Deal Comps

Precedent Transaction Comps

先载入 xlsx-author 技能，再动手建表。 出处载体（Class A 的 Source: 批注 / Class B 的 来源 表加 来源编号 列）、四色含义、「填机构不填接口名」、URL 的两种诚实 写法、交付时的 ::zcode-file-citation，都在那份 SKILL.md 里。只照路径调它的 recalc.py 不等于读过它——2026-08-24 那批里有一份工作簿正是这样，五项全漏。

What comparable deals actually paid, assembled from the announcements that disclosed them. Two artifacts are possible — a Markdown snapshot or an .xlsx workbook — and they carry provenance differently. Choose once, up front, and say which you chose.

Read this before you start: there is no precedent-transaction database

This plugin has no deals database, no league-table feed, and no private-deal terms source. No tool here answers the question "give me comparable transactions". Anyone who has used a banking terminal expects one to exist; it does not, and pretending otherwise produces a table of invented rows.

A precedent set here is assembled by hand, deal by deal:

1. Candidates are surfaced through get_stock_events (并购指标, 借壳上市 status, 增发 used as acquisition currency) and search_stocks for a same-industry universe, then narrowed by reading.
2. Each candidate's terms are extracted from the announcement full text via wind-docs.get_company_announcements.
3. Each row cites the announcement it came from. A row you could not tie to an announcement does not go in the table.

Say this to the user in the delivery message, in plain words: the set is hand-built, here is how candidates were found, here is what the search covered, and here is what it will therefore have missed — unlisted acquirers, deals where neither side is an A-share issuer, and deals whose terms were never disclosed are structurally invisible to this method.

A precedent set you could not assemble is 源不可用, not an empty table. If the event and announcement searches did not run, or ran and could not be read, say so. 检索范围内未发现 is only honest when the searches actually executed and returned nothing, and even then it means "our searches found none", never "no comparable transactions exist".

Reference prices: filter by volume first. get_stock_performance does not skip suspended-trading days — it fills all four OHLC values with the previous close and nulls only the volume. Taking a "prior trading day close" therefore yields a carried-over price: the number is right, the base date is wrong, and nothing looks odd. A row with null volume is not a trading day. See references/stock-price-traps.md.

Language of the deliverable

Authoring prose here is English. The deliverable follows the user's language (the language policy). Provenance tags follow the deliverable — [披露] [测算] [预期] [推断] [媒体] in a Chinese one, [Reported] [Est.] [Consensus] [Inferred] [Media] in an English one. One style per document, never mixed.

Pick the artifact, and know which citation vehicle governs it

This skill carries both citation vehicles because it has two output forms (citation vehicles: sources_section, cell_comments).

Both are right about their own medium (the house formatting policy arbitrates). Do not carry a workbook's borders into a document, and do not strip a workbook of them because a document does without. In both media, numbers are right-aligned and text is left-aligned, and every deal states its own source exactly once.

Everything below about candidate identification, term extraction, 口径 discipline, premium reference dates, and coverage honesty applies to both artifacts. Only the artifact changes.

Step 1 — Define the screen, and write it down

Before searching, agree and record:

• The subject deal or subject company the set is for.
• Industry perimeter — the 行业分类 from get_stock_info, plus the business description, because an industry code and a business comparability judgement are not the same thing.
• Window — announcement dates from and to. State it; a 2019 multiple and a 2026 multiple are not members of one population.
• Deal-size floor and ceiling, if any.
• Deal type — control acquisition, minority stake, asset purchase, 借壳上市, merger of equals. Never mix control and minority deals in one median: a minority stake carries no control premium and its multiple is not comparable.
• What counts as included — announced only, or completed only, or both with a status column. Say which.

The screen definition goes in the deliverable. A precedent set whose selection rule is undocumented cannot be reviewed, and its median means nothing.

Step 2 — Surface candidates

• search_stocks for the same-industry A-share universe, giving a code list to work through.
• get_stock_events per code (one call per name), reading 并购指标 for 备考财务数据 and 借壳上市 status, and 增发 where shares were issued as acquisition currency.
• wind-docs.get_company_announcements per candidate, searching the disclosure record for 交易预案, 重组草案, 重大资产购买报告书, 收购报告书, 权益变动报告书 within the window.
• wind-docs.get_financial_news may surface a deal you would otherwise miss — but a deal found only there is [媒体]/[Media] and cannot enter the multiples table until an announcement confirms its terms. It may be listed separately as a known deal with undisclosed terms.

Record how many candidates each step produced and how many survived. That count is part of the coverage block.

Step 3 — Extract the terms, from the announcement

For each surviving deal, read the announcement and take:

Every one of these is [披露]/[Reported] and cites the announcement. A figure you computed from them is [测算]/[Est.]. A figure only a news article carries is [媒体]/[Media], and it is not silently promoted.

Step 4 — Compute the transaction multiples

Transaction EV = equity value of the transaction
                 (grossed to 100% where a partial stake was acquired — say so)
               + target debt (有息负债)
               − target cash
               + minority interest
               + other bridge items

EV / Revenue   = transaction EV / target revenue    (报告期 stated)
EV / EBITDA    = transaction EV / target EBITDA     (报告期 and construction stated)
P / E          = equity value / target net profit   (归母 or 全口径 — stated)

Discipline that decides whether the set is usable:

• Gross up partial stakes. A 51% stake bought for X implies an equity value of X/0.51 only if there is no control premium embedded — say which convention you used and flag it as [测算]/[Est.].
• One 口径 per column. 归母 in a P/E column throughout, or 全口径 throughout, never a mix, never an average of the two.
• One period convention per column. LTM throughout or last-full-year throughout. Where a deal only discloses a 备考 figure, put it in its own column or flag the row.
• Negative or near-zero denominators produce no multiple. Write n.m. and exclude the row from the median; do not print a 300x that drags the mean.
• State N for every statistic. A median of three deals is three observations, not a level. Show min / 25th / median / 75th / max only where N supports it, and show the observations themselves where it does not.

