Turns real public Australian data (ABS / RBA / PTRS) into the macro layer a credit team applies before it models individual borrowers — the current macro conditions to assess against, the macro drivers to stress, and a 1–5 industry credit-risk score with a ready-to-use PD overlay per ANZSIC division.
Real public data only. Sources are ABS, RBA and the Payment Times Reporting Scheme (PTRS). All synthetic / staged / illustrative data has been removed; the only non-data inputs are clearly-labelled methodology assumptions (stress shocks, scoring weights, and four CRE variables with no clean free public series).
The three sections below are also rendered to a single board report (Markdown / HTML / Word) straight from the live contracts on every pipeline run, for HR or a credit committee to open: md · html · docx.
The model-ready data sits in eight CSV "contracts" in outputs/contracts/
— the stable interface any downstream PD / LGD / ECL model reads.
Every figure is the Q1 2026 vintage — pipeline DATA_AS_OF = 2026-06-16. Every quarterly /
monthly ABS series now reports through the March 2026 quarter (the end of Q1 2026); the cash
rate is refreshed live. The reference period each real source actually reports for:
| Source / series | Catalogue | Reference period reported |
|---|---|---|
| Australian Industry — financial ratios, structure | ABS 8155.0 | FY 2023–24 (annual — FY 2024–25 not yet released) |
| Business Indicators — profit / margins | ABS 5676.0 | Mar 2026 quarter |
| Business Indicators — inventories | ABS 5676.0 | Mar 2026 quarter |
| Labour Force, Detailed — industry employment | ABS 6291.0 | Mar 2026 |
| Building Approvals, non-residential — property cycle | ABS 8731.0 | Mar 2026 |
| Cash rate | RBA F1 | Live — latest available |
| GDP, unemployment, CPI, WPI | ABS 5206 / 6202 / 6401 / 6345 | Observed — Mar 2026 (live-fetched + parsed each run) |
| House prices, exchange rate (TWI), sector-output proxy | ABS 6416 (discontinued 2021) · RBA F11 · ABS 8155+5676 | Stated — no clean current free series |
The four economy-wide headline levels (GDP, unemployment, CPI, WPI) are now fetched live from the
ABS releases and parsed (download_macro_indicators.py),
joining the live cash rate; the rest stay stated (the ABS residential price index was discontinued
after Dec-2021, and there is no clean free series for the TWI change or the sector-output proxy).
The macro stress shocks in Section 2 are scenario-design assumptions (config vintage 2026-Q1),
not dated observations. The committed cache + full per-file attribution are in
data/cache/ATTRIBUTION.md.
A lender's loss rate is driven as much by the environment it lends into as by any single borrower. Before scoring an industry or a deal, the engine reads where the economy and the property markets currently sit — the conditions a credit team assesses against. The headline rates are live-fetched from the latest ABS/RBA releases; a few series with no clean current public series are stated (and flagged as such). Together they form the base of every stress scenario in Section 2.
The broad drivers that condition PD and LGD across the whole book, regardless of product.
| Macro condition | Current level | What it signals for credit risk | Source |
|---|---|---|---|
| GDP growth (real, YoY) | 2.5% | Around trend — moderate revenue growth | ABS 5206 (Mar 2026, live) |
| Unemployment rate | 4.3% | Low — supports household & SME debt-servicing | ABS 6202 (Mar 2026, live) |
| Cash rate | 4.35% (+0.5pp YoY) | Restrictive — debt-servicing pressure on leveraged borrowers | RBA F1 (live) |
| Inflation (CPI, YoY) | 2.4% | Back inside the 2–3% band — easing cost-of-living pressure | ABS 6401 (Mar 2026, live) |
| Wage growth (WPI, YoY) | 3.3% | Now outpacing CPI — modest positive real-income growth | ABS 6345 (Mar 2026, live) |
| Exchange rate (TWI, change) | 0.0% | Stable — neutral for FX-exposed corporates | RBA F11 (stated) |
| Industry / sector output (YoY) | 2.0% | Modest — sector-revenue channel for SME / corporate | ABS 8155 + 5676 (stated) |
Property-secured lending splits into residential and commercial — they sit on different cycles, so the engine reads them separately.
Residential property. The residential signal is house-price growth: +4.0% YoY — the collateral channel for residential mortgages, where a fall lifts mortgage LGD (Section 2b). This one is stated, not live: the ABS Residential Property Price Index (Cat. 6416) was discontinued after the Dec-2021 quarter and has no clean free replacement index, so it stays a labelled reading pending a new public source.
