What an LLM for financial analysis can—and cannot—do大模型金融分析能做什么、不能做什么
Large language models are strong at transforming language: finding relevant passages, extracting disclosures, comparing narrative changes, summarizing filings, classifying risks and drafting an analyst-readable explanation. They do not inherently know the latest filing, guarantee correct arithmetic, preserve units or distinguish a restated value unless the surrounding system supplies and checks those facts.
大语言模型擅长处理语言：寻找相关段落、提取披露、比较叙述变化、总结监管文件、分类风险并起草便于分析师阅读的解释。但模型不会天然掌握最新文件，也不保证算术正确、单位一致或识别重述数据；这些事实必须由外围系统提供并验证。
Good fit适合
Document search, extraction, summarization, topic comparison and cited narrative synthesis.文档检索、信息提取、摘要、主题比较和带引用的叙述综合。
Needs tools需要工具
Ratios, period alignment, currency conversion, screening and any repeatable calculation.比率、期间对齐、汇率换算、筛选和任何需要重复执行的计算。
Needs human approval需要人工批准
Forecast assumptions, material conclusions, compliance-sensitive output and investment decisions.预测假设、重大结论、合规敏感输出和投资决策。
Financial analysis tasks: model, tools or analyst?金融分析任务应该交给模型、工具还是分析师
| Task任务 | LLM role大模型角色 | Required support必要支持 | Approval审批 |
|---|---|---|---|
| Summarize a 10-K总结10-K | Draft cited summary起草带引用摘要 | Filing retrieval and section parser文件检索与章节解析 | Analyst review分析师复核 |
| Compare financial statements比较财务报表 | Explain changes解释变化 | Normalized tables and code标准化表格与代码 | Material items重大项目 |
| Calculate ratios计算财务比率 | Select and explain选择并解释 | Deterministic formula engine确定性公式引擎 | Definition check口径检查 |
| Forecast cash flow预测现金流 | Describe scenarios描述情景 | Model, assumptions and sensitivities模型、假设与敏感性 | Required必须 |
| Recommend an investment给出投资建议 | Not autonomous不得自主决定 | Research, valuation and risk controls研究、估值与风险控制 | Required必须 |
Evidence-first LLM financial analysis architecture证据优先的大模型金融分析架构
RAG helps retrieval; it does not guarantee truthRAG改善检索，但不能保证事实正确
Retrieval-augmented generation can supply current, relevant documents and citations. It can still retrieve the wrong period, miss a table footnote, mix consolidated and segment values, or cite a passage that does not support the generated claim. Evaluate retrieval relevance, citation entailment, numeric consistency and answer completeness separately.
检索增强生成可以提供当前相关文档和引用，但仍可能检索错误期间、遗漏表格脚注、混合合并与分部数据，或引用并不支持生成结论的段落。因此需要分别评估检索相关性、引用是否真正支持主张、数字一致性和回答完整性。
Make LLM financial statement analysis numerically reliable让大模型财务报表分析在数字上可靠
The model may choose a formula or explain a result, but a deterministic service should perform arithmetic. Inputs need explicit periods, units, signs, currencies, restatement status and source locations. The output should expose both the formula and the exact values used.
模型可以选择公式或解释结果，但算术应由确定性服务执行。输入需要明确期间、单位、正负号、货币、是否重述和来源位置；输出则应同时展示公式与实际使用的数值。
Keep calculations outside free-form generation让计算脱离自由文本生成
A strong pattern is tool → structured result → explanation. The calculation tool returns named inputs, formula, result and provenance; the LLM converts that object into a readable narrative. If the tool fails or required data is missing, the system should return “not verified” rather than approximate a number.
稳健模式是“工具→结构化结果→文字解释”。计算工具返回具名输入、公式、结果与来源，大模型再把该对象转化为易读叙述。如果工具失败或缺少必要数据，系统应返回“未验证”，而不是估算数字。
Controls for auditable AI financial analysis可审计AI金融分析的控制措施
Require claim-level citations with document, date and location.重大主张必须带有文档、日期和位置级引用。
Use tested, versioned functions and reconcile totals.使用经过测试、版本化的函数并核对总数。
Assign owners, escalation rules and approval thresholds.分配负责人、升级规则和审批阈值。
- Create a golden test set.建立黄金测试集。 Include tables, footnotes, restatements, missing data and contradictory disclosures.覆盖表格、脚注、重述、缺失数据和矛盾披露。
- Measure components separately.分别评估各组件。 Track retrieval, extraction, calculation, citation and narrative errors.分别跟踪检索、提取、计算、引用和叙述错误。
- Log every material step.记录每个重大步骤。 Store model, prompt, tool input, output, source version and approval.保存模型、提示词、工具输入输出、来源版本与审批。
- Fail safely.安全失败。 Block publication when evidence is missing or validations disagree.当证据缺失或验证不一致时阻止发布。
Use a financial analysis prompt contract使用金融分析提示词契约
A production prompt should define the task, permitted sources, accounting period, entity scope, units, output schema and refusal conditions. Tell the model to distinguish disclosed facts from calculations, assumptions and interpretations. Require every material sentence to carry a source identifier or calculation identifier, and require “not found” when the supplied evidence does not answer the question.
