| name | ai4finance-foundation-fingpt |
| description | Build FinGPT-inspired financial language intelligence datasets and evidence-backed sentiment or forecasting briefs. Use for financial news curation, sentiment analysis, RAG context assembly, event labeling, and market-reaction studies that must obtain all live financial data through QVeris. |
FinGPT-Inspired Intelligence with QVeris
Use the upstream FinGPT project as a methodology reference. Do not imply that this skill bundles, trains, or hosts a FinGPT model.
Data boundary
Acquire every external financial datum through QVeris. Use native QVeris tools when available; otherwise use the official QVeris HTTP or CLI flow. Require QVERIS_API_KEY for real calls.
Use Discover → Inspect → Call for each capability. Write discovery queries as English API capability descriptions, select tools by coverage and reliability, inspect their parameters, and only then call them. Do not call provider APIs directly.
Use the capability plan in qveris.manifest.json. Treat social data and structured filings as optional enrichments; never fill missing records with invented text or labels.
Workflow
- Normalize the requested ticker or universe, market, language, and UTC date range.
- Discover and call tagged financial news, text-sentiment, adjusted market-history, and financial-ratio tools.
- Record
search_id, tool_id, provider, retrieval time, market timezone, and original event timestamp for every batch.
- Deduplicate news by canonical URL, headline similarity, and event time. Keep conflicting reports as separate evidence.
- Align each document only with market and fundamental data that was public at that timestamp. Prevent future leakage.
- Produce one or more local outputs:
- a normalized financial-language dataset;
- a source-grounded sentiment brief;
- RAG context grouped by issuer and event;
- forward-return labels for offline forecasting research.
- Separate provider sentiment from locally inferred sentiment. Explain aggregation and confidence rules.
- Report coverage gaps, stale records, duplicate rate, class balance, and any missing market reaction window.
Output contract
Return:
- request scope and as-of time;
- QVeris capability and tool trace;
- normalized evidence records with timestamps and URLs when returned;
- sentiment or event labels with provenance;
- market-reaction features calculated locally;
- missing-data and leakage warnings;
- a clear statement that results are research inputs, not investment advice.
Validation
From this skill directory, run:
node scripts/qveris_candidate_runner.mjs validate .
node --test tests/smoke.test.mjs
node scripts/qveris_candidate_runner.mjs preflight .
QVERIS_API_KEY=... node scripts/qveris_candidate_runner.mjs live .
