Earnings Analysis

Equity Research Earnings Update

Create professional EARNINGS UPDATE REPORTS analyzing quarterly results for companies already under coverage, following institutional standards (JPMorgan, Goldman Sachs, Morgan Stanley format).

Key Characteristics:

• Length: 8-12 pages
• Word Count: 3,000-5,000 words
• Tables: 1-3 summary tables (NOT comprehensive)
• Figures: 8-12 charts
• Turnaround: 1-2 days (within 24-48 hours of earnings)
• Audience: Clients already familiar with the company
• Focus: What's NEW - beat/miss, updated estimates, thesis impact
• Font: Times New Roman throughout (unless user specifies otherwise)

When to Use

Use when the user requests:

• "Create an earnings update for [Company] Q3 2024"
• "Analyze [Company]'s quarterly results"
• "Post-earnings report for [Company]"
• "Q1/Q2/Q3/Q4 update for [Company]"

Do NOT use if:

• User requests "initiation report" → Use different skill
• User requests "flash note" or "quick take" → Different format
• Company is not already covered → Need initiation first

Critical Requirements

0. Data Integrity — Never Fabricate

This is the highest-priority requirement in this skill. It overrides the desire for a complete-looking report. A blank cell is always better than an invented number.

Every number, date, and factual claim in the report carries exactly one provenance tag, rendered as a visual chip so the reader can see which class it belongs to:

• [披露] — a primary record: the company disclosed it (earnings release, 10-Q/6-K, transcript, investor deck) or an exchange/regulator recorded it. The default and most trusted class.
• [测算] — a value YOU computed or assumed (margin = EBITA/revenue, price target, forward estimates, like-for-like growth). Must be derivable from [披露] inputs, and the derivation must be stated or reproducible.
• [预期] — a third-party estimate, with a named provider, and the provider must be a data vendor's aggregated consensus. The wired route is 同花顺 一致预期 via hexin-stock.get_stock_financials (预测每股收益平均值, 目标价(综合值), 评级机构家数 — ask with the forecast year inline); LSEG / Visible Alpha only where 同花顺 does not cover it. An unattributed consensus number does not exist. Note that 目标价(综合值) is an aggregate and is citable; a named peer 券商's own target price is not — this is sell-side research, so engage with a competitor's reasoning and redo the arithmetic yourself as [测算].
• [推断] — our analytical inference with no corresponding record: an attribution hypothesis ("the beat came mostly from pricing"), a read-across to peers, a likely cause.
• [媒体] — reported by media and NOT corroborated by a disclosed record. Stays [媒体] until a record confirms it, at which point it becomes [披露] and cites that record.

When two tags could apply, take the weaker one: a [披露] figure you then adjusted is [测算].

When the request asks "how much of it", build the bridge — see the decomposition rule below. 「毛利率变化里多少来自原材料 价格、多少来自自身降本」 is arithmetic, not a characterisation: open → signed steps → close, with a named 其他/未拆分 residual, every step [测算] with its basis. Where the split cannot close (segment volumes or unit costs not disclosed at the granularity asked for), name the missing input, then give the qualitative read tagged [推断] so the reader can see they got a view rather than the numbers. Answering an unhedged 「价格联动与降本消化了大部分增量」 to a "how much" question is the failure this rule exists to prevent.

[测算], [推断], and [媒体] are never optional in any format — nothing lets a reader tell our arithmetic from a disclosure, our judgement from a finding, or an unconfirmed report from a record. An earnings update puts all these classes side by side in every beat/miss table, so chip every number in tables, variance columns, valuation summaries, and KPI strips.

[n] does not carry the class, and in an earnings update [披露] is not optional either. A citation says which document a figure came from; it cannot say whether the figure was printed there or computed by us off it. The exemption for uniformly-sourced flowing prose (a sector primer, a market landscape) does not reach this deliverable: a 业绩点评 is wall-to-wall mixed classes, and the moment a [n] stands in for a chip the reader loses the one distinction the tags exist for. Measured on two 业绩点评 of near-identical length: 134 tags against 64, and the thinner one had swapped [n] for [披露] through its 摘要 and 分部 sections and left all thirteen of its 环比 figures — every one computed off the prior quarter — reading as disclosures.

