zai-org Index Valuation
Two questions, and the second is the one that gets skipped: where is the multiple inside its own history, and how much of the period's move was earnings rather than the multiple. Both answers are worthless without their window, so the window is stated everywhere the number appears.
The two failure modes this skill exists to prevent
A percentile with no window. 「PE 处于历史 12% 分位」 over three years and over ten years are different claims, and they routinely disagree in direction. Every percentile in the output carries its lookback window and its start date, in the same cell or the same sentence. If hexin-index.index_data returns a percentile whose default window is ~1 year (its metadata shows 截止日前一交易日), and asking for 「5 年」/「10 年」 in the query does NOT change it — a 5-year phrasing came back empty. So for any window beyond 1 year: fetch the raw PE/PB daily series and compute the percentile yourself, label [测算] with the window you chose. Presenting a percentile whose window you do not know is the failure — it looks like the most precise number on the page.
A decomposition presented as a fact. The 盈利/估值 split is arithmetic on two retrieved series, so it is [测算], and it states the identity it used and the inputs it used.
Workflow
Step 1: Resolve the index
wind-index.get_index_basicinfo — full name, 发布机构, 基日, 基点, 计算方法, 成分数量, 分类。hexin-index.index_data 的指标字典只有价格与 PE/PB,没有这些编制信息。 Two indices with similar names (全收益 vs 价格指数, 全指 vs 成分指数) have different levels and different valuation histories; say which one every number belongs to. Confirm the window and the percentile lookback with the user if they did not state one; default to a stated 10 年 lookback and say that it is the default.
Step 2: Valuation and its history
Levels from 同花顺, percentiles from Wind. hexin-index.index_data gives PE / PB levels and
index-level 每股收益; wind-index.get_index_fundamentals gives the percentile with 排名 /
最大排名, so the reader can recompute it, and it serves multi-year windows. Report the rank
and sample size beside every percentile.
- Never take a percentile from 同花顺 — its windows are broken in two different ways, and a 分位数 of 100.0 is the tell. Traps §1 / §8.
- Never mix the two vendors' percentiles in one table. They disagree materially — 12 points apart on 沪深300 on the same day (traps §9). Name the vendor, the window and the 口径 on each one, and use a single vendor throughout any comparison.
- 隐含 ROE = PB / PE,
[测算]— there is no 成份股加权 ROE field on either vendor (traps §2).
Record for each multiple:
- The 口径 — PE(TTM) / PE(静态) / PE(动态) are three different numbers, never compared.
- 报告期 of the earnings and
检索于for the price. Different dates; both in the header. - The percentile, its window, and the window's start date.
板块-level valuation — three routes, not interchangeable, say which you used:
| 需要 | 走哪条 |
|---|---|
| 板块当期估值水平 | hexin-index.sector_data(整体法,一次调用给 PE/总市值/成份股个数) |
| 板块估值分位 | 该板块对应的指数 → wind-index.get_index_fundamentals |
| 非整体法加权,或无对应指数 | 枚举成份股 hexin-stock.get_stock_financials 自行加权 |
sector_data 的主体必须写成 <板块名>板块(<分类体系>),例如 食品饮料板块(申万行业);空表是
解析失败而非无数据(traps §7)。它没有分位数指标,问它会凭空作答(traps §6)。
Step 3: Price series for the period
K 线序列走 hexin-stock.get_stock_performance（codes 收指数代码，
如 000300.SH；indicators 收 收盘价/成交量/涨跌幅/换手率；start_date+end_date
给窗口）。它按交易日历过滤，非交易日不会被填进序列。
PE/PB 只有单日截面，没有日频序列。 hexin-index.index_data 一次一个日期，
可取 收盘价/开盘价/最高价/市盈率/市盈率中位数/市净率。它的「市盈率」口径上游未说明
（返回体会带 口径提醒），要明确 TTM 口径或要分位数时不要用它。
分位数走 wind-index.get_index_fundamentals，它返回 排名/最大排名/分位数
三者，所以那个百分位是可复现的（分位 = 排名 ÷ 最大排名，实测两者能对上）。
自己按日频 PE 序列算分位数这条路走不通——没有那个序列。
Step 4: Decompose the period return — [测算]
For window t0 → t1, with P the index level, E the index earnings per unit, and M the multiple (P = M × E):
P1/P0 = (M1/M0) × (E1/E0)
盈利贡献 = E1/E0 − 1
估值贡献 = M1/M0 − 1
交叉项 = (E1/E0 − 1) × (M1/M0 − 1)
State, every time: which multiple (PE(TTM) unless said otherwise), which earnings series it implies, the two endpoint dates, and that the three components multiply rather than add — reporting them as an additive split without the 交叉项 is the common error. A log decomposition (ln(P1/P0) = ln(M1/M0) + ln(E1/E0), additive by construction) is an acceptable alternative when it is named as the method used. Dividends are outside a price-index decomposition; say so rather than letting the residual absorb them.
