Regulatory data sources

China Banking Penalty Data: PBOC Administrative Penalty Notices and Court Records

China banking penalty data starts with the regulator's own publication series, and a court judgment is a different evidence object. The People's Bank of China publishes administrative penalty decisions in a structured disclosure section: each entry carries a decision number, the party, the violation type, the penalty content, the deciding authority and the decision date. A legal AI product that wants to answer "what has this institution been penalised for" needs that register, not a case-law corpus interpreted as if it were one.

This guide identifies the official starting points, shows what a real publication looks like, and proposes a collection and acceptance design for overseas data teams. It does not offer a national penalty count, a risk score or a legal opinion on any institution. The source check is dated October 20, 2026. A single worked example illustrates document structure; it is not a statement about enforcement intensity or about any market participant.

1. Decide which question the dataset has to answer

Three requests sound similar and need different evidence. "Show me every published penalty imposed on this bank in the last two years" needs the regulator's publication series with complete date handling. "Explain why customer due-diligence failures are penalised and how the amounts differ" needs violation typing plus the underlying rule text. "Tell me whether this counterparty is currently subject to a restriction" needs current official status, checked today, which a historical publication list cannot supply on its own.

Write the target output before collecting a single row. A dataset scoped to "published administrative penalty decisions, by regulator, by decision date" is bounded and testable. A dataset described as "all punishment a financial institution has received" quietly assumes criminal judgments, supervisory measures, internal disciplinary actions and unpublished warnings, none of which the publication series contains. The gap between those two descriptions is where credibility is lost first.

Our administrative penalty source guide sets out the general distinction between a penalty publication and a judicial document. Banking adds a specific complication: several authorities publish comparable material in different formats, and the institutional map changed in 2023.

2. Where the official publications actually sit

The People's Bank of China publishes administrative enforcement information through its government-information disclosure section. Within it, the administrative enforcement information page lists administrative review catalogues, approval disclosures, administrative penalty disclosures and inspection item lists. The administrative penalty disclosure page is the relevant series for penalty decisions, and it is paginated by publication rather than by decision.

The National Financial Regulatory Administration (NFRA) is the second anchor. It was established in 2023, absorbing the former China Banking and Insurance Regulatory Commission, and it publishes penalty information for banks, insurers and related institutions through its own disclosure sections. Treat PBOC and NFRA as two authorities with two numbering systems and two publication rhythms, not as one merged feed. Securities enforcement sits with the China Securities Regulatory Commission and foreign-exchange enforcement with the State Administration of Foreign Exchange; our CSRC guide and SAFE guide cover those tracks separately because merging them produces double counting.

3. A real publication: penalty decisions 68 to 70 of 2026

A checked example is the PBOC page Yin Fa Jue Zi [2026] Nos. 68-70, published on 16 October 2026. The page tabulates three decisions issued by the People's Bank of China on 3 June 2026 against one institution and two natural persons attached to it. The institutional row states that the bank failed to perform customer identity verification obligations and failed to submit large-value and suspicious transaction reports as required, and records a fine of RMB 1.06 million. The two individual rows state that the named officers are held responsible for those same failures, with fines of RMB 55,000 and RMB 20,000.

Four structural facts follow from that one page. First, the publication date and the decision date are not the same clock: this page was published in October for decisions dated four months earlier. Second, a single publication identifier can cover several decisions against several parties. Third, natural persons appear alongside the institution, with a parent-child relationship expressed in words rather than in a field. Fourth, the violation type is expressed as a plain-language list, not a controlled code, and the same wording recurs across institutions.

What the page does not establish is equally important. It does not state whether the institution appealed, whether any related civil litigation followed, whether the underlying customer accounts were remediated, or whether other publications cover the same institution on other dates. Reading those conclusions into a single notice is the most common extraction failure in this domain.

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

The disclosure table carries a fixed set of columns, which is a gift for a data team because it makes the extraction target explicit. It also makes the modelling decisions visible early.

Keep the original Chinese text and any English label side by side. An English rendering should describe scope rather than replace the source, especially for violation descriptions that carry legal elements. Store a text span or row locator with every extracted field so a reviewer can verify a value without re-reading the whole publication, and version the parser so a schema change is distinguishable from a source correction.

