GRA · AI and Financial Model Risk
Spatial financeWhere capital meets hazard, exposure and transition
Stress and scenario models
Stress and scenario models project how losses, capital and liquidity would change under adverse conditions. They are used for supervisory stress tests, internal capital assessment and climate scenarios. Because they model rare events with little data, they carry more uncertainty than everyday models. This page sets out how stress models fail, how they are validated, and who decides how their results are used.
Stress and scenario models on the record.
Stress and scenario models: how the risk develops.
Verification and uncertainty. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.
Few data points for extremes
Severe recessions and market crashes are rare, so the relationships linking scenarios to losses are estimated from few observations and may not hold in the next crisis.
Chains of models compound error
Scenario variables pass through satellite models for each portfolio, then aggregation. Errors at each link compound, and the end result looks more precise than it is.
Results drive capital
Where stress results set capital buffers, there are incentives to choose favourable assumptions. Supervisory models and challenge counter this.
Behaviour is often fixed
Many stress models hold balance sheets and management actions static, missing the reactions that amplify or dampen losses.
Transition and physical risk on a portfolio.
Three scenario families after the NGFS framework move a sector portfolio through carbon costs and warming damages; the transmission view follows the losses to lenders, investors, public finances and households. Scenario: named assumptions about the future, not a forecast.
Controls
- Carbon price at the horizon, orderly path
- 200 per tCO2
- Delay before a disorderly transition
- 50% of horizon
- Cost passed to customers
- 50%
- Warming today above pre-industrial
- 1.30°C
- Warming per decade, current policies
- 0.25°C
- Damage coefficient
- 0.24% per °C²
- Years ahead
- 25 years
Readouts
- Worst scenario at the horizon
- −33.6% of value
Numbers, methods and sources
| Year | Orderly transition (% value) | Delayed, disorderly transition (% value) | Current policies (% value) |
|---|---|---|---|
| 0 | −0.38 | −0.38 | −0.38 |
| 3 | −2.87 | −0.53 | −0.67 |
| 6 | −5.36 | −0.68 | −0.97 |
| 9 | −7.84 | −0.84 | −1.27 |
| 12 | −10.33 | −0.99 | −1.57 |
| 15 | −12.82 | −7.12 | −1.87 |
| 18 | −15.31 | −15.05 | −2.17 |
| 21 | −17.80 | −23.00 | −2.48 |
| 24 | −20.29 | −30.94 | −2.79 |
Metodes
- Climate scenario families NGFS (2023), Climate Scenarios for central banks and supervisors, Phase IV
- Climate stress test of the financial system Battiston and others (2017), Nature Climate Change 7
- Damage function Nordhaus (2017), Revisiting the social cost of carbon, PNAS 114
- Transmission channels BCBS (2021), Climate-related risk drivers and their transmission channels
- Uncertainty fan (parameter Monte Carlo) Spiegelhalter, Pearson and Short (2011), Visualising uncertainty about the future, Science 333
- One-at-a-time sensitivity (tornado) Saltelli and others (2008), Global Sensitivity Analysis: The Primer, Wiley
Data
The model runs on the teaching values set by the controls. To calibrate it for a place, read it against these admitted layers and series, which this page’s map can draw:
- Urban extents Natural Earth
- Seaports Natural Earth
- Past year, daily temperature (ERA5 via Open-Meteo) Open-Meteo.com (CC BY 4.0), Copernicus ERA5 (C3S/ECMWF) · register open-meteo.public-pages
- Surface air temperature, daytime (AIRS) NASA AIRS · register nasa.gibs.context-live-layers
Across the six pillars, led by Nexus Consortium
- Experiments Labs
- Built in AI and Financial Model Risk in Foundry
- Tested at AI and Financial Model Risk in Universe
- Codified in Digital, Cyber and AI in Standards
- Published in AI and Financial Model Risk in Reports
- Taught in Faculty of Technology and Digital Systems in Academy
What is measured, and how well.
Each indicator carries its unit, its cadence, the kind of source behind it and the evidence label a value would carry. No values are shown: a programme supplies them under its own authority.
| Indicator | Unit | Cadence | Source type | Label |
|---|---|---|---|---|
| Stress loss projection against realised loss in past downturns | ratio | per back-test | Back-testing | ESTIMATED |
| Sensitivity of results to key assumptions | change in capital ratio | per exercise | Sensitivity analysis | MODELLED |
| Management overlays on stress results | currency | per exercise | Stress documentation | DECLARED |
| Supervisory model versus firm model differences | difference in projected losses | per exercise | Supervisory results | Published statistics |
| Stress models validated within cycle | share | annual | Validation records | DECLARED |
| Scenario variables used per model | count | per exercise | Model documentation | DECLARED |
What stress and scenario models means for exposure, loss and capital.
The exposure, loss, scenario and capital views a risk officer, supervisor or ministry of finance would ask for, with the frameworks they report against. Structures follow supervisory and actuarial practice. Values are illustrative: never prices, ratings or advice.
