Cookbook



back

Foundation Cookbook

A systematic comparison framework for physical-risk models — what they measure, how they translate climate hazards into financial loss, and where they disagree.

Version 2.4Updated April 2026
Maintained by PARC Foundation

Introduction

Why a cookbook? · 4 min read

Physical-risk models are heterogeneous by design. Each is built for a specific business question — pricing reinsurance, stress-testing a loan book, screening a sovereign portfolio — and quietly bakes in assumptions about hazard intensity, exposure granularity, vulnerability, and financial translation. Comparing their outputs without understanding these choices leads to misleading conclusions.

This cookbook breaks every model in the Pladifes Models Base into a common four-layer recipe so you can read across them. Each section explains what the layer should answer, what choices model authors typically make, and how those choices propagate into the final loss estimate.

How to use it. Pair this guide with the Models Base: the model card for every entry mirrors the four layers below, so you can move from theory to practice in two clicks.

Taxonomy

A common four-layer recipe

We decompose every model into four layers. The hazard layer says what climate signal is used; the exposure layer says what assets are at risk; the vulnerability layer relates intensity to physical damage; the financial layer translates damage into a P&L impact.

Layer What it answers Typical inputs
Hazard How intense, how often, where? CMIP6, ERA5 reanalysis, peril catalogues
Exposure What's at risk and where is it? Asset registers, coordinates, market values
Vulnerability How much damage at a given intensity? Damage curves, engineering studies
Financial What does the damage cost the holder? Cash flows, insurance terms, recovery

Hazard layer

From climate signal to peril intensity

The hazard layer turns a climate scenario (typically an SSP from CMIP6) into a spatial field of intensities at a given return period. Models differ on three axes: which climate ensemble they read, which downscaling they apply, and which return periods they expose to the financial layer.

Common choices

Most asset-management-oriented models lean on bias-corrected CMIP6 outputs at 25–50 km resolution. Insurance peril models often use proprietary downscaled catalogues calibrated to historical loss data, which improves short-return-period accuracy at the cost of opacity.

I(x, t, p) = D ( μ_GCM(x, t) ) , p ∈ {1, 5, 10, 50, 100, 250} years

Exposure & vulnerability

Linking assets to damage

An exposure dataset is only useful if it resolves to the spatial scale of the hazard. Coarse exposures (country-level GDP) under fine hazard fields produce smooth, biased aggregate losses; fine exposures (asset-level coordinates) under coarse hazards produce confident-looking but spurious tails.

Vulnerability is where models diverge most. Engineering damage curves anchor losses in physical reality but extrapolate poorly outside calibration ranges. Statistical curves fit to insurance data are robust within their domain but proprietary, jurisdiction-specific, and silent on adaptation.

Financial translation

From damage to P&L

Translating physical damage into a financial loss depends on the holder. A property insurer applies policy terms (deductibles, limits, reinsurance). A bank lender applies LTV and recovery assumptions. An asset manager projects damage onto issuer cash flows and revaluates discounted equity. Models that conflate these channels are not portable across mandates.

Holder Translation Key uncertainty
P&C insurer Policy terms · Reinsurance treaties Demand surge · Cat treaties
Bank lender LTV impact · Recovery rates Sovereign backstop · Insurance penetration
Asset manager Cash-flow shock · Equity revaluation Pass-through · Adaptation capex

Cross-model comparison

Where consensus ends

For the same SSP2-4.5 portfolio, top-down and bottom-up models can disagree by a factor of 3 on 2050 expected loss — and by an order of magnitude on the 99.5th percentile. The disagreement is structural, not numerical: top-down models miss spatial concentration, bottom-up models miss macro feedbacks.

Rule of thumb. Use top-down models for headline figures and stress narratives. Use bottom-up models for site-specific decisions. Never average them — average the assumptions, not the outputs.

Known limitations

What none of these models do well — yet

Cascading hazards (drought triggering wildfire triggering flood on burned soil), tipping points, and adaptation feedbacks are largely absent from production models. Compound losses are linearly summed. Adaptation is exogenous. Sovereign default in the wake of multi-peril years is not modelled. Read your outputs with these blind spots in mind.