5. The judicial track is a separate pool, not a substitute

For a court-data pilot, start from Chinese retrieval terms and treat every count as a candidate pool rather than a validated filter. In our own corpus of 160,284,678 records (benchmark entries excluded), the cause of action 金融借款合同纠纷 alone returns 4,569,943 records covering 3,709,261 distinct case numbers. A wider union of six financial causes — financial loan contracts, credit card, financial non-performing debt recovery, insurance, property insurance contracts and life insurance contracts — returns 6,883,844 records covering 5,584,471 distinct case numbers. Within that union, phrase matching finds 中国人民银行 in 951,474 records, 金融监督管理总局 in 1,699, 银保监会 in 15,355 and 行政处罚 in 2,683.

The decisive number is the smallest one. The PBOC penalty decision prefix 银罚决字 appears in one record inside that financial-cause pool, and in 42 records across the whole corpus. Administrative penalty decisions are published by the regulator and, as a body of text, are almost entirely absent from the judicial corpus. A team that tries to reconstruct a penalty register from court data will produce a small, unrepresentative sample and will not know it.

Two method notes matter here. Phrase behaviour depends on the analyzer: a phrase query against the wrong analyzer can return a false zero, which is why our counts take the union of two analyzer modes rather than a single match. And a co-occurrence figure is not a finding about outcomes — a judgment mentioning 中国人民银行 may concern a payment dispute, an account freeze or a supervisory document produced as evidence. Review a sample and describe the document role before publishing any rate. The analyzer mismatch note and the full-text coverage guide explain why field presence is not the same as readable reasoning.

6. Natural persons, partial names and a three-year window

Banking penalty publications routinely include individuals: officers held responsible for the institution's failure. Names are partially anonymised in the published table, and the same person may appear across several decisions. Two consequences follow. First, a person-level record needs its own entity type, linked to the institution, and it must not inherit the institution's violation as if the person had been separately charged in court. Second, personal data handling deserves an explicit policy: what is stored, how long it is retained, how it is displayed to end users and whether it may be used for training. Our licensing guide treats these as rights questions to be confirmed, not as defaults.

The publication period column is the second half of the same problem. A disclosure that is to be displayed for three years is scheduled to leave the register on a date that is often not restated on the page itself. If your product answers "is this institution currently penalised", it is answering a status question, and a status question needs a re-observation cadence plus a recorded last-checked date. Absence after a window expires is not evidence that the penalty was withdrawn.

7. Match and split identifiers deliberately

Use separate fields for the regulator's decision number, the publication page, a court case number and any internal record id. The decision number contains a year and a sequence, and a single publication can list a contiguous range, so a one-publication-one-decision assumption breaks immediately. Retain the original form of each identifier and store a normalised copy for matching; full-width brackets, spacing and Chinese numerals in ranges all affect string comparison.

Entity resolution is the harder half. Bank names appear with branch suffixes, the same group may be penalised at head office and at regional level on different dates, and 中国人民银行 as a deciding authority is not the same object as a bank whose name contains a similar phrase. Prefer explicit identifiers and reviewed rules; keep name-only matches as candidate links with a reason and a review status, and never let candidate links feed an automated risk score.

8. A pilot that can fail for the right reasons

Build the acceptance set before the pipeline. Include a publication containing several decisions, an institution row paired with two natural-person rows, a branch-level decision, a decision whose publication date differs from the decision date, and a court record that merely mentions the regulator. A pipeline that quietly merges the first four, or that treats the fifth as a penalty, should fail acceptance.

Measure retrieval and extraction separately, and keep them honest. A decision that exists but is never retrieved is a collection failure; a retrieved decision with an invented penalty amount is an extraction failure. Report the denominator, the labelling method and the reviewed error rate for each. A convenience sample cannot support a national enforcement rate, and a small sample must be described as such.

Agree in advance what stops the pilot: unclassified publication formats, missing decision identifiers, person-level rows without review, or a corpus that turns out to contain almost none of the target documents. For the wider handover process, our delivery acceptance checks apply directly to the judicial component.

9. Commercial scope and language are separate questions

Public accessibility and permitted commercial use are different tests. Record where each item came from, and obtain review for storage, display, redistribution, translation and model training rather than inferring a grant from a successful download. Regulatory publications and a licensed judicial collection can carry different permissions, and person-level records deserve the strictest review of the three.

SinoVerdict's 160M+ figure describes court-data records; it is not a banking penalty register size and implies nothing about administrative coverage. Request a scoped trial API key and corpus coverage report for the judicial component, and discuss bulk, REST and MCP delivery against an agreed scope through our delivery overview.

Test language behaviour on the endpoint you will actually call. Chinese source text, an English user interface, English query support, generated English summaries and a fully translated corpus are five different capabilities. The English-access capability matrix sets out those boundaries. A banking pilot should run its own known Chinese and English queries rather than borrowing a general statement about coverage.

Frequently asked questions