FinVerse: An Autonomous Agent System for Versatile Financial Analysis

TL;DR

FinVerse addresses the challenge of deploying stable, data-grounded autonomous agents for finance by integrating a large-scale, domain-specific API ecosystem (over 642 APIs across ten domains) with an embedded code interpreter and an SFT-tuned workflow. The system relies on a memory-enabled Profile and Overall Plan, plus modules for Summary&Reflexion and a specialized tool pipeline (api-select, api-details, code-exec) to execute executable data-analysis tasks. Key contributions include (i) a FinVerse architecture tailored to finance, (ii) fine-tuning open-source LLMs via four sub-tasks to substitute for heavy-duty models, and (iii) an open, real-world dataset (14,107 questions) for agent development, with demonstrated improvements in task planning, tool usage, code generation, and end-to-end robustness. Empirical results show that FinVerse with SFT outperforms baselines in task-specific and end-to-end metrics, achieving strong robustness and a favorable efficiency-accuracy balance, signaling practical impact for real-time financial analysis and decision support. The work also contributes a reproducible dataset and a scalable API-based framework that can be extended beyond finance to other specialized domains, with a YouTube demo validating the end-to-end pipeline under realistic scenarios.

Abstract

With the significant advancements in cognitive intelligence driven by LLMs, autonomous agent systems have attracted extensive attention. Despite this growing interest, the development of stable and efficient agent systems poses substantial practical challenges. In this paper, we introduce FinVerse, a meticulously crafted agent system designed for a broad range of financial topics. FinVerse integrates over 600 financial APIs, enabling access to more accurate and extensive financial information compared to generalist agents. To enhance financial information processing capabilities, FinVerse is equipped with an embedded code interpreter, enabling the execution of complex data analysis tasks with precision and efficiency. Our work includes an empirical comparison of several LLMs in driving FinVerse. Specifically, we propose our own scheme for training LLMs using SFT to optimize LLM performance within FinVerse. Recognizing the scarcity of specialized datasets to build LLMs for agents, we have constructed a dataset and plan to make it open-source, providing a valuable resource for peer application developers. The demo video has been released on YouTube at https://www.youtube.com/watch?v=sk8L9_Wv7J4

Figures (6)

• Figure 1: A comparative illustration of the anticipated outputs across various systems
• Figure 2: The overall architecture of FinVerse
• Figure 3: Demo showcase of agent meta
• Figure 4: Demo showcase of code-exec
• Figure 5: Prompt to generate agent meta.