中文精读
判断怎样才有资格变成行动：吴浩哲谈AI Trading的可审计通路
42章经单期精读 · 金融研究与Agent责任
嘉宾：吴浩哲，Vibe-Trading作者及维护者，香港大学在读博士、HKUDS成员｜主持：曲凯
发布：2026年9月27日｜时长：40分46秒｜整理：2026年10月18日
官方单集与节目介绍
三万多个Star，证明不了交易收益
官方介绍称Vibe-Trading开源半年有3.4万Star；嘉宾回顾项目从专业量化工具转向多视角研究工作台。Star反映关注，不是完成率、实盘业绩或收益验收。
“决策便宜，行动很贵”里的决策，主要指市场中大量判断与信号，不是说审慎拍板没有价值。吴浩哲想在材料、假设、验证、有限行动与复盘之间建立通路。本文关注这种产品机制，保留当期观点；回测、真实交易和获利判断分别讨论，不从访谈生成新的交易策略。
1. 专业指标不等于用户理解了自己在承担什么
线索：01:33–03:18、05:14–07:31。 项目最初有四百多个因子、九个市场回测引擎及统计检验。普通用户若看不懂因子、IC或夏普率，就难以解释为什么相信结果。功能丰富，未必形成可检查的判断。
后来工作台更强调多种材料和解释。模型负责读资料、提假设、写代码；Harness即运行与约束框架，决定哪些结果可进入下一步；人决定允许范围、规模与例外责任。
这套分工不是把人放在最后随手点确认。新判断不断出现，检查也要持续，不能只在最初写一次规则。编辑推论：界面应让用户看到出处、假设、未解决项和动作状态，而不是把一段自信结论紧贴执行按钮。理解和授权各有要求，不能互相代替。
2. 事实、解释和行动，中间至少隔着一次验证
线索：08:29–09:56、12:06–15:41。 现金流改善是具体例子：可能经营变好，也可能只是回收应收或延迟付款。研究需要原始数字、出处、时间、同行变化及相反证据；证据不足应停在不能判断，不用含混建议营造完成感。
专用工作台的理由，是计算可重跑、过程可审计，确定数字由工具算，模型负责组织材料与假设。但“全部行业know-how已在模型里”是嘉宾过强概括，本期没有证明，领域规则与错误仍需检查。
他还讲自己研究存储行业的经历，区分内存、闪存和机械硬盘，再决定真实买入。80%至90%工作由AI完成是自述比例，没有完整样本、成本和收益分布。它只能说明一次研究分解，不能验证工具胜率。
编辑推论：研究得出合理解释、历史回测成立、真实执行完成、长期结果较好，是四个不同层次。每层都要补证据，不能把一个成功故事当成后面所有步骤的保证。
3. 数据对齐不是清洗干净，而是知道当时能知道什么
线索：19:56–25:35。 吴浩哲把系统拆为模型、独家数据、数据对齐、Harness及人的责任，强调后两者守住判断进入行动的时刻。收益是需观察的结果，不会因为系统完备就自然保证。
对齐有三种具体失败：第一是口径，价格是否复权、金额单位、M代表月还是分钟；第二是时点，回测不能拿后来修订的数字当作当时已公开信息；第三是缺失，数据源失败时应暂停并说明，不能静默用另一种口径替代。
嘉宾举过降级后金额差六七个数量级的bug，这是项目经验，不是已复现测试。它说明看起来流程继续运行，反而可能最危险，因为后续推导都建立在错误输入上。
编辑推论：更多数据只是让命题可检验，不自动带来超额收益。数据错与假设错要分别记录，不能让“数据还不够”永远保护原假设；也不能用缺失值冒充零，或把当前版本偷放进历史评估。
4. 更便宜的概率，不一定补上行动所缺的理由
线索：27:23–31:53。 吴浩哲对Jev的评价是，它降低了已经很便宜的答案成本，却没有提供完整理由。他认为数字、日期和噪声材料仍是问题，适合一些大批初筛，不足以直接承担核心链。这是当期个人判断，本篇未读其文档或独立比较。
初筛也不是无代价：漏掉关键材料会影响研究，不能照搬“少看几条无所谓”。需要看误报、漏报及下游影响，而不只比较速度与调用费。
概率还有相关性问题。他举两条各90%判断，提醒直接相乘为81%暗含独立性。若都引用同一条错误新闻，整条推理会一起失败。这个例子说明模型自信与校准不同，不能变成仓位或下注公式。
编辑推论：把来源相关、事实不明和模型不确定分别展示。多个Agent引用同一材料，并不构成多份独立证据；语言越确定，也不证明概率越可信。
5. 工具让信息更快，也可能让错误更快
线索：09:59–11:24、16:13–19:42。 嘉宾谈社媒情绪时提醒，方法公开后，别人可能专门投放误导材料。论坛反向指标与交易零和的概括都是其观点，本篇不据此提炼获利方法；可迁移的是输入会受到策略性污染。
