Master Prompts for Investment Analysis: Structured Output for Portfolio Reports and Market Summaries

Investment analysis fails when facts, calculations, assumptions, and opinions blur together.

A master prompt can impose that separation. It can organize approved portfolio data, summarize supplied research, compare scenarios, and produce a draft report with source references. It cannot verify a market claim by repeating it, predict returns, or decide what a person should buy or sell.

Establish an As-Of Boundary

Every analysis input needs an as-of date and source identity. Prices, holdings, ratings, estimates, corporate actions, and economic data change at different times.

{
  "dataset_id": "authorized-source-id",
  "as_of": "YYYY-MM-DDTHH:MM:SSZ",
  "source_type": "approved-enum",
  "coverage": [],
  "currency": "approved-code",
  "verification_status": "verified"
}

The application should reject stale or incompatible datasets under rules defined by the investment team. The model should not fill a gap from memory.

Separate Output Classes

Require every material statement to be a verified fact from an identified source, a deterministic calculation with a method ID, a supplied assumption, a scenario result, an analyst interpretation, or an unresolved question.

This classification makes review faster and prevents a polished narrative from hiding an unsupported leap.

Portfolio Reports

Code should calculate allocation, concentration, performance, contribution, exposure, and benchmark comparisons from approved data. The model can explain verified outputs in language matched to the intended audience.

Do not ask the model to recreate performance calculations from prose. Fees, flows, time weighting, benchmarks, and corporate actions require controlled methods.

Market Summaries

Use licensed or otherwise authorized sources, preserve dates and identifiers, and require direct support for each claim. A summary should show conflicting evidence and missing coverage rather than force a confident conclusion.

The SEC, FINRA, and NASAA warn that AI-generated investment information may be faulty and recommend confirming underlying sources. See the official Investor.gov alert on AI and investment fraud.

Scenario Analysis

Scenarios are conditional, not forecasts. Supply assumptions, method, horizon, and limitations. Let deterministic services calculate scenario values where possible, then have the model structure the explanation.

{
  "scenario_id": "approved-scenario",
  "assumptions": [],
  "calculation_version": "method-version",
  "results": [],
  "limitations": [],
  "probability_assigned": false
}

Do not let the model assign probabilities unless an authorized methodology and responsible reviewer explicitly provide them.

Control Advice and Distribution

The output should state its audience, permitted use, review status, and whether it is research, internal analysis, marketing, or client communication. Those categories can trigger different requirements.

Code should block distribution outside the approved channel. A prompt cannot establish investor suitability, licensing, supervision, records retention, or required disclosures.

Respect Data Rights

Market, research, index, estimate, and portfolio data may carry contractual limits on use, storage, transformation, and redistribution. Authorization to view data in one product does not automatically authorize sending it to a model or reproducing it in a client document.

The application should associate each dataset with permitted users, purposes, destinations, and retention rules. Retrieval must filter by those rights before generation, and logs should avoid duplicating licensed or confidential content.

Show What the Analysis Did Not Cover

Coverage fields are as important as findings. Record unavailable securities, excluded asset classes, missing history, stale estimates, unsupported currencies, and sources that could not be retrieved.

Require the final draft to surface material exclusions near the relevant conclusion. A separate limitations appendix is useful, but it should not hide a gap that changes the meaning of the summary.

When analysts disagree, preserve the competing interpretations and their supporting source IDs. The model may organize the disagreement; it should not average distinct judgments into false consensus.

Human Review Is Substantive

Analysts verify sources, assumptions, material omissions, interpretation, and uncertainty. Portfolio and compliance professionals determine suitability, permitted use, disclosures, and distribution. Developers enforce entitlements, dataset identity, calculations, schema, retention, and release controls.

Read Master Prompts for Finance for the data boundary and Financial Document Automation for controlled report assembly.

Test the Analysis Contract

Test stale prices, a missing holding, mixed currencies, benchmark changes, split-adjustment errors, conflicting sources, an unavailable source, an unsupported recommendation, prompt injection in retrieved research, and an unauthorized account.

The system should refuse to turn uncertainty into a recommendation. Structured output is useful because it exposes what is known and what still needs a qualified person.

Explore analysis workflow contracts in the CyWire marketplace.

This article is technical information, not investment, accounting, tax, or legal advice. It is not a recommendation to buy or sell any security.