At 5:46am on 17 January 1995, a magnitude 6.9 earthquake ruptured a fault directly beneath Kobe, Japan. It lasted fourteen seconds. When it ended, 6,434 people were dead, 300,000 were displaced, and approximately USD 100 billion of economic damage had been inflicted on one of the world's most earthquake-prepared nations. The fire following earthquake β triggered by ruptured gas lines across the wooden residential districts β burned for two days and destroyed entire neighbourhoods that the shaking itself had left standing.
What made Kobe so devastating was not the magnitude β a M6.9 event is significant but not extreme. It was the combination: a blind fault that no one had adequately mapped, a shallow rupture directly beneath a dense urban area, older timber construction that predated Japan's post-1981 seismic code, poor soil conditions amplifying the ground motion in low-lying districts, and the secondary fire that the emergency services could not contain because water mains had fractured. The cat model that existed in 1995 captured the physical shaking. It did not capture the fire. It did not adequately capture the soil effects in the older districts. The insured loss was USD 3 billion β roughly 3% of the economic loss. The protection gap was staggering.
Earthquakes are the sudden release of elastic strain energy stored in the Earth's crust along fault zones. Tectonic plates move continuously β millimetres to centimetres per year β driven by convection in the mantle below. At plate boundaries and within plates, faults accumulate stress as rocks on either side of the fault are dragged in opposite directions. When the accumulated stress exceeds the frictional resistance on the fault plane, the fault ruptures. The stored elastic energy releases as seismic waves that radiate outward in all directions from the rupture zone.
The point where rupture initiates is the hypocenter (focus), which lies at depth within the crust. The point on the Earth's surface directly above is the epicenter. For large earthquakes, rupture propagates along a fault segment that may extend for tens to hundreds of kilometres β the geometry, direction, and speed of this propagation all shape the ground motion pattern at the surface.
Earthquake magnitude measures the total energy released at the source. The Moment Magnitude Scale (Mw) is the standard used in modern seismology and cat modelling. It is logarithmic: each unit increase corresponds to approximately 32 times greater energy release. A Mw 7.0 earthquake releases roughly 1,000 times more energy than a Mw 5.0. Magnitude is a source property β a single number for the entire event. It is not what damages buildings. Intensity is.
Intensity is the severity of shaking at a specific location β and it is what the hazard module calculates. The same magnitude earthquake can produce very different intensities at different sites depending on distance from the fault, depth of the rupture, fault mechanism, local geology, and soil conditions. Two buildings in the same city can experience intensities differing by a factor of three or more from the same earthquake. The primary intensity metric in cat modelling is Peak Ground Acceleration (PGA) β the maximum acceleration the ground experiences during shaking, expressed as a fraction of gravitational acceleration (g).
The type of faulting shapes the pattern and severity of ground motion. Three primary mechanisms occur:
Two sides of the fault move horizontally past each other. Produces relatively moderate ground motion relative to magnitude. Dominates in transform boundaries.
One side rides up over the other β compressional. Produces stronger ground motion than strike-slip at the same magnitude. Subduction zone earthquakes are thrust events.
Rock on one side drops relative to the other β extensional. Generally produces lower ground motion than thrust at the same magnitude.
The earthquake hazard module has a more complex scientific foundation than most other perils. It must integrate information from multiple disciplines β seismology, geology, geodesy, and geotechnical engineering β to produce a stochastic event catalogue and assign ground motion intensities at every point on the hazard grid.
The stochastic catalogue for earthquake is built by characterising every significant fault and seismogenic zone in the model domain. For each source, the model estimates three things: the maximum magnitude the fault can produce, the rate at which earthquakes of different magnitudes occur (the magnitude-frequency distribution), and the spatial distribution of rupture scenarios.
The magnitude-frequency distribution typically follows the Gutenberg-Richter relationship β an empirical regularity observed in virtually every seismically active region: the logarithm of earthquake frequency decreases linearly with magnitude. There are roughly ten times as many magnitude 5 earthquakes as magnitude 6, and ten times as many magnitude 6 as magnitude 7. This relationship, calibrated to the historical and geological record for each fault system, produces the annual rates assigned to each synthetic event in the catalogue.
Rate estimates come from three data sources used together:
Once the stochastic catalogue defines what events occur and at what rates, the hazard module must calculate the ground motion intensity those events produce at every location on the grid. This is done using Ground Motion Prediction Equations β mathematical relationships that estimate PGA (or spectral acceleration at different periods) as a function of earthquake magnitude, source-to-site distance, fault mechanism, and local site conditions.
GMPEs are derived from regressions on large databases of recorded ground motions from past earthquakes. They are the primary source of epistemic uncertainty in earthquake hazard modelling β different scientific teams using similar datasets derive meaningfully different equations, and the choice of GMPE is one of the largest single drivers of vendor model divergence for earthquake.