Commercial property (CRE). Read two ways:
(i) Price & yield levels — labelled assumptions (no clean free quarterly public series; anchored to RBA FSR commentary). These feed the CRE stress paths in Section 2a:
| CRE condition | Current level | Reading |
|---|---|---|
| Commercial-property prices | 0.0% | Flat |
| Office vacancy rate | 12.0% | Elevated — office segment under pressure |
| CRE rents | 0.0% | Flat |
| CRE cap rates | 6.0% | — |
(ii) Building-type cycle — real ABS data (Cat. 8731 non-residential building approvals, Mar 2026). This is where CRE risk actually concentrates right now: offices remain the softest segment (a clear downturn), now joined by agricultural buildings, while warehouses are the firmest (still in growth). The softness score runs 1 (firm) → 5 (soft); "PD if standalone" is the overlay that segment would carry on its own.
| Building type | Cycle stage | Softness (1 firm → 5 soft) | Approvals YoY | PD if standalone |
|---|---|---|---|---|
| Offices | downturn | 4.30 | −19.8% | 1.15× |
| Agricultural & aquacultural buildings | downturn | 3.90 | −27.0% | 1.15× |
| Retail & wholesale trade buildings | slowing | 3.40 | +6.5% | 1.15× |
| Education buildings | slowing | 3.25 | −31.3% | 1.15× |
| Aged care facilities | slowing | 3.20 | −22.0% | 1.10× |
| Short-term accommodation buildings | neutral | 3.10 | +112.0% | 1.10× |
| Health buildings | neutral | 2.40 | −36.0% | 1.00× |
| Warehouses | growth | 1.70 | +108.5% | 0.95× |
The table lists the eight specific building types; the chart additionally shows three "Total …"
bars, which are aggregate reference lines only. The five-segment rollup these building types feed
(CRE / Retail / Industrial / Construction / Residential) is in
property_market_overlays.csv. Approvals are a
forward-looking proxy for building activity (commencements/completions are not yet staged), and
the high positive YoY figures reflect volatile low-base months, so they are read for direction,
not magnitude.
Latest (as of 2026-06-15): cash-rate regime restrictive / rising, arrears
Low / Improving, overall macro-regime flag base
(macro_regime_flags.csv). The cash-rate regime is
real RBA F1; the arrears level/trend is a labelled qualitative assumption from the RBA FSR. A
"base" regime means no recession overlay is applied at current readings — but the engine
pre-computes the stress dials in Section 2 so a downturn can be costed instantly.
Stress testing answers one question a credit team is always asked: if the economy turns bad, how much more does the bank lose?
A recession doesn't create a loss directly. It first hurts the borrower (less income) and the collateral (lower property prices), and that shows up three ways — more borrowers default, the bank loses more on each one, and bigger balances are owing when they do:
weak economy → borrowers & property hurt → higher PD, LGD, EAD → higher loss
- PD (probability of default) — the chance a borrower stops paying.
- LGD (loss given default) — after the collateral is sold, the share of the money owed that the bank still loses.
- EAD (exposure at default) — how much is owed at the moment of default.
- Loss = PD × LGD × EAD — push all three up and the loss multiplies quickly.
What this layer builds. It turns one economic scenario into a set of "dials" — how many times worse PD, LGD and EAD get for each type of loan. A downstream model multiplies those dials onto its own loans to get the stressed loss. There are three standard ways to build such dials, simplest first:
| Method | What it is | Used here? |
|---|---|---|
| Multiplier | judgement-based "×2 in a severe recession" factors | ✅ this layer — fits simple / sector-level work |
| Rating migration | re-grade each borrower under the stressed scenario | a loan-level model would |
| Statistical model | a regression trained on years of real default data | a big, data-rich bank would |
This layer uses the multiplier method: it works at the public, sector level and needs no borrower-by-borrower default history. So these are illustrative dials, not a fitted model — a ready-to-use starting overlay, not a calibrated PD. A bank with the data would replace them with a statistical model and validate it every year.
