy and robustness.