Step 5 — The control premium, and its reference date

A premium without its base date is not a number. This is the most abused figure in deal work.

For each deal, pull the target's price history with get_stock_performance (一次一个标的 per call, dates yyyyMMdd) around the announcement date and compute:

Every premium column states, on the face of the table:

• which reference it uses (1-day, 30-day, or other),
• the reference date,
• the reference price, and
• whether the stock was suspended (停牌) before the announcement — an A-share target is frequently suspended ahead of a 重组, which makes the last close stale and the 1-day premium misleading. Where the stock was suspended, say so in the row and prefer a pre-suspension reference, stating the change.

Premia are [测算]/[Est.] — we computed them — even though both inputs are [披露]/[Reported].

Step 6 — Present it

Markdown snapshot. One table of deals with the key terms, one table of multiples and premia with a statistics block, the screen definition, then ## 来源 (or ## Sources) and the coverage block. Every deal row carries an inline [n] marker pointing at its announcement. The count of distinct [n] markers must equal the number of entries. Entry schema:

[n] 〔一手|二手〕发布主体 · 文档或系统名 · 日期(发布日; 检索于) · URL
[n] 〔Primary|Secondary〕 Publisher · Document or system · Date (published; retrieved) · URL

An exchange announcement retrieved through wind-docs.get_company_announcements is 一手 / Primary. A term relayed by a news article is 二手 / Secondary and must name what it relays. A generic attribution blob — "资料来源：公司公告、同花顺" — is not a citation: it names no document, no date, and no mapping from claim to source.

Workbook. Four sheets:

Cell comments in the workbook take the standard form, written as each cell is created:

Source: <System or Document>, <Date>, <Reference>, <URL if applicable>

e.g. Source: XX股份发行股份购买资产暨关联交易预案, 2025-11-04, 第四节「标的资产评估 情况」, https://…. A calculated cell carries a formula instead — a source comment on a calculated cell means a hardcode is hiding in it.

Checks asserts: EV bridge ties for every row; every multiple equals its numerator over its stated denominator; every premium equals offer over its stated reference price; the statistics block covers exactly the rows not marked n.m. or excluded; and the N shown equals the number of rows in the statistic.

Common errors in a precedent set

Verification before delivery

Workbook. Run python3 ../xlsx-author/scripts/recalc.py "<the workbook you just wrote>" 30 (path relative to this skill's directory). Exit 0 clean · 2 fix and re-run · 3 recalc_unavailable is NOT a pass — no formula was evaluated; run xlsx-author's substitute protocol (reference check, independent recompute in Python, assert the identities on Checks) and record it in the coverage block · 1 hard failure. Then apply the model-scope audit conventions from the audit-xls skill (it ships with the write-research plugin; work its checklist by hand if it is not loaded), and sweep the four font colours and the cell comments.

Snapshot. Distinct [n] markers equal Sources entries; every entry declares 一手/二手 (Primary/Secondary) and every secondary names its relay chain; every premium carries its reference date; every statistic carries its N.

Delivery — the coverage block

Close with this block, verbatim heading — ## 覆盖范围与局限 for a Chinese deliverable, ## Coverage and Limitations for an English one — even when everything passed. It goes before or after ## 来源 / ## Sources.

## 覆盖范围与局限
检索于: <timestamp>  ·  口径/委托用途: <e.g. 交易定价参考 / 董事会材料>

筛选口径: <industry perimeter, window, size range, deal type, included statuses>

| 检查项 | 结论 | 源 | 检索于 |
|---|---|---|---|
| 同行业候选交易识别 | 有记录(N 家筛出, M 笔入选) | 同花顺 hexin-stock get_stock_events | 2026-07-25 |
| 交易条款提取 | 有记录(M 笔全部取自公告) | 万得 wind-docs.get_company_announcements | 2026-07-25 |
| 停牌前参考价 | 有记录(M 笔) / 部分源不可用 | 同花顺 hexin-stock get_stock_performance | 2026-07-25 |
| 未披露条款交易 | 有记录(K 笔仅见媒体,未入表) | 万得 wind-docs.get_financial_news | 2026-07-25 |

本次未能覆盖: 本插件没有可比交易数据库,交易集为逐笔人工搜集。非上市收购方、
双方均非 A 股主体的交易、以及未披露对价的交易在本方法下不可见。<plus any tool
that failed, named as 源不可用 and what it would have covered>
数据滞后性: 公告披露与事件收录存在滞后;终止交易的条款为公告当时口径。
溢价基准: <1-day / 30-day, and the reference dates used>
公式核验: <recalc.py evaluated all N formulas, zero errors (exit 0) / recalc.py
could NOT evaluate the formulas (exit 3) — static lint plus openpyxl
reference/recompute/identity checks only; the workbook is NOT verified>

For the Markdown snapshot the same block compresses to prose, but must still name what was searched, what was not, how many deals survived, the premium reference, and the retrieval time.

检索范围内未发现 means the source was queried and returned nothing — never rendered as "there are no comparable transactions". 源不可用 means it could not be queried, and is a finding in its own right. Placeholders stay placeholders: never invent an example premium or a representative multiple to make a table look complete.

This set stages market evidence for the deal team. It is not a fairness opinion, it does not set a price, and a median is not a recommendation.