他预测信息更快进入价格，也讨论假新闻与多Agent同向行动。财报后漂移研究的年份、适用范围与“消失”时间是模糊回忆，合谋研究也未给出处，不能写成现实市场已有普遍规律。
编辑推论：自动化缩短读材料到执行的链，也缩短纠错机会。系统要留下原始来源、版本、确定计算和停止原因；速度改善应与错误传播一起衡量。只让Agent互相同意，无法完成反证。
6. 金融工作成果与跑赢市场，是不同产品目标
线索：32:34–38:57。 吴浩哲区分AI Trading的收益问题，与AI for Finance的估值复核、材料准备、对账等工作。后者更直接的交付是同事能接手、复算、修改，第二天还能继续用，而非一次炫目输出。
模型更强未必解决数据接入和团队流程。不同技能的人得到何种增强，也是其经验分层，不是已测能力分数。人参与还涉及决定、承诺和责任的意义感，这是一种价值判断，不能证明机器投资已超人。
结尾嘉宾明确，工具不是致富密码，风险与责任不能消失。本篇保留这个边界，不复述具体开户、仓位或市场尝试操作。
编辑推论：人在环路同时包含授权责任与主体感，两者不宜混淆。即使用户愿意承担，也不能替系统免除记录、核验和解释；即使AI能做，也应明确哪些决定仍由负责人作出。
用一份研究报告检查交付，不触发真实动作
下面是编辑练习。选一个不连接交易执行的历史研究样例，逐栏检查：
| 层次 | 留下什么 |
|---|---|
| 材料 | 原始出处、公开时间、版本、单位和缺失项 |
| 解释 | 哪些是事实，哪些是模型假设，反证在哪里 |
| 计算 | 可重跑的确定工具、输入和输出 |
| 验收 | 同事是否能复算、修改并接手，未解决什么 |
| 动作边界 | 本次只完成研究，下一步需要谁确认及什么证据 |
这个表检验研究可交接性，不检验盈利，更不是交易策略。将建议、执行和最终回读分成状态，也适用于一般生产Agent：工具调用成功与用户目标完成，始终需要分别证明。
回听导航与关联阅读
- 01:26：从量化工具到可理解工作台。
- 04:35：可验证通路与角色分工。
- 13:43：一次研究案例的证据边界。
- 19:57：数据对齐与行动约束。
- 27:23：便宜初筛与核心判断。
- 30:26：概率与相关性。
- 32:34：稳定工作成果与收益目标。
相关主题：可审计研究、数据时点、Agent责任。可对照RC的协作上下文和Jenny的企业复核成本，检查建议怎样进入可交接成果。
资料边界：完整阅读所供转写、旧精读与官方介绍，未逐句复听、运行项目、回测或核交易记录。Star与收益分开，个人案例不验证胜率；研究、真实执行与收益各自独立。Jev局限、论文与市场预测保留当期归因，片尾歌词不作论据。
English edition
Which Judgments May Become Actions? Wu Haozhe on an Auditable AI-Trading Process
42 Zhangjing episode analysis · Financial research and agent responsibility
Guest: 吴浩哲, author and maintainer of Vibe-Trading, doctoral student at the University of Hong Kong and HKUDS member | Host: Qu Kai
Published: September 27, 2026 | Duration: 40 minutes 46 seconds | Prepared: October 18, 2026
Official episode and introduction
Thirty thousand stars cannot establish trading returns
The introduction reports 34,000 stars within half a year of Vibe-Trading becoming open source. Wu recalls moving from specialist quantitative tools toward a multi-perspective research workstation. Stars measure attention rather than completion, live performance or verified returns.
