FinTeam: A Multi-Agent Collaborative Intelligence System for Comprehensive Financial Scenarios

1 Introduction

• FinTeam assigns financial tasks to a document analyzer, analyst, accountant, and consultant, reflecting specialist teams in financial companies.The agents collaborate to provide comprehensive solutions across macroeconomic, industry, and company analysis.Source: §1 · p. 2, Figure 1 · p. 2
• The agents process financial texts, analyze materials using knowledge bases, calculate numerical results with tools, and answer finance questions.The system trains agents with specific financial expertise and a constructed dataset.Source: §1 · pp. 2–3
• 62.00% human preference evaluations produced a winning rate for FinTeam, whose overall score was 4.86 out of 5 and exceeded baselines by 0.03 to 0.82 points.The evaluation used 150 real investor queries spanning three financial scenarios.Source: §1 · p. 2
• The authors evaluate the system on comprehensive real-world scenarios, beyond the simple financial tasks emphasized in prior evaluations.Both automated evaluations and human assessments are used to assess the workflow.Source: §1 · p. 3

2.1 LLMs for Finance

• Financial LLMs target distinct tasks, including structured risk assessment and entity linking, chart interpretation, and deep financial reasoning.Examples include PIXIU, FinVis-GPT, and InvestLM.Source: §2.1 · p. 4
• Single-agent financial models are described as limited in decomposing and solving complex, multi-step financial tasks.This motivates exploring modular and collaborative multi-agent systems.Source: §2.1 · p. 4

2.2 Multi-Agent Collaboration

• Multi-agent approaches use role-playing, collaboration, and task decomposition; examples span general LLM frameworks and structured legal and medical collaborations.AutoGen and MetaGPT provide collaboration frameworks, while SMART, MASER, MDAgents, and AI Hospital apply coordination in other domains.Source: §2.2 · p. 4
• Financial agent studies include systems for investment and trading decisions, but most remain trading-focused and lack broader macroeconomic, industry, and company analysis.Examples include TradingGPT, FinMem, and FinAgent.Source: §2.2 · p. 4

3 FinTeam

• FinTeam defines document analyst, analyst, accountant, and consultant roles, trained for specific financial skills and deployable individually or collaboratively.Collaboration is organized around macroeconomic, industry, and company analysis.Source: §3 · p. 4

Document Analyzer

• The document analyzer performs intent recognition, financial entity extraction, and sentiment analysis on financial texts.Training draws on labeled open-source datasets and financial-report text automatically annotated by ChatGPT.Source: Document Analyzer · p. 5

Analyst

• The analyst uses retrieval-augmented generation to analyze current financial materials retrieved from multiple knowledge bases.The Chain-of-Retrieval dataset generates questions from financial contexts, retrieves reference documents, and creates answers grounded in those contexts.Source: Analyst · p. 5

Accountant

• The accountant uses computational tools to address calculation challenges, generating tool commands that are executed and integrated into its response.Its tools support quantitative-analysis queries with commands, inputs, and outputs.Source: Accountant · pp. 5–6

Consultant

• The consultant interprets and answers finance-related queries, with training data drawn from Chinese-language FiQA, finance-term questions, and forum queries.The dataset also includes multi-turn dialogues created from seed topics.Source: Consultant · p. 6

3.2 Scenario Settings

• Scenario workflows assign macroeconomic analysis to term extraction, explanation, supplementary data gathering, and synthesis; industry analysis adds sector research and requested recent news.Company analysis combines PEST and SWOT assessment, optional sentiment analysis, financial-statement retrieval, ratio calculation, and consultant synthesis.Source: §3.2 · p. 7

4.1 Evaluation Data and Setups

• The test set contains 150 investor inquiries, with 50 questions each for macroeconomic, industry, and company analysis.Questions came from NGA Grand Era and were rule-filtered for diversity; industry questions cover 27 sub-sectors.Source: §4.1 · p. 7
• GPT-4o rates accuracy, thoroughness, clarity, and professionalism on a 1-to-5 scale, while finance undergraduates choose the best answer anonymously.The human-evaluation acceptance rate measures how often a model's output is selected as the top answer.Source: §4.1 · pp. 7–8