For the same earthquake scenario, different GMPEs can predict PGA values differing by a factor of 2β3 at distances beyond 50km from the fault β precisely the range that matters most for urban exposure in high-seismic regions. Two vendors calibrating to the same instrumental catalogue but using different GMPEs will produce meaningfully different hazard surfaces for the same portfolio. This is not error β it is epistemic uncertainty: both GMPEs are defensible representations of the physical process given available data. The divergence is most pronounced at long distances, in regions with thin ground motion records, and for subduction zone events where the data is most sparse.
Seismic waves behave very differently in different soil types. Hard rock transmits ground motion with relatively little modification. Soft unconsolidated soils β alluvial sediments, reclaimed land, river deltas β can amplify ground motion dramatically relative to bedrock. The amplification factor depends on the soil's shear wave velocity (characterised as Vs30, the average shear wave velocity in the top 30 metres), the depth to bedrock, and the frequency content of the incoming waves.
// Vs30 = average shear wave velocity in top 30m. Amplification varies with frequency and input intensity. Mexico City lakebed represents an extreme case of soft soil resonance amplification.
The vulnerability module converts the ground motion intensity calculated by the hazard module into a damage ratio for each building. Earthquake vulnerability is more sensitive to construction characteristics than almost any other peril β the difference between an unreinforced masonry building and a modern seismically designed reinforced concrete frame can produce damage ratios differing by a factor of 5 or more at the same ground motion intensity.
Earthquake vulnerability functions take the form of fragility curves β probability distributions that describe the likelihood of reaching or exceeding specific damage states as a function of ground motion intensity. For a given PGA or spectral acceleration value, the damage function returns the mean damage ratio and its associated uncertainty distribution. The shape of the curve is characteristic of the building type: brittle materials like unreinforced masonry have steep fragility curves that rise quickly with intensity; ductile systems like modern RC shear wall have flatter curves that rise more gradually and plateau at lower damage ratios.
The direct structural damage from ground shaking is only part of the earthquake loss. Several secondary perils can generate losses comparable to or exceeding the primary shaking loss, yet are either not modelled or only partially modelled in standard commercial cat model releases.
Saturated sandy soils lose shear strength during shaking, behaving temporarily like a fluid. Buildings settle, tilt, or sink. Infrastructure β pipelines, utilities, roads β fails catastrophically. Widespread in river deltas, coastal reclaimed land, and areas with high water tables.
Large subduction zone earthquakes can generate ocean waves that arrive at coastlines with little warning. Inundation depths can reach 10β40m in channelling bays and fjords. Coastal infrastructure and low-lying communities face total loss. The hazard zone extends far inland along low-gradient coastlines.
Ruptured gas mains and downed electrical lines ignite fires that spread through areas where water pressure has failed due to broken mains. In densely built wooden construction areas, fire-following-earthquake can cause more damage than the shaking itself.
Ground shaking triggers slope failures in mountainous terrain. Roads and infrastructure are severed, communities are buried, and rescue access is blocked. Economic losses can be significant in Andean, Himalayan, and Alpine seismic zones.
Earthquake insurance penetration varies more dramatically than any other major peril. In Japan, most commercial property carries earthquake coverage, and a government-backed residential earthquake scheme exists β yet the protection gap from the 2011 TΕhoku disaster still exceeded 70% of economic loss. In Turkey, despite mandatory catastrophe insurance through TCIP, large swaths of commercial and industrial property remain uninsured against earthquake. In many developing markets β Bangladesh, Pakistan, Nepal, Central America β earthquake insurance penetration is negligible despite very high physical hazard.
The coverage structure for earthquake varies significantly by market:
Demand surge is a significant non-modelled item for earthquake. After a major urban earthquake in a concentrated geography, the simultaneous demand for structural engineering assessments, demolition, and construction materials drives costs well above normal market rates β typically 20β40% above pre-event levels for large concentrated events. Standard cat model outputs are calibrated to normal repair cost schedules. The demand surge component of the final insured loss consistently causes actual losses to exceed initial modelled estimates.
