01
概览Overview
读财报、搜新闻、写研报，已经成了大模型的拿手活。可一旦把时间拨回财报发布前，要求它预测下一季收入、判断公司的资金压力，再给出一份能经得起事后验证的投资意见，事情还会这么顺利吗？Reading financial reports, searching the news and writing research notes have become familiar strengths of large language models. But move the clock back to before an earnings release, ask a model to forecast next quarter’s revenue, assess a company’s funding pressure and deliver an investment view that can withstand later scrutiny. Will the work still go as smoothly?
近期，Prophetlab团队推出 Investor’s Last Exam（ILE），由资深买方投资人和基金经理主导构建，给金融Agent准备了一场“投资人的最后一场考试”。团队把海量金融资料、专业研究任务和结果验证放进同一个环境，让AI沿着真实分析师的工作流程，完成从搜集证据到形成投资判断的整套研究。Prophetlab introduces Investor’s Last Exam (ILE), a benchmark led by experienced buy-side investors and fund managers: a final exam for financial agents. It brings a large financial database, professional research assignments and outcome verification into one environment, asking AI to follow the analyst’s workflow from gathering evidence to forming an investment judgment.
第一，让AI回到决策当天。 ILE建立了包含343万个文档版本、总量1.52 TB的金融数据库，覆盖九类信息，并配套结构化市场数据。每项任务都有明确的时间截止点，检索、阅读和计算共同遵守这条时间边界，Agent只能拿到当时已经可获得的资料。First, take AI back to the decision date. ILE provides a financial database with 3.43 million document versions and 1.52 TB of material across nine source categories, alongside structured market data. Every task has a fixed information cutoff. Retrieval, reading and computation share that boundary, giving the agent access to the information available at the time.
第二，让买方专家来定义什么叫做好投研。 86项任务覆盖机会筛选、证据研究、预测建模、估值决策和持续跟踪，并延伸到综合研究与市场交易。专家把关键判断拆到具体的数据、假设和计算关系上，模型需要说明自己的结论究竟是怎样推出来的。Second, let buy-side experts define good research. The 86 assignments span opportunity screening, evidence gathering, forecasting, valuation and continuing coverage, extending to integrated research and market trading. Experts connect each key judgment to specific data, assumptions and calculations. Models must explain how they arrived at their conclusions.
第三，让研究结论接受后续结果检验。 ILE同时考察研究质量、预测准确性和模拟投资表现。分析写得是否有依据、预测能否兑现、转化为投资信号后表现如何，都有各自的检验方式。Third, put research conclusions to the test of subsequent events. ILE evaluates research quality, forecast accuracy and simulated investment performance. The evidence behind an analysis, the accuracy of its predictions and the payoff of the resulting signals each face their own test.
11个前沿模型与系统交出的答卷颇有看点：GPT-6 Astra和Claude Opus 5.5以68.1分、66.9分领先，但Astra的预测建模得分只有52.5。再把视线转向预测结果和模拟收益，模型的排名又发生了变化。从读懂资料到判断未来，金融Agent还有一道硬关要过。The eleven evaluated systems deliver some revealing results. GPT-6 Astra and Claude Opus 5.5 lead overall with 68.1 and 66.9 points, yet Astra scores just 52.5 on Forecast & Model. Switch to forecast outcomes or simulated returns, and the ranking changes again. There is another demanding step between understanding the evidence and anticipating the future.
02
怎样让AI回到真实的历史投研现场？How can AI return to a real historical research setting?
做历史预测，最容易被忽略的问题是：AI到底看到了哪个时点的资料？一篇网页可能保留着早期发布日期，正文却已经更新；一条回顾性报道，也可能直接写出当时尚未公布的结果。In historical forecasting, an easy question to overlook is: which moment does the information actually belong to? A web page may retain an early publication date even after its content changes. A retrospective article may directly reveal an outcome that had not yet been announced.
论文中的特斯拉案例就出现了这种情况。模型在搜索词里明确加上了2026年3月24日之前的日期限制，返回结果中却出现了7月2日公布的第二季度实际交付量。在GPT-6 Astra的公开网页搜索对照实验中，86条研究轨迹有49条出现了截止点之后的来源，占57%。搜索框里写了日期，拿到的资料依然可能“穿越”。The Tesla case in the paper shows this problem in practice. Although the model explicitly searched for information before March 24, 2026, a result contained actual second-quarter deliveries announced on July 2. In the GPT-6 Astra public-web comparison, 49 of 86 research trajectories contained post-cutoff sources: 57%. A date in the search query still let future information through.
