Why NGFS scenarios matter for climate risk modelling
The Network for Greening the Financial System (NGFS) scenarios have become the reference framework for climate risk modelling across banks, insurers and asset managers. They provide a common vocabulary — Orderly, Disorderly, Hot House World, Fragmented World — for pricing physical and transition risk into balance sheets.
Translating those macro pathways into concrete financial impacts is where most institutions struggle. This guide sets out a pragmatic pipeline that separates hazard-specific physical risk functions from transition pathways, and layers explainable AI on top for governance.
Step 1: Anchor on the scenario, not the score
Pick an NGFS phase and vintage explicitly. Version the following inputs together:
- GDP, inflation and policy paths for each region
- Carbon price trajectories by sector
- Sector emissions and technology mix projections
- Global mean temperature and hazard drivers feeding physical models
Storing these as a versioned scenario bundle avoids the classic error of mixing macro variables from one vintage with hazard maps from another.
Step 2: Hazard-specific physical risk functions
Physical risk cannot be collapsed into a single elasticity. Build a separate loss function per hazard, calibrated against published damage curves:
| Hazard | Data anchor | Loss function shape |
|---|---|---|
| River and coastal flood | Depth-damage curves by asset class | Piecewise linear in flood depth |
| Wind | Peak gust vs. structural vulnerability | Power-law with sector coefficients |
| Heat | Wet-bulb temperature vs. productivity | Sigmoid with sector thresholds |
| Wildfire | Burn probability and proximity | Threshold with severity multiplier |
Each function should ship with its calibration dataset, fit statistics and an uncertainty envelope. Regulators expect to see the working, not just the number.
Step 3: Transition impacts by sector
Transition risk is best modelled as a sector-level cash-flow adjustment driven by carbon price and demand shifts. For each sector:
- Map production units to Scope 1 and Scope 2 emissions intensity.
- Apply the NGFS carbon price to derive a per-unit cost delta.
- Adjust demand using the scenario's technology mix (for example, EV penetration or renewables share).
- Roll the resulting margin impact into a discounted cash-flow revaluation.
Keep sector coefficients transparent — a single opaque neural net that maps scenarios to P&L will not survive validation.
Step 4: Explainable AI as the differentiator
Generic catastrophe modelling software treats climate as a black box. An explainable AI approach for climate risk modelling adds three things:
- Feature attribution for every loss estimate, so users can see which hazard, sector or region drove the result
- Champion-challenger benchmarking between the AI-driven model and a transparent baseline (for example, a linear damage function)
- Drift monitoring across scenario vintages, flagging when re-calibration is needed
This is where AI adds value without breaching model risk expectations under SR 11-7, the EBA model risk guidelines and the EU AI Act.
Step 5: Validation regulators will accept
Every NGFS translation pipeline should be shipped with:
- Backtesting against historical loss events where feasible
- Sensitivity tables across scenario choice and key parameters
- Cross-checks against academic and peer benchmarks
- A documented override log for any expert-judgement adjustments
Closing thought
NGFS scenarios are the input, not the answer. A defensible climate risk modelling stack combines hazard-specific physical functions, transparent transition pathways and explainable AI — with governance artefacts that hold up under supervisory review.