Home → Insights → Securities enforcement data

China Securities Enforcement Data: CSRC Penalties and Court Records

China securities enforcement data starts with the regulator's publications. Court records answer a related question: what happened when conduct, a regulatory action or an investor claim reached the courts? A useful research product preserves both tracks and the evidence connecting them. It does not turn every mention of the securities regulator into a penalty, or every penalty into a completed litigation history.

This guide is for data teams building China coverage for legal research, regulatory intelligence and enterprise search. It identifies official starting points, shows fresh judicial candidate counts, and proposes an acceptance plan for combining sources. The document model below is a recommended research design, not a claim that SinoVerdict supplies a complete CSRC register. For the wider litigation topic, see our securities and capital-markets court-data guide.

1. Start with the official CSRC source categories

The China Securities Regulatory Commission website presents administrative penalties, market-entry bans, administrative reconsideration and supervisory measures as separate publication categories. Preserve that distinction in your source inventory. A shared regulator name does not make the document classes interchangeable. Regional-office publications also need their issuing office recorded, rather than being flattened into a generic central-CSRC label.

The official document search interface exposes controls including document name, document type and the name of the sanctioned subject. These are useful discovery fields. Their presence does not establish an export facility, unrestricted automation, historical completeness or permission for every commercial reuse. Assess those requirements separately before making a collection commitment.

2. Keep published totals and collected records separate

In its 2024 enforcement overview, published on May 15, 2025, CSRC reported 739 securities and futures enforcement cases handled and 592 penalty decisions. It also reported 1,327 responsible-subject person/entity instances, 118 people subject to market-entry bans, and 178 suspected-crime cases or leads referred to public-security authorities. These are different measures within that official reporting scope.

Use those figures as source-attributed context. Do not add them into one enforcement total, treat subject instances as unique entities, or assume the handled cases and decisions form matched cohorts. An official annual total also does not prove that a website search currently exposes the same number of accessible documents. Your collection count needs its own retrieval date, document-class definition and duplicate policy. A difference is a reconciliation question, not immediate evidence that either source is wrong.

For a procurement pilot, define a small comparison set that can actually be checked. Specify the offices, decision years and instrument categories. Ask the supplier to return matching publications, missing items and uncertain matches against that set. A pass on that exercise supports the tested slice. It does not authorize extrapolation to every regional office, historical year or related court proceeding.

3. Inspect an instrument before designing extraction fields

The official CSRC Decision [2025] No. 49 illustrates why extraction must retain more than a topic tag. The document discusses market manipulation, separately states confiscation and fines, allocates amounts between two subjects, and describes five-year market-entry restrictions. It also contains review and litigation language. Those components should remain distinguishable in a structured representation.

That example supports a schema design, not a statement about the subjects' current status. This article has not independently established subsequent review outcomes, execution, payment or later changes. An original decision is evidence of what that instrument states. A current-status field requires additional dated evidence and a defined update process; it cannot be populated merely because the original page remains online.

Retain raw numeric expressions alongside parsed values. Store currency, unit, monetary category, liable subject and whether a figure is an individual allocation or a combined amount. A document-level total repeated in each subject row can multiply the apparent sanctions. Where the source wording does not resolve the allocation, mark it unresolved and keep the supporting passage for review.

4. What a court corpus can contribute

Judicial records can contain litigation about regulatory actions, civil investor claims and criminal proceedings involving related conduct. Start by identifying the procedural object of each document. A regulator's name may appear in a factual history, a quoted rule, an argument or a reference to an earlier instrument. Retrieval alone does not establish that the court reviewed a CSRC penalty, or that it endorsed every fact described by a party.

On October 16, 2026, we ran three reproducible queries against the judicial corpus. Each body-phrase query combines ik_smart and ik_max_word with Boolean OR and excludes records labeled dataset=bench. The accompanying query and count evidence preserves the requests, analyzer output, timestamps and shard results. These are overlapping candidate record pools, not verified enforcement populations.

No semantic precision audit or distinct-case calculation was performed for these pools. Their intersections were not measured, so the counts must not be added. They are not unique defendants, issuers, violations or penalties. Using both analyzers addresses a known retrieval mismatch; it does not prove exhaustive recall. Abbreviations, alternative wording, missing bodies and ingestion boundaries can still exclude relevant records.

