Regulatory data sources

China Insurance Regulatory Penalties: NFRA, CBIRC and CIRC Notices vs Court Records

China insurance penalty data lives in the regulator's own publication series. A court judgment that mentions the regulator is a different evidence object, and the difference matters more in insurance than in almost any other Chinese regulatory domain. Three authorities have run insurance supervision since 1998, each with its own name, its own document numbering and its own disclosure page. Two of those three official domains no longer resolve. A team that treats "insurance enforcement in China" as one feed will build a dataset whose coverage claim cannot survive its first audit.

This guide maps the official starting points as of October 23, 2026, sets out the three-regulator naming lifecycle with measured corpus evidence, and proposes a data model and acceptance test list for overseas legal AI teams. It is informational only and is not legal advice. Every court-side number below is an aggregate query result over our own corpus, not an enforcement statistic: a corpus record is a document, not a penalty decision, not a distinct case and not a compliance finding.

1. Three requests that sound alike and need different evidence

"How many times was this insurer fined last year?" needs the regulator's publication series with exact decision dates and party resolution. "Which conduct types does the supervisor penalise, and how do the amounts cluster?" needs violation typing plus the rule text behind each type. "Is this counterparty currently subject to a supervisory measure?" needs today's official status, which no historical publication list can supply.

There is a fourth request that arrives disguised as data work and is really due diligence: "show me the court record that proves the penalty happened." A judgment can quote a penalty notice, restate the regulator's findings, or review the decision on appeal. It can also mention the regulator for reasons entirely unrelated to enforcement, and it can be a 2009 contract dispute in which the defendant's supervisor happens to be named. Treating that pool as an enforcement register inverts the direction of the evidence.

Write the target answer before collecting a first row, then decide which of the two tracks owns it. Our administrative penalty source guide sets out the general split; our banking penalty guide covers the parallel case where PBOC publishes and NFRA supervises. Insurance adds a complication neither of those has: the supervising authority itself was renamed twice, and the naming is observable inside the judicial corpus.

2. Where the official publications actually sit

The National Financial Regulatory Administration (国家金融监督管理总局, NFRA) is the current anchor. Established in 2023, it absorbed the former China Banking and Insurance Regulatory Commission and publishes administrative penalty information for banks, insurers and related institutions through its own government-information disclosure sections. On October 23, 2026 the portal homepage returned HTTP 200. The administrative penalty listing is a client-side page: a plain HTTP fetch of the list URL returns a small script shell rather than the rows, so extraction requires rendering or an approved interface rather than a static GET.

The two predecessor domains are gone. On the same date, www.cbirc.gov.cn (the China Banking and Insurance Regulatory Commission) and www.circ.gov.cn (the China Insurance Regulatory Commission) both failed DNS resolution from our network — no HTTP status, no redirect, no parked index. That is a practical sourcing fact, not a technicality: an overseas team building a historical insurance enforcement dataset cannot point at the old regulator domains, and any pipeline that still references them is silently collecting zero rows while reporting success. Historical publications have to be sourced from the successor portal, from archived snapshots, or from a licensed feed, and whichever route is chosen has to be recorded per source.

3. Three regulators, three numbering systems

Penalty documents carry a prefix that identifies the issuing era. The China Insurance Regulatory Commission published under a 保监罚 (Bao Jian Fa) prefix; the China Banking and Insurance Regulatory Commission published under 银保监罚决字 (Yin Bao Jian Fa Jue Zi); the NFRA publishes under 金罚决字 (Jin Fa Jue Zi). These prefixes are not cosmetic. They are the cheapest available way to date a document by its own text, independent of whatever the harvesting pipeline believes about the publication date.

We measured how far each prefix actually travels into the judicial corpus. Out of the full 160,291,678-row collection, the strings appear as body phrases in these numbers of court documents:

The absolute counts are small and that is the finding: judicial quotation of an insurance penalty decision number is a thin slice, on the order of low hundreds of documents across the entire corpus. A product that offers "find the court case behind any insurance penalty" is therefore promising something the corpus does not contain at scale. A product that offers "trace how courts reason about insurance penalties when they appear" is offering something measurable, provided the coverage report states the slice size rather than implying it is the whole.

4. The naming lifecycle is visible inside the corpus

Regulator names behave like a versioned identifier, and the abbreviations are not interchangeable with the formal names. Measured as body phrases over the same collection:

Two conclusions follow. First, a query that lists only formal names misses the abbreviation-heavy documents: the two formal predecessor names together are exceeded by the single abbreviation 银保监会 on its own, so a name list that omits abbreviations loses a large share of hits. Second, the current authority has barely entered older judgments — 金融监督管理总局 appears in 3,083 documents against 32,019 for the abbreviation of its 2018 predecessor — which is what a 2023 rename should look like in a corpus weighted toward earlier filings, and which will drift upward as post-2023 decisions are published and cited.