“Decisions are cheap; actions are expensive” refers mainly to abundant judgments and signals, not the worthlessness of considered decisions. Wu seeks a route from materials through hypotheses and validation to bounded action and review. This article examines that product mechanism. Backtests, actual trading and profitability remain separate, and no new trading strategy is derived from the interview.
1. Specialist metrics do not establish that users understand responsibility
Episode reference: 01:33–03:18 and 05:14–07:31. The original project contained over four hundred factors, backtesting engines for nine markets and statistical tests. A user unable to interpret factors, IC or Sharpe ratios cannot easily explain trust in a result. Extensive features do not automatically create an inspectable judgment.
The workstation increasingly emphasizes materials and explanations. Models read, propose hypotheses and write code. A harness—the execution and constraint framework—controls which results proceed. People define permitted scope, scale and exceptions.
This does not place a person at the end for a casual confirmation. Continuing new judgments need continuing checks beyond one initial rule. Editorial interpretation: Interfaces should expose provenance, assumptions, unresolved issues and action states rather than attach a confident paragraph directly to execution. Understanding and authorization have separate requirements.
2. Facts, explanations and actions require intervening validation
Episode reference: 08:29–09:56 and 12:06–15:41. Improved cash flow may reflect better operations, collected receivables or delayed payments. Research needs original figures, sources, dates, peer changes and contradictory evidence. Insufficient evidence should end in an inability to judge, not an ambiguous recommendation simulating completion.
A dedicated workstation offers reproducible calculations and auditable processes. Deterministic tools calculate figures while models organize material and hypotheses. Wu’s claim that models already contain all industry know-how is stronger than the episode establishes. Domain rules and errors still require checks.
He recounts investigating storage, separating memory, flash and hard drives before making a real purchase. His estimate that AI did eighty to ninety percent of the work lacks a complete sample, costs and return distribution. It illustrates research decomposition without validating the tool’s win rate.
Editorial interpretation: A reasonable explanation, a successful historical backtest, completed live execution and good long-term results are four claims. Each requires evidence. One success story cannot guarantee subsequent stages.
3. Data alignment includes knowing what was knowable then
Episode reference: 19:56–25:35. Wu discusses models, proprietary data, data alignment, the harness and human responsibility, emphasizing the last checks before action. Returns are outcomes to observe, not something a complete system automatically guarantees.
Three failures make alignment concrete. Definitions include adjusted prices, currency units and whether M means month or minute. Timing means a backtest cannot use later revisions as though they were public then. Missing data should cause an explicit pause rather than silent substitution with a different definition.
A fallback allegedly created amounts six or seven orders of magnitude wrong. This is project experience rather than a replicated test. Continuing apparently normal operation can conceal bad inputs underlying every subsequent inference.
Editorial interpretation: More data may make a proposition testable without producing excess return. Record data errors separately from failed hypotheses so insufficient data cannot indefinitely protect an idea. Missing values are not zero, and current versions do not belong silently inside historical evaluation.
4. Cheaper probabilities may not supply the reasons action requires
Episode reference: 27:23–31:53. Wu argues that Jev reduces the cost of an already cheap answer without supplying a full explanation. He regards numbers, dates and noisy material as remaining problems, permitting some bulk screening rather than direct ownership of the core chain. This is his contemporary assessment; documentation and comparisons were not independently checked here.