Scope — the engine-room step only. This layer turns a scenario into PD/LGD/EAD dials and stops there; the rest of a real stress test (projecting capital, setting management actions, reporting to the board) stays with the modelling and capital teams. Regulatory anchors honoured:
mild= two quarters of ~zero growth (Basel CRE36.51);severeis GFC-like but plausible (APS 220 §72); the roll-up takes no diversification benefit (APG 113 §92); and everything is flagged illustrative, to be independently validated in production (APS 220 §76).
Twelve drivers. base is where each sits today (Section 1); mild / moderate / severe
add a worsening shock on top — mild is a textbook mild recession (two quarters of ~zero
growth), severe is a GFC-like path. Source:
macro_scenario_paths.csv.
| Macro driver | Base | Mild | Moderate | Severe |
|---|---|---|---|---|
| GDP growth (real, YoY) | 2.5% | 0.7% | −0.8% | −2.8% |
| Unemployment rate | 4.3% | 5.5% | 6.8% | 8.3% |
| Cash rate | 4.35% | 4.6% | 4.85% | 5.1% |
| Inflation (CPI, YoY) | 2.4% | 3.2% | 3.9% | 4.9% |
| Wage growth (WPI, YoY) | 3.3% | 2.8% | 2.3% | 1.5% |
| House-price growth (YoY) | 4.0% | −4.0% | −11.0% | −21.0% |
| Exchange rate (TWI, change) | 0.0% | −5.0% | −10.0% | −15.0% |
| Industry / sector output (YoY) | 2.0% | 0.0% | −2.0% | −5.0% |
| Commercial-property prices* | 0.0% | −7.0% | −15.0% | −28.0% |
| Office vacancy rate* | 12.0% | 14.0% | 16.0% | 19.0% |
| CRE rents* | 0.0% | −4.0% | −9.0% | −16.0% |
| CRE cap rates* | 6.0% | 6.4% | 6.9% | 7.6% |
* labelled assumption — no clean free quarterly public series.
Different loans break for different reasons, so each product reacts to its own short list of
drivers. The number next to each driver is its share of the blame for that product — the
shares add up to 100%. It says which drivers matter and in what proportion — not how big the hit
is. The size of the hit is set separately, by how sensitive that product is (Section 2c). The
arrow shows the bad direction (↑ = worse when it rises, ↓ = worse when it falls). Source:
portfolio_macro_sensitivity.csv.
Two disciplines the textbooks insist on, both built in. (1) A driver can legitimately hit PD, LGD and EAD through different channels — e.g. rising unemployment means more defaults and weaker sale prices and more arrears — but you must never double-count the same effect twice. (2) These shares are illustrative, not fitted regression coefficients; in a severe recession every driver is at its worst at once, so the shares wash out and the product's overall sensitivity does the work — the shares matter most in milder scenarios, where some drivers move more than others.
Worked example — residential mortgages. The point is that the three quantities respond to different drivers, so the engine maps them separately:
- PD — will they default? A jobs-and-rates story: unemployment ↑ (0.45 share), then cash rate ↑ (0.25), wage growth ↓ (0.20), inflation ↑ (0.10).
- LGD — how much is lost if they do? Almost entirely house prices ↓ (0.70), because collateral recovery falls; unemployment and cash rate are minor.
- EAD — how much is owed? Mostly cash rate ↑ (0.50) and unemployment ↑ (0.30).
