Financial Stress and Insurability Gap Index — Methodology
Version: v0.1 (draft) · Index date: 30 Nov 2026 · Geography: Australia SA3 (ASGS 2021), New Zealand SA2 (2023)
What this index measures
Where households are under statutory financial stress and where the cost or availability of home insurance is becoming a second, compounding pressure. It is built entirely from administrative and regulatory open data, so every number can be traced to a government table.
It is published quarterly with the AFSA insolvency release. Each quarter is a new version; earlier versions are never overwritten.
What this index is not
- Not a survey. It contains no self-reported stress or sentiment.
- Not a credit score, not a property valuation, not a hazard model.
- Not address-level. The finest unit is SA3 (Australia, ~340 areas, typical population 30,000–130,000) and SA2 (New Zealand, ~2,400 areas).
- Not a prediction. It describes the most recent quarter of observed administrative data.
The four pillars
| Pillar | Measure | Source | Cadence |
|---|---|---|---|
| Insolvency stress | Personal insolvencies per 10,000 residents, trailing 12 months (4 quarters) | AFSA quarterly personal insolvency statistics, SA3 | Quarterly |
| Cashflow (gaming) stress | Electronic gaming machine net expenditure per resident, trailing 12 months | Vic (VGCCC via DataVic), NSW (Liquor & Gaming NSW premises list), Qld (OLGR) | Monthly / quarterly |
| Small-business contraction | Rolling 12-month net ABN cancellations (companies and trading trusts) per 1,000 GST-registered businesses, ATO bulk-sweep weeks excluded | Australian Business Register bulk extract | Weekly |
| Insurance affordability gap | Average combined home building + contents premium as a share of annual gross household income, at SA3 | ACCC Insurance Monitoring Report 2026 supplementary table T1 (average premium band per SA3); APRA CVA appendix was not published as a data file | Baseline (series discontinued) |
New Zealand (provisional)
New Zealand (v0.1) is a provisional two-pillar composite at SA2 2023, published separately from the Australian index and labelled as such — it is not comparable to the AU composite. The two pillars, equally weighted, are:
| Pillar | Weight | Measure | Source |
|---|---|---|---|
| Deprivation | 0.50 | NZDep2023 socioeconomic deprivation score (higher = more deprived) | University of Otago NZDep2023, SA2 |
| Bond turnover | 0.50 | Trailing-12-month rental-bond lodgements ÷ mean active bonds (housing churn proxy) | MBIE Tenancy Services bond data, SA2 |
Each pillar is transformed and percentile-ranked exactly as the Australian pillars are, and the composite is their renormalised weighted mean — but both pillars must be present for a composite to be published (the minimum-coverage rule above). An SA2 with deprivation but no bond-turnover reading keeps its deprivation percentile and is flagged insufficient_coverage with no composite, because a one-pillar NZ composite is just NZDep relabelled. Statutory NZ insolvency data at SA2 is still being assessed, so it is not yet a pillar.
The MBIE bond series is keyed on the pre-2023 SA2 vintage — Stats NZ calls it SA2 2018, MBIE labels the same codes "2019". To reach the SA2 2023 spine it is apportioned through a meshblock-count-weighted SA2 2018→2023 correspondence (geo.corr_sa2_2018_2023), built from the Stats NZ Geographic Areas Table 2023 (one row per 2023 meshblock carrying both the 2018 and 2023 SA2 code). For each 2018 SA2 the ratio to a 2023 SA2 is the share of its meshblocks that fall in that 2023 SA2; 379 of 2,253 SA2-2018s split across more than one 2023 SA2, so the weighting matters. Bond lodgements and active-bond counts are apportioned by these ratios and the pipeline asserts the apportioned total equals the source total within 0.1%. Meshblock-count weighting is a documented approximation — the Geographic Areas Table carries no meshblock population; it should be upgraded to meshblock-population weighting once Census 2023 meshblock counts are loaded (TODO(source)). The NZ output is written as index_nz.parquet beside the Australian Parquet; it is generated only once the concordance (geo.corr_sa2_2018_2023) is loaded.
