Resilience — Direct Physical Loss
How the Resilience module prices the direct physical impact of climate and geophysical hazards on a site — property damage and business interruption — and unifies it with the nature-dependency Stress Test into one site-, company- and portfolio-level risk model.
Resilience — Direct Physical Loss
Work in Progress — Pilot Phase. The Direct Physical Loss engine is under active development alongside the Nature Stress Test. The methodology below reflects the current build; damage curves and hazard coverage are progressively extended and validated, and outputs may change.
Purpose
The Nature Stress Test prices the loss a site faces when the ecosystems it depends on degrade. It says nothing about the direct physical hit of a hazard on the asset itself — a flood damaging the building, a cyclone tearing off the roof, an earthquake cracking the structure. Direct Physical Loss (DPL) closes that gap: for each hazard, at a site's location, it crosses the hazard intensity with a vulnerability (damage) curve specific to the asset type, turning intensity into a percentage of loss.
The Resilience module unifies the two. The nature-dependency risks (Nature Stress Test) and the direct physical risks (DPL) become one risk model, expressed on the same quantities — a fraction of the site's revenue, and for physical damage a fraction of its asset value — so both can be read, ranked and aggregated together at site, company and portfolio level.
At a glance
| Unit of analysis | Site (asset), aggregated to company and portfolio |
| Risk drivers | 11 physical hazards (DPL) + 25 ecosystem services (Nature Stress Test) = 36 |
| Loss channels | Property damage (% of asset value) · Business interruption (% of revenue) — two figures, never summed |
| Output | Average Annual Loss (AAL) per driver, as a level and a delta to baseline |
| Scenarios / horizons | Optimistic / Pessimistic × 2035 · 2050 · 2080 |
| Sign convention | negative = loss ; positive = a reduction in loss vs baseline |
How it works — from intensity to an annual loss
For each hazard, at the site's coordinates:
1. Intensity. The hazard intensity is sampled from a georeferenced hazard raster — flood water depth (m), cyclone wind (m/s), peak ground acceleration (g), and so on — read per return period and per scenario × horizon.
2. Vulnerability curve. The intensity is passed through a calibrated damage curve, keyed on the site's asset type (SiteType — offices, warehouse, factory, cropland, hydropower…), giving a conditional loss: the fraction lost given that this event occurs. Curves are re-implemented from published science (see The eleven hazards).
3. Annualisation — the AAL. A single event's damage is not the risk; the risk is the expected loss per year. The conditional losses are integrated over the hazard's annual exceedance probability p = 1/RP to give the Average Annual Loss:
AAL = ∫ D(intensity(p)) dp over p = 1/RP, p ∈ [0, 1]
The result is dominated by frequent, moderate events, not the damage of one rare design event — a ~30 % conditional loss on a 1-in-100-year flood becomes a small percent-per-year AAL. For a hazard with a single published hazard-curve anchor (earthquake), the exceedance curve is reconstructed analytically (PSHA log-linear). For chronic hazards with no discrete event (drought, extreme heat, subsidence), a frequency × severity bridge replaces the integral.
Flood defences. The open flood hazard is undefended — it floods a site at every return period. Where real defences exist (dikes, sea walls), the AAL integral is truncated at the defence's design exceedance probability (a defence holds everything more frequent than its design event), using the global FLOPROS protection-standard database. Each arm reads its own standard — the merged riverine standard truncates the fluvial arm, the coastal one the coastal arm, never crossed: St. Petersburg reads RP10 000 coastal against RP1 000 riverine, so using the riverine field there would understate the sea defence tenfold. A FLOPROS 0 means "unknown", not a zero-year defence, and degrades to undefended — as does an absent polygon, which is the common case on the coastal field (612 polygons of 4 647), so most coastal sites still integrate undefended. A single knob controls how conservatively an existing standard is read.
Two loss channels, never summed
Every hazard produces two distinct figures on two different denominators:
- Property damage (PD) — a fraction of asset value: the one-off value physically destroyed. A stock.
- Business interruption (BI) — a fraction of one year's revenue: output lost while the site is stopped or throttled. A flow, derived from the damage ratio through a restoration-time (loss-of-function) relation (
revenue lost = downtime days / 365, capped at one year), gated to the revenue-bearing occupancies.
Adding a stock to a flow mixes units, so PD and BI are kept separate; any combination happens downstream in the risk engine, never here.