The same split for every product (shares in brackets):
| Product / portfolio | PD — main drivers (share) | LGD — main drivers (share) | EAD — main drivers (share) |
|---|---|---|---|
| Residential mortgages | unemployment ↑ (0.45), cash rate ↑ (0.25), wage growth ↓ (0.20), inflation ↑ (0.10) | house prices ↓ (0.70), unemployment ↑ (0.20), cash rate ↑ (0.10) | cash rate ↑ (0.50), unemployment ↑ (0.30), GDP ↓ (0.20) |
| Credit cards | unemployment ↑ (0.50), wage growth ↓ (0.30), inflation ↑ (0.20) | unemployment ↑ (0.50), inflation ↑ (0.30), wage growth ↓ (0.20) | GDP ↓ (0.40), cash rate ↑ (0.35), unemployment ↑ (0.25) |
| SME lending | GDP ↓ (0.35), unemployment ↑ (0.25), cash rate ↑ (0.20), sector output ↓ (0.20) | CRE prices ↓ (0.45), house prices ↓ (0.35), sector output ↓ (0.20) | cash rate ↑ (0.40), GDP ↓ (0.35), sector output ↓ (0.25) |
| Corporate lending | GDP ↓ (0.40), sector output ↓ (0.30), cash rate ↑ (0.15), FX ↓ (0.15) | sector output ↓ (0.45), CRE prices ↓ (0.30), FX ↓ (0.25) | cash rate ↑ (0.45), GDP ↓ (0.35), sector output ↓ (0.20) |
| Commercial property (CRE) | CRE prices ↓ (0.30), vacancy ↑ (0.30), rents ↓ (0.20), cash rate ↑ (0.20) | CRE prices ↓ (0.40), cap rates ↑ (0.35), rents ↓ (0.25) | cash rate ↑ (0.50), vacancy ↑ (0.30), CRE prices ↓ (0.20) |
| Development finance | CRE prices ↓ (0.35), GDP ↓ (0.25), vacancy ↑ (0.20), cash rate ↑ (0.20) | CRE prices ↓ (0.55), cap rates ↑ (0.25), rents ↓ (0.20) | cash rate ↑ (0.45), CRE prices ↓ (0.30), GDP ↓ (0.25) |
The pattern: household products (mortgages, cards) pivot on the labour market; business products (SME, corporate) on GDP and sector output; property products (CRE, development finance) on property values, vacancy and cap rates. (Development finance's textbook drivers — presales and construction costs — have no free public series, so GDP and office-vacancy stand in as proxies here.)
Combining the scenario (2a) with each product's driver mix (2b) — then scaling by how sensitive the
product is and capping the result — gives the deliverable: a PD / LGD / EAD stress multiplier
("dial") for every product, in every scenario
(macro_stress_segment_multipliers.csv).
A monitoring model multiplies these onto its own base numbers; stressed
loss = stressed PD × stressed LGD × stressed EAD.
Severe-scenario multipliers (mild and moderate are in the contract; PD rises most because default is the most cyclical parameter):
| Segment | PD × | LGD × | EAD × |
|---|---|---|---|
| Residential mortgages | 2.35 | 1.48 | 1.20 |
| Credit cards | 2.65 | 1.42 | 1.24 |
| SME lending | 2.65 | 1.54 | 1.28 |
| Corporate lending | 2.50 | 1.54 | 1.28 |
| Commercial property (CRE) | 2.95 | 1.72 | 1.32 |
| Development finance | 3.25 | 1.78 | 1.36 |
Property-secured segments (CRE, development finance) carry the largest multipliers — consistent
with the office downturn in Section 1b. For a quick portfolio-wide cross-check, a simpler headline
overlay (PD ×1.2 / 1.5 / 2.0 across mild / moderate / severe, plus LGD, CCF and a property-value
haircut) is in downturn_overlay_table.csv.
Applied to a committed demo book, these multipliers give an illustrative portfolio expected-loss
uplift of 1.9× mild, 3.2× moderate, 5.6× severe (exposure-weighted, no diversification
benefit per APG 113 §92; macro_stress_demo_summary.csv).
A real monitoring model would apply the same multipliers to its actual facilities.
The same input has an industry axis. Each division's current-conditions inputs (employment,
margins, demand) set how hard the sector-output driver in 2b bites for SME and corporate books,
and feed the per-industry PD overlay in Section 3. Source:
business_cycle_panel.csv.
| Industry | Employment YoY | EBITDA margin | Demand YoY | Macro (current-conditions) score |
|---|---|---|---|---|
| Agriculture, Forestry and Fishing | −5.1% | 14.6% | −27% | 4.20 |
| Mining | −5.1% | 47.3% | — | 3.20 |
| Manufacturing | −0.9% | 9.2% | +95% | 3.40 |
| Retail Trade | −0.5% | 7.8% | +6% | 3.40 |
| Wholesale Trade | −8.7% | 6.1% | +108% | 3.40 |
| Arts and Recreation Services | −5.8% | 13.5% | −46% | 3.60 |
A sector with falling employment and thin margins is exactly where the sector-output shock bites
hardest, so a downstream SME / corporate model can scale its sector-output sensitivity by the
industry's score. Demand growth is a volatile approvals/indicator proxy (base effects), so it is
read alongside employment and margins, not alone. The full per-industry panel (all components and
source rows) is the business_cycle_panel.csv contract.
Each ANZSIC division gets a single 1 (low risk) → 5 (high risk) score that blends a structural view with a current-conditions view, then maps to a risk level and a PD overlay. The whole calculation is in src/panels/macro_signals.py (scoring) and src/overlays/build_industry_risk_scores.py (ladder).
