Methodology Thesis & Independent Review Source

0 · Abstract, review status & nomenclature

Objective. PortfolioCheck is an asset-level climate and environmental screening system for Ghanaian secured lending. It converts geolocation evidence into six dimensionless prioritisation indices, an overall vulnerability index and a conditional financial sensitivity overlay. This thesis documents the complete chain from input data to output, including assumptions, equations, uncertainty and validation gaps.

Abbreviations and symbols

Assertion classes

1 · Purpose & scope

PortfolioCheck screens a Ghanaian bank's property / mortgage book for physical climate risk at the level of each individual asset and translates it into the collateral and credit terms a lender and the Bank of Ghana need. It is screening-grade decision-support that feeds, and does not replace, a bank's IFRS 9 / PD-LGD models.

This document is the technical specification. It is generated from the running engine, so the formulas and constants below are exactly those in production.

Deliberate boundaries

• Not a property valuation, an insurance determination, or the bank's IFRS 9 ECL / PD-LGD model.
• Covers PHYSICAL climate risk; transition risk is out of scope.
• Forward figures are scenario estimates, not forecasts.
• Financial constants are illustrative until calibrated to a bank's own loss data.
• Does not constitute or replace an Environmental Impact Assessment / EPA permit (LI 1652, Act 1124).

2 · Processing pipeline

• Ingest — a CSV book (location + exposure + LTV; optional type, adaptations, insurance, construction).
• Geolocate — coordinates, GhanaPostGPS digital address, or geocoded free-text → a point.
• Screen — six hazards scored 0–1 from the location SEED model (+ live engines where enabled).
• Adapt — recorded, evidence-gated adaptation measures reduce the relevant hazard scores.
• Composite — combine the six hazards into a 0–100 vulnerability score and a band.
• Money — AAL → collateral repricing → climate ECL overlay (PD × LGD × EAD).
• Project — pull each asset through IPCC AR6 / NGFS scenarios and horizons.
• Output — dashboard, board report, BoG semi-annual return, supervisor aggregation.

3 · Hazard model

Six hazards are scored 0–1 per asset. The SEED model is a transparent, deterministic proxy keyed to Ghana's physical geography (flood basins, the coastal-erosion belt, river lines, urban-heat and savanna-drought gradients, protected/Ramsar land). LIVE mode overlays the SafeGround / FloodGuard engines and climate reanalysis.

3.0 Engines, execution modes and fallback behaviour

SafeGround and FloodGuard are processing and integration engines, not primary datasets. They query or serve the named source layers, apply the documented transformations and return source metadata. PortfolioCheck never treats an unavailable endpoint or absent layer as proof of low risk: that hazard retains its SEED result and the mode remains disclosed.

3.1 Exact SEED calculations

SEED scores are deterministic geographic heuristics. They preserve workflow when live services are unavailable, but they are not probabilities, return periods or measured hazard intensities. Distances use the haversine great-circle equation:

3.2 Live transformations and evidence gates

3.2a Hazard-by-hazard data and decision lineage

The following register answers the reviewer questions behind every risk number: which engine ran, which underlying information it used, how it became a score, what uncertainty remains, and whether it is allowed to affect a monetary result.

Flood / surface water

Coastal / sea-level

Extreme heat

Riparian proximity

Land / environmental constraint

Drought / water stress

3.3 Coastal decision hierarchy

• Coordinate quarantine: a water/suspect point is not scored until corrected.
• Grade A: dated observed shoreline retreat covering the site → engineering assessment; financial channel may be enabled.
• Grade B: low-lying terrain context → site survey and storm-tide assessment; no automatic loss.
• Grade C: published sector evidence → enhanced due diligence; no automatic loss.
• Grade D: proximity or unresolved gap → context only; no automatic loss.
• A defence receives zero benefit unless footprint, frontage, crest, design, condition and maintenance are verified.