Do not run the decomposition when either endpoint's earnings series is missing — report 源不可用 for the split and keep the price return, which is still a fact.
Step 5: Read it
- Where the multiple sits versus its own history, with the window.
- What the weighted earnings series has been doing — a low percentile on collapsing earnings is not the same signal as a low percentile on flat earnings, and the decomposition is what tells them apart.
- Composition drift: if 成分数量 or 计算方法 changed inside the lookback, the early part of the percentile window is a different index. Flag it; this is
[推断]unless the index provider documented the change, in which case cite it.
Step 6: Assemble
Short-form is Markdown in-session. A valuation section inside a strategy report goes to PDF via report-render; a multi-index percentile table goes to .xlsx via xlsx-author. State the choice in one clause. That workbook is xlsx-author's Class B case — many retrieved rows, few formulas — so its provenance vehicle is the 来源 worksheet plus a 来源编号 column on every data sheet, not a comment per cell, and the 口径与局限 block on the 来源 sheet carries the coverage states. A roster delivered without those has no provenance at all. 用户要 Word 时同样走 report-render（DocxReport，与 Report 同一套调用）——要出 Word 就先载入 report-render 技能，再动手建；手搓 python-docx / docx-js 会丢掉标签配色、[n] 跳转与封面页，机制与实测记录在那个技能里，不在这里重述。
# [指数全称]([代码]) 估值位置
检索于: [date] · 盈利报告期: [date] · 估值分位回溯窗口: [X 年](起始 [date])
## 结论一句话
[当前倍数 + 分位 + 窗口 + 盈利方向]
## 估值与历史分位
| 指标 | 口径 | 当前值 | 分位 | 回溯窗口(起始日) | 报告期 | 标签 |
|---|---|---|---|---|---|---|
| PE | TTM | [值] | [X%] | [X 年]([date]) | [date] | [披露][n] |
| PB | — | [值] | [X%] | [X 年]([date]) | [date] | [披露][n] |
| PS | TTM | [值] | [X%] | [X 年]([date]) | [date] | [披露][n] |
## 成份股加权盈利
| 项目 | 最新 | 同比 | 报告期 | 标签 |
|---|---|---|---|---|
| 加权净利润 | [值] | [%] | [date] | [披露][n] |
| 加权 ROE | [值] | [pp] | [date] | [披露][n] |
## 区间收益拆分 [测算]
窗口: [t0] → [t1] · 方法: P = M × E,乘法拆分(含交叉项) · 倍数口径: PE(TTM)
| 分项 | 贡献 | 说明 |
|---|---|---|
| 区间价格收益 | [%] | 来自 hexin-index.index_data [n] |
| 盈利贡献 | [%] | E1/E0 − 1 |
| 估值贡献 | [%] | M1/M0 − 1 |
| 交叉项 | [%] | 两者乘积 |
不含股息;价格指数口径。
## 读数
[分位的含义、盈利方向、成分或算法变更提示(如有,标 [推断] 或引用指数公司公告)]
## 覆盖范围与局限
检索于: [date] · 口径/委托用途: 指数估值研判
| 检查项 | 结论 | 源 | 检索于 |
|---|---|---|---|
| 指数基础信息(基日/算法/成分数) | 有记录 / 检索范围内未发现 / 源不可用 | hexin-index | [date] |
| PE/PB 当期水平 | ... | 同花顺 hexin-index.index_data | [date] |
| 历史分位(含排名/最大排名) | ... | Wind wind-index.get_index_fundamentals | [date] |
| 分位回溯窗口口径 | 有记录 / 源不可用(自算,窗口见表头) | hexin-index | [date] |
| 隐含ROE(自算 PB/PE) | ... | hexin-index [测算] | [date] |
| 区间 K 线 | ... | hexin-index | [date] |
本次未能覆盖: [取不到的序列,以及它本应回答的问题]
数据滞后性: 盈利端报告期 [date],价格端 检索于 [date];季报披露前分位包含的是上一期盈利。
## 来源
[n] 〔一手|二手〕发布主体 · 文档或系统名 · 日期(发布日; 检索于) · URL
## 来源: the index provider's own publication (中证指数、上交所、深交所、Wind 指数说明) is 一手; a Wind index field sourced from the provider is 一手 and names the field and the index code, with 检索于. A broker note quoting a percentile is 二手 and names what it relays — and is not a substitute for pulling the number. Distinct [n] markers must equal the entry count.
Guardrails
- No invented percentiles, no remembered multiples, no example numbers left in the template. Placeholders stay placeholders until the tool returns a value.
- Never compare PE(TTM) against PE(静态), a 全收益 index against a 价格 index, or a percentile computed over one window against one computed over another. Where the user's question forces such a comparison, restate it on one basis and say what you changed.
- A negative or near-zero earnings base makes PE meaningless rather than high; say so instead of reporting a large multiple.
- 分位 is a description of history, not a forecast. This skill does not conclude 买入/卖出 and does not set a target level.