为此，ILE从资料的历史版本入手，为每个版本记录可获得时间，保留修订前后的内容。到2026年10月15日，数据库已收录3,430,336个文档版本，覆盖财报、电话会、公司公告、分析师预期、研报和新闻等九类信息，涉及美股、A股、港股等市场。后续数据继续补充，已有历史版本持续保留。ILE starts with historical source versions. Each version records when it became available, and earlier content is retained after revisions. By October 15, 2026, the database contained 3,430,336 document versions across nine categories, including financial reports, earnings calls, issuer events, analyst estimates, research reports and news, covering US, mainland Chinese, Hong Kong and other markets. New data continues to arrive while historical versions remain available.
这条时间边界贯穿整个研究过程。Agent可以搜索原文、阅读表格、下载文件、查询结构化数据，也可以编写代码完成计算和建模；它用到的资料和查询结果，都来自任务时点允许访问的数据。这样，面对需要跨多份财报、反复计算才能完成的研究，模型有足够的工具，也必须遵守当时的信息条件。That time boundary follows the entire research process. Agents can search original sources, read tables, download files, query structured data and write code for calculations and modeling. The material and query results all come from the information available at the task’s historical cutoff. Agents have the tools to work across multiple reports and repeated calculations within the information conditions of the time.
03
买方专家怎样出题？How do buy-side experts design the tasks?
ILE的任务来自买方分析师和基金经理的研究实践。什么变化值得关注？公司的增长假设是否站得住？新的披露会让原有判断改变多少？这些问题被整理为五类、二十个细分方向，再加入综合研究和市场交易任务，组成86项专业作业。ILE tasks come from the research practice of buy-side analysts and fund managers. Which changes matter? Do the company’s growth assumptions hold up? How much should a new disclosure change the prior view? These questions form five categories and twenty subtypes, with integrated research and market trading bringing the suite to 86 professional assignments.
出题时，专家先确定研究对象、时间范围和需要作出的判断，再搭建参考分析。每个关键事实保留来源、期间、单位和文档位置，事实之间的会计关系、收入拆解和同业影响也要讲清楚。团队将这些关系组织成金融依赖图，让评分能够追到支撑结论的具体证据、假设和计算。AI协助寻找遗漏资料、检验假设、提出反例，多模型参与复核，最终由专家确定题目和评分要求。Experts first fix the research subject, time horizon and judgment to be made, then build a reference analysis. Every material fact retains its source, period, unit and location. Accounting relationships, revenue decompositions and peer effects must also be explicit. The team organizes these links into financial dependency graphs, so each score can be traced to the evidence, assumptions and calculations supporting the conclusion. AI helps find missing evidence, test assumptions and propose counterexamples; multiple models review drafts, and experts make the final decisions on tasks and scoring requirements.
以债务再融资为例，把现金、未来现金流和未用授信加起来，看上去可能足够偿债。但贷款契约可能限制提款，账面上尚未使用的额度，未必是当前状态下真正能动用的钱。专家会把这条决定结论的关系写进任务和评分标准。模型能否识别这样的约束，直接影响它的风险判断。Consider debt refinancing. Adding cash, future cash flow and unused credit may appear sufficient to repay the debt. But a borrowing covenant can restrict access to the credit line: an unused limit on paper may not be drawable in the company’s current state. Experts put that decisive relationship into the task and rubric. Recognizing the constraint changes the risk judgment.
04
模型拿到一道题后，会怎样做研究？How does an agent work through a research assignment?
模型拿到一道题后，研究要怎样一步步往下做？AEP这道题要求它回到2026年第二季度报告，将两期现金流放到同一张表里，再把资金来源、资本开支和融资需求连起来。检索负责找资料，阅读负责定位与核对，终端操作处理文件和计算，最后提交完整的研究结论。What happens after an agent receives a task? The AEP assignment takes it into the second-quarter 2026 report to place two periods of cash flows on the same footing, then connect funding sources, capital expenditure and financing needs. Search finds the material, reading locates and checks the evidence, terminal operations handle files and calculations, and the final submission delivers the research conclusion.
