srinandan equity-research
Purpose
Answers the security-only half of "should I buy/sell this stock?" — is the
name attractive on its own merits, independent of any user. It produces an
EquityAssessment:
a DCF-derived intrinsic value and upside/downside versus the current price,
fundamental quality ratios, the subject's trading multiples, a standalone
valuation verdict (undervalued / fairly_valued / overvalued / unknown),
and factual drivers/risks.
Its output feeds the suitability skill, which pairs this security-only view
with the user's IPS, holdings, and risk tolerance to produce an advisory
recommendation. This skill has no notion of the user.
Methodology adapted from the Apache-2.0 anthropics/financial-services
reference skills (model-builder/dcf-model, model-builder/comps-analysis) —
see the repository NOTICE.
When this skill runs
- The user asks whether a specific ticker is worth buying or selling (e.g. "is AAPL a good buy right now?"), or asks to analyze/value a name.
- The root planner selects it before
suitabilitywhen a recommendation is requested.
It does not run for imperative trade commands with an explicit quantity
("buy 10 shares of AAPL") — those go to action-drafting.
Deterministic core + advisory narrative
Per the reviewer pattern (ADR-0014), the numbers of record are deterministic:
primitives/equity_research.py computes the DCF, ratios, and multiples from a
FundamentalsSnapshot. The LLM layer only narrates and contextualizes those
numbers — it never invents figures. Any divergence is auditable.
Inputs
| Field | Source | Required |
|---|---|---|
FundamentalsSnapshot |
executors.EdgarFundamentalsProvider (SEC EDGAR, primary/free), behind the TTL cache |
Yes |
| Latest price | executors.get_quote (Alpaca free market data; deterministic mock fallback offline) |
Recommended — needed to compute upside and a verdict |
ValuationAssumptions |
Defaults in the primitive; overridable (discount rate, terminal growth, projection years) | No |
Data access scope: read:external_market_data, read:fundamentals. This skill
reads only public company data — no Firestore/BigQuery, no user data — so its
tool surface can never leak private financial information (same isolation
rationale as research).
Output: EquityAssessment
Transient (not persisted), like the Drift Report and Research Brief. Key fields:
| Field | Meaning |
|---|---|
dcf.intrinsic_value_per_share_usd, dcf.upside_pct |
Two-stage DCF value and its gap to the current price |
quality |
Net/FCF margin, revenue CAGR, ROE, debt/equity from the latest annual period |
multiples |
Subject's P/E, P/FCF, EV/EBIT, market cap |
valuation_verdict |
undervalued / fairly_valued / overvalued / unknown |
confidence |
high / medium / low by data completeness |
key_drivers, key_risks |
Factual, model-derived points |
disclaimers |
Always present — this is not investment advice |
Failure modes
- No fundamentals / unknown ticker → the provider raises; the skill reports that it cannot assess the name rather than fabricating figures.
- Missing FCF or price → the DCF is skipped and the verdict is
unknown; quality/multiples are still returned where computable. Guessing is never substituted for missing data.
Registry metadata
- Registered as:
projects/{project}/locations/{location}/skills/private-equity-research - Skill revision: 0.1.0 (draft — not yet registered)
- Approval scope:
read:external_market_data,read:fundamentals