Large Language Models and Stock Investing: Is the Human Factor Required?

Battle Count: 68

Highly relevant to quantitative trading as it directly evaluates LLM-generated stock signals in a live trading environment with monthly rebalancing. The paper provides concrete performance metrics (excess returns, Information Ratios) for LLM-based long-short portfolios, identifies specific failure modes that affect trading decisions, and demonstrates that human oversight is critical for reliable signal generation. The findings on prompt engineering, regulatory filing grounding, and reasoning failure taxonomy are directly actionable for practitioners building AI-assisted trading systems. However, the paper focuses on a single market and relatively simple portfolio construction.

The core methodology (querying LLMs with structured prompts, constructing long-short portfolios, computing performance metrics) is moderately complex but well-documented. The multi-factor scoring framework with six categories and specific weights is clearly specified. However, implementing the full pipeline requires: (1) access to multiple LLM APIs with web-browsing capabilities, (2) real-time market data feeds, (3) CNMV regulatory filing retrieval, (4) iterative chain-of-thought human review processes, and (5) proper out-of-sample evaluation protocols. The human-in-the-loop component adds operational complexity.

The paper provides detailed prompt templates (Annex 2), a standardized protocol, and specifies the multi-factor scoring framework with exact weights and metrics. However, it relies on proprietary LLM APIs (ChatGPT, Gemini, DeepSeek, Perplexity) whose versions change frequently, making exact replication difficult. The evaluation uses live market data and real-time model outputs, which cannot be reproduced after the fact. The IBEX-35 universe and CNMV filings are publicly available.