The coverage floor, which is where this actually goes wrong:

• Every 同比 / 环比 / 占比 / pct figure carries a class. A move you computed off a prior period is [测算]; one the issuer printed in the release is [披露]. Untagged they are the same figure to the reader. report-render's verify.py reports untagged derivations on the built file, and build() refuses a table whose 同比/环比 column has no chip in its header or any cell under it.
• One chip per run, not per figure. 「同比 +102.7%、环比 +76.6%[测算]」 is the correct shape for a run of one class; splitting at a class boundary — 「毛利率 23.15%[披露]，环比 −1.67pct[测算]」 — is the correct shape where it crosses. Chipping each figure separately is noise, and so is a note under the table explaining the 口径 in prose: the reader cannot map a sentence onto cells.
• Route every chip through rep.chip(标签) / rep.tagged(值, 标签). Do not define your own helper, copy the hex values, or hand-write the <font> markup: colour, size and the legal tag set live in theme.CHIP, and a local copy is how one batch shipped three different renderings of the same five tags.

Hard rules — violating any of these is a serious failure:

1. Never invent a consensus number. If you only have consensus for total revenue and EPS (the usual case), then the beat/miss table shows consensus for ONLY those rows. Leave segment-level consensus cells blank — do NOT back-fill them with your own estimate dressed up as "consensus." Then say which line items consensus was retrievable for in the coverage block (Requirement 7); a blank cell without that statement reads as an oversight rather than a finding.
2. Never present an [测算] value as [披露] or [预期]. A forward estimate, a margin you divided out, a "prior estimate" you never actually published — all are [测算] and must be labeled so, or omitted.
3. No fake "old vs new estimate" columns. Only show an estimate-revision table if you genuinely have a prior estimate on record. Otherwise show disclosed historical actuals only, and say so.
4. Reconcile before you write. Cross-check that prior-year × (1+YoY) ≈ current value; if segment pieces don't tie to the total, find the real disclosed figure rather than plugging a gap.
5. Two conflicting disclosed figures → show both with their basis. (E.g. CMR "+1% reported / +8% ex-subsidy" — report both, don't cherry-pick the flattering one.)

When a reader cannot tell whether a number is fact, your estimate, the Street's, your inference, or an unconfirmed press report, the report has failed regardless of how polished it looks.

1. Speed & Timeliness

• Publish within 24-48 hours of earnings release
• Focus on NEW information only
• Don't rehash company background extensively

2. Beat/Miss Analysis

• Lead with whether company beat or missed estimates
• Quantify variances (e.g., "Revenue beat by $120M or 3%")
• Explain WHY results differed from expectations

3. Summary Format

• Keep tables to 1-3 (summary only, not comprehensive)
• No full P&L/Cash Flow/Balance Sheet (just key metrics)
• Assume reader has seen initiation report

4. Citations & Source Attribution — mandatory

CRITICAL: Properly cite all data with SPECIFIC sources and CLICKABLE HYPERLINKS.

Use numbered inline citations [n] in body text, figure captions, and table notes, mapped to a numbered ## 来源 section at the end. Where the output format supports links (PDF, HTML), make [n] a clickable internal link that jumps to the matching entry, and make each entry's title + URL a clickable external link. Pair citations with the provenance chips from Requirement 0 so each number shows BOTH its class and its source — a chip without an [n] asserts that a record exists without saying which one.

Marker rules:

• [n] ascends on first use and is reused for repeat citations of the same source.
• The count of distinct [n] markers must equal the number of Sources entries. A marker with no entry, or an entry no marker points to, is a defect.
• A generic attribution blob ("Source: company filings, Street estimates") is not a citation: it names no document, no date, and no mapping from claim to source.

Include specific citations in every figure and table:

Source: Q3 FY24 10-Q filed 2024-11-08 [1]; company earnings release [2].

HOW CLICKABLE LINKS WORK (per format):

• PDF (reportlab): inline [n] → internal links (<a href="#refN">[n]</a>) jumping to the entries; entries → external links (<a href="https://...">), whole title line clickable for a larger hit area. Verify by counting /Link annotations in the built file (external URI + internal GoTo) and confirming the internal jumps resolve. report-render automates this count.
• HTML/Markdown: <a href="#refN"> anchors and <a href="..."> external links.
• In every case: plain-text URLs that are not real links do NOT count as done.