The whole-corpus figure of 160M+ describes records, not a count of unique judgments with complete text. The size of that corpus establishes neither complete CSRC coverage nor superiority over another supplier. For a selected slice, inspect actual text, identifiers and metadata. Our full-text coverage guide explains why an existing field can still contain an unusable value.

5. Model documents, subjects and relations separately

A practical pilot can use three linked tables. The document table describes each publication or court record. The subject table stores the names and roles actually observed in that document. The relation table captures a proposed or verified connection between documents, with evidence and review status. This prevents an uncertain name match from silently becoming an asserted legal history.

For documents, request the source track, issuing office or court, original identifier, title, publication date, instrument date where stated, retrieval timestamp and text availability. Preserve identifiers in their original form before adding normalized search keys. A number meaningful within one issuing office or document category need not be globally unique. A court's case number belongs to a different identifier system from an administrative decision number.

For subjects, keep original Chinese names and distinguish the sanctioned person, company, issuer, intermediary, claimant and respondent where the text supports those roles. English translations and aliases should be separate derived fields. Masked names should stay masked. Common names, corporate renaming and incomplete identifiers call for review, not confident automatic merging.

For relations, record the link type, supporting passage, method, reviewer state and observation date. An express decision-number reference is stronger evidence than a shared topic. Even an explicit reference needs context: a judgment might describe the instrument as background rather than decide its validity. Preserve that role instead of assigning every linked document the same outcome.

For example, suppose a pilot returns an administrative instrument, a civil pleading summarized in a judgment, and a later appeal document. Treat these as three document objects until their relationships are checked. The civil claim may concern some of the same conduct while seeking a different remedy. The appeal might belong to that civil proceeding rather than challenge the administrative instrument. A shared company and overlapping dates help find candidates; they do not settle those distinctions.

English search adds another review layer. Keep the original Chinese term, the English label and the evidence passage together. A convenient umbrella label such as enforcement action can help navigation, but the reader should still see the original document class. Test whether a summary preserves who asserted a fact, which authority made a finding, and what the operative part actually says. Mark unavailable passages instead of completing a narrative from related documents.

6. Ask a precise question of each date and amount

A research timeline should distinguish alleged conduct, decision issuance, website publication, judicial events and your retrieval. Leave dates unavailable when the source does not state them. Choosing one date because it is consistently populated can change the research question: a publication-year chart measures publication under your collection rules, not necessarily the timing of underlying conduct or enforcement activity.

The same principle applies to money. A fine, confiscated proceeds, claimed civil compensation, awarded compensation and an amount discussed in the facts are separate quantities. Require a category and a supporting passage before aggregating. An amount asserted by a claimant must not become an award merely because both appear near the word compensation.

Do not turn these candidate pools into success rates, investor recovery rates or firm-risk rankings. Those products require explicit populations, review of outcomes, treatment of multiple proceedings and assessment of missing data. A mixed set of civil, criminal and administrative documents cannot supply a common denominator simply because the same entity name appears across it.

7. Define a coverage pilot that can fail clearly

Start with a buyer-supplied set of known administrative instruments and a separate set of known judicial documents. Test each track against its own requirements. A working search endpoint is only an access check. Coverage acceptance should identify which expected items were returned, what usable fields they contain, and which relationships remain unsupported after review.

Add negative examples: a court document that merely quotes a securities rule, a name collision, an amount mentioned only in a pleading, and an instrument with no confirmed later proceeding. The product should keep those distinctions visible. If every query produces an apparently complete enforcement story, the pilot may be rewarding unsupported inference rather than useful retrieval.

Agree the correction and refresh process before production. Ask how amended publications, withdrawn pages, broken references and newly discovered duplicates are represented. Define whether the delivered object is a fixed research snapshot or an updated service. For legal reuse, personal-data handling and downstream distribution, assess the applicable terms and permissions independently; this page provides research-design information rather than legal advice.

SinoVerdict can be evaluated for the judicial component through a scoped trial API and coverage report. Describe your subject matter, date window, required text, identifier checks and acceptable unknowns. Review the delivery options and delivery acceptance checklist, then request a trial. Keep a complete administrative register and current sanctions-status monitoring as separately specified requirements.

Frequently asked questions