Population floor. New Zealand SA2s with a 2023 usual-resident population below 500 (and the non-residential SA2s — inlets, forests, parks — that carry no population at all) are given no composite and missing_pillars = ['small_population']; they are held out of the percentile ranking so a handful of people in a tiny area cannot produce an extreme score. Population is summed to SA2 from the NZDep2023 SA1 file (URPopnSA1_2023) into base.nz_sa2_population. At v0.1 this removes 214 of 2,379 SA2s (including small industrial zones such as Kawerau Industrial that would otherwise top the table on a few residents). On the map these areas are solid grey, "No data (small population)".
Geography and apportionment
Postcode and LGA series are apportioned to SA3 with the ABS 2021 population-weighted correspondence files — the direct CG_POA_2021_SA3_2021 and CG_LGA_2021_SA3_2021 grids. Targets are restricted to the spatial SA3 set and ratios renormalised per source unit; in practice this only removes the special-purpose units (ABS ZZZZZ sentinel, migratory/offshore and no-usual-address SA3s) that carry no geometry, so no real postcode or LGA share is lost. Area-weighted joins are not used anywhere: in regional Australia a postcode's land area and its population sit in different places, and area weighting misallocates urban losses into rural SA3s. After every apportionment the pipeline asserts that the apportioned total equals the source total within 0.1%.
New Zealand geography is Stats NZ Statistical Area 2 (SA2) 2023 (generalised), stored in EPSG:2193; non-digitised SA2s are held out of the spatial layer, and deprivation attaches via NZDep2023 at SA2, with offshore/inlet SA2s left unscored by Otago (no decile).
Structural baseline: Census and SEIFA
Census 2021 (General Community Profile tables G01, G02, G37) and SEIFA 2021 are loaded at SA2 and SA3 as a slow-moving baseline (for per-capita denominators and socio-economic context). Tenure shares come from G37 (occupied private dwellings by tenure). Census tables are restricted to the spatial geography spine, which drops the special-purpose codes (migratory/offshore/no-usual-address, ~0.2% of national population); loaded SA3 population reconciles to the national Census count within 0.5%.
ABS publishes SEIFA 2021 at SA2 but not at SA3. SA3 index scores are therefore derived as the Usual-Resident-Population-weighted mean of the SA2 scores beneath each SA3, for all four indexes (IRSD, IRSAD, IER, IEO). SA3 deciles are recomputed by ranking the derived SA3 scores into ten equal-count groups (decile 1 = lowest score = most disadvantaged). This is a derived product, not an ABS output: a handful of reserve-only SA3s with negligible population (e.g. Illawarra Catchment Reserve, Blue Mountains – South) have no scored SA2 beneath them and so carry no SEIFA value. Derived SA3 scores should be read as indicative; where an SA2-level SEIFA figure exists it is preferred.
Transformation and scoring
- Each pillar is transformed with a Yeo-Johnson power transform (standardised), fitted on the current quarter. Where a pillar has fewer than 50 observations it falls back to a rank-normal transform, Φ⁻¹(rank / (N+1)). Raw z-scores are not used because every pillar is heavily right-skewed; untransformed, a handful of mining towns and high-density gaming LGAs would compress the other 80% of areas into a narrow band.
- The transformed pillar is converted to a percentile rank 0–100 across all areas in the country. Higher means more stress.
- The composite is the weighted mean of the four pillar percentiles:
| Pillar | Weight |
|---|---|
| Insolvency stress | 0.30 |
| Cashflow (gaming) stress | 0.25 |
| Small-business contraction | 0.20 |
| Insurance affordability gap | 0.25 |
Where a pillar is missing for an area, weights are renormalised over the pillars present and the missing pillar is named in missing_pillars.
