SMVaddadi/propensity-data-apra-ncpd

apra_ncpd — mirror

A convenience mirror of a public dataset, pushed by propensity.data.hf_mirror so that benchmark runs do not depend on a flaky upstream origin.

This is a mirror, not a grant of rights

Copyright and licence remain with the original source named below. Mirroring moves bytes; it does not and cannot change the terms they arrived under. If you intend to redistribute this further, or to use it commercially, read the original terms — not this card, and not the license: field above, which is our best reading and may be wrong.

Provenance

• Source: APRA National Claims and Policies Database, masked Claims and Policy report bundles, December 2024 publication. (c) Australian Prudential Regulation Authority (APRA) 2024, licensed CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) per the publication page's own Disclaimer and Copyright block. APRA DOES NOT ENDORSE this repository or any use of this data -- non-endorsement is a condition of the grant, not a courtesy. APRA accepts no responsibility for accuracy, completeness or currency. Note this is CC BY 4.0 while the sibling apra_life_claims and apra_gi_institution mirrors are CC BY 3.0 AU: APRA's own instruments differ between its workbooks and this publication page, and each was read on its own. WHAT THIS IS: eight cross-tab aggregate tables, 720,006 rows total [measured], covering Australian public and product liability and professional indemnity, reporting years 2004-2024. Largest arm 158,404 x 20; arms are {claims,policy} x {by_occupation,by_state} x {excess-band,limit-band}. Measures are claims reported/finalised and gross payments/incurred (claims arms), and risks-in-force, risks written and gross earned/written premium (policy arms). Covariates include 9 product types, 8 states plus unknown, and 70 occupation/industry groups across TWO coexisting taxonomies (ANZSIC numeric for liability, letter-coded professions for indemnity). WHY IT IS HERE: the only occupation-resolved liability experience table in this collection, and the only APRA product carrying a policy-design covariate (excess and limit-of-indemnity bands). Good for frequency and severity RATES. IT IS NOT CLAIM GRAIN, contrary to what this programme expected of it. The underlying DATABASE is claim-level, but this PUBLICATION is cross-dimensional bucket aggregates, as APRA's own explanatory notes state. Measured proof: the key is unique across all 158,404 rows of the default arm, n_claims_finalised runs 0 to 1,442, and 115,335 of 158,404 rows report ZERO finalised claims. A claim-grain table has neither a count column nor zero-claim rows. SIX MEASURED CAVEATS. (1) 'Masked' removed the INSURER ENTIRELY -- no ABN, no company name, no insurer code in any of the eight files -- so there is NO ENTITY KEY and entity-disjoint cross-validation is impossible on this asset. That is the real cost of the masking, and it is a modelling limitation rather than a privacy one. (2) Masking also withholds cells, and it hits KEY columns not just measures: raw product_type reads as 10 values where there are 9, class_of_business as 3 where there are 2, and four year columns arrive as object dtype mixing ints with ''. That also breaks key uniqueness on three arms, with 91.7% of one arm's duplicate keys carrying a ''. (3) Suppression uses FOUR encodings and the publisher's own footer is wrong about one: the footer says '<5' marks counts too small to publish, but '<5' appears NOWHERE in the claims bundle, so claims-arm small counts are masked as '*' and are indistinguishable from genuine aggregation; and state_jurisdiction suppresses to a BLANK, alone among the dimensions, while separately carrying a publisher 'unknown' category that means something else. (4) THE ARMS DO NOT RECONCILE: gross claims incurred totals 37.10 / 36.85 / 39.47 / 36.79 billion AUD across the four claims arms, a 2.68bn (7.3%) spread on one underlying quantity, because suppression is applied at different aggregation levels. Pick one arm; never union or average them. (5) Band vocabularies differ between the claims and policy arms, so they CANNOT be joined into a loss ratio. (6) gross_claims_incurred is negative on 23.9% of rows (reserve releases) and claim counts are zero on 71.4%. RE-IDENTIFICATION ASSESSED LOW, and the asset is not restricted: a k=5 suppression floor holds exactly -- minimum non-zero count is 5.0 and zero cells report a count in (0,5), in all eight files -- there is no age, sex, name, postcode or date finer than a year, and geography is state/territory (8 values for a continent). APRA has formally determined these publications non-confidential. WHAT THIS IS NOT: not claim-grain records; not an insurer panel (no entity); not motor or personal lines (the nine product types are CON, CYB, D&O, MAL, MAN, PIO, PLO, PRO, PUB -- no motor product exists); and it contains NO adjudication outcome, so nothing records a claim accepted or declined.
• Origin URL: https://www.apra.gov.au/news-and-publications/national-claims-and-policies-database-statistics
• Loader: propensity.data.public_loaders.load_apra_ncpd
• Extract: full pull from origin
• Reported source: apra_ncpd
• Rows pushed: 158,404
• Columns: 20