REQUIRED SOURCES LIST:

Cite in every earnings update:

• ✅ Earnings release (with date and URL)
• ✅ 10-Q / 10-K / 6-K filing (with filing date and EDGAR link)
• ✅ Earnings call transcript (with date)
• ✅ Investor presentation/supplemental materials (if available)
• ✅ Consensus estimates source (同花顺 一致预期 via hexin-stock.get_stock_financials, or LSEG etc. where it does not cover — with retrieval date) — only for metrics you actually have consensus on (see Requirement 0)
• ✅ Prior guidance (from previous quarter's materials)

THE SOURCES SECTION:

Heading is exactly ## 来源. One entry per marker, one shape:

[n] 〔一手|二手〕发布主体 · 文档或系统名 · 日期(发布日; 检索于) · URL

• 〔一手|二手〕 is mandatory, not implied. Primary is the issuing document itself: the earnings release, the periodic filing, the transcript, the investor deck, the regulator's decision, or a database field sourced from those. Everything else is Secondary.
• A Secondary entry must name what it relays — 〔Secondary〕 <outlet> · relaying <the report or filing it quotes> · …. Writing "Secondary" without naming the relay chain hides the chain it was supposed to expose. If a figure resolves only to a news article quoting a report, that is Secondary, and say in the coverage block that the underlying report was not obtained.
• An estimates aggregator is Primary for the consensus figure it publishes, and Secondary if you took that consensus from an article quoting it.
• Both dates when both exist. Publication date establishes vintage, retrieval date establishes what we could see. For a live database query publication may be absent — then retrieved <date> alone. Where a figure depends on time of day (a pre-earnings consensus snapshot, the session's move), carry time and session.
• In a paginated report, ## 来源 starts on its own page and is the only hard page break besides the one after the cover.

Placeholder shape — substitute the real documents, never ship the placeholders:

## 来源

[1] 〔Primary〕 <Company> · Q<N> FY<YY> earnings release · <YYYY-MM-DD> (published; retrieved <YYYY-MM-DD>) · https://…
[2] 〔Primary〕 SEC EDGAR · <Company> Form 10-Q, Q<N> FY<YY> · <YYYY-MM-DD> (filed; retrieved <YYYY-MM-DD>) · https://www.sec.gov/…
[3] 〔Primary〕 <Company> · Q<N> FY<YY> earnings call transcript · <YYYY-MM-DD> (published; retrieved <YYYY-MM-DD>) · https://…
[4] 〔Primary〕 <consensus provider> · consensus for revenue and EPS, pre-earnings close · <YYYY-MM-DD HH:MM close> (retrieved) · https://…

VERIFICATION CHECKLIST:

• Every figure has source with specific document and date
• Every table has source with document reference
• Beat/miss analysis cites consensus source with its retrieval date
• Guidance changes cite current and prior guidance sources
• Key statistics have footnotes
• ## 来源 lists every material, each with 〔一手|二手〕, dates, and URL
• Distinct [n] marker count equals the entry count
• ALL URLs are CLICKABLE HYPERLINKS (not plain text)
• All SEC filings hyperlinked to EDGAR viewer

5. Updated Estimates

• Update forward estimates based on results
• Show old vs. new estimates clearly — but only if a genuine prior estimate exists (see Requirement 0). If there is no prior estimate on record, do not manufacture an "old" column; instead present disclosed historical actuals and your new estimates, each tagged [披露] / [测算].
• Explain what changed and why

6. Output Format & Format-Agnostic Quality Bar

The deliverable format follows the user's request — if they asked for PDF, produce PDF. If unspecified, the house formatting policy decides: an 8-12 page earnings update is long-form, so deliver a PDF, and state that choice in one clause when you make it. The report-render skill builds the PDF (reportlab + matplotlib) and automates the layout and /Link checks below. 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。

Whatever the format, these quality conditions are NON-NEGOTIABLE. They are format-agnostic — the technique differs per format, but the outcome must hold in all of them:

A. Charts and their source notes must never overlap. Each chart needs a "Figure X — title" caption above and a Source: line below, and the source line must sit clear of the plot, axis labels, and data labels.

• In matplotlib/PDF: render the source as a figure-level footnote (fig.text() in figure coordinates) with reserved bottom margin (subplots_adjust(bottom=...)), NOT ax.text() at hand-tuned data coordinates — the latter collides with x-axis labels when the data range shifts.
• In HTML/PPTX: put the caption and source as separate text elements above/below the image, not overlaid on it.
• State a figure's or table's source exactly once, in its note line — never also inside the figure canvas and never also as a row of the table. For a chart built through report-render, that means rep.figure(..., note=...) and not charts.source().