Step 1 — macro (current-conditions) score. Five components, each scored 1 → 5 from ABS business-indicator series, then averaged:
| Component | Higher risk when… | Source |
|---|---|---|
| Employment score | industry employment YoY growth is falling | ABS 6291 Labour Force |
| Margin level score | profit-to-sales / EBITDA margin is thin | ABS 5676 / 8155 |
| Margin trend score | margin is deteriorating YoY | ABS 5676 / 8155 |
| Inventory score | inventory days / stock-build risk is rising | ABS 5676 |
| Demand score | demand-proxy YoY growth is weak | ABS approvals / indicators |
macro_risk_score = mean(employment, margin level, margin trend, inventory, demand)
Step 2 — blend with structural risk. classification_risk_score (1–5) captures the
structural risk of the division (cyclicality, capital intensity, revenue concentration):
industry_base_risk_score = 0.55 × classification_risk_score + 0.45 × macro_risk_score
Step 3 — map to level + PD overlay via a five-band ladder:
| Base score | Level | PD multiplier |
|---|---|---|
| < 1.60 | Low | 0.90× |
| 1.60 – 2.00 | Moderate-low | 0.95× |
| 2.00 – 2.80 | Medium | 1.00× |
| 2.80 – 3.23 | Moderate-high | 1.10× |
| ≥ 3.23 | Elevated | 1.15× |
| Input | Value | Why |
|---|---|---|
| Classification (structural) score | 4.12 | Structurally cyclical — weather- and commodity-exposed, concentrated revenue |
| → Employment YoY | −5.1% | Falling employment → high employment score |
| → EBITDA margin | 14.6% | Mid → moderate margin score |
| → Demand proxy YoY | −27% | Weak (volatile base effect) → high demand score |
| Macro (current-conditions) score | 4.20 | Average of the five components above |
Blend: 0.55 × 4.12 + 0.45 × 4.20 = 4.16 → Elevated band (≥ 3.23) → 1.15× PD overlay.
That 1.15× is the deal-level industry overlay a PD model multiplies through. These are
point-in-time, illustrative current-conditions overlays — not calibrated PD estimates.
Headline: 5 industries score Elevated, 2 Moderate-high, 10 Medium, 1 Moderate-low (as of
2026-06-16; industry_risk_scores.csv). No industry
currently sits in the Low or High band.
| Industry | Classification | Macro | Base score | Level | PD overlay |
|---|---|---|---|---|---|
| Agriculture, Forestry and Fishing | 4.12 | 4.20 | 4.16 | Elevated | 1.15× |
| Mining | 3.88 | 3.20 | 3.57 | Elevated | 1.15× |
| Manufacturing | 3.50 | 3.40 | 3.46 | Elevated | 1.15× |
| Retail Trade | 3.25 | 3.40 | 3.32 | Elevated | 1.15× |
| Wholesale Trade | 3.12 | 3.40 | 3.25 | Elevated | 1.15× |
| Construction | 2.75 | 3.20 | 2.95 | Moderate-high | 1.10× |
| Arts and Recreation Services | 2.38 | 3.60 | 2.93 | Moderate-high | 1.10× |
| Accommodation and Food Services | 2.75 | 2.60 | 2.68 | Medium | 1.00× |
| Public Administration and Safety | 1.62 | 3.60 | 2.51 | Medium | 1.00× |
| Education and Training | 1.75 | 3.40 | 2.49 | Medium | 1.00× |
| Rental, Hiring and Real Estate Services | 2.38 | 2.60 | 2.48 | Medium | 1.00× |
| Transport, Postal and Warehousing | 2.50 | 2.40 | 2.46 | Medium | 1.00× |
| Administrative and Support Services | 2.12 | 2.80 | 2.43 | Medium | 1.00× |
| Information Media and Telecommunications | 2.12 | 2.60 | 2.34 | Medium | 1.00× |
| Other Services | 2.38 | 2.20 | 2.30 | Medium | 1.00× |
| Electricity, Gas, Water and Waste Services | 2.25 | 2.00 | 2.14 | Medium | 1.00× |
| Health Care and Social Assistance | 1.50 | 2.80 | 2.08 | Medium | 1.00× |
| Professional, Scientific and Technical Services | 1.75 | 2.00 | 1.86 | Moderate-low | 0.95× |
Per-industry component detail and the source rows behind every number are the
business_cycle_panel.csv contract.