4.2 Implementation Details

• Agents are fine-tuned with LoRA on Qwen2.5-7B-Chat for two epochs, using a maximum sequence length of 4096.Training uses four NVIDIA V100-32G GPUs and a learning rate of 5 × 10−5.Source: §4.2 · p. 8

4.3 Main Results

• FinTeam outperforms GPT-3.5-turbo and Xuanyuan-13B across all dimensions, with significant improvements in thoroughness, professionalism, and overall score.The evaluation uses GPT-4o ratings and pairwise significance tests; finance undergraduates also anonymously select the best responses.Source: §4.3 · p. 8, §4.1 · p. 8
• Table 3 reports FinTeam's highest overall score, 0.08 above the baseline, with gains of 0.25 in thoroughness and 0.16 in professionalism.The reported improvements in thoroughness, professionalism, and overall score are statistically significant.Source: Table 3 · p. 8, Table 4 · p. 8
• 62.00% of human evaluation selections went to FinTeam, which was chosen 93 times across 150 test cases.The authors report that this preference result is consistent with the GPT-4o assessment.Source: Table 5 · p. 9, §4.3 · p. 9

5 Analysis

• The authors investigate how individual agents perform across three aspects to understand FinTeam's generation quality.Source: §5 · p. 9

5.1 Evaluation Setup

• The analysis examines NLP, financial knowledge, and data-analysis abilities using FinCUGE, FinEval, and a manually created calculation dataset.FinCUGE uses accuracy, F1, and ROUGE; FinEval uses multiple-choice accuracy, and calculation tasks assess formula and result accuracy.Source: §5.1 · p. 9

5.2 Results on Financial NLP Tasks

• The Document analyzer scores 47.20 on average, 7.43 points above Qwen2.5-7B-Instruct on FinCUGE's six financial NLP tasks.It achieves the highest performance in all task-specific metrics.Source: §5.2 · pp. 9–10, Table 6 · p. 10

5.3 Results on Financial Knowledge Tests

• The Consultant outperforms Qwen2.5-7B-Instruct by 2.06 points on the out-of-distribution FinEval benchmark.FinEval contains 1,151 multiple-choice questions across 34 financial subcategories.Source: §5.1 · p. 9, Table 7 · p. 10, §5.3 · p. 10

5.4 Results on Data Analysis

• Accountant is 0.09 higher than the baseline in Formula&Result accuracy when computational plugins are added to the model.The 100-problem evaluation measures accuracy in formula construction and result calculation.Source: §5.1 · p. 9, Table 8 · p. 10, §5.4 · p. 10

• FinTeam connects multiple subtasks through LLM-agent interaction to handle complex financial tasks.Source: §6 · p. 11
• Four agents trained on different sub-datasets collaborate in macroeconomic, industry, and company analysis workflows.Source: §6 · p. 11
• The authors establish multidimensional financial benchmarks and report robust capabilities across the three scenarios.Source: §6 · p. 11

• The scenario design leaves out many financial tasks, which the authors identify as a direction for future exploration.Source: Limitations · p. 11
• The system is tailored to Chinese financial contexts, and its effectiveness in global markets remains untested.Source: Limitations · p. 11
• Because the system may generate investment-related advice, users should be cautious; financial outcomes are not guaranteed.Source: Limitations · p. 11

7.1 Information of NLP datasets

• The document analyzer is trained on sentiment analysis, entity extraction, and summarization datasets; Figure 9 lists the open-source sources.Subsets exceeding 10,000 samples are randomly sampled down to 10,000 each.Source: §7.1 · p. 13
• Table 9 reports data statistics for the financial NLP datasets.Source: Table 9 · p. 13