Licence and redistribution

• Licence as found: cc-by-4.0
• Redistribution verdict: open
• Basis: APRA, CC BY 4.0. Read 2026-11-10 VERBATIM from the publication page that serves these exact ZIPs, https://www.apra.gov.au/news-and-publications/national-claims-and-policies-database-statistics (HTTP 200, 356,262 bytes), in its "Disclaimer and Copyright" block: "(c) Australian Prudential Regulation Authority (APRA) 2024. This work is licensed under the Creative Commons Attribution 4.0 Licence (CCBY 4.0). This licence allows you to copy, redistribute and adapt this work, provided you attribute the work and do not suggest that APRA endorses you or your work. To view a full copy of the terms of this licence, visit: https://creativecommons.org/licenses/by/4.0/" -- re-verified independently by the orchestrator, and the page carries NO 3.0 reference. Names copy, REDISTRIBUTE and adapt, in prose, by the rights holder, on the page serving the files. WHY NOT THE SITE-WIDE NOTICE: apra.gov.au/copyright offers a CC BY 4.0 safe-harbour expressly CONDITIONED on the material "containing the following copyright statement: (c) Commonwealth of Australia 2024" -- and these files contain no copyright statement at all, so that hook does not attach. The publication-page notice needs no hook, being already about this publication. WHY NOT UNVERIFIED despite the silent ZIPs: see the comment above. Attribution AND non-endorsement both travel; non-endorsement is a condition of the CC BY grant, not a courtesy. APRA's disclaimer travels too. PERSONAL DATA cleared independently, and it is the bar that binds first. APRA's own explanatory notes state: "APRA determined through public consultation that the NCPD publications are not confidential. NCPD publications do however include privacy masking, with unmasked data available only on request and subject to APRA approval and acceptance of privacy terms and conditions." The GATED artefact is the unmasked one; this loader reads the ungated masked one. Measured: a k=5 suppression floor holds EXACTLY -- minimum non-zero count is 5.0 and zero cells report a count in (0,5), in all eight files -- and there is no age, sex, name, postcode or date finer than a year. The age+sex+small-geography+exact-date combination that disqualified sibling assets cannot be formed here.

cc-by-4.0 requires attribution, and that duty is not discharged by a parquet file. Any use, publication or onward copy must carry:

