Overview

TradeAgents is a research desk made of language-model agents. You give it one ticker and one date. It seats a team of analysts to study that company as of that date, makes a bull and a bear argue over what they found, has a trader turn the verdict into a plan and a risk team attack the plan, and ends with a portfolio manager signing a rating.

The point is not to replace your judgment. It is to put a whole desk's worth of argument in front of you — every turn in the open, with the evidence each role worked from — so you can see where a conclusion came from, and where it is weakest.

TradeAgents 是一个由大语言模型智能体组成的投研台。你给它一只股票和一个日期，它会组建一支分析师团队，研究这家公司在那一天的状况；让看涨与看跌研究员就研究结果展开辩论；由交易员把结论变成交易计划，再由风险团队对计划发起质询；最后由投资组合经理签发评级。

它的目的不是替你做判断，而是把一整个投研团队的论证摆在你面前——每一轮发言都公开，并附上每个角色所依据的证据——让你看清结论从何而来，又在哪里最站不住脚。

Loading the cached snapshot…

正在读取缓存快照…

The broad US model combines trend (40%), breadth (35%) and volatility (25%). Trend uses price against MA50 and MA200, with RSI adjustments; breadth uses the share of constituents above MA50 and MA200; volatility uses VIX and its three-month term ratio.

Score = Σ(signal score × available weight) / Σ(available weight). Missing inputs are excluded and coverage is shown; an unavailable signal is never a zero. A high score with partial coverage is less complete evidence than a fully covered score. The score is an environment indicator out of 100, not the probability of a gain or an expected return.

The per-index view uses each index's own trend. Nasdaq Composite breadth may use a labelled QQQ / Nasdaq-100 proxy. Outside the S&P 500, volatility is ranked against the index's own history, rather than applying VIX levels to a different index. A-share and Hong Kong index-only views currently lack constituent breadth and valuation, and show partial coverage. Scores using different inputs are not directly comparable.

美股整体模型由趋势（40%）、市场宽度（35%）和波动环境（25%）组成。趋势比较现价与50日、200日均线，并用 RSI 调整；宽度观察成分股站上50日、200日均线的比例；波动环境使用 VIX 及其三个月期限比率。

评分＝Σ（信号分数×可用权重）÷Σ可用权重。缺失输入不计入，并显示覆盖率；缺数据不会记为零。覆盖不全的高分，其证据完整度低于数据齐全的高分。分数是满分100的环境指标，不是上涨概率或预期收益率。

各指数视图使用自身走势。纳斯达克综合指数的宽度可能使用明确标注的 QQQ／纳斯达克100代理。标普以外指数按自身历史给波动率排序，不把 VIX 水平直接套用到另一指数。A股和港股的指数独立视图目前缺少成分股宽度及估值，会显示覆盖不全。不同输入产生的评分不宜直接比较。

Company news: a dated evidence trail

The Bull & Bear tab gathers available news over the recent three-month window and sorts the extracted arguments on a time series. Each argument links its sources and publication dates; a snapshot records the requested period, generation time and processing coverage. Incomplete collection is labelled rather than presented as exhaustive.

Price-impact ranges describe conditional scenarios inferred from the evidence. They are not measured causal effects, guarantees or additive forecasts. Several headlines may describe the same event; do not sum their percentage ranges. Read the final summary alongside its supporting sources and date range.

ETF: distinguish NAV from fundamental value

Supported issuer adapters: DIA, MDY, QQQ, SPY, XBI, XLB, XLC, XLE, XLF, XLI, XLK, XLP, XLRE, XLU, XLV, XLY, XRT. The panel separates official dated NAV, a holdings-based NAV estimate and the market premium or discount. The estimate rolls a same-date official NAV forward using disclosed weights, constituent returns and distributions, assuming unchanged weights and a fixed residual cash balance. It is not official intraday NAV.

Premium / discount = (market price / NAV − 1) × 100%. A fundamental scenario range uses covered constituents' cached DCF estimates and requires at least 80% weight coverage. Missing prices, stale holdings, currency mismatches or insufficient coverage can suppress an estimate. Always compare the dates and coverage before interpreting a gap.