生产级提示词应明确任务、允许使用的来源、会计期间、实体范围、单位、输出结构和拒答条件。要求模型区分已披露事实、计算结果、假设与分析解释；每个重大句子必须带来源标识或计算标识，当所给证据无法回答问题时必须返回“未找到”，而不是补全看似合理的答案。
Return structured claims before prose先返回结构化主张，再生成叙述
A useful intermediate object contains claim text, claim type, source location, period, currency, unit, formula, confidence and validation status. Validators can reject missing citations, mismatched periods, impossible signs or totals that do not reconcile before the LLM turns approved claims into prose.
有用的中间对象应包含主张文本、主张类型、来源位置、期间、货币、单位、公式、置信度和验证状态。验证器可以在大模型把获批主张转化为叙述之前，拒绝缺少引用、期间不匹配、符号异常或总数无法勾稽的结果。
Separate abstention from failure区分主动拒答与系统失败
“The evidence does not support an answer” is a successful safety outcome. A timeout, parser error or unavailable API is an operational failure. Record them differently so evaluation rewards honest abstention and engineering teams can diagnose actual system defects.
“现有证据不足以回答”属于成功的安全结果；超时、解析错误或API不可用则属于运行故障。两者需要分别记录，使评估体系奖励诚实拒答，同时让工程团队能够定位真正的系统缺陷。
Connect an LLM for financial analysis with QVeris and FMP用QVeris与FMP连接大模型金融分析
QVeris can help an agent discover and invoke authorized FMP APIs for financial statements, company information and other available datasets, then pass structured responses into calculation and retrieval tools. The current FMP documentation and subscription determine endpoint availability, fields, history, entitlements and usage limits.
QVeris可以帮助智能体发现并调用已授权的FMP API，获取财务报表、公司信息及其他可用数据集，再把结构化响应传入计算和检索工具。端点可用性、字段、历史范围、权限与使用限制取决于当前FMP文档和订阅。
QVeris is a discovery and orchestration layer, not a substitute for licensed data, model governance, financial expertise or accountable decision-making. Preserve every tool response and source identifier so analysts can trace the final answer.
QVeris是发现与编排层，不能替代数据许可、模型治理、金融专业知识或负责决策。应保存每次工具响应和来源标识，使分析师能够追溯最终答案。
Financial RAG guide金融RAG指南 · Financial statement analysis财务报表分析 · QVeris documentationQVeris文档
LLM for financial analysis FAQ大模型金融分析常见问题
Can an LLM analyze financial statements?大模型能分析财务报表吗？
It can extract, compare and explain financial information, but reliable work requires verified documents, structured tables, deterministic calculations, citations and human review.
它可以提取、比较和解释财务信息，但可靠工作需要已验证文档、结构化表格、确定性计算、引用和人工复核。
Is RAG enough to prevent hallucinations?RAG足以防止幻觉吗？
No. RAG can retrieve current evidence, but retrieval, citation and generation can each fail. Validate claim support and numeric consistency.
不足以。RAG可以检索当前证据，但检索、引用与生成都可能失败，仍需验证主张支持关系和数字一致性。
Should the LLM calculate financial ratios?应该让大模型直接计算财务比率吗？
Use the LLM to select or explain formulas, but use tested code for arithmetic and expose inputs, definitions and sources.
可让大模型选择或解释公式，但算术应使用经过测试的代码，并展示输入、定义和来源。
What financial tasks always need human review?哪些金融任务始终需要人工复核？
Material conclusions, forecasts, valuation assumptions, compliance-sensitive outputs and investment decisions require accountable review.
重大结论、预测、估值假设、合规敏感输出和投资决策都需要明确负责人复核。
How do you evaluate an LLM financial workflow?如何评估大模型金融工作流？
Measure retrieval relevance, extraction accuracy, calculation correctness, citation support, completeness, reproducibility and safe failure.
应衡量检索相关性、提取准确率、计算正确性、引用支持度、完整性、可复现性和安全失败能力。
How does QVeris fit into the workflow?QVeris在工作流中扮演什么角色？
QVeris helps agents discover and call authorized APIs, providing structured external data for controlled calculations and cited synthesis.
QVeris帮助智能体发现和调用授权API，为受控计算和带引用的综合分析提供结构化外部数据。
References and methodology sources参考资料与方法来源
- U.S. GAO — AI use and oversight in financial services
- SEC — AI, finance and hallucination risks
- FMP — Financial data API documentation
Educational content only; not investment advice. Verify every material output before relying on it.本文仅供教育用途，不构成投资建议。依赖任何重大输出前都应完成验证。