Resolve ticker, exchange, reporting period, event time, currency, and whether the user wants explanation, comparison, or screening.
确认股票代码、交易所、报告期、事件时间、币种，以及用户需要解释、对比还是筛选。
AI Stock Research Workflow
AI 股票研究工作流
AI Stock Research Assistant for Agents面向 Agent 的 AI 股票研究助手
Build an AI stock research assistant that starts with industry valuation, reads market context, creates a screener candidate pool, and reviews ROE, PE, and share float before generating a research memo.
构建一个 AI 股票研究助手：先看行业估值和市场环境，再用 screener 形成候选池，随后检查 ROE、PE 和流通股结构，而不是直接给出买卖结论。
5research steps个研究步骤
ROEfundamental review基本面复核
Floatliquidity check流动性检查
Why an AI Stock Research Assistant Needs a Workflow
为什么 AI 股票研究助手需要工作流
The original QVeris blog shows a practical lesson: a stock research AI should not jump from a ticker to an opinion. A useful assistant first frames the industry, checks sector behavior, reviews market movers, builds a candidate pool, and then asks whether the fundamentals and share structure support further review.
QVeris 原始博客展示了一个很实际的经验：股票研究 AI 不应该从一个 ticker 直接跳到观点。真正有用的助手应该先建立行业背景、检查板块表现、观察市场异动、形成候选池，再判断基本面和股本结构是否值得继续研究。
The AI Stock Research Assistant Workflow
AI 股票研究助手的工作流
Industry
Compare industry valuation before reading company-level moves.
先比较行业估值，再看个股变化。
Market
Read sector performance, gainers, losers, and most active stocks.
读取板块表现、涨幅榜、跌幅榜和活跃股。
Screener
Build a candidate pool using market cap, volume, sector, and trading status.
用市值、成交量、行业和交易状态构建候选池。
Fundamentals
Review ROE, earnings yield, and implied PE for selected companies.
复核 ROE、盈利收益率和隐含 PE。
Float
Check share float so liquidity-driven moves are not misread.
检查流通股，避免把流动性异动误读成基本面机会。
Worked Example: Investigating a Post-Earnings Selloff
完整案例：调查财报后的股价下跌
A useful stock research assistant needs more than a generic workflow. The following example shows how the agent should turn one ambiguous market question into a bounded investigation with evidence, calculations, and explicit stop conditions.
真正有用的股票研究助手不能只有一套抽象流程。下面这个案例展示 Agent 应该如何把一个模糊的市场问题，转化为范围清晰、带证据和计算过程、并且具备停止条件的研究任务。
Compare the move with prior close, intraday range, normal volatility, volume, sector ETF, index, and selected peers.
把本次波动与前收盘、日内区间、正常波动、成交量、行业 ETF、指数和选定同行比较。
Collect the earnings release, filing, presentation, guidance, transcript, and any corporate action published around the event.
收集财报公告、监管文件、演示材料、业绩指引、电话会记录及事件前后发布的公司行动。
Separate reported versus adjusted metrics, calculate growth and margins consistently, and compare results with consensus and prior guidance.
区分报告口径与调整口径，统一计算增长和利润率，并与市场一致预期及此前指引比较。
Label facts, interpretations, and unknowns. State what would confirm or invalidate each explanation instead of forcing one confident narrative.
标记事实、解释和未知项，并说明什么证据能够确认或推翻每种解释，而不是强行给出唯一叙事。
Stop condition:停止条件：
If the quote timestamp, filing period, consensus source, or peer definition cannot be verified, the assistant should return a partial memo and request the missing evidence. It should not fill gaps with plausible-sounding market commentary.
如果无法核实行情时间戳、文件报告期、一致预期来源或同行定义，助手应返回一份明确标注缺口的部分备忘录，并请求补充证据，而不是用听起来合理的市场评论填补空白。
What QVeris Adds to Stock Research AI
QVeris 能为 Stock Research AI 增加什么
DISCOVER
Find the right financial capability
发现合适的金融能力
An agent can search for industry valuation, screener, market mover, key metric, or share float capabilities before calling anything.