先找什么？What comes first?
找到AEP报告，定位两期经营、投资与融资现金流。Find AEP’s report and locate operating, investing and financing cash flows for both periods.
怎样核对？How is it checked?
核对期间与单位，将投资现金支出减去经营现金流入。Align periods and units, then subtract operating inflow from investing cash use.
怎样形成判断？How does judgment follow?
追溯现金变化的来源，把经营改善与资本需求放到一起。Trace the sources of cash movements and connect operating improvement to capital needs.
Qwen 3.8 Max在这道题中进行了6次检索、52次阅读和10次终端操作。它先按公司名找到报告，再下载原文、查找现金流表，反复核对经营、投资与融资项目，最后提交报告。多轮取证之后，真正拉开差距的，是怎样把这些数字接成一个金融判断。On this task, Qwen 3.8 Max makes six searches, 52 reads and ten terminal calls. It finds the report by company name, downloads the original text, locates the cash-flow statements and repeatedly checks operating, investing and financing items before submitting its report. After those rounds of evidence gathering, the decisive step is turning the figures into a financial judgment.
同一份披露、同一组数字，经营现金流改善是一层结论，融资缺口是否缩小、改善能否持续则是下一层。模型需要把这两层接起来，才能回答公司的资金压力究竟发生了什么变化。An increase in operating cash flow is one finding. Whether the funding gap narrows, and whether that improvement persists, are further questions. Connecting them establishes how the company’s funding pressure has changed.
经营现金流增加，融资压力就小了吗？Does stronger operating cash flow reduce funding pressure?
美国电力公司AEP的一道任务，把这种难度展现得很直观。2026年上半年，公司的经营现金流同比增加7.5亿美元。看起来，手头更宽裕了。但同期投资现金支出增加得更多，融资前的资金缺口从26.76亿美元扩大到31.81亿美元，反而增加了5.05亿美元（图2右）。An American Electric Power task makes this difficulty concrete. In the first half of 2026, operating cash flow rose by $750 million. The company appears to have more cash at hand. Yet investing cash use grew by more, widening the pre-financing shortfall from $2.676 billion to $3.181 billion—an increase of $505 million (Figure 2, right).
进一步追问，经营现金流的改善又有多少能持续？净利润下降、非现金调整、客户押金和退款时点，都影响着这个答案。把数字抄全、把加减法算对之后，仍然需要判断这笔现金究竟从哪里来，以及它能否覆盖公司的资本需求。Then comes the next question: how much of the cash-flow improvement can persist? Lower net income, non-cash adjustments, customer deposits and the timing of refunds all affect the answer. After recovering the figures and checking the arithmetic, the analyst still has to establish where the cash came from and whether it covers the company’s capital needs.
结果很有代表性：11个系统全部找到了八项披露数据，也都对上了两条现金流恒等式；只有5个算出了两期融资缺口，5个给出了有依据的综合融资判断，3个分清了燃料成本回收中的净额调整与总额回收。 ILE把这些差别落实到评分里，研究者可以看到模型具体卡在了哪一步。The results are revealing. All eleven systems recover the eight reported inputs and reconcile both cash-flow identities. Only five calculate both funding shortfalls, five reach a supported overall funding judgment, and three distinguish the net fuel-cost adjustment from gross fuel collections. ILE captures these differences in its scores, showing where each model’s research gets stuck.
05
11个前沿模型交出了什么答卷？How do eleven frontier systems perform?
在86项任务上，GPT-6 Astra以68.1分位居总榜首位，Claude Opus 5.5以66.9分紧随其后，DeepSeek V4.1 Flash得到63.0分。细看各类任务，模型已经表现出鲜明的能力差异。Across the 86 assignments, GPT-6 Astra leads with 68.1 points, followed by Claude Opus 5.5 at 66.9 and DeepSeek V4.1 Flash at 63.0. The task categories reveal distinct strengths across models.
研究做得不错，为什么预测仍然吃力？Why does strong research still struggle with forecasts?
Astra在研究与证据类任务上得到78.9分，在评估与决策类任务上得到77.6分，到了预测与建模类任务，得分降到52.5。Opus的研究与证据得分为75.7，预测与建模得分为52.6。资料整理得好，推导写得顺，离预测得准仍有相当一段距离。Astra scores 78.9 on Research & Evidence and 77.6 on Evaluate & Decide, but falls to 52.5 on Forecast & Model. Opus scores 75.7 on Research & Evidence and 52.6 on Forecast & Model. Organizing evidence and producing a fluent chain of reasoning leave a substantial distance to accurate prediction.