B. Chinese / CJK text must actually render — never blank, and must wrap correctly. Embed a real TTF (Regular AND Bold; built-in CID fonts such as STSong-Light render blank in viewers without Adobe CJK packs), and set wordWrap='CJK' on every paragraph style that can hold Chinese, or lines wrap early and stop short of the right margin. These are silent failures — nothing raises. The full catalogue (font-family mapping, variable fonts, matplotlib tofu, table cells that ignore wordWrap) is in report-render's references/pitfalls.md; read it before hand-rolling a builder.

C. Links must be genuinely clickable (where the format supports links). Inline [n] markers jump to the Sources entries; entries open the source. In PDF these are real /Link annotations — count them in the built file — in HTML real <a> tags. Never plain blue text.

D. No overflow, orphan, or half-empty pages/slides. After building, render every page/slide and check: no element overlapping another, no source/caption clipped, no page left mostly empty because one block got stranded.

• Bind every figure/table to its caption (reportlab: one KeepTogether([caption, image, note])), or the caption strands at a page bottom while the image flows to the next page.
• Let content flow; do NOT force breaks. A PageBreak() before every heading is the #1 cause of half-empty pages. The only legitimate hard breaks are after the cover and before ## 来源.
• If a figure "almost fits," shrink it so it fits (e.g. 174mm → ~138mm) rather than stranding a half-empty page. Cap figure height at roughly one third of the text column.
• Fix remaining gaps by reordering, merging orphans, or resizing — then re-render and re-check. (A modest gap is fine on the cover and on the standalone Sources page; mid-document content pages should be full.)

The self-check is mandatory regardless of format: build it, render/open it, look at every page.

7. Coverage and Limitations

Every earnings update closes with a coverage block, headed exactly ## 覆盖范围与局限. It is not optional and not abbreviated when everything succeeded — a reader cannot otherwise distinguish "we checked seven things and found nothing" from "we checked four and three never ran."

Each check reports exactly one of three states:

检索范围内未发现 is a statement about our search, not about the world: never render it as "the company did not disclose it." 源不可用 is a finding, not a gap to paper over.

For an earnings update the block covers, at minimum:

• Which filings, transcripts, releases, and supplemental materials were obtained, and which were not yet available at the time of writing.
• Whether consensus was retrievable, from which provider, and for which line items — this is where the blank segment-consensus cells from Requirement 0 get their explanation ("consensus retrievable for total revenue and EPS only; no segment-level consensus available from ").
• Whether a prior published estimate exists (if not, say so — it is why there is no old-vs-new column).
• The retrieval timestamp, and the as-of time of the consensus snapshot.

## 覆盖范围与局限
Retrieved: <YYYY-MM-DD HH:MM TZ>  ·  Scope: Q<N> FY<YY> results for <Company> (<ticker>)

| 检查项 | 结论 | 源 | 检索于 |
|---|---|---|---|
| Earnings release | 有记录 [1] | <publisher> | <date> |
| Form 10-Q | 检索范围内未发现 (not yet filed) | SEC EDGAR | <date> |
| Consensus, revenue + EPS | 有记录 [4] | <provider> | <date> |
| Consensus, segment level | 源不可用 | <provider> (not covered) | <date> |

本次未能覆盖: <sources that failed, and what they would have covered>

High-Level Workflow

The earnings update process follows 5 phases:

Phase 1: Data Collection (30-60 minutes)

🚨🚨🚨 CRITICAL: TRAINING DATA IS OUTDATED 🚨🚨🚨

BEFORE STARTING - COMPLETE THESE 4 STEPS IN ORDER:

1. CHECK TODAY'S DATE - Write down the current date
2. GO TO THE STRUCTURED SOURCE FIRST, NOT WEB SEARCH - the disclosure record, in this order:
   • A shares: wind-docs.get_company_announcements (业绩预告/快报/定期报告;无日期参数,窗口写进 query 后自筛) and hexin-stock.get_stock_financials for the reported lines
   • HK/US: hexin-global-stock.global_stock_financial; where the form type or filing date matters, sec-search.sec_full_text_search with form_types and a date window
   • finance-search.finance_search sits between the MCP tools and general web search — 先走 finance-search.finance_search(金融垂搜,按档位召回并标注来源等级),它没有命中再落通用搜索。媒体档默认带日期窗口;官方档(weight>=3)反过来 —— 日期窗口会丢掉 publish_time 为空的记录(该档多数如此),取原始文件时不带窗口。 General web search remains the last fallback for what none of these carry (a transcript, a call replay, a management quote) — not the entry point. Pulling an earnings number off the web when a tool answers is how an undated figure enters the note.
3. VERIFY THE DATE - Confirm the release is within the last 3 months, from the source's own date field
4. CHECK TRANSCRIPT DATE - Verify transcript date matches release date

COMMON MISTAKE: Using outdated earnings calls from training data instead of searching for the latest.