4 · Composite vulnerability

The composite blends the weighted average hazard load with the single worst hazard, so a lone severe hazard cannot be averaged away:

Scaled to 0–100 and banded:

5 · Hazard → money

5.1 Average Annual Loss (physical damage)

Per-hazard AAL is a ceiling × a convex function of the hazard score; the per-hazard ceilings are INFORMED BY the JRC Africa flood depth-damage curve (Huizinga et al. 2017 — depth-damage × an assumed event frequency) and First Street / UNDRR AAL tiers, and are ILLUSTRATIVE until back-solved from a bank's own loss data. This is a transparent screening proxy, not a full per-parcel depth-damage integration. Hazards combine as independent damages:

Construction quality scales the damage an asset suffers (structural vulnerability):

5.1a Agricultural / farmland asset class

Where the collateral is a FIELD, not a building, PortfolioCheck routes it to a farmland model (classified from the asset's property type — see core/agri.py). Three things change, because a field is not a building: (1) the perils are agronomic — drought leads (a rainfall-failure year wipes much of the season), then flood, then heat stress, so a drought-led weighting and a production-loss AAL table replace the structural ones; (2) the value at risk is a flow (one season's production, hitting the borrower's cash-flow / PD) not a stock (the land, the collateral, survives a single event), so the collateral repricing / LGD is land-resilient and capped lower; (3) the remedies are agronomic (irrigation, drought-tolerant varieties, field drainage, crop insurance), not building retrofits. The observed crop-health signal (NDVI / VCI vs the field's own green-season normal) is supplied by FloodGuard and folded into the drought hazard. This is the classification the Directive implies for a mixed book (Para 43(f)(iv) counterparty activity + 48(b) vulnerable sectors — agriculture the archetypal one).

Farmland collateral repricing is damped by land resilience (×0.5) and capped at 20% (building cap 40%). Illustrative — calibrate to the bank's crop-loss / weather-index data.

A field is an AREA, not a pin. Where an asset carries a parcel boundary (a GeoJSON polygon, WKT, or a coordinate list in a geometry column, or via /api/screen?geometry=), the live screen reads the hazards area-weighted across the whole polygon rather than at a single centroid: it samples every ~100–250 m grid cell inside the field and returns the mean plus the AFFECTED FRACTION — the share of the field in the high-flood band, or the share of the field's cropland under NDVI/VCI stress (= the share of the season's production at risk). This corrects the centroid blind spot where half a field floods while the pin sits on the dry half. A sub-cell parcel falls back to the centroid reading. Powered by FloodGuard's /api/{floodrisk,droughtchange,cropchange}/poly.

5.1b Agricultural source-to-variable and evidence register

5.1c Minimum data required before agricultural loss calibration

5.1d Agricultural data-quality hierarchy

5.1e Output status and permitted interpretation

5.1f Calibration, back-testing and independent-validation gate

• At least 5–10 years of crop/season/region observations where available.
• Observed yield, production value and loss/claim outcomes linked to parcel and hazard date.
• Train/validation separation with spatial and temporal cross-validation.
• Crop-specific discrimination and calibration diagnostics, including uncertainty intervals.
• Sensitivity, specificity, false-positive/false-negative and stability analysis.
• Bank PD/LGD/EAD integration only after bank model governance and independent validation.
• Independent agronomist, geospatial specialist and quantitative model-validator sign-off.

Production gate: calibrated agricultural monetary loss remains disabled as a claim until the minimum input fields are present, Ghana crop/season baselines are versioned, performance is tested out of sample, and independent model validation signs off. Until then the numerical loss overlay is shown only as an illustrative sensitivity with an explicit label.

5.2 Market repricing (collateral devaluation)

A forward collateral haircut as the market prices in a share of the AAL, plus an insurability penalty an insured asset does not carry:

5.3 Stressed value & LTV

5.4 Loss Given Default

A smooth, band-free downturn liquidation recovery, Basel-style floored so LGD is never zero:

5.5 Probability of Default

PD rises with negative equity (strategic default) and uninsured-damage cash-flow stress:

5.6 Expected Credit Loss overlay

6 · Forward-looking scenarios (IPCC AR6 / NGFS)

Each hazard score is projected with a bounded, saturating odds map, so a projection never exceeds 1.0 and a high-risk asset saturates while mid-range assets move most:

The forcing f converts an intensification factor to a unit forcing: multiplier hazards (flood, riparian, heat, drought) f = factor − 1; coastal f = sea-level-rise(cm) / 50.