把总分再拆开，优势落在不同的地方。Qwen 3.8 Max的证据得分达到81.1，居于首位；Opus的数值计算得到90.1。Astra则在金融逻辑、勾稽核验、决策一致性和不确定性上领先。找到依据、算清关系、判断未来，是几道相互衔接的关。Breaking down the overall score reveals different leaders. Qwen 3.8 Max leads Evidence at 81.1, while Opus earns 90.1 on Calculation. Astra leads Financial logic, Reconciliation, Decision consistency and Uncertainty. Finding support, checking the relationships and judging the future are connected challenges.
换一把尺子，排名会怎样？How does a different yardstick change the ranking?
研究总分：86题 · 预测结果：38道含结果验证的任务 · 满分100。Overall research: 86 tasks · Forecast outcomes: 38 tasks with outcome checks · Scores out of 100.
特斯拉的一项预测任务中，3个系统拿到了全部支撑分析分，交付量、收入和利润率的预测却都没拿到结果分。需求、价格和成本假设可以组成一套看上去很完整的模型，最终还是要接受实际披露的检验。In one Tesla forecast task, three systems earn all the supporting-analysis points but none of the outcome points for deliveries, revenue or margin. Demand, price and cost assumptions can produce a complete-looking model; the final test still comes from the actual disclosure.
特斯拉O80要求模型在2026年6月22日预测第二季度交付量、GAAP收入和汽车业务毛利率。Astra、Qwen Max和Opus都低估了交付量与收入，又高估了毛利率。三个模型、三个指标，九个误差全部超过了各自的零分阈值。支撑分析拿到50分，预测结果却是0分，这个反差把“算得通”和“预测准”之间的距离摆到了台面上。Tesla O80 asks for second-quarter deliveries, GAAP revenue and automotive gross margin as of June 22, 2026. Astra, Qwen Max and Opus all underestimate deliveries and revenue while overestimating margin. Three models, three metrics: all nine errors exceed their respective zero-credit limits. Fifty points for supporting analysis and zero for forecast outcomes put the gap between a coherent calculation and an accurate prediction in plain sight.
预测榜和收益榜，谁又走到了前面？Who moves ahead on forecast outcomes and returns?
如果单独考察预测与实际结果的吻合程度，DeepSeek V4.1 Flash以40.3分领先，高于Astra的37.7分和Opus的37.6分。DeepSeek V4 Pro在市场类任务中得到64.6分，同样展现出总分之外的优势。When the measure is how well predictions match realized outcomes, DeepSeek V4.1 Flash leads with 40.3 points, ahead of Astra at 37.7 and Opus at 37.6. DeepSeek V4 Pro scores 64.6 on Market Trading, showing another strength beyond its overall rank.
团队还把预测转成投资信号：将模型对基本面的预测与当时分析师预期比较，形成多空方向，在财报披露后结算。2026年3月至8月，覆盖11家公司、20项任务的模拟投资实验中，计入交易与借券成本后，Hunyuan 4 Preview以8.50%的收益领先，DeepSeek V4.1 Flash和Astra分别为7.98%、7.18%，同期始终做多基准为1.18%。Kimi K3虽然研究总分低于Opus，模拟收益却达到7.15%，超过Opus的6.39%。The team also turns forecasts into investment signals. A model’s fundamental forecast is compared with the analyst estimate available at the time to choose a long or short position, which is settled after disclosure. Across eleven companies and twenty assignments from March to August 2026, Hunyuan 4 Preview leads with an 8.50% simulated return after trading and borrowing costs. DeepSeek V4.1 Flash and Astra follow at 7.98% and 7.18%, versus 1.18% for the always-long baseline. Kimi K3 earns 7.15%, above Opus’s 6.39%, despite its lower overall research score.