Screening still has consequences. Missing a critical document can alter research. Evaluate false positives, false negatives and downstream impact rather than accepting that overlooking some material does not matter or comparing speed and price alone.
Probabilities also involve dependence. Multiplying two ninety-percent judgments into eighty-one percent silently assumes independence. If both rely on one false story, the chain can fail together. This distinguishes confidence from calibration rather than providing a position-size or betting formula.
Editorial interpretation: Display shared sources, uncertain facts and model uncertainty separately. Several agents citing the same material do not constitute independent evidence. More certain language does not establish a better probability estimate.
5. Faster information can mean faster error
Episode reference: 09:59–11:24 and 16:13–19:42. Discussing social sentiment, Wu notes that public methods can invite deliberately misleading inputs. Claims about inverse forum signals and zero-sum trading remain his views rather than profitable methods extracted here. The transferable point is strategic contamination of inputs.
He predicts faster incorporation of information into prices and discusses false news and agents moving together. Dates, scope and the disappearance of post-earnings drift are recalled imprecisely, while the collusion research has no supplied reference. They do not establish universal behavior in real markets.
Editorial interpretation: Automation shortens the path from reading to execution and also the opportunity to correct mistakes. Preserve sources, versions, calculations and reasons to stop. Evaluate error propagation alongside speed. Agreement among agents does not complete disconfirmation.
6. Financial work and beating markets are different product objectives
Episode reference: 32:34–38:57. Wu distinguishes trading returns from AI-for-finance tasks such as valuation checks, materials and reconciliation. The latter deliverables need colleagues to inherit, reproduce, revise and continue them tomorrow, beyond one impressive output.
Stronger models may not fix data connections or team workflows. His descriptions of improvements for differently skilled users are experiential categories rather than measured scores. Human participation also involves meaning through decisions, commitments and responsibility. That value position does not prove machines have surpassed people in investment.
Wu explicitly rejects the tool as a formula for wealth. Risk and responsibility remain. This article preserves that boundary without repeating account, position or market-experiment instructions.
Editorial interpretation: Human involvement includes authorization and a sense of agency, which should not be confused. Willingness to accept risk cannot remove recording, checking and explanation. Capability also does not decide which choices a responsible person should retain.
Inspect a research report without triggering real action
For an editorial exercise, choose a historical research example disconnected from trading execution:
| Layer | Retain |
|---|---|
| Materials | Original sources, publication time, versions, units and missing items |
| Explanation | Facts, model hypotheses and contradictory evidence |
| Calculation | Reproducible deterministic tools, inputs and outputs |
| Acceptance | Whether colleagues can reproduce, revise and inherit it, plus unresolved issues |
| Action boundary | Research completion only; who must confirm any next step and on what evidence |
This tests handover quality, not profit, and supplies no trading strategy. Separating recommendation, execution and final readback also applies to ordinary operational agents. A successful tool call and a completed user goal require different evidence.
Listening guide and related reading
- 01:26: From quantitative tools to understandable workstations.
- 04:35: Verifiable processes and roles.
- 13:43: Evidence limits of one research case.
- 19:57: Data alignment and action constraints.
- 27:23: Inexpensive screening versus core judgment.
- 30:26: Probability and dependence.
- 32:34: Stable work deliverables versus returns.
Related themes: auditable research, point-in-time data and responsibility. Compare RC’s collaboration context and Jenny’s enterprise review costs, examining how recommendations become transferable outputs.
Source boundary: The complete supplied transcript, earlier note and official introduction were read without sentence-by-sentence audio, running the project, backtesting or checking trading records. Stars and returns differ; a personal case establishes no win rate. Research, execution and profitability require independent evidence. Jev limitations, papers and forecasts retain contemporary attribution; closing lyrics are not evidence.