公司新闻：带日期的证据链

“多空观点”页收集近期三个月范围内可获得的新闻，提取观点并按时间序列排列。每条观点关联来源和发布日期；快照记录请求期间、生成时间及处理覆盖情况。采集未完成时会明确标注，不会宣称已穷尽所有新闻。

股价影响范围是根据证据推演的条件情景，不是已测量的因果影响、承诺或可以直接相加的预测。多篇新闻可能描述同一个事件，不应把各条百分比范围累加；最后的总结需要结合来源与时间范围一起读。

ETF：分清净值与基本面价值

已接入发行商适配器的ETF：DIA, MDY, QQQ, SPY, XBI, XLB, XLC, XLE, XLF, XLI, XLK, XLP, XLRE, XLU, XLV, XLY, XRT。面板分别展示带日期的官方净值、按持仓估算的净值，以及市场溢价／折价。持仓估算以同日官方净值为锚，用披露权重、成分股收益与分配调整，假设权重和剩余现金余额不变；它不是官方盘中净值。

溢价／折价＝（市场价格÷净值−1）×100%。基本面情景区间使用已覆盖成分股的缓存 DCF，要求至少80%的权重覆盖。缺少价格、持仓过旧、币种不匹配或覆盖不足时，估算可能不显示。解读价差前，请同时核对日期和覆盖率。

• A report written as a conversation — one turn per role, in the order the desk spoke, each with its own link.
• A decision — exactly one of Buy, Overweight, Hold, Underweight or Sell, with the portfolio manager's reasoning and the price levels and sizing behind it.
• A PDF of the whole report, charts included, and an email to you when the run finishes.
• A link you can share with somebody who has no account; it expires after 30 days.

• 一份以对话形式呈现的报告——每个角色一轮发言，按投研台的发言顺序排列，每一轮都有独立链接。
• 一个决策——在买入、增持、持有、减持、卖出中五选一，附投资组合经理的理由，以及背后的价位与仓位建议。
• 一份 PDF，包含完整报告与图表；分析完成后还会发邮件通知你。
• 一个可分享的链接，对方无需账号即可打开，30 天后失效。

TradeAgents is built for research. It places no orders, connects to no brokerage account and holds no money. A report is a model's argument, not investment advice: it can be wrong, and it can be confidently wrong. Nothing here is a recommendation to buy or sell anything — the rating a run ends on is the desk's conclusion, and the decision is yours.

TradeAgents 为研究而生。它不下单、不连接任何券商账户，也不持有任何资金。报告是模型的论证，而不是投资建议：它可能出错，而且可能错得很自信。本站的任何内容都不构成买入或卖出任何资产的建议——一次分析最后给出的评级只是投研台的结论，决定权在你。

• 6 analysts on every run — market, sentiment, news, earnings, quality and valuation — and 9 for China A-shares, which add policy, hot-money and lock-up specialists.
• Structured debate — 3 bull-versus-bear rounds and 3 risk rounds by default, each argued from the analysts' reports.
• Dated to the run — the data tools are asked for what was known on the run's date, so a report dated in the past is written from that day's information, as far as each source can supply it.
• Streams as it works — each section appears on the page the moment its author finishes, so a long run can be read while it is still going.
• Your choice of model and thinking depth, recorded on every report so a cheap-tier run never passes for the most capable one.
• English or Chinese — the report, its PDF and the completion email are written in the language the run was started in.

• 每次分析 6 位分析师——市场、情绪、新闻、盈利、质量与估值；A 股分析为 9 位，额外加入政策、游资与解禁三位专项分析师。
• 结构化辩论——默认进行 3 轮多空辩论与 3 轮风险辩论，每一轮都以分析师报告为依据。
• 以分析日期为准——数据工具只被询问在分析日期当天已知的信息，因此一份日期在过去的报告，会在各数据源所能提供的范围内，基于当天的信息写成。
• 边写边看——每个章节在作者写完的那一刻就出现在页面上，耗时较长的分析也可以边运行边阅读。
• 可自选模型与思考深度，并记录在每份报告上，低价档位的分析不会被误认为最高档位。
• 中文或英文——报告、PDF 和完成通知邮件都使用发起分析时选择的语言。

• Quick start — your first report in five steps.
• How a run works — every stage, and what each one hands the next.
• The desk — every role on the team.
• Research Lab — what we have measured, including where we were wrong.

• 快速开始——五步完成第一份报告。
• 一次分析如何运作——每个阶段，以及它交给下一阶段什么。
• 投研团队——团队中的每个角色。
• 研究实验室——我们测量过的东西，包括我们曾经错在哪里。