| 模型 / 基准Model / baseline | 模拟收益Simulated return | 超额收益Excess return | 最大回撤Max. drawdown |
|---|---|---|---|
| Hunyuan 4 Preview | +8.50% | +7.32个百分点 pp | 2.72% |
| DeepSeek V4.1 Flash | +7.98% | +6.80个百分点 pp | 3.65% |
| GPT-6 Astra | +7.18% | +6.00个百分点 pp | 1.50% |
| Kimi K3 | +7.15% | +5.97个百分点 pp | 4.02% |
| Claude Opus 5.5 | +6.39% | +5.21个百分点 pp | 2.94% |
| GLM 5.3 | +6.19% | +5.01个百分点 pp | 3.53% |
| Qwen 3.8 27B | +5.20% | +4.02个百分点 pp | 1.90% |
| DeepSeek V4 Pro | +3.64% | +2.46个百分点 pp | 2.03% |
| Ling 3.0 Flash | +2.51% | +1.33个百分点 pp | 2.64% |
| Qwen 3.8 Max | +1.67% | +0.48个百分点 pp | 2.05% |
| 始终做多Always long | +1.18% | +0.00个百分点 pp | 7.63% |
| 始终做空Always short | -2.27% | -3.45个百分点 pp | 8.86% |
2026年3—8月 · 20项机会 · 11家公司 · 收益已计入交易与借券成本。March–August 2026 · 20 opportunities · 11 companies · Returns include trading and borrowing costs.
这些结果让金融Agent的评价更有意思了。同一个模型，可能擅长组织证据，也可能更善于捕捉预期差。ILE把研究质量、预测结果和投资表现放在一起，方便研究者检验自己的改进究竟作用在了哪里。This makes financial-agent evaluation more interesting. A model may excel at organizing evidence, or at identifying a gap against market expectations. ILE puts research quality, forecast outcomes and investment performance alongside one another, helping researchers see where their improvements take effect.
06
研究花得更多，效果就更好吗？Does spending more on research produce better results?
长程投研也很考验Agent如何安排工作。DeepSeek V4.1 Flash每项任务平均进行73.8轮交互、113.8次工具调用，在检索和计算上投入最多；总分领先的Astra，调用次数反而更少。Opus与Qwen 3.8 Max的阅读和计算调用量相近，总分仍相差4.8分。Long research workflows also test how agents organize their work. DeepSeek V4.1 Flash averages 73.8 interaction rounds and 113.8 tool calls per task, with the most retrieval and computation. The overall leader, Astra, makes fewer calls. Opus and Qwen 3.8 Max have similar reading and computation counts, yet their overall scores differ by 4.8 points.
成本差距也相当醒目。DeepSeek V4.1 Flash平均每项任务的模型API成本为0.36美元，得到63.0分；Opus的成本为7.96美元，约为前者的22倍，得分高出3.9分。怎样用好每一次检索、每一轮计算，让多步研究真正形成有效判断，是金融Agent值得继续打磨的地方。Costs differ sharply too. DeepSeek V4.1 Flash scores 63.0 at an average model API cost of $0.36 per task. Opus costs $7.96—roughly twenty-two times as much—for another 3.9 points. Using each search and each calculation to build a useful judgment remains an important challenge for financial agents.
07
怎样带你的Agent来测？How can you bring your agent to ILE?
如果你在训练金融模型、搭建投研Agent，或研究长程任务中的检索、规划与推理，欢迎来挑战ILE。可以从一份经营预测开始，也可以尝试完整的公司研究：在同样的历史信息条件下，看看你的Agent能把证据用到什么程度，能否发现关键约束，又能形成怎样的预测和投资判断。If you train financial models, build investment-research agents, or study retrieval, planning and reasoning over long tasks, come test your agent on ILE. Start with an operating forecast or try a complete company research assignment. Under the same historical information conditions, how well can your agent use evidence, identify decisive constraints and form forecasts and investment judgments?
你的Agent，能通过投资人的最后一场考试吗？Can your agent pass the investor’s last exam?
08
谁在共同构建这场考试？Who is building this exam together?
1 Prophetlab · 2 RethinkLab · 3 上海交通大学安泰经济与管理学院 · 4 中国科学技术大学 · 5 睿远基金 · 6 常春藤资产管理1 Prophetlab · 2 RethinkLab · 3 Antai College of Economics and Management, SJTU · 4 USTC · 5 Foresight Fund · 6 IvyRock Asset Management
* 共同一作 · † 通讯作者* Equal contribution · † Corresponding author
研究与合作：Research and collaboration: henryjiang@prophet-lab.com