Signal or Noise in Multi-Agent LLM-based Stock Recommendations?

Battle Count: 109

Highly relevant to quantitative trading. The paper provides the first portfolio-level Monte Carlo validation of a deployed multi-agent LLM equity system, demonstrating statistically significant stock selection skill (p=0.003) on S&P 500. The strong-buy signal functions as an effective universe filter that can sit upstream of any portfolio-construction process. The NNLS attribution reveals adaptive agent integration mechanisms. The IC/ICIR framework connects to classical factor-model practice. The beta robustness check (β=0.865, α=+1.18%/month) addresses a key concern for quant practitioners. The practical framing as a universe filter rather than a point-in-time alpha forecast is directly actionable for systematic equity strategies.

High implementation complexity. Requires: (1) a multi-agent LLM system with four specialist agents and a synthesis agent, each producing structured text analyses; (2) live execution at each observation date to avoid look-ahead bias; (3) text embedding pipeline using text-embedding-3-small; (4) NNLS optimization for per-stock-date attribution; (5) Monte Carlo simulation framework with 10,000 draws; (6) cross-sectional IC/ICIR computation with proper date-level aggregation; (7) fixed-cohort universe management. The system is proprietary (MarketSenseAI), and replicating the full pipeline requires access to the underlying LLM agents, prompts, and financial data feeds. The statistical methodology (MC test, NNLS, IC) is well-described and implementable, but the signal generation layer is the primary barrier.

The paper uses a fixed random seed for Monte Carlo simulations (K=10,000) and specifies the embedding model (text-embedding-3-small). However, MarketSenseAI is a proprietary deployed system by Alpha Tensor Technologies, and the underlying LLM agents, prompts, and full pipeline are not open-sourced. The fixed-cohort design (467 S&P 500 stocks, 94 S&P 100 stocks) and observation dates are specified, but the actual agent outputs and thesis texts are not publicly available. The NNLS methodology is fully described mathematically.