Large Language Models for Financial and Investment Management: Applications and Benchmarks

In one sentence

Summary

The paper addresses the urgent need for a systematic examination of large language models in finance, bridging the gap between cutting-edge AI technology and practical implementation in investment management. It aims to synthesize recent developments to help researchers and practitioners understand the diverse applications, methodologies, and impacts of LLMs, which are transforming traditional financial practices through contextual understanding and content generation. The authors position this work as a holistic, application-driven review that fills gaps left by existing surveys, which often lack deep dives into practical challenges or broader implications for financial decision-making. The methodology involves a comprehensive categorization of existing literature into five key application areas: linguistic tasks, sentiment analysis, financial time series, financial reasoning, and agent-based modeling. For each area, the authors analyze specific methodologies such as textual analysis, knowledge-based analysis, forecasting, data augmentation, planning, decision support, and simulations. The paper also compiles a collection of datasets, benchmarks, and code associated with mainstream applications, offering valuable resources for the community. It reviews the evolution from pre-LLM methods to transformer-based architectures, highlighting advancements in handling long documents, multimodal data, and domain-specific adaptations. Key findings indicate that LLMs excel in summarizing complex financial narratives, extracting entities, and quantifying market sentiment, which can enhance investment performance. They are also shown to support financial planning, generate investment recommendations, and simulate market behaviors through agent-based modeling. However, the paper notes that consensus has not yet formed on which applications will most impact current practices, and many efforts are at an early stage. Limitations include challenges with multimodal document structures, high computational costs, and the need for robust evaluation benchmarks. The work serves as a foundational resource for understanding the current state and future potential of LLMs in finance.

Research problem

The rapid evolution of large language models has created a surge in financial applications, but there is a lack of systematic examination of their diverse applications, methodologies, and impact. Existing surveys often lack deep dives into practical challenges or focus primarily on technical aspects without addressing broader implications for financial decision-making. This paper aims to bridge the gap between AI technology and practical implementation in finance.

Core contributions

• Provides a holistic, application-driven framework categorizing LLM applications in finance into five key areas: linguistic tasks, sentiment analysis, time series, reasoning, and agent-based modeling.
• Synthesizes methodologies across textual analysis, knowledge-based analysis, forecasting, data augmentation, planning, decision support, and simulations.
• Compiles a comprehensive collection of datasets, benchmarks, and code associated with mainstream financial LLM applications.
• Highlights the transformative potential of LLMs in tailoring financial advice, enhancing sentiment analysis, and constructing asset allocation programs.
• Identifies key challenges such as handling multimodal documents, long-context processing, and the need for robust evaluation benchmarks.

Method

• Categorization of LLM applications into five key areas: linguistic tasks, sentiment analysis, time series, reasoning, and agent-based modeling.
• Review of methodologies including textual analysis, knowledge-based analysis, forecasting, and data augmentation.
• Analysis of specific models and frameworks such as DocLLM, FinBERT, and knowledge graph approaches.
• Compilation of datasets, benchmarks, and code resources for mainstream financial LLM applications.
• Evaluation of LLM capabilities in summarization, entity recognition, sentiment analysis, and financial reasoning.
• Discussion of challenges in handling multimodal documents and long-context processing.

Results & findings

• LLMs can tailor financial advice to specific individuals and organizations to improve investment goals.
• Investment performance can be enhanced by speedily deploying sentiment analysis from financial news and social media.
• LLMs can construct programs for asset allocation and asset-liability management through reasoning capabilities.
• Knowledge graphs constructed from LLM extractions can enhance question-answering systems and industry classification.
• Multimodal document understanding remains a challenge, with layout-aware models like DocLLM showing promise.
• Consensus has not yet formed on which LLM applications will most impact current financial practices.
• Many LLM efforts in finance are at an early stage, necessitating further research and robust benchmarks.

Data & experiments

• Comprehensive review of existing literature on LLMs in finance, categorizing studies by application area.
• Analysis of methodologies including textual analysis, knowledge-based analysis, forecasting, and agent-based modeling.
• Compilation of datasets, benchmarks, and code resources for mainstream financial LLM applications.
• Discussion of specific models and frameworks such as DocLLM, FinBERT, and various knowledge graph approaches.
• Evaluation of LLM capabilities in summarization, entity recognition, sentiment analysis, and financial reasoning.

Limitations & caveats

• Consensus has not yet formed on which LLM applications will most impact current financial practices.
• Many LLM efforts in finance are at an early stage, necessitating further research and robust benchmarks.
• Challenges remain in handling multimodal documents and long-context processing.
• High computational costs and the need for robust evaluation benchmarks are significant limitations.
• The paper is a survey and does not present new empirical results or models.
• The rapid evolution of the field means that some applications and benchmarks may become outdated quickly.

Curator notes

Discovery status: round2_promoted. Source links: A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist; Dólares or Dollars? Unraveling the Bilingual Prowess of Financial LLMs Between Spanish and English; TradingGPT: Multi-Agent System with Layered Memory and Distinct Characters for Enhanced Financial Trading Performance