Climate-Risk Modelling for Banks and Insurers

Two channels, different mechanics

Physical risk arises from hazards: flood, windstorm, wildfire, drought, heat and sea-level rise, in both acute event form and chronic gradual form. It affects insurance claims, collateral values, borrower cash flows, business continuity and the insurability of particular locations.

Transition risk arises from policy, technology, market preference and litigation as economies decarbonise. It affects asset values, borrower creditworthiness in exposed sectors, stranded-asset potential and the viability of specific business models. Unlike physical risk, it is driven substantially by decisions rather than by physics, which makes it faster-moving and harder to project.

The horizon problem

Financial planning operates over three to five years; most lending is shorter; most insurance contracts are annual. Climate effects accumulate over decades. This mismatch produces a genuine analytical difficulty and a frequent rhetorical error — concluding either that climate risk is irrelevant to short-horizon decisions or that it dominates them.

• Long-dated exposures such as mortgages, infrastructure finance and life liabilities extend into the horizon where physical effects are material.
• Transition effects can repricing rapidly in anticipation, so the market impact can arrive long before the physical change.
• Annual insurance contracts reprice frequently, but the portfolio strategy, geographic footprint and reinsurance relationships behind them do not.
• Adaptation, land-use change and building standards can alter the trajectory materially, in either direction.

Scenario analysis and its interpretation

Climate scenario frameworks describe internally consistent pathways for emissions, temperature, policy and technology. They are conditional narratives: if the world follows this path, these conditions follow. They do not come with probabilities, and assigning probabilities to them is a judgement the user makes, not a property of the scenario.

This matters for how the output can be used. A percentile-based capital measure requires a probability distribution. A scenario suite without probabilities supports comparison, vulnerability identification and strategic discussion, but cannot be converted directly into a capital number without an additional and clearly labelled judgement.

1. Select a small number of scenarios spanning orderly transition, disorderly transition and limited action.
2. Translate scenario variables into portfolio-relevant drivers: hazard frequency changes, carbon prices, sector demand, energy costs.
3. Apply those drivers to counterparty and exposure data at the greatest granularity the data supports.
4. Report relative results — which portfolios, sectors and locations are most affected — before reporting absolute loss estimates.
5. State explicitly what is modelled, what is assumed and what is omitted.

Data is the binding constraint

Hazard models have improved substantially. Exposure data generally has not. Address-level geocoding accuracy, construction type, elevation relative to local flood defences, and the actual operating locations of corporate borrowers are frequently unknown or recorded at a granularity too coarse for hazard modelling to be meaningful.

• Geocoding to postcode rather than to address can move a property across a flood boundary.
• Counterparty emissions data is often estimated rather than reported, and indirect value-chain emissions are estimated with wide error.
• Sector classifications are too coarse to distinguish a firm actively transitioning from a direct competitor that is not.
• Improving exposure data usually yields more than refining the hazard model, and should be prioritised accordingly.