This working paper evaluates four state-of-the-art LLMs (ChatGPT, Gemini, DeepSeek, Perplexity) on IBEX-35 stock prediction tasks and finds that LLMs can deliver economically meaningful outperformance only when combined with structured prompts and active human oversight; recurring failures (data retrieval, financial misinterpretation, arithmetic and meta-reasoning) limit autonomous deployment, while grounding recommendations in regulatory filings substantially improves accuracy and risk-adjusted returns
速读内容
结论摘要 — 人工监督是必要条件
LLMs can outperform the benchmark, but only under supervised chain-of-thought (CoT) workflows that expose intermediate calculations for human correction; unsupervised naïve queries show weak or no excess returns .
Regulatory filings (CNMV disclosures) materially improve accuracy and Information Ratios, especially for naïve and structured prompts, by providing standardized, high-integrity grounding data .
Recurring taxonomy of reasoning failures (live data, financial interpretation, computation, meta-reasoning) explains why autonomous use is risky and why human-in-the-loop (HITL) mitigations are effective .
Trading strategy & evaluation framework
Item
Specification
Investment universe
35 IBEX-35 stocks
Construction
Long 5 highest-scoring, Short 5 lowest-scoring (equal weight)
Holding period
Monthly (rebalance first trading day)
Evaluation window
Apr 2025 — Jan 2026 (10 monthly cycles)
Performance measured by excess return (alpha), Information Ratio (IR), directional accuracy and weighted F1-score .
Each sub-metric has sub-weights (e.g., within Valuation: P/E 60% / P/B 40%) and normalization to [0,1] based on typical IBEX-35 ranges .
LLM reasoning failures — taxonomy and implications
Live data retrieval: entity mismatches, fabricated figures, temporal mixing of periods lead to incorrect inputs .
Financial interpretation: misreading ratio directionality (e.g., treating low P/E as negative), currency/accounting conflation .
Computational processing: aggregation/weighting mistakes, carryover of stale outputs, missing-value shortcuts .
Meta-reasoning & transparency: failed self-correction, opaque justifications, inconsistent processing depth—these cause fluent but unreliable outputs .
Performance summary & model ranking (aggregate metrics)
Model
Excess Return (avg)
Information Ratio
Accuracy
F1-Score
ChatGPT
0.013
0.29
0.560
0.526
DeepSeek
0.019
0.35
0.555
0.531
Gemini
0.022
0.64
0.568
0.546
Perplexity
0.021
0.48
0.589
0.556
Perplexity and Gemini lead across many metrics; Perplexity particularly strong in classification and structured/CoT settings, while DeepSeek shows the most conservative / inconsistent improvement pattern .
Performance over time (subperiod improvement)
Successive model updates (Apr 2025 → Jan 2026) improved arithmetic accuracy and reduced hallucinations; performance gains are visible across prompting strategies but reasoning failures persist .
CoT + regulatory filings yields the highest cumulative excess return (Perplexity CoT from 3.5% → 4.0% monthly; ~40.2% cumulative excess in 10 months) conditional on model ability to parse attachments .
Practical mitigations recommended
Enforce "show your work", verify data provenance (citations), perform iterative validation, and embed consistent Human-in-the-Loop oversight to prevent silent propagation of errors
思维导图
深度阅读
下面是对“Large Language Models and Stock Investing: Is the Human Factor Required?”(CNMV Working Paper No.96,March 2026)一篇研究报告的逐章、逐表、逐图的详尽解析与解构。为了便于溯源,凡直接基于报告内容的陈述,句末都标注了来源页码标识(如 )。总体分析结构包括:元数据与概览、逐节深度解读、图表/图片逐一解读、估值与评价指标说明、风险与缓解建议、批判性视角与结论性综合。
一、元数据与概览(引言与报告概览)
报告标题与作者:Large Language Models and Stock Investing: Is the Human Factor Required?,作者为 Ricardo Crisóstomo 和 Diana Mykhalyuk,发布机构为西班牙证券市场监管机构的研究系列(CNMV Working Paper No.96),发布日期为 2026 年 3 月。该信息载于题头与摘要部分。