Climate risk is becoming a software product. Banks, insurers, governments, and supply chain teams all want to know: what could climate change do to these assets, these people, this region? If you're an engineer moving into climate tech, you'll meet this vocabulary in your first week.
The core equation
Almost every physical climate risk model is some version of:
Risk = Hazard × Exposure × Vulnerability
- Hazard: how likely and intense is the event? (A 1-in-100-year flood reaching 1.2 m.)
- Exposure: what's in the way? (Buildings, roads, crops, people.)
- Vulnerability: how much damage does the hazard do to what's exposed? (A single-storey mud-brick house vs a raised concrete building.)
Engineers often focus on the hazard model because it's the most technical. In practice, exposure and vulnerability data are frequently the weakest links.
Physical vs transition risk
| Physical risk | Transition risk | |
|---|---|---|
| Source | Climate hazards | Policy, technology, market shifts |
| Examples | Flood, heat, drought, wildfire, sea level | Carbon pricing, regulations like EUDR, stranded assets |
| Data | Climate models, hazard maps, asset locations | Policy scenarios, emissions data, sector models |
Compliance work like EUDR is a form of managing transition risk: regulation changes what supply chains can sell.
Scenarios, not predictions
Climate risk models don't predict the future. They test several:
- IPCC Shared Socioeconomic Pathways (SSPs) describe emissions and development pathways.
- NGFS scenarios (Network for Greening the Financial System) are widely used in finance for both physical and transition risk.
A good model shows results across scenarios and time horizons (e.g. 2030, 2050) so users see the range, not one number.
Downscaling
Global climate models are coarse. Turning them into street-level flood depth or farm-level drought risk requires downscaling:
- Statistical downscaling maps coarse outputs to local observations.
- Dynamical downscaling runs higher-resolution regional models.
- ML downscaling is an active research area, learning fine-scale patterns from data.
Every step adds uncertainty. Be careful about showing a precise-looking map at 10 m resolution built on 100 km inputs.
How a risk platform is built
From a software perspective, most platforms look like:
- Data ingestion: hazard layers (rasters), climate model outputs (NetCDF/Zarr), asset data (geospatial vectors).
- Geospatial join: which assets sit in which hazard cells.
- Damage functions: vulnerability curves translate hazard intensity into loss.
- Aggregation: by portfolio, region, scenario, horizon.
- Presentation: maps and reports. See Building a Climate Data Dashboard.
Tools you'll meet: xarray, Zarr, PostGIS, GDAL, and cloud object storage.
Where engineers get it wrong
- False precision: decimal places that hide wide uncertainty.
- Ignoring vulnerability: treating exposure as risk.
- One scenario: presenting a single future as "the" answer.
- Static models: not updating as hazard data and assets change.