Agent 可以先发现行业估值、screener、市场异动、关键指标和 share float 等能力，再决定调用什么。
INSPECT
Check parameters before execution
调用前检查参数
The workflow avoids pretending a screener supports ROE or PE directly when the schema does not provide those filters.
如果 screener 的 schema 不支持直接按 ROE 或 PE 筛选，工作流不会假装它可以做到。
CALL
Return structured research inputs
返回结构化研究输入
The output is a research memo, candidate list, valuation context, and next questions, not investment advice.
输出应该是研究备忘录、候选列表、估值背景和下一步问题，而不是投资建议。
What a Production-Ready Research Memo Should Contain
生产级研究备忘录应该包含什么
The final answer should be a reviewable research object, not one uninterrupted paragraph. A stable output contract also makes it easier to compare model versions, audit tool calls, and decide whether a human analyst has enough evidence to continue.
最终答案应该是一份可复核的研究对象，而不是一整段没有层次的文字。稳定的输出契约还能帮助团队比较模型版本、审计工具调用，并判断人工分析师是否已经获得足够证据继续研究。
Scope and identity
范围与实体身份
State the company, ticker, exchange, reporting currency, research date, event window, and the exact question being answered. Record any assumptions used to resolve an ambiguous symbol.
写明公司、股票代码、交易所、报告币种、研究日期、事件窗口和需要回答的具体问题。若股票代码存在歧义，还要记录实体解析时使用的假设。
Required: identifiers and timestamps before analysis begins.
必须包含：分析开始前确认实体标识和时间戳。
Market reaction
市场反应
Show price, volume, volatility, sector, index, and peer context over a declared interval. Explain whether the move is company-specific, sector-wide, or partly driven by market conditions.
在明确的时间区间内展示价格、成交量、波动率、行业、指数和同行背景，并解释波动是公司特有、行业普遍，还是部分由市场环境推动。
Avoid: describing an 8% move without a benchmark or timestamp.
避免：在没有基准和时间戳的情况下描述 8% 的波动。
Fundamental change
基本面变化
Separate revenue, margins, cash flow, balance-sheet changes, guidance, and management commentary. Distinguish reported values from adjusted values and calculated ratios.
分别呈现营收、利润率、现金流、资产负债表变化、业绩指引和管理层评论，并区分报告值、调整值与自行计算的比率。
Required: source period, units, and calculation formula.
必须包含：来源报告期、单位和计算公式。
Valuation and peers
估值与同行
Compare valuation only after defining a defensible peer set. Explain differences in growth, margins, geography, capital structure, and accounting that make a simple PE comparison incomplete.
只有在定义合理同行范围后才比较估值，并解释增长、利润率、地域、资本结构和会计口径差异为何会让简单的 PE 对比失真。
Good practice: show median, range, and peer inclusion rules.
良好实践：展示中位数、区间和同行纳入规则。
Catalysts, risks, and unknowns
催化、风险与未知项
List dated catalysts, material risks, missing data, and competing explanations. Each claim should identify what observable evidence would strengthen or weaken it.
列出带日期的催化因素、重大风险、缺失数据和相互竞争的解释，并说明哪些可观察证据会增强或削弱每项判断。
Never hide: unavailable transcripts, stale estimates, or unresolved corporate actions.
不得隐藏：缺失的电话会记录、过期预期或尚未确认的公司行动。
Source ledger and confidence
来源台账与置信度
Attach the source URL, provider, publication time, retrieved time, and fields used for each important claim. Give confidence by section rather than assigning one unsupported score to the whole memo.
为每项重要判断附上来源 URL、供应商、发布时间、获取时间和使用字段。应按章节给出置信度，而不是给整份备忘录打一个没有依据的总分。
Review-ready: another analyst can reproduce the key numbers.