REQUIREMENTS:

• ✅ Search for latest earnings - do NOT rely on training data
• ✅ Write down today's date and the release date found
• ✅ Verify release date is within 3 months of today
• ✅ Verify transcript date matches release date
• ✅ If dates don't match or are old (>3 months), search again

See references/workflow.md for detailed search procedures and verification steps.

Phase 2: Analysis (2-3 hours)

• Beat/miss analysis for each key metric
• Segment/geographic/product breakdown
• Margin and guidance analysis
• Update financial model and estimates

See references/workflow.md for detailed analysis framework.

Phase 3: Chart Generation (1-2 hours)

Create 8-12 charts focusing on quarterly trends and what's new:

• Quarterly revenue progression
• Quarterly EPS progression
• Quarterly margin trends
• Revenue by segment/geography
• Key operating metrics
• Beat/miss summary
• Estimate revisions
• Valuation charts

See references/workflow.md for chart specifications.

Phase 4: Report Creation (2-3 hours)

Create the 8-12 page report with the structure below.

See references/report-structure.md for complete page-by-page templates and formatting requirements.

High-level structure:

• Page 1: Earnings summary with draft rating and price target (草稿,待分析师确认)
• Pages 2-3: Detailed results analysis
• Pages 4-5: Key metrics & guidance
• Pages 6-7: Updated investment thesis
• Pages 8-10: Valuation & estimates
• Pages 11-12: Appendix (optional)
• Closing: ## 覆盖范围与局限, then ## 来源 on its own page

Phase 5: Quality Check & Delivery (30 minutes)

Verify content, formatting, accuracy, and timeliness before delivery.

See references/best-practices.md for quality checklist and common mistakes to avoid.

Output Specification

Primary Deliverable: an 8-12 page earnings update in the format the user requested. If the user didn't specify, deliver PDF (long-form, per the house formatting policy) and say so in one clause; report-render builds it with reportlab + matplotlib. File Name: [Company]_Q[Quarter]_[Year]_Earnings_Update.[ext] Example: Nike_Q2_FY24_Earnings_Update.pdf Regardless of format, satisfy the format-agnostic quality bar in Requirement 6 (charts/source not overlapping, CJK visible, links clickable, no orphan/overflow pages), the data-integrity rules in Requirement 0, and the coverage block in Requirement 7.

Contents:

• Page 1: Summary with draft rating, draft price target (草稿,待分析师确认; stance [推断], target arithmetic [测算]) and key takeaways
• Pages 2-3: Detailed results analysis
• Pages 4-5: Key metrics and guidance
• Pages 6-7: Updated thesis assessment
• Pages 8-10: Valuation and estimates
• Pages 11-12: Appendix (optional)
• ## 覆盖范围与局限
• ## 来源 — on its own page, every entry 〔一手|二手〕, dated, and clickable
• 8-12 embedded charts
• 1-3 summary tables

Optional Deliverable: XLS model update (optional for earnings updates)

Key Differences from Initiation Report

Resources

references/workflow.md

Detailed Phase 1-5 instructions with step-by-step procedures for data collection, analysis, chart generation, and report creation.

references/report-structure.md

Complete page-by-page templates, table formats, and formatting requirements for the report.

references/best-practices.md

Examples of good/bad headlines, tips for success, common mistakes to avoid, and comprehensive quality checklist.

Dependencies

Required:

• Python: matplotlib/pandas for charts
• The report-render skill for PDF output — it builds the document and runs the render-and-inspect self-check, including the /Link annotation count. Its references/pitfalls.md is the canonical list of silent font, wrapping, figure, and pagination failures; read it before hand-rolling a builder.
• For Chinese output: a CJK font that actually renders — an embedded TTF (e.g. Noto Sans SC), Regular and Bold, with fontTools if static weights must be instantiated from a variable font

Optional:

• pptx-author — when the user wants slides
• XLS skill for model updates (not required for earnings updates)