Intensification factors by scenario × horizon (vs today)

Flood / riparian — extreme-rainfall (Rx1day) multiplier

Heat — hot-day-count multiplier

Drought — consecutive-dry-days multiplier

Coastal — sea-level rise

NGFS pathways map to the AR6 physical forcing consistent with their warming outcome:

7 · Adaptation measures & evidence gating

Recorded completed measures reduce the relevant hazard score only after supervisory evidence acceptance; credits combine on residual risk and are capped. A bank declaration alone receives no credit.

A +1 m raised plinth subsumes the +0.3 m credit; where elevated ≥1 m, floodproofing / PFR credits are damped ×0.5 (they protect against the same residual). Measure credits:

8 · Data quality (PCAF)

Every asset carries a PCAF-style 1–5 data-quality flag; a screening tool is never better than 3. This is the honesty layer a supervisor weights submissions by.

9 · Validation

Accra 3-June-2015 flood hindcast. At a flood-score threshold of 0.50, the flood layer flags 12/12 documented-flooded localities (sensitivity 100%) and spares 6/6 higher-ground controls (specificity 100%).

Parameter sensitivity. Each finance constant varied ±20%; % change in the demo book's total climate-ECL uplift (base GHS 428,223), ranked by impact:

9.1 Validation status by component

9.2 Minimum calibration protocol

• Freeze a development sample and a later or out-of-region validation sample.
• Define outcomes before fitting: inundation/depth, damage ratio, claim, arrears/default, realised recovery and sale discount.
• Assess discrimination (ROC/AUC or precision-recall), calibration/Brier score, spatial residuals and confidence intervals.
• Fit hazard-to-damage relationships by construction and property type; never infer damage directly from an ordinal traffic light.
• Estimate PD/LGD overlays inside the bank's IFRS 9 governance with conservatism and overrides documented.
• Back-test annually, monitor drift and disparate impact, version every parameter change, and obtain independent validation.

9.3 Governed loss-data onboarding and approval

LOSS-1.0 registers events, assets, agricultural seasons, insurance claims, loan outcomes, valuation history and adaptation evidence using pseudonymous identifiers. Uploads are validated row by row, linked to the selected institution portfolio and stored as drafts. A separate reviewer from the same institution must approve; makers cannot approve their own submissions.

• Observed gross loss, verified loss, insurance recovery and modelled annual loss remain separate fields.
• Grade D modelled or remotely sensed values and Grade E unverified/synthetic values cannot independently calibrate observed loss.
• The development gate requires at least 30 matched Grade A-C events, three event years, two hazards and 30 verified loss/exposure pairs.
• Passing the development gate is not model approval: holdout testing, uncertainty estimation, stability analysis and independent validation remain mandatory.
• Approved observations can be displayed beside modelled AAL but never overwrite or silently tune it.

10 · Regulatory framework mapping

Active parameter set: BoG baseline v1 (2026-06) [baseline] (stamped on every report and return). Outputs map to:

• Bank of Ghana Climate-Related Financial Risk Directive (2024) — the semi-annual return (Para 48); the physical-risk disclosure method (Para 43(f)); the example physical metrics (Table 1). The forthcoming Standardized Disclosure Template is Para 46.
• IFRS S2 / TCFD — risk-management and metrics disclosure.
• NGFS — the prescribed scenario family.
• PCAF — per-asset data-quality flagging.
• Ghana Sustainable Banking Principles — asset-level E&S risk + green-loan origination.

The full clause-by-clause alignment is a living page at /directive (14 core outputs, 4 partial (physical half), 6 feed the bank's process, 3 need the bank's wider book, 6 out of scope). Paragraph references verified against the BoG CRFR Directive, November 2024. The coverage block on every physical-risk-section return:

BoG semi-annual return (Para 48) coverage

11 · Limitations & gap register

Every board report carries this gap register — what the screen does not replace and who must close it:

11.1 Known methodological limitations

• Hazard scores mix observed, modelled, ordinal and contextual evidence; they are harmonised indices, not equal physical units.
• The independent-damage product assumes separable marginal damages; compound flood/coastal/riparian events can violate this assumption.
• The tail model and scenario projection are screening constructs, not catastrophe-model event sets.
• Property value inferred from exposure/LTV inherits valuation-date, seniority and data-quality errors.
• Missing live data degrades to SEED rather than statistical imputation; absence of a flag is never proof of safety.
• Resolution mismatch can create parcel errors near coastlines, rivers, lagoons, defences and steep terrain.
• Model uncertainty is disclosed qualitatively; predictive intervals require calibration data.