The year profile of the 银保监会 phrase makes the same point quantitatively: 740 documents in 2019, 2,456 in 2020, 6,222 in 2021, 12,722 in 2022, then 3,061 in 2023 and 5,071 in 2024. The 2023 drop coincides with the institutional transition; it does not mean supervisory activity stopped, because the successor publishes under a different name and a different prefix.

5. The insurance disputes pool is not an enforcement register

The largest insurance-related object in a Chinese case-law corpus is contract litigation, not regulation. Combining twenty civil insurance causes — property, liability, guarantee, life, health, personal-accident, subrogation, reinsurance, marine, agency, broker and others, excluding the social-insurance and work-injury benefit causes that belong to the labour track — the pool contains 849,971 records, of which 678,888 carry a case number ending in 号. The criminal track is separate and much smaller: combining 保险诈骗罪 and 保险诈骗 gives 3,037 records and 2,770 numbered case numbers, and those documents concern fraud by insured parties, not supervision of insurers.

Read that table as a corpus property, not as litigation volume. Three artefacts are visible and each one is a trap for a downstream model. The 2021–2022 bulge is far larger than the surrounding trend, and the records-to-case-number gap widens at the same time (102,833 records against 77,845 numbered numbers in 2021; 243,181 against 155,530 in 2022), which is what batch arrival and multi-document case clusters look like rather than a jump in disputes. The 2019–2020 trough sits exactly where nationwide publication volumes collapse, and a chart that normalises by nothing will read it as a decline in insurance litigation. And the year field itself is not clean: across the whole collection, year keys below 1990 or above 2025 account for 1,469 rows, including keys such as 200, 201 and 1013, so any time series needs an explicit validity rule rather than a raw bucket read. See our coverage and recency analysis and field census for the field-level view.

Within that pool, the phrase 行政处罚 (administrative penalty) appears in 1,901 records (1,743 with numbered case numbers). That is 0.2% of the pool, and it is the honest size of the overlap between insurance contract litigation and anything described as a penalty. Meanwhile 保险法 (Insurance Law) appears in 413,679 records and 投保人 (policyholder) in 267,520, which is what one expects from contract disputes and shows why keyword presence is not a proxy for subject matter. Field coverage is uneven too: 102,112 rows in this pool carry an empty province string, and the procedure field mixes Chinese labels with numeric codes such as 1 and 0301 alongside 民事一审 and 执行实施 — see the procedure code crosswalk before you group by stage.

6. Where the court track does help

Administrative litigation is the right pool when the question is judicial review of supervisory action. Combining eight administrative causes — 行政处罚, 行政复议, 行政强制, 行政确认, 行政执行, 行政撤销, 行政处理, 行政赔偿 — gives 770,960 records, dominated by 行政处罚 (511,708) and 行政复议 (101,455). Intersecting that pool with the phrase 保险监督管理委员会 yields 399 records / 395 numbered case numbers, split across 行政复议 (247), 行政处罚 (134), 行政处理 (8), 行政撤销 (6) and 行政确认 (3). A separate statutory phrase is worth tracking on its own: 保险监督管理机构, the term the Insurance Law uses for the supervising body, appears in 4,752 documents and is not interchangeable with either regulator's name.

Note the gap years. In the intersection above, the year buckets run 2008 through 2025 but contain no records for 2019 and 2020, while 2016 holds 94 and 2024 holds 68. The most plausible explanation is publication coverage rather than an absence of supervisory litigation, and the correct response is to disclose the gap rather than interpolate through it. Our placeholder-body audit covers the related failure mode where a document exists but its usable text does not.

7. Let the register's columns define the data model

Insurance penalty publications follow a tabular disclosure pattern similar to the banking series: a party, a document number, a violation description, a penalty content, a deciding authority and a decision date, with natural persons often published alongside the institution they were responsible within. The columns are the data model; the modelling decisions are what break projects.

8. Acceptance tests before you sign

The publication series and the judicial corpus fail in different ways, so the acceptance test should ask different questions of each. This list is deliberately adversarial: each item is designed so that a partial delivery fails it.

Run the same test twice, six weeks apart, and compare the deltas. Chinese publication practice changes with the institution that owns it, so a single acceptance run proves the pipeline works on one snapshot; only a second run shows whether the source is stable. Our pilot scope and exit criteria note and vendor coverage verification note set out a wider version of this process, and the delivery-side checks live in the acceptance checks guide.

9. Commercial scope and language are separate questions

Two decisions are often bundled and should not be. The commercial decision is what the licence permits: redistribution, derivative works, model training, internal research and external display are five different permissions, and a public disclosure page does not settle any of them. The language decision is how much of the Chinese source is carried through: extracted fields only, extracted fields plus verbatim snippets, or extracted fields plus full translated text. A scope that promises English-language penalty reasoning implies translation of free-text violation descriptions, which is a different quality problem from translating structured fields. Our licensing guide, termination-survival note and English-native research note cover those tracks; the insurance dispute corpus itself is described in the insurance dispute dataset overview.

Frequently asked questions