可复核标准：另一位分析师能够复现关键数字。
AI Stock Research Assistant vs Stock Picker
AI 股票研究助手 vs 荐股工具
| Dimension维度 | Stock Picker荐股工具 | QVeris-style Research AssistantQVeris 式研究助手 |
|---|---|---|
| Goal目标 | Rank or recommend stocks quickly快速排序或推荐股票 | Organize verifiable data before a human decision在人工决策前组织可验证数据 |
| Data flow数据流程 | Often starts with a ticker or signal通常从股票代码或信号开始 | Starts with industry valuation, market context, screener, fundamentals, and float从行业估值、市场背景、筛选器、基本面和流通股结构开始 |
| Best output最佳输出 | Buy/sell style conclusion买入或卖出式结论 | Candidate pool, research memo, evidence, and follow-up questions候选池、研究备忘录、证据和后续问题 |
| Risk主要风险 | Can over-explain market heat as company quality可能把市场热度过度解释为公司质量 | Separates story, valuation, liquidity, and fundamental quality分别分析叙事、估值、流动性和基本面质量 |
Useful Data Sources for This Workflow
这个工作流适合连接的数据源
A reliable AI investment research assistant should cite and compare multiple sources. For filings and float context, teams often reference SEC EDGAR. For financial metrics and screeners, developer teams may compare vendors such as Financial Modeling Prep, then use QVeris to route calls through a unified capability workflow.
可靠的 AI 投资研究助手应该能引用和比较多个来源。涉及文件和股本结构时，团队通常会参考 SEC EDGAR。涉及财务指标和筛选器时，开发者也会比较 Financial Modeling Prep 等数据源，再通过 QVeris 用统一能力工作流完成调用。
How to Evaluate an AI Stock Research Assistant
如何评估 AI 股票研究助手
A strong AI stock research assistant should make its reasoning path visible. The practical test is not whether it can write a fluent stock summary, but whether it can preserve the user’s question, select appropriate data, expose calculation choices, and produce a memo another analyst can reproduce.
好的 AI 股票研究助手应该让研究路径可见。真正的测试不是它能否写出流畅的股票摘要，而是能否保持用户问题不变、选择合适数据、公开计算口径，并生成另一位分析师可以复现的备忘录。
Evidence coverage证据覆盖
All material claims use primary or clearly identified secondary sources.
所有重大判断均使用一手来源或明确标识的二手来源。
Freshness and timing时效与时间口径
Quotes, filings, estimates, and news carry event time and retrieval time.
行情、文件、预期和新闻都带事件时间与获取时间。
Traceability可追溯性
The reviewer can move from each conclusion back to fields and source URLs.
复核者可以从每项结论追溯到字段和来源 URL。
Calculation integrity计算完整性
Units, periods, denominators, adjusted metrics, and formulas are explicit.
单位、期间、分母、调整指标和公式均明确说明。
Intent preservation意图保持
The agent does not change the ticker, time range, peer set, or task silently.
Agent 不会悄悄改变股票、时间范围、同行集合或任务。
Uncertainty handling不确定性处理
Unknowns, conflicting data, and low-confidence explanations remain visible.
未知项、冲突数据和低置信度解释始终保持可见。
Suggested acceptance threshold:建议验收门槛：
A memo should score at least 85/100, with no zero in evidence coverage, freshness, or traceability. Teams should also maintain a fixed regression set covering ticker ambiguity, stock splits, restated filings, missing transcripts, delayed quotes, foreign currencies, and conflicting provider values.
备忘录建议至少达到 85/100，且证据覆盖、时效和可追溯性不得出现零分。团队还应维护固定回归集，覆盖股票代码歧义、拆股、重述文件、电话会记录缺失、延迟行情、外币口径和供应商数据冲突等情况。
Data Checklist for an AI Stock Research Assistant
AI 股票研究助手的数据检查清单
Before generating a memo, the assistant should confirm that every required data area is present, comparable, and fresh enough for the question. The checklist below is intentionally broader than a ticker quote because most research failures happen when context is missing rather than when prose is weak.
