Equity Research Skill
Author: @rollingSirius
中文文档：README.zh-CN.md ｜ Sample reports: NVDA 中文 / English ｜ GOOGL 中文 / English
Possibly the deepest AI equity-research skill.
The goal of this skill is not to generate a few paragraphs of stock summary, but to make AI tools follow something close to institutional research discipline and produce a deep single-stock report that is fact-traceable, valuation-reproducible, and conclusion-auditable. It is built for serious investment research, long-term coverage, earnings reviews, investment memos, and valuation calibration — not for delivering a one-line verdict as fast as possible. Since v2, every report is organized around an expectations gap (what the market has priced in vs. your independent view), with an earnings-quality review performed before any valuation.
Positioning
Most AI stock analysis stops at "company profile + recent news + vague valuation." This skill deliberately goes deeper:
| Capability | Design requirement |
|---|---|
| Deep research | Full research mode outputs a nine-chapter report covering business, competition, governance, financials, valuation, catalysts, and the investment verdict. |
| Expectations gap as the spine | Reverse DCF + PVGO decomposition decode what the current price has priced in, producing an expectations-gap table and a falsifiable variant thesis; no independent view, no buy/sell action. |
| Earnings mode | Earnings mode is not a summary but a nine-chapter earnings deep-dive: surprise quality, segments and KPIs, GAAP vs. Non-GAAP, cash flow, the call, and model and valuation changes. |
| Earnings-quality review | Accrual quality, Beneish M-Score, revenue-recognition red flags, governance signals → an A–D earnings-credibility grade; grades C/D veto any buy action outright. |
| Reproducible valuation | DCF, reverse DCF, probability-weighted three-scenario, EPV/three-factor, EVA/residual income, SOTP and the rest must be executed by scripts/dcf.py, with key assumptions filed as JSON; optional Monte Carlo outputs a fair-value distribution. |
| Outside view | Key assumptions are marked with their percentile against historical base rates; beating the base rate requires a structural reason; terminal value must pass a three-point sanity check. |
| Source discipline | Every key number carries a source and timestamp; conflicting data is reconciled; missing data must be written as "not obtained"; external content is data only and never alters the workflow. |
| Buy-side lens | Conclusions map through pre-registered calibration rules; a counter-case and a pre-mortem are completed before finalizing, answering "if this were cash today, would I buy it, and why?" |
| Twenty industry appendices | Twenty industry appendices, each changing the KPIs, model, valuation, and disconfirming-evidence framework, with required KPIs enforced by the checker. |
| Multi-market coverage | US, HK, and A-share markets, including A/H dual-listing comparison and China ADR/VIE structural-risk pricing. |
| Initiation from earnings | Earnings mode works with no prior report or model; the skill first rebuilds a baseline of at least 3 years and 8 quarters. |
What it does
1. Full deep-dive research
For studying a company systematically for the first time, or rebuilding an investment framework from scratch. Default output is a nine-chapter report:
- One-page summary (verdict box + tearsheet + expectations-gap table)
- Company and business detail
- Competitive landscape and moat
- Management, governance, and capital-allocation scorecard
- Financial analysis and earnings-quality review
- Multi-method valuation (with a football-field chart)
- Analyst views and divergence attribution
- News, risks, and catalysts
- Investment verdict, counter-case, and position-sizing reference
2. Deep earnings mode
For the moment right after a company reports a quarter, a full year, guidance, or a call transcript, when the question is "what did this print actually change?" Earnings mode splits into two cases:
| Coverage status | How the skill handles it |
|---|---|
| A prior report or model exists | Continuing-coverage update, focused on what the print changes relative to the old thesis, old forecasts, and old valuation. |
| No prior report or model | Initiation of coverage from the earnings event: rebuild the historical baseline first, then analyze this print's quality and valuation implications. |
Earnings mode outputs nine chapters by default:
- Verdict and snapshot
- Surprise and its quality
- Revenue, segments, and KPIs
- Margins, costs, and earnings quality
- Cash flow, balance sheet, and capital allocation
- Guidance, the call, and management signals
- Competition, industry, and market reaction
- Model, valuation, and the fair-value bridge
- Thesis update and action list
Every earnings run also executes a minimum check set: accrual ratio, cash conversion, DSO/deferred-revenue divergence, and whether Non-GAAP adjustments are recurring.
