Open datasets and methodological notes

Datasets: evidence-led reference

This page identifies official patent-data collections and the documentation needed before drawing statistical conclusions. It does not announce a GIPResearch.org dataset that has not actually been created or distributed. An open-source index, a free graphical search portal, a downloadable database and a licensed application programming interface are different forms of access. The EPO documents PATSTAT as a research-oriented structured database assembled from bibliographic, legal-event, register and citation sources. Researchers must examine coverage and data terms before claiming a global trend or redistributing extracted records.

Data sets have different units of observation

Patent applications, publication documents, patent families, priority filings and granted rights are not interchangeable counts. One invention may generate multiple publications, related applications or national filings, and an analytical table that counts every document as a separate invention will answer a different question from a family-level chart. Define whether each row is a filing, publication, application, patent family, cited document or legal event. Record any harmonization of applicant names and cross-border office codes. If the research changes its unit after downloading data, retain both definitions so a later reader can understand the difference in totals.

PATSTAT should be treated as a snapshot

The EPO describes PATSTAT Global and PATSTAT EP Register as structured resources for patent statistics, not simply as a real-time patent search screen. PATSTAT combines source streams, and its documentation emphasizes coverage, codes and release editions. A chart based on one PATSTAT release should cite that release, date of extract, table fields and selection rule; later releases can add backfile corrections. Do not compare results from different snapshots without reconciliation. The EPO's Technology Intelligence Platform and other access modes have separate capabilities, so a public page should describe the current official options without promising identical access to every user.

Primary reference: EPO PATSTAT documentation and coverage. Interpret only within the source’s published jurisdiction and scope.

Legal-event data are not a final legal opinion

INPADOC collects legal events from multiple authorities and is useful for event-pattern research. However, a recorded event may need interpretation under the issuing jurisdiction's law, and absence of a recorded event does not prove a right is in force. Record event code, source authority, event date, database ingestion date when available and relevant limitation. For any high-consequence clearance decision, the national or regional register and suitably qualified advice may be required. A statistical dataset can count recorded event codes, but it should not label all records as enforced or expired without carefully validated legal rules.

Access rights and reproducible retrieval

Before collecting records, identify whether the interface is intended for human search, programmatic access or bulk distribution. The EPO directs automated queries away from ordinary Espacenet pages toward designated REST services. OPS requires registration and is governed by data terms and fair-use limits. A public methodology may describe retrieval methods without redistributing a licensed database. If a CSV is published, include its author, origin, aggregation rules, license status, machine-readable schema, update date, and whether fields were transformed or suppressed. Do not characterize a third-party collection as an original open dataset merely because its metadata can be viewed.

Primary reference: EPO INPADOC legal event information. Interpret only within the source’s published jurisdiction and scope.

Normalize identifiers without losing provenance

Retain original publication numbers, country codes, kind codes and publication dates. Then derive standardized identifiers in separate fields; do not overwrite the source values. Applicant names may require disambiguation, and one entity can appear under several spellings or subsidiaries. Technology categories should be tagged with the scheme edition and query rule. Family consolidation needs an explicit definition because simple and extended families represent different relationships. A reproducible pipeline publishes both a transformation map and validation tests, such as detecting duplicate keys and counting missing dates. If data corrections are manual, record who made them and why.

Describe missingness, exclusions and quality

A convincing dataset card reports what period and territories are present, whether full text or only bibliographic fields are available, how citations are collected and which records are excluded. It should distinguish unavailable data from zero activity. For instance, an office with poor backfile coverage can appear to have fewer patent publications than another even where underlying innovation activity is not lower. State if a dataset contains machine translations and whether those were reviewed for technical terms. Avoid precise claims about market share or patent ownership unless the population and matching decisions are documented and independently traceable.

Publish data documentation before charts

A useful original contribution is a blank data dictionary showing mandatory fields: source URI, record ID, jurisdiction, document type, query, retrieval timestamp, scheme edition, transformation function, deduplication status and confidence note. A second deliverable is a coverage ledger listing each authority's inclusion period and unknowns. Only after filling those with observed facts should the publisher calculate counts or charts. No downloaded table or patent landscape should be inferred from the presence of this documentation page. Link readers toward the /sources/ directory for selecting a portal and /methodology/ for choosing an analytical question.

Primary reference: EPO Open Patent Services. Interpret only within the source’s published jurisdiction and scope.

Illustrative research scenario: applying the distinctions

A hypothetical patent-activity table contains 1,000 raw publication records collected from a specified database snapshot. Before plotting them, inspect duplicate publication IDs, repeated family members, missing dates, and unmatched applicant aliases. Define whether the study will count documents, applications or families. Keep a column recording which family rule was applied. Separately calculate the portion of records with missing publication dates and any jurisdictions absent from the source's stated coverage; do not convert missing records into zeros. If classifying artificial intelligence, record the official CPC or IPC version used and test a sample of classified hits for relevance. Keep the unmodified raw identifiers separate from transformed fields, and document the code or manual rules used in cleaning. Because these numbers are illustrative, not collected research data, do not publish them as empirical findings. The useful public output is the reproducible data dictionary, validation checklist, and explanation of why alternative counting rules could yield different trends.

Practical decision and evidence matrix

What this reference establishes — and what it does not

The material above is a source-bounded reference concerning datasets. It distinguishes official documentation from editorial method and provides practical decision boundaries; it does not claim an exhaustive global search, an empirical dataset that was not collected, a legal opinion, or certified implementation of publisher policy. Readers seeking to apply a source to a current filing, claim or professional obligation should inspect the applicable original authority, record its effective date, and obtain the necessary qualified advice. A page is updated substantively only when its evidence, method or answer materially changes.