生成备忘录之前，助手应确认每个必要数据区域都已获取、具有可比性，并且时效足以回答当前问题。下面的检查清单刻意超出单一行情范围，因为大多数研究失败源于背景缺失，而不是文字不够流畅。
| Data area数据区域 | Why it matters为什么重要 | Agent behaviorAgent 行为 |
|---|---|---|
| Price and volume价格与成交量 | Shows whether the market reaction is unusual or normal noise.判断市场反应是异常还是普通波动。 | Compare price, range, adjusted history, volume, volatility, sector, and index over the same interval.在同一时间区间比较价格、区间、复权历史、成交量、波动率、行业和指数。 |
| Fundamentals基本面 | Prevents the assistant from ranking companies on headlines alone.避免助手只根据新闻标题排序公司。 | Inspect revenue, margins, cash flow, debt, ROE, growth, and valuation using consistent periods.用一致报告期检查营收、利润率、现金流、债务、ROE、增长和估值。 |
| Filings and events文件与事件 | Explains what changed and gives the memo a primary-source trail.解释发生了什么，并为备忘录建立一手来源链路。 | Cite filings, earnings releases, transcripts, guidance, corporate actions, and event dates.引用监管文件、财报公告、电话会、业绩指引、公司行动和事件日期。 |
| Expectations and guidance市场预期与指引 | A strong quarter can still disappoint if expectations were higher.即使业绩本身不错，只要低于更高预期，市场仍可能失望。 | Record consensus source and timestamp; compare reported, prior guidance, and new guidance separately.记录一致预期来源和时间戳，并分别比较实际值、此前指引和最新指引。 |
| Ownership, float, and liquidity持股、流通股与流动性 | Thin float, lockups, or concentrated ownership can amplify price moves.低流通股、锁定期或持股集中可能放大价格波动。 | Check shares outstanding, float date, institutional ownership, short interest, splits, and issuance.检查总股本、流通股日期、机构持股、空头仓位、拆股和增发情况。 |
Production Risks for Stock Research Agents
股票研究 Agent 的生产风险
Stock research agents can sound confident even when the data path is weak. Production quality depends on detecting the failure mode before it reaches the memo and giving the system a safe behavior for missing or contradictory evidence.
股票研究 Agent 即使数据路径很弱，也可能表现得很自信。生产质量取决于能否在问题进入备忘录之前识别失败模式，并为缺失或冲突证据设置安全行为。
Stale or mismatched time
数据过期或时间错配
A delayed quote may be described as live, a prior-quarter filing may be combined with current news, or a consensus estimate may have been captured after the event.
延迟行情可能被写成实时数据，旧季度文件可能与当前新闻混用，一致预期也可能是在事件发生后才获取。
Safe behavior: compare event time, published time, and retrieved time before synthesis.
安全行为：综合分析前比较事件时间、发布时间和获取时间。
Entity or period mismatch
实体或报告期错配
Tickers can map to multiple exchanges, fiscal years can differ from calendar years, and providers may label trailing and forward metrics differently.
同一股票代码可能对应不同交易所，财年可能与自然年不同，各供应商对历史指标和前瞻指标的标注也可能不同。
Safe behavior: validate identifiers, currency, fiscal period, and metric definition before calculation.
安全行为：计算前验证实体标识、币种、财务期间和指标定义。
False peer comparability
错误的同行可比性
Companies in the same sector may have different revenue models, geographies, leverage, growth stages, or accounting policies. A neat valuation table can still be analytically wrong.
同一行业的公司可能具有不同商业模式、地域、杠杆、增长阶段或会计政策。估值表即使排版整齐，分析也可能是错的。
Safe behavior: record peer selection rules and explain material differences before ranking.
安全行为：记录同行选择规则，并在排序前解释重大差异。
Tool failure hidden by fluent prose
流畅文字掩盖工具失败
A provider can return an empty field, stale cache, rate-limit error, or partial response while the model still produces a polished explanation.
供应商可能返回空字段、过期缓存、限流错误或不完整响应，但模型仍然可以生成看似完整的解释。
Safe behavior: validate required fields and surface tool errors; never treat missing data as zero.
安全行为：验证必需字段并公开工具错误，绝不能把缺失数据当成零。
Why capability routing matters:为什么能力路由很重要：
A QVeris-style agent can discover the needed capability, inspect its input and output contract, call the tool, and validate returned fields before writing. If evidence remains insufficient, it should clarify the request, add a source, or stop—not manufacture a conclusion.
QVeris 式 Agent 可以先发现所需能力，检查输入输出契约，再调用工具并验证返回字段。如果证据仍然不足，就应该澄清请求、增加来源或停止，而不是制造结论。
Frequently Asked Questions
常见问题
These questions cover the practical choices teams face when designing, operating, and reviewing an AI stock research assistant.
