Researchers aggregate zero-shot LLM outputs for better stock return prediction

Researchers have developed a lightweight supervised aggregator to combine outputs from multiple zero-shot Large Language Models (LLMs) for classifying corporate disclosures. This method aims to improve prediction accuracy by leveraging diverse perspectives from LLMs that can analyze financial disclosures without task-specific fine-tuning. The aggregator demonstrated superior performance compared to individual LLM classifiers and baseline models, particularly for disclosures with mixed signals where different models produced conflicting predictions. The study suggests that while zero-shot LLMs capture valuable financial signals, a supervised aggregation approach is key to maximizing these gains. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Enhances LLM utility in financial analysis by improving accuracy in corporate disclosure classification.

RANK_REASON Academic paper on a novel method for LLM output aggregation.

Read on arXiv cs.AI →

• LLM
• arXiv
• Kemal Kirtac
• FinBERT
• U.S.

• paper
• product