4. Minimum coverage. A composite is only computed where enough pillars are present — at least 3 of the 4 for Australia, and both for New Zealand. An area below that threshold keeps its available pillar percentiles but has composite = null and insufficient_coverage appended to missing_pillars; it is drawn as "No composite" on the map. This stops a composite from resting on one pillar (for New Zealand that would just be NZDep relabelled) or on two in Australia. At v0.1 the rule withholds a composite from 28 of 340 Australian SA3s (including small areas that otherwise ranked on a single pillar) and 207 of the scored New Zealand SA2s that have only one pillar.
5. All pillar percentiles are published alongside the composite. A composite is never shown without its components.
Weights are a judgement, documented here so they can be argued with. Changing any weight produces a new index version.
ABN sweep filtering
The ABR periodically cancels large numbers of inactive ABNs in bulk. These weeks would otherwise register as sudden business failure. A week is flagged as a sweep when national cancellations exceed the 52-week median by more than three standard deviations; flagged weeks are excluded from the rolling contraction rate. Flagged weeks are listed in each version's _meta.json.
Contraction denominator floor. The contraction pillar is a rate per 1,000 GST-registered businesses. In SA3s with a very small business base a few cancellations swing the rate wildly, so the pillar is set to null (and contraction added to missing_pillars) wherever the GST-business denominator is below 1,000. At v0.1 this floors 32 of 340 SA3s — mostly offshore territories (Christmas, Cocos, Norfolk, Lord Howe, Jervis Bay), reserve-only SA3s, and small remote areas including Huon – Bruny Island (denominator ≈ 772), whose composite was otherwise led by an unreliable contraction reading.
Versioning and reproducibility
- Raw downloads are kept at
data/raw/<source>/<date>/with SHA-256 sidecars. - Each index version is a folder
data/out/index/v<date>/containing the Parquet file and a_meta.jsonwith sources, licences, attribution strings, weights, transform, sweep weeks and generation time. - Code is public. Re-running the pipeline against the same raw files reproduces the version byte-for-byte except for the generation timestamp.
Sources and attribution
| Source | Licence | Attribution |
|---|---|---|
| ABS ASGS 2021 boundaries and correspondences | CC BY 4.0 | Australian Bureau of Statistics, ASGS Edition 3, 2021 |
| ABS Census 2021 GCP and SEIFA 2021 | CC BY 4.0 | Australian Bureau of Statistics, Census 2021; SEIFA 2021 |
| AFSA quarterly personal insolvency statistics | CC BY 2.5 AU | Australian Financial Security Authority |
| Victorian gaming expenditure by LGA | CC BY 4.0 | Victorian Gambling and Casino Control Commission, via DataVic |
| NSW liquor licence premises list | CC BY 3.0 AU | Liquor & Gaming NSW, via Data.NSW |
| Queensland gaming machine data | CC BY 4.0 | Queensland Office of Liquor and Gaming Regulation, via publicdata.au |
| ABN Bulk Extract | CC BY 3.0 AU | Australian Business Register |
| APRA Insurance Climate Vulnerability Assessment 2026 | CC BY 3.0 AU | Australian Prudential Regulation Authority |
| ACCC Insurance Monitoring Report 2026 | CC BY 3.0 AU | Australian Competition and Consumer Commission |
| Stats NZ SA2 2023 | CC BY 4.0 | Stats NZ |
| NZDep2023 | CC BY 4.0 | University of Otago |
| MBIE rental bond data | CC BY 3.0 NZ | Ministry of Business, Innovation and Employment |
Exact attribution strings as required by each licence are in data/sources.yaml and reproduced in every output's _meta.json.
Known limitations (v0.1)
- The insurance pillar is a baseline, not a time series: APRA's CVA was published once and the ACCC series has ended. Movement in the composite between quarters comes from the other three pillars until a recurring affordability source exists.
- Gaming expenditure at sub-state level in open data exists only for Victoria and Queensland. The NSW source (Liquor & Gaming NSW premises list) carries electronic gaming machine counts, not net expenditure, so in v0.1 NSW SA3s and the states without any open sub-state series (WA, SA, Tas, NT, ACT) carry
cashflowinmissing_pillars. The cashflow pillar is present for 144 of 340 SA3s; the composite for the rest is a renormalised mean of the pillars that are present. - AFSA suppresses small quarterly counts by SA3. Annualising the insolvency pillar over four quarters fills most of this (334 of 340 SA3s scored); six SA3s suppressed across all four recent quarters carry
insolvencyinmissing_pillars. - ABN postcode is the registered business address, which for some entities is an accountant's office.