这些问题覆盖团队在设计、运行和复核 AI 股票研究助手时最常遇到的实际选择。
Can an AI stock research assistant make buy or sell decisions?AI 股票研究助手可以直接做买卖决策吗？
It can organize evidence, calculate metrics, compare scenarios, and identify unanswered questions, but a production system should not present an unsupported buy or sell instruction as fact. The safer contract is decision support: show the evidence, assumptions, risks, and confidence so an authorized human or downstream policy can decide.
它可以组织证据、计算指标、比较场景并识别未解决问题，但生产系统不应把缺乏依据的买卖指令当成事实。更安全的契约是提供决策支持：展示证据、假设、风险和置信度，再由获得授权的人或下游策略作出决定。
Does every stock research workflow need real-time data?每个股票研究工作流都需要实时数据吗？
No. Long-horizon fundamental research may work with end-of-day prices and the latest verified filings, while intraday event analysis needs fresher quotes and timestamps. The assistant should match data freshness to the user’s decision horizon and label delayed data clearly.
不需要。长期基本面研究通常可以使用日终价格和最新核实文件，而日内事件分析则需要更及时的行情与时间戳。助手应让数据时效与用户决策周期匹配，并明确标记延迟数据。
Which metrics should the assistant use for screening?助手进行筛选时应该使用哪些指标？
The answer depends on the stated research objective. A quality screen may combine ROE, margins, leverage, cash conversion, and earnings stability; a valuation screen may use PE, EV/EBITDA, free-cash-flow yield, and growth. The assistant must inspect the screener schema instead of assuming every provider supports every filter.
这取决于明确的研究目标。质量筛选可以组合 ROE、利润率、杠杆、现金转化和盈利稳定性；估值筛选可以使用 PE、EV/EBITDA、自由现金流收益率和增长指标。助手必须先检查筛选器 schema，而不能假设每个供应商都支持所有过滤条件。
How do you reduce hallucinations in financial research?如何减少金融研究中的幻觉？
Use structured tool outputs, required source fields, freshness checks, deterministic calculations, and a claim-to-source ledger. The agent should distinguish retrieved facts from model interpretation and refuse to invent values when a required field is absent.
应使用结构化工具输出、必填来源字段、时效检查、确定性计算和“判断—来源”台账。Agent 应区分检索事实与模型解释，并在必要字段缺失时拒绝编造数值。
How often should a research memo be refreshed?研究备忘录应该多久更新一次？
Refresh when a material event changes the evidence set: earnings, guidance, filings, corporate actions, major news, estimate revisions, or an abnormal market move. Store the original memo and produce a dated delta so reviewers can see exactly what changed.
当重大事件改变证据集合时就应更新，例如财报、指引、监管文件、公司行动、重大新闻、预期调整或异常市场波动。应保留原始备忘录并生成带日期的差异说明，让复核者清楚看到发生了什么变化。
Does an AI assistant replace a financial analyst?AI 助手会替代金融分析师吗？
Its strongest role is reducing collection and normalization work while making the evidence trail easier to inspect. Human analysts still define the question, judge comparability, challenge assumptions, interpret incomplete information, and own decisions that carry financial or regulatory consequences.
它最适合减少数据收集和口径统一工作，并让证据链更容易复核。人工分析师仍需定义问题、判断可比性、挑战假设、解释不完整信息，并对具有财务或监管后果的决策负责。
Related Guides for Stock Research Agents
股票研究 Agent 相关指南
If your workflow starts with price data, compare free and paid data sources first. If it needs live monitoring, review real-time stock price APIs. If the assistant must explain market movement, add financial news and market data capabilities before asking the model to write commentary. For implementation details, use the QVeris documentation to inspect capability schemas before wiring production calls.
如果你的工作流从价格数据开始，应先比较免费和付费数据源；如果需要实时监控，应重点评估实时股价 API；如果助手要解释市场波动，则应在生成评论前加入金融新闻和市场数据能力。进入实现阶段后，应通过 QVeris 文档检查能力 schema，再接入生产调用。