ETFAgents: A Multi-agent System with a Single LoRA-fine-tuned Agent

Exchange-Traded Funds (ETFs) provide diversified and liquid exposure to broad markets and sectors, yet retail ETF timing decisions still depend heavily on rule-based indicators and often underuse unstructured sentiment information. This paper proposes ETFAgents, a neuro-symbolic, role-specialized multi-agent framework for ETF decision support. The system adapts institutional-style separation of duties into five agent teams covering analysis, research, risk management, trading, and managerial arbitration. Unlike purely LLM-driven workflows, ETFAgents combines probabilistic language-model reasoning with deterministic risk governance, including liquidity checks, no-leverage retail constraints, NAV premium or discount adjustment, and a Hard Veto mechanism that can override unsafe candidate actions generated by LLM-assisted agents. For ETF sentiment analysis, we fine-tune Llama 2-7B using LoRA to build ETFlora, which achieves a weighted F1 score of 87.43%. Backtests on the 510210 ETF and 159901 ETF cover January–March 2024 and the full year of 2024. To isolate the effect of the sentiment backbone, we further conduct a 2025 controlled replacement test in which only the Analyst Team Agent’s sentiment model is changed while the downstream agent workflow, risk rules, execution model, transaction cost, slippage, and liquidity constraints remain fixed. The 2025 stress test also includes a Hard Veto conflict case that connects a risk override to investor review. Interpretability is evaluated in the 2026 Q1 audit window through Agent-SHAP, global Role-SHAP summaries, monthly stability diagnostics, and role neutralization. The results suggest that the proposed architecture is associated with improved risk-adjusted ETF decision metrics in the evaluated settings. They should not be interpreted as evidence of universal investment outperformance.