Emergent Abilities of Large Language Models (LLMs) in Finance

This dissertation advances the application of Large Language Models (LLMs) in financial domains through three essays exploring novel approaches to enhance their performance and adaptability. Essay 1, "A comparative analysis of instruction fine-tuning LLMs for financial text classification," explores the effectiveness of instruction fine-tuning smaller-scale models (Mistral-7B, Llama3-8B, and Phi3-mini) for adapting general-domain LLMs to specialized financial tasks. While instruction fine-tuned models demonstrated robust performance on financial classification tasks, they showed performance degradation on unseen tasks. This degradation was effectively addressed through a novel model merging technique that integrated domain-specific fine-tuned models with the base model, resulting in enhanced zero-shot performance on unseen tasks. Essay 2, "Enhancing financial question answering with a multi-agent reflection framework," addresses the challenges of numerical reasoning in financial question answering by developing a multi-agent framework with specialized critic agents that reflect on reasoning steps and final answers. This framework demonstrates significant performance improvements over single-agent reasoning and performs competitively with larger models while offering a cost-effective solution for financial QA tasks. Essay 3, "FinVision: A multi-agent framework for stock market prediction," introduces a multi-modal multi-agent system for financial trading that integrates specialized LLM-based agents to process diverse financial data types. The framework incorporates a reflection module for analyzing historical trading patterns, with ablation studies confirming that the visual reflection component significantly enhances the system's decision-making capabilities.