Eight CSV "contracts" in outputs/contracts/ — the stable interface
any downstream PD / LGD / ECL model reads: industry_risk_scores, downturn_overlay_table,
macro_regime_flags, macro_scenario_paths, portfolio_macro_sensitivity,
property_market_overlays, industry_financial_benchmarks (APG 220 §64 ratios per industry),
plus the explainability panels (business_cycle_panel, property_cycle_panel,
property_market_overlays_by_building_type).
A board report in outputs/reports/ — Industry_Analysis_Q1_2026 in
.md / .html / .docx, rendering these three sections from the live contracts (links at the top
of this README).
| Area | In this project |
|---|---|
| PD overlays | Per-industry base risk scores and a pd_multiplier per ANZSIC division, ready to condition a PD model or scorecard. |
| Stress testing | Base / mild / moderate / severe macro paths routed to per-product and per-industry PD / LGD / EAD drivers, plus portfolio-wide multipliers and property haircuts. |
| LGD / collateral | Property-market overlays (cycle stage, softness, region risk) feeding LGD and collateral assumptions. |
| APRA APG 220 grounding | Per-industry medians of the financial ratios APG 220 §64 names as standard credit-assessment benchmarks. |
| Australian regulatory landscape | Works directly with ABS industry / building-approval / labour releases, the RBA cash-rate table and FSR, and the Payment Times Reporting Scheme. |
| Data engineering & governance | ETL from XLSX / CSV / PDF into validated CSV contracts; live ABS/RBA fetch with committed-cache fallback; reproducible outputs; pytest suite; real-data-only discipline with assumptions labelled, never presented as data. |
Written in Python (pandas, openpyxl, pdfplumber, python-docx, matplotlib). The skills transfer directly to SAS / SQL / R model-development and validation work.
python -m venv .venv
.venv\Scripts\activate # Windows; macOS/Linux: source .venv/bin/activate
pip install -r requirements.txt
python run_pipeline.pyrun_pipeline.py auto-downloads the real public data (ABS / RBA / PTRS); if any source is
unreachable it falls back to a committed real-data cache, so it always produces the reports —
online or offline. The flow runs end to end: fetch → build panels + overlays → export the 8 CSV
contracts → validate → render the board report. The data vintage is pinned
(DATA_AS_OF = 2026-06-16); the committed cache and its source attribution live in
data/cache/. Methodology notes:
cash-flow lending and
property-backed lending.
| Source | Provides |
|---|---|
| ABS — Australian Industry (8155.0), Business Indicators (5676.0), Labour Force (6291.0) | Industry financial ratios, activity, employment |
| ABS — Building Approvals, non-residential (8731.0) | Property-cycle and segment signals |
| RBA — F1 cash-rate table, Financial Stability Review, SMP, Chart Pack | Rate regime and stress context |
| Payment Times Reporting Scheme (PTRS) | Payment-stress signal |
This repo is the macro / industry / property overlay layer of a single commercial credit-risk stack — the upstream context a lender applies before modelling individual borrowers:
| Layer | Repo | What it does |
|---|---|---|
| Macro & industry overlays | this repo | ABS/RBA/PTRS → industry risk scores, property overlays, downturn stress |
| External benchmarks | external-benchmark | Bank & regulator Pillar 3 disclosures → PD / LGD / EAD reference points |
| Consumer modelling | consumer-credit-pd-ead-scorecard | Borrower-level PD / EAD scorecards |
| Mortgage modelling | mortgage PD/LGD/EAD repo (link pending) | Property-backed PD / LGD / EAD |
The flow: macro/industry overlays + external benchmarks → modelling → validation. This repo is
top-down (sector data from ABS/RBA/PTRS); external-benchmark is bottom-up (bank/regulator
disclosures). The modelling repos consume both.
The engine assembles and scores public, sector-level signals into reusable overlays. It deliberately does not set a borrower's final PD, an internal-portfolio LGD, a regulatory capital number, or any loan-level model — those belong to the modelling repos that consume these overlays. Keeping that boundary sharp — public reference data here, modelling judgement there — is what makes the outputs trustworthy as an industry benchmark.
Built by Jane Wu. Real public data only (ABS/RBA/PTRS); a sector-level reference layer, not a firm-level credit model. Released under the MIT License.