Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An Agent Framework for Real-Time Financial Information Searching with Large Language Models

Jinzheng Li, Jingshu Zhang, Hongguang Li, Yiqing Shen

TL;DR

FinSearch tackles the problem of real-time financial information retrieval by integrating LLM reasoning with diverse, up-to-date data sources and a temporal context-aware search process. It introduces a DAG-based multi-step pre-planner, an adaptive search executor with a dynamic query rewriter, a 72-hour temporal weighting mechanism, and an LLM-driven response generator to produce coherent, temporally contextual analyses that include visualizations. The authors validate the approach on FinSearchBench-24, a 1,500-question benchmark spanning stock, rate, policy, and industry domains, and demonstrate superior accuracy across multiple LLM backbones compared to baselines, with acceptable latency. Key findings show that temporal weighting provides substantial gains and that combining temporal awareness with adaptive query refinement yields the best overall performance, suggesting strong practical impact for finance professionals requiring timely, precise insights from integrated data streams.

Abstract

Financial decision-making requires processing vast amounts of real-time information while understanding their complex temporal relationships. While traditional search engines excel at providing real-time information access, they often struggle to comprehend sophisticated user intentions and contextual nuances. Conversely, Large Language Models (LLMs) demonstrate reasoning and interaction capabilities but may generate unreliable outputs without access to current data. While recent attempts have been made to combine LLMs with search capabilities, they suffer from (1) restricted access to specialized financial data, (2) static query structures that cannot adapt to dynamic market conditions, and (3) insufficient temporal awareness in result generation. To address these challenges, we present FinSearch, a novel agent-based search framework specifically designed for financial applications that interface with diverse financial data sources including market, stock, and news data. Innovatively, FinSearch comprises four components: (1) an LLM-based multi-step search pre-planner that decomposes user queries into structured sub-queries mapped to specific data sources through a graph representation; (2) a search executor with an LLM-based adaptive query rewriter that executes the searching of each sub-query while dynamically refining the sub-queries in its subsequent node based on intermediate search results; (3) a temporal weighting mechanism that prioritizes information relevance based on the deduced time context from the user's query; (4) an LLM-based response generator that synthesizes results into coherent, contextually appropriate outputs. To evaluate FinSearch, we construct FinSearchBench-24, a benchmark of 1,500 four-choice questions across the stock market, rate changes, monetary policy, and industry developments spanning from June to October 2024.

An Agent Framework for Real-Time Financial Information Searching with Large Language Models

TL;DR

FinSearch tackles the problem of real-time financial information retrieval by integrating LLM reasoning with diverse, up-to-date data sources and a temporal context-aware search process. It introduces a DAG-based multi-step pre-planner, an adaptive search executor with a dynamic query rewriter, a 72-hour temporal weighting mechanism, and an LLM-driven response generator to produce coherent, temporally contextual analyses that include visualizations. The authors validate the approach on FinSearchBench-24, a 1,500-question benchmark spanning stock, rate, policy, and industry domains, and demonstrate superior accuracy across multiple LLM backbones compared to baselines, with acceptable latency. Key findings show that temporal weighting provides substantial gains and that combining temporal awareness with adaptive query refinement yields the best overall performance, suggesting strong practical impact for finance professionals requiring timely, precise insights from integrated data streams.

Abstract

Financial decision-making requires processing vast amounts of real-time information while understanding their complex temporal relationships. While traditional search engines excel at providing real-time information access, they often struggle to comprehend sophisticated user intentions and contextual nuances. Conversely, Large Language Models (LLMs) demonstrate reasoning and interaction capabilities but may generate unreliable outputs without access to current data. While recent attempts have been made to combine LLMs with search capabilities, they suffer from (1) restricted access to specialized financial data, (2) static query structures that cannot adapt to dynamic market conditions, and (3) insufficient temporal awareness in result generation. To address these challenges, we present FinSearch, a novel agent-based search framework specifically designed for financial applications that interface with diverse financial data sources including market, stock, and news data. Innovatively, FinSearch comprises four components: (1) an LLM-based multi-step search pre-planner that decomposes user queries into structured sub-queries mapped to specific data sources through a graph representation; (2) a search executor with an LLM-based adaptive query rewriter that executes the searching of each sub-query while dynamically refining the sub-queries in its subsequent node based on intermediate search results; (3) a temporal weighting mechanism that prioritizes information relevance based on the deduced time context from the user's query; (4) an LLM-based response generator that synthesizes results into coherent, contextually appropriate outputs. To evaluate FinSearch, we construct FinSearchBench-24, a benchmark of 1,500 four-choice questions across the stock market, rate changes, monetary policy, and industry developments spanning from June to October 2024.

Paper Structure

This paper contains 13 sections, 3 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Overview
  4. Search Pre-Planner
  5. Search Executor with Dynamic Query Rewriter
  6. Temporal Weighting Mechanism
  7. Response Generator
  8. Experiments
  9. Benchmark Construction
  10. Implementation Details
  11. Results
  12. Ablation Study
  13. Conclusion

Figures (2)

  • Figure 1: Overview of the FinSearch. The user's financial query is first interpreted by a pre-planner that constructs a directed graph where nodes represent atomic search tasks connected to specific data sources. The search executor then traverses this graph, with a dynamic rewriter refining queries at each step based on emerging insights. The temporal mechanism evaluates the time-sensitivity of gathered information before the response generator creates a unified analysis incorporating both textual and visual elements.
  • Figure 2: Case study for the ablation study. We visualize FinSearch's analytical process across different configurations when analyzing Chinese stock market trends. It compares four distinct scenarios denoted by different colors: optimal configuration (green) showing accurate analysis aligned with market data, partial configuration without dynamic query rewriter (red) demonstrating divergent conclusions, and complete configuration with detailed temporal context (blue) providing comprehensive market insights. Each panel illustrates FinSearch's query decomposition, information processing, and response generation, highlighting how different components contribute to analysis quality. The comparison reveals that the full configurations with all components enabled produce more nuanced, temporally-aware, and well-supported financial search results compared to configurations missing any features.