The eleven hazards
Each hazard's intensity source, its published damage curve, and the channels it prices. Each links to its detailed layer fact sheet.
| Hazard | Intensity source | Damage curve (public source) | Channels |
|---|---|---|---|
| Flood (river + coastal) | GIRI / Deltares–Aqueduct depth | JRC continental depth–damage (Huizinga 2017) | PD + BI |
| Cyclone | STORM return-period wind | Emanuel (2011) excess-wind sigmoid | PD + BI |
| Earthquake | GSHAP peak ground acceleration | HAZUS lognormal fragility | PD + BI |
| Wildfire | Fire Weather Index × observed burn probability | FireLossRate (Nicoletta 2023 / Abo El Ezz 2022), on the Van Wagner (1987) FWI→intensity relation | PD + BI |
| Hail | Hail climatology | Schmid (2024) | PD + BI |
| Extreme heat | WBGT (from CMIP6 temperature) | ILO / Kjellström work-capacity | BI |
| Extreme precipitation | RX5day excess over drainage capacity | Cubic-in-excess (pluvial proxy) | PD + BI |
| Landslide | NASA susceptibility | Susceptibility × activation; the class→loss shape is calibrated on the NASA Global Landslide Catalog (Uzielli 2008, V = I × S) | PD + BI |
| Clay shrink-swell | Expansive-soil susceptibility | Susceptibility × wet–dry cycle | PD + BI |
| Subsidence | Herrera-García susceptibility | Susceptibility (chronic creep) | PD + BI |
| Drought | WRI Aqueduct water stress | Shortfall probability × conditional loss, scaled by water dependence; a site that reports its water indicators (m³/yr) is priced on a quantitative water-criticality instead (FAO-33 Ky, van Vliet 2016) | BI |
Calibrated vs proposal. Six hazards run on a published curve end to end and are tagged built — flood, cyclone, earthquake, hail, wildfire and extreme heat. Five carry at least one first-cut bridge and are tagged proposal — extreme precipitation, the geotechnical trio (landslide, clay shrink-swell, subsidence) and drought — so they can be excluded from a strictly-calibrated total. The tag sits on the weakest link, not the whole chain: the landslide class→loss shape is measured against the NASA Global Landslide Catalog, and it is the absolute scale and the activation frequencies that remain posed.
Scenarios, horizons and the baseline delta
Each figure is a snapshot at each horizon (2035 · 2050 · 2080), per scenario — never cumulated over time and never financially discounted. Each snapshot is a deviation from the baseline (today):
DPL(horizon) = AAL(horizon) − AAL(baseline)
expressing the loss as a delta keeps it homogeneous with the nature-dependency path, which is 0 at baseline by construction. But a delta reads zero on already-saturated or time-invariant hazards — a site already at extreme flood depth, or earthquake / hail / subsidence, which have no forward trigger (Δ ≈ 0). So alongside the delta the module also surfaces the level (absolute AAL at each horizon), which keeps those high-risk sites visible. Both views are delivered. Time-varying hazards resolve the requested scenario onto the nearest published forcing at or above it, so a hazard published on a coarser scenario set still answers the optimistic/pessimistic toggle rather than silently falling back to baseline.
The Resilience unification
The Resilience module carries 36 risk drivers — the 11 DPL hazards plus the 25 SEEA-EA ecosystem services. The nature side is re-keyed onto the service (never the ecosystem component): a site depends on the service the components deliver, not on the raster input.
Where a physical hazard and an ecosystem-service failure are two estimates of the same loss on the revenue channel, they are arbitrated per channel — the DPL estimate wins (a loss priced through a calibrated vulnerability curve takes precedence over one reached by a dependency proxy), never summed and never double-counted; the nature estimate is retained as corroboration. Where the DPL prices nothing on a channel, the nature estimate stands alone. On the asset-value channel there is no overlap — only the DPL produces property damage.
Arbitration applies only where the two price the same loss, and that is narrower than it looks. Water was the real overlap — the drought curve and the nature water-stress axis were two transforms of a single Aqueduct field — and it was resolved by taking water out of the nature side entirely, not by arbitrating it. Conversely the four ground-and-rain hazards (landslide, clay shrink-swell, subsidence, extreme precipitation) each carry a business-interruption channel that is not arbitrated against a neighbouring nature axis: the physical engine prices a discrete event (the slope lets go, the site stops for the works) while the dependency path prices the chronic degradation of a service (erosion, siltation, soil loss), and the two read entirely disjoint inputs. Two measurements of two distinct processes coexist; only two readings of the same field are arbitrated.
Scope & limitations
- Public literature only. Every damage curve is re-implemented from published science; no proprietary curves.
- Undefended by default. Flood defences are applied through FLOPROS design standards (a truncation of the AAL integral), per arm and read conservatively via a single knob; taken at face value until calibrated. Coverage is partial on the coastal standard, so most coastal sites are still read undefended — and therefore over-read.
- Proposal bridges. Drought, extreme precipitation and the geotechnical activation terms are first-cut and tagged, not yet calibrated. Extreme heat is not among them — it runs on the published ILO / Kjellström work-capacity relation.
- Sites only. The loss is priced at the site; value-chain (upstream/downstream) diffusion is not modelled.
- Asset typology. Curves key on Darwin's
SiteType; the classification can be extended in later versions.
Methodological grounding
The approach mirrors regulator-grade physical-risk practice — AAL over return periods as in catastrophe modelling, HAZUS fragility and loss-of-function, JRC continental depth–damage, Emanuel cyclone vulnerability — and is aligned with ECB, NGFS and TNFD climate- and nature-risk framing. Per-hazard sources and derivations are on each layer's fact sheet.