Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Customized FinGPT Search Agents Using Foundation Models

Felix Tian, Ajay Byadgi, Daniel Kim, Daochen Zha, Matt White, Kairong Xiao, Xiao-Yang Liu Yanglet

TL;DR

Experiments show that FinGPT Search Agent outperform existing models in accuracy, relevance, and response time, making them promising for real-world financial applications.

Abstract

Current large language models (LLMs) have proven useful for analyzing financial data, but most existing models, such as BloombergGPT and FinGPT, lack customization for specific user needs. In this paper, we address this gap by developing FinGPT Search Agents tailored for two types of users: individuals and institutions. For individuals, we leverage Retrieval-Augmented Generation (RAG) to integrate local documents and user-specified data sources. For institutions, we employ dynamic vector databases and fine-tune models on proprietary data. There are several key issues to address, including data privacy, the time-sensitive nature of financial information, and the need for fast responses. Experiments show that FinGPT agents outperform existing models in accuracy, relevance, and response time, making them practical for real-world applications.

Customized FinGPT Search Agents Using Foundation Models

TL;DR

Experiments show that FinGPT Search Agent outperform existing models in accuracy, relevance, and response time, making them promising for real-world financial applications.

Abstract

Current large language models (LLMs) have proven useful for analyzing financial data, but most existing models, such as BloombergGPT and FinGPT, lack customization for specific user needs. In this paper, we address this gap by developing FinGPT Search Agents tailored for two types of users: individuals and institutions. For individuals, we leverage Retrieval-Augmented Generation (RAG) to integrate local documents and user-specified data sources. For institutions, we employ dynamic vector databases and fine-tune models on proprietary data. There are several key issues to address, including data privacy, the time-sensitive nature of financial information, and the need for fast responses. Experiments show that FinGPT agents outperform existing models in accuracy, relevance, and response time, making them practical for real-world applications.

Paper Structure

This paper contains 32 sections, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Web Agents
  4. Information Retrieval (IR)
  5. Large Language Models (LLMs)
  6. LLM Finetuning Methods
  7. Retrieval-Augmented Generation (RAG)
  8. LLM Customization Methods
  9. Local Customizable FinGPT Agent for Individual Users
  10. Overview
  11. User Preferences and Search Engine
  12. Refined Prompt and Long Context Window
  13. Generating Responses & Updating Context
  14. Local Customizable FinGPT Agent for Institutions
  15. Overview
  16. ...and 17 more sections

Figures (5)

  • Figure 1: The overall dataflow of FinGPT agents. A user submits a query, while an agent retrieves relevant data from preferred sources, search engines, local files, or a dynamic vector database. The LLM or fine-tuned financial foundation model (FFM) processes the prompts, refines them, and utilizes Retrieval-Augmented Generation (RAG) to generate responses. Both users' feedback and generated responses are then stored into the local dynamic vector database for future interactions.
  • Figure 2: An overview of customized FinGPT agent for individual users.
  • Figure 3: An overview of customized FinGPT agent for institutional users.
  • Figure 4: Screenshots of the customized FinGPT search agents for individual users (left & middle) and institution users (right).
  • Figure 5: Performance on the FinanceBench dataset islam2023financebench.