Scenario pathways
How stress and scenario models could be stressed under the scenario families central banks and supervisors use.
| Scenario family | Transition risk | Physical risk | What a stress test asks |
|---|---|---|---|
| Orderly | Early, gradual policy | Laer | Can the book absorb a steady repricing? |
| Disorderly | Late, abrupt policy | Lower to medium | What breaks under a sudden repricing and credit migration? |
| Hot house world | Limited policy | High and rising | Where do physical losses concentrate in collateral and insurance? |
Where the evidence comes from.
Named public datasets and institutional records, each with its licence class. Sources by agreement are used only under their terms and never republished.
| Bron | What it gives | Licence class |
|---|---|---|
| Federal Reserve stress test methodology documentation | Supervisory model descriptions | Openbare domein |
| EBA EU-wide stress test methodology | Constrained bottom-up approach and rules | Free, EU reuse terms |
| Bank of England stress testing documentation | Methods and model approach | Free, Bank of England terms |
| BCBS stress testing principles | International principles for stress testing | Free, BIS terms |
| NGFS scenarios technical documentation | Climate scenario models and limits | Free, NGFS terms |
Who decides, and what the evidence serves.
AI and Financial Model Risk prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.
| Wie besluit | The question | Evidence this page serves |
|---|---|---|
| Boards and risk committees | Whether stress results support capital and risk appetite | Results, sensitivities and validation |
| Model validators | Whether stress models are fit for purpose | Back-testing and benchmark evidence |
| Supervisors | How firm and supervisory models are compared and used | Exercise results and model reviews |
| Capital planning teams | Which overlays are justified | Overlay rationale and evidence |
Outputs this topic produces
- Guide to stress model chains and where error enters
- Method note on back-testing stress models against past downturns
- Brief on overlays and their governance
- Questions for boards using stress results
What experts ask.
The questions concern how far to trust a model of the unprecedented.
- Are the scenarios severe enough, and do they capture the risks that matter?
- How are macroeconomic or physical scenarios translated into losses, and through which sub-models?
- How much uncertainty accumulates along that chain?
- Are second-round effects, such as fire sales and feedback to the economy, included?
- How are results verified when the scenario has never occurred?
Mechanisms and drivers.
Stress tests link a scenario to outcomes through a series of models, for example from economic variables to default rates to losses to capital. Each link adds assumptions and uncertainty, and models calibrated on normal periods may not capture behaviour in extreme conditions. Climate scenarios extend horizons far beyond historical data, compounding uncertainty further.
Many stress tests consider institutions in isolation, missing interactions in which many firms’ responses amplify stress. Machine learning components add new verification challenges. Presenting results as precise numbers can create false confidence; ranges and sensitivity analysis communicate limits more honestly.
How it is studied, and where the methods stop.
| Method or data | What it shows | Limits |
|---|---|---|
| Scenario design | Plausible severe conditions | Choice of scenario shapes the answer |
| Satellite and loss models | Translation of scenarios into losses | Calibrated on limited stress data |
| Sensitivity and uncertainty analysis | Range of plausible outcomes | Cannot capture unknown unknowns |
| Reverse stress testing | Conditions that would cause failure | Identifies vulnerabilities, not probabilities |
| System-wide models | Interactions between institutions | Require behavioural assumptions |
How it is studied, and what the methods cannot say.
- 01Uncertainty quantification and model validationPrediction intervals for loss models · Model ensembles · Model risk capital or buffersApplied here
- 02Probability and statisticsCalibration testing · Drift statistics · Group outcome comparisonApplied here
- 03AI evaluation and assurance engineeringValidation of machine learning models · Red-teaming generative AI · Fairness assessmentApplied here
- 04Artificial intelligence and machine learningGradient-boosted credit models · Language models in operations · Anomaly detectionApplied here
Questions and answers.
Why are stress models riskier than everyday models?
What is a satellite model?
How are stress models validated?
Is a stress test a forecast?
Why report ranges rather than single numbers?
Does the platform run stress tests for firms?
Dra bewyse vanaf hierdie bladsy by.
Reports is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.
- EmailGateway GW-GRASend it from this pageEvidence, a correction, a case, a method, a review, a translation or profile input. It arrives tagged with this page, so the editor who owns it receives it.
- ZenodoOpen depositDeposit it and submit to the communityPublish your own record on Zenodo with your ORCID iD, then submit it to the Reports community there. Curators review it; the DOI stays yours.
- ReportsGlobal Risks AllianceSee every way to contributeThe contribution types, how each is reviewed, what credit you receive, the licence and what is never accepted.
Stuur nooit persoonlike data, sensitiewe infrastruktuurbesonderhede of enigiets wat u geen reg het om te deel nie. 'n Bydrae word voor gebruik beoordeel; die stuur daarvan verleen geen outeurskap of redaksionele regte nie, en borgskap koop nooit 'n bevinding nie.
Bring an exposure or capital question to the Alliance.
Information and enquiries are open today. An enquiry receives a reference and names the next step; applications, payments, compute allocations and registrations open only when their team, terms and review capacity exist.
Legal documents.
Maps, data and figures on this site are shown for context and general information. They are not advice, forecasts or warnings; the limits of the information are set out in the Disclaimers and Data Notices.