APRA National Claims and Policies Database, masked Claims and Policy report bundles, December 2024 publication. (c) Australian Prudential Regulation Authority (APRA) 2024, licensed CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) per the publication page's own Disclaimer and Copyright block. APRA DOES NOT ENDORSE this repository or any use of this data -- non-endorsement is a condition of the grant, not a courtesy. APRA accepts no responsibility for accuracy, completeness or currency. Note this is CC BY 4.0 while the sibling apra_life_claims and apra_gi_institution mirrors are CC BY 3.0 AU: APRA's own instruments differ between its workbooks and this publication page, and each was read on its own. WHAT THIS IS: eight cross-tab aggregate tables, 720,006 rows total [measured], covering Australian public and product liability and professional indemnity, reporting years 2004-2024. Largest arm 158,404 x 20; arms are {claims,policy} x {by_occupation,by_state} x {excess-band,limit-band}. Measures are claims reported/finalised and gross payments/incurred (claims arms), and risks-in-force, risks written and gross earned/written premium (policy arms). Covariates include 9 product types, 8 states plus unknown, and 70 occupation/industry groups across TWO coexisting taxonomies (ANZSIC numeric for liability, letter-coded professions for indemnity). WHY IT IS HERE: the only occupation-resolved liability experience table in this collection, and the only APRA product carrying a policy-design covariate (excess and limit-of-indemnity bands). Good for frequency and severity RATES. IT IS NOT CLAIM GRAIN, contrary to what this programme expected of it. The underlying DATABASE is claim-level, but this PUBLICATION is cross-dimensional bucket aggregates, as APRA's own explanatory notes state. Measured proof: the key is unique across all 158,404 rows of the default arm, n_claims_finalised runs 0 to 1,442, and 115,335 of 158,404 rows report ZERO finalised claims. A claim-grain table has neither a count column nor zero-claim rows. SIX MEASURED CAVEATS. (1) 'Masked' removed the INSURER ENTIRELY -- no ABN, no company name, no insurer code in any of the eight files -- so there is NO ENTITY KEY and entity-disjoint cross-validation is impossible on this asset. That is the real cost of the masking, and it is a modelling limitation rather than a privacy one. (2) Masking also withholds cells, and it hits KEY columns not just measures: raw product_type reads as 10 values where there are 9, class_of_business as 3 where there are 2, and four year columns arrive as object dtype mixing ints with ''. That also breaks key uniqueness on three arms, with 91.7% of one arm's duplicate keys carrying a ''. (3) Suppression uses FOUR encodings and the publisher's own footer is wrong about one: the footer says '<5' marks counts too small to publish, but '<5' appears NOWHERE in the claims bundle, so claims-arm small counts are masked as '*' and are indistinguishable from genuine aggregation; and state_jurisdiction suppresses to a BLANK, alone among the dimensions, while separately carrying a publisher 'unknown' category that means something else. (4) THE ARMS DO NOT RECONCILE: gross claims incurred totals 37.10 / 36.85 / 39.47 / 36.79 billion AUD across the four claims arms, a 2.68bn (7.3%) spread on one underlying quantity, because suppression is applied at different aggregation levels. Pick one arm; never union or average them. (5) Band vocabularies differ between the claims and policy arms, so they CANNOT be joined into a loss ratio. (6) gross_claims_incurred is negative on 23.9% of rows (reserve releases) and claim counts are zero on 71.4%. RE-IDENTIFICATION ASSESSED LOW, and the asset is not restricted: a k=5 suppression floor holds exactly -- minimum non-zero count is 5.0 and zero cells report a count in (0,5), in all eight files -- there is no age, sex, name, postcode or date finer than a year, and geography is state/territory (8 values for a continent). APRA has formally determined these publications non-confidential. WHAT THIS IS NOT: not claim-grain records; not an insurer panel (no entity); not motor or personal lines (the nine product types are CON, CYB, D&O, MAL, MAN, PIO, PLO, PRO, PUB -- no motor product exists); and it contains NO adjudication outcome, so nothing records a claim accepted or declined.

Citation

Australian Prudential Regulation Authority (2026). National Claims and Policies Database statistics -- masked Claims and Policy report bundles, December 2024. Licensed CC BY 4.0; APRA does not endorse this use.

Loading

from propensity.data.hf_mirror import load_mirrored

result = load_mirrored("apra_ncpd", org="SMVaddadi")
result.frame.head()

load_mirrored falls back to the original loader (origin, then its own offline replica) if this mirror is unreachable, so it is safe to depend on.