- SA3 boundaries are 2021; Census 2026 geographies (mid-2027) will trigger a re-base.
- About 52,900 persons recorded against ABS special-purpose codes (migratory, offshore, and no usual address) are excluded from the SA3 spine, which is why the SA3 population sums to 25.37m against the national Census count of 25.42m.
- The insurance pillar uses the ACCC T1 average-premium band midpoint (1,250-wide bands; "$6,250+" → $6,875) over Census median household income — a coarse ratio, not a modelled premium, pending a finer premium source.
- Gaming vintage gaps: Victoria's
Merri-bek(2022 rename of Moreland) does not match the 2021 ABS LGA name; Queensland gaming is keyed by ASGS 2011 SA2 names and ~14% of spend sits in post-2011 growth-area SA2s absent from theCG_SA2_2011_SA2_2021concordance. Both are logged, not silently dropped. - NZ bond data is keyed on Stats NZ SA2 2018/2019 ids; it is apportioned to the SA2 2023 spine through the meshblock-count-weighted
geo.corr_sa2_2018_2023concordance (see the New Zealand section). Four MBIE bond SA2 ids have no mapping in the Geographic Areas Table and drop out; 368 of 2,379 SA2s have no bond pillar (no bond activity or no mapped source), and the NZ composite there rests on deprivation alone (flagged inmissing_pillars).
Output schema
Each version is published as data/out/index/v<date>/index.parquet with a sidecar _meta.json. The Parquet has one row per geography:
| Column | Type | Description |
|---|---|---|
version |
text | Index version, e.g. v0.1 |
country |
text | AU (v0.1 is Australia only; the NZ composite is published separately) |
geo_level |
text | SA3 for Australia |
geo_code |
text | ASGS 2021 SA3 code |
period |
date | Reference quarter (first day of its last month), e.g. 2026-06-01 |
p_insolvency |
float | Insolvency pillar percentile, 0–100 (null if in missing_pillars) |
p_cashflow |
float | Gaming cashflow pillar percentile, 0–100 |
p_contraction |
float | Small-business contraction pillar percentile, 0–100 |
p_insurance |
float | Insurance affordability pillar percentile, 0–100 |
composite |
float | Weighted composite percentile, 0–100 |
missing_pillars |
list[text] | Pillars unavailable for this area (empty when all four are present) |
_meta.json carries the source/attribution/licence list, the weights, the transform, the reference period, the per-pillar missing-area counts, and the ABN bulk-sweep weeks excluded from the contraction pillar.
The provisional New Zealand composite, when generated, is written beside it as index_nz.parquet with the same row shape but NZ pillar columns p_deprivation and p_bond_turnover in place of the four Australian pillars; its sources, weights and caveat sit under the nz_provisional key in _meta.json. In the index.stress table both countries share one schema: Australian rows populate p_insolvency…p_insurance and leave p_deprivation/p_bond_turnover null; New Zealand rows do the reverse.
Changelog
- v0.1 (30 Nov 2026): initial release. Australia SA3, reference quarter = latest AFSA quarter (2026-06). All four pillars annualised (trailing 12 months). Insolvency and contraction cover all/nearly all SA3s; cashflow covers Victoria and Queensland (expenditure, now fully apportioned after fixing the Merri-bek rename and the post-2011 Queensland SA2 names); insurance covers 330 of 340 SA3s. New Zealand provisional two-pillar composite (NZDep2023 + bond turnover) now generated at SA2 (2,379 areas; 2,005 with both pillars), with the MBIE bond series apportioned from SA2 2018 to SA2 2023 via the meshblock-count-weighted Geographic Areas Table 2023 concordance.