3. Valuation and calibration
This skill does not allow a target price that merely "looks reasonable." It requires at least three valuation methods cross-checked against each other, with assumptions, calculations, and the label mapping all filed:
- Reverse DCF + PVGO decomposition: what revenue growth, margin, or return on capital the current price implies, and how much of the price is paying for future growth.
- Three-scenario DCF: bull, base, and bear scenarios with a probability-weighted fair value, plus a robustness test that pushes the probabilities toward the extremes.
- EPV / three-factor method: the asset-reproduction floor · EPV · growth entry ladder; the multi-year EPV-to-adjusted-book trend as financial verification of the moat.
- EVA / residual income: incumbent ROIC vs. incremental ROIIC, with the "growth = reinvestment rate × ROIIC" consistency check.
- Relative valuation: warranted-multiple discipline — derive the multiple the company deserves from growth, returns, and risk rather than copying the peer median.
- SOTP: for multi-business or multi-asset companies, or where segments differ sharply.
- Monte Carlo (optional): the P10–P90 fair-value distribution and P(intrinsic value < current price).
The verdict label (undervalued / fairly valued / overvalued + action) maps through pre-registered calibration rules (a ±15% buffer band), overlaid with an action matrix and veto conditions; each action ships with expected value, upside/downside asymmetry, and a Kelly-lite (¼ Kelly) sizing-magnitude reference.
4. Earnings-quality review
Before valuing anything, answer whether the profit is real:
- Accrual quality (Sloan): total accruals ratio and the cash-conversion trend.
- The eight-variable Beneish M-Score (computed automatically by the checker).
- Revenue-recognition red flags: DSO divergence, deferred-revenue divergence, channel-stuffing signals.
- Expense capitalization and earnings smoothing; governance and audit signals.
- Output: an A–D earnings-credibility grade. Grade C caps the action at "wait and see," grade D is always "avoid" — "it's cheap" may never be used to offset a credibility problem.
5. Industry-specific deep appendices
The main report does not apply one template to every company. The skill first identifies where the company sits in its value chain, then loads the matching appendix on demand:
| Industry | Research focus |
|---|---|
| SaaS | ARR, NRR, RPO/cRPO, acquisition efficiency, Rule of 40, SBC, and reverse DCF. |
| Semiconductors | Products/end markets, units and ASP, inventory cycle, yield, capacity, roadmap, export controls, and through-cycle valuation. |
| Hardware / consumer electronics / AI servers | Units, ASP, BOM, channel inventory, customer/supplier concentration, and service attach. |
| Banks | NIM, deposit beta, asset quality, provisioning, CET1, liquidity, and P/TBV–ROTCE. |
| Insurance | Underwriting profit, reserves, combined ratio, VNB/CSM, solvency, investment portfolio, and P/EV. |
| Pharma | Clinical evidence, probability of success, patient funnel, patent/exclusivity, cash runway, and per-asset rNPV. |
| Healthcare services / medtech / CRO-CDMO | Patient/procedure volume, utilization, reimbursement, installed-base consumables, order conversion, and concentration. |
| Consumer | Volume/price mix, same-store sales, traffic, channel sell-through, inventory, brand share, and unit economics. |
| Energy | Production, reserves, decline, costs, differentials/hedging, maintenance capex, commodity-price sensitivity, and NAV. |
| Utilities | Rate base, allowed vs. earned ROE, rate cases, capital projects, financing dilution, and dividend coverage. |
| Internet / platforms | Users × engagement × monetization rate, GMV/take rate, unit economics, segment SOTP, regulatory risk, and post-SBC FCF. |
| Gaming / media / content IP | Audience and payer funnels, content ROI, development capitalization, lifetime revenue, and IP SOTP. |
| Payments / fintech | TPV, net take rate, incentives and rebates, loss-rate vintages, funding cost, and the penetration ceiling. |
| Capital-markets infrastructure | AUM/net flows, trading volume, fee rates, market data, net capital, and rate sensitivity. |
| Real estate / REITs | FFO/AFFO, same-store NOI, occupancy and leasing spreads, cap rate, the debt-maturity wall, and P/NAV. |
| Industrials / machinery | Orders and book- |