12 · Hazard & data sources (with vintage)

13 · Current version

14 · References and authoritative links

• Bank of Ghana (2024), Climate-Related Financial Risks Directive: https://www.bog.gov.gh/wp-content/uploads/2024/11/BOG-Climate-Related-Financial-Risks-Directive.pdf
• IPCC (2021), AR6 WGI Africa Regional Fact Sheet: https://www.ipcc.ch/report/ar6/wg1/downloads/factsheets/IPCC_AR6_WGI_Regional_Fact_Sheet_Africa.pdf
• IPCC (2021), AR6 WGI Chapters 11–12 and Interactive Atlas: https://www.ipcc.ch/report/ar6/wg1/
• IPCC (2022), AR6 WGII Chapter 9 Africa: https://www.ipcc.ch/report/ar6/wg2/chapter/chapter-9/
• Huizinga, de Moel & Szewczyk (2017), JRC depth-damage functions: https://publications.jrc.ec.europa.eu/repository/handle/JRC105688
• NGFS climate scenarios portal: https://www.ngfs.net/ngfs-scenarios-portal/
• NASA POWER methodology and data: https://power.larc.nasa.gov/
• NASA POWER API resolution and replacement of near-real-time meteorology: https://power.larc.nasa.gov/docs/tutorials/service-data-request/api/
• EC JRC Global Surface Water v1.4 (30 m; 1984–2021): https://developers.google.com/earth-engine/datasets/catalog/JRC_GSW1_4_GlobalSurfaceWater
• ESA WorldCover 2021: https://esa-worldcover.org/
• Copernicus DEM: https://dataspace.copernicus.eu/explore-data/data-collections/copernicus-contributing-missions/collections-description/COP-DEM
• Permanent Service for Mean Sea Level, Takoradi station 331: https://psmsl.org/data/obtaining/stations/331.php
• Ghana coastal infrastructure inventory (2023): https://doi.org/10.1007/s44218-023-00026-6
• FEMA/USACE/NIBS; Kreibich et al. (2005); Defra FD2657 and CIRIA C790 — transferred adaptation priors, not Ghana calibration.

14.1 Agricultural references and implementation status

• BOG-CRFR-2024 — Bank of Ghana, Climate-Related Financial Risks Directive (2024): https://www.bog.gov.gh/wp-content/uploads/2024/11/BOG-Climate-Related-Financial-Risks-Directive.pdf. Implementation: used in current screening.
• IPCC-AR6-WGII-AFRICA — IPCC, AR6 WGII Chapter 9: Africa (2022): https://www.ipcc.ch/report/ar6/wg2/chapter/chapter-9/. Implementation: used in current screening.
• NASA-POWER — NASA, POWER meteorology and rainfall (runtime): https://power.larc.nasa.gov/. Implementation: used in current screening.
• S2-NDVI-VCI — European Commission / ESA, Copernicus Sentinel-2 crop condition observations (runtime): https://sentinels.copernicus.eu/web/sentinel/missions/sentinel-2. Implementation: used in current screening.
• FAO-WAPOR-3 — FAO, WaPOR v3 water and land productivity database (2009–present): https://www.fao.org/aquastat/en/geospatial-information/wapor/index.html. Implementation: required enhancement; not yet used in scoring.
• CHIRPS-3 — UCSB Climate Hazards Center, CHIRPS v3 rainfall (1981–present): https://www.chc.ucsb.edu/data/chirps. Implementation: required enhancement; not yet used in scoring.
• SOILGRIDS-2 — ISRIC — World Soil Information, SoilGrids 2.0 (2021): https://docs.isric.org/globaldata/soilgrids/index.html. Implementation: required enhancement; not yet used in scoring.
• MOFA-SRID — Ghana Ministry of Food and Agriculture, SRID, Agricultural production estimates and Facts & Figures (annual): https://srid.mofa.gov.gh/node/31. Implementation: required enhancement; not yet used in scoring.
• FAO-AQUACROP — FAO, AquaCrop crop-water productivity model (current): https://www.fao.org/aquacrop/. Implementation: required enhancement; not yet used in scoring.