Data sources
EDINET: Japan's official corporate disclosure system
All structured filing data comes from EDINET, the Financial Services Agency's electronic disclosure system. EDINET is the regulatory source of truth for every listed company in Japan. It carries securities reports, large shareholding filings, tender offers, extraordinary reports, treasury stock filings, and the other disclosures companies and their owners are required to file.
Japan Finsight ingests new filings from EDINET on an ongoing basis. How far back the archive holds each document type, and how deeply each is extracted, is published per type on the coverage page. The short version: annual reports from January 2018, most other types from 2020 or 2021, and the types EDINET only serves for a short statutory window from the point ingestion of that type began. Every EDINET-sourced value carries its filing identifier (edinet_document_id), so you can trace it to the source on EDINET's portal. Where a value is parsed from filing text, anything that cannot be verified against the filing is stored as NULL, never guessed; several extractions also keep the filing's verbatim figure beside the parsed number.
Filings are in Japanese. Typed fields, company identities, and extracted signals are returned in English; full narrative sections are returned as filed, in Japanese, for your AI to read or translate.
Parser layer
The parsing that turns EDINET's machine-readable filing data (XBRL) into typed, structured fields is handled by edinet-tools, an MIT-licensed Python library, developed and maintained as a separate open-source project. Typed financial figures come from the filings' own machine-readable data, not from an AI reading PDFs. pip install edinet-tools gives you the same parsers Japan Finsight uses internally, by design.
As of edinet-tools 0.8.0, parser output is verified against a corpus of 176,460 real filings. One check: the equity-ratio identity agrees with the filer's own stated ratio in 30,338 of 30,340 J-GAAP filings, 2,302 of 2,304 IFRS filings, and 77 of 77 US GAAP filings.
LLM-extracted fields
Several Japan Finsight tools add LLM-extracted fields on top of the deterministic XBRL parsing:
get_capital_policy: target payout ratios and dividend skip reasons extracted from annual report text. The company's stated reason, in its own terms, with the source passage preserved.get_related_party_transactions: counterparties, transaction types, and amounts structured from filing text. Every parsed amount keeps the filing's verbatim figure beside it; unparseable amounts stay null rather than guessed.get_holding_actions: activist-intent annotations on large shareholding filings (Doc 350). Known activist filers are flagged from a maintained registry, and each filing's stated change proposals and desired outcome are extracted from the filing text.get_shareholder_register: the top-10 shareholder table in each annual report. Holder names, share counts, and ownership percentages are extracted from the filing text by a language model, then every value is deterministically verified against that text. A value that fails verification is stored as NULL, never guessed. Coverage runs back to fiscal 2018 across the listed universe.
Not LLM: get_material_events classifies extraordinary reports (Doc 180) by deterministic rule over the FSA ordinance citation in each filing. No language model is involved.
Extraction policy: we extract facts the company stated in the filing. We do not derive judgments about the company. Empty results mean "the company didn't state this," not "the company doesn't do this." Every extracted row carries its filing identifier, and parsed amounts either keep the verbatim source figure or are verified against the filing text, with unverifiable values stored as NULL.
Companion projects
edinet-tools: the parser library (PyPI).kabu-agent: an example AI agent built on EDINET data (open source).
What we don't source from
- Bloomberg / FactSet / Refinitiv: none. All structured data is sourced directly from FSA EDINET.
- Live market data: none. Japan Finsight is the disclosure record, not a terminal. No live prices, no estimates, no news feed.
- TDNet: earnings releases (決算短信, tanshin) are filed through TSE, not EDINET, and are not ingested. If tanshin coverage matters to your workflow, email data@japanfinsight.com; extraction priority is driven by subscriber demand.