| Region | Hazard Character | Insurance Significance | Modelling Maturity |
|---|---|---|---|
| Japan | Complex multi-source hazard β Pacific and Philippine plate subduction, plus crustal faults. Near-field and far-field events both significant. Tsunami risk on Pacific-facing coasts. | Largest single earthquake insurance market globally. High insured asset density across a compressed geography. Annual earthquake premiums among the highest per capita in the world. | Very high. Extensive ground motion records, detailed fault databases, frequent model updates post-major events. Best-calibrated earthquake models in the world. |
| California (US) | Multiple parallel fault systems β San Andreas, Hayward, Newport-Inglewood, and others. Shallow crustal events dominate. Thrust faults beneath the LA Basin create blind fault risk. | Very high insured value concentration in San Francisco Bay Area and Greater Los Angeles. Low residential penetration (typically 10β15%) creates adverse selection in the commercial book. | Very high. UCERF3 fault model, extensive strong motion network, well-calibrated GMPEs. One of the two or three most mature earthquake model domains globally. |
| Turkey | North Anatolian Fault dominates β a right-lateral strike-slip fault that has produced a westward-propagating sequence of major earthquakes across the 20th century. East Anatolian Fault system produces a separate hazard domain. | Growing insured market. TCIP provides mandatory residential coverage but commercial penetration is uneven. Urban concentration in Istanbul creates extreme accumulation risk for global (re)insurers. | Good and improving. Major investment in model development post-1999 Marmara and 2023 KahramanmaraΕ events. Vendor model divergence at the tail remains significant. |
| New Zealand | Complex plate boundary with both subduction and strike-slip faulting. Alpine Fault capable of Mw 8+ events. Multiple crustal fault systems throughout both islands. | EQC government scheme provides first NZD 300K of residential earthquake cover. Commercial market exposed above EQC cap. Canterbury 2010β11 sequence reshaped the domestic market permanently. | Good. Canterbury sequence provided calibration data that significantly improved vulnerability functions for New Zealand building stock. Ongoing seismic network investment. |
| Chile / Peru / Ecuador | Nazca Plate subduction under South America β one of the world's most seismically active convergent margins. Capable of Mw 9+ megathrust events along the full length of the coast. | Growing insured market, particularly in Chile which has higher insurance penetration than its neighbours. Significant industrial and mining exposure. Lima represents a major uninsured accumulation risk. | Moderate. Chile well-modelled. Peru and Ecuador less mature. Historical instrumental record shorter than needed for confident tail calibration. |
| Middle East / Iran | Active collision zone between Arabian and Eurasian plates. Multiple fault systems throughout Iran, with Tehran sitting on several active faults. Gulf of Aqaba and Dead Sea Transform also active. | Low insurance penetration in Iran. Growing energy and infrastructure exposure across the Gulf states. Tehran represents one of the most significant uninsured earthquake accumulation risks globally. | Low to moderate. Sparse instrumental network, limited loss data. Significant epistemic uncertainty in hazard models. |
Some of the most damaging urban earthquakes β the 1994 Northridge event in Los Angeles, the 1995 Kobe event in Japan β ruptured faults that were unknown or poorly characterised before the earthquake. A fault that is not in the model's source database contributes nothing to the hazard at nearby locations. Urban development has historically occurred without complete knowledge of subsurface fault geometry, and even well-studied regions continue to discover new active fault segments. This is an irreducible limitation of the current state of geological knowledge.
Soil amplification can modify ground motion intensity by a factor of 3β10, yet site-specific Vs30 measurements are unavailable for most locations outside the United States, Japan, and a handful of other well-instrumented countries. Models apply proxy Vs30 data derived from topographic slope or geological mapping β coarser and less accurate than direct measurement. For urban areas on soft soils β coastal cities, river delta cities, reclaimed land β this gap introduces systematic underestimation of hazard at the most vulnerable locations.
Standard stochastic catalogues assume seismic events are statistically independent β the occurrence of one earthquake does not change the probability of another in the near term. In reality, major earthquakes increase stress on adjacent fault segments and trigger aftershock sequences that can include damaging mainshock-level events. The Canterbury 2010β11 sequence in New Zealand β five damaging events over 18 months β illustrates that standard independence assumptions can materially underestimate the probability of clustered large losses within a single policy year.
Tsunami, fire-following-earthquake, and landslides are either excluded or only partially modelled in most standard commercial cat model releases. For portfolios with coastal Japan exposure, the tsunami component of a major subduction zone event can exceed the ground shaking component. For portfolios concentrated in densely built urban areas with older wooden construction, fire-following-earthquake adds meaningful uncertainty. Presenting earthquake model output without clearly noting what secondary perils are and are not included is a significant presentation gap.
Damage functions are calibrated to post-earthquake damage surveys from countries with adequate data collection infrastructure β primarily the US, Japan, and Europe. In markets like Iran, Bangladesh, Pakistan, Nepal, and much of Central America, local building stock has significantly different construction practices, material quality, and code compliance than the buildings these functions were calibrated to. Applying Western-calibrated vulnerability functions to non-Western building stock introduces unknown directional bias β most likely underestimating damage for markets with lower construction quality than the calibration data represents.
The standard scale for earthquake magnitude in modern seismology and cat modelling. Logarithmic: each unit increase represents ~32Γ more energy release. A source property β one number for the whole event, not a measure of local shaking severity.
The primary earthquake intensity metric in cat modelling. Measured in fractions of gravitational acceleration (g). A site-specific property β varies across the affected area depending on distance, local geology, and soil conditions. The input to earthquake damage functions.
The hypocenter (or focus) is the point within the crust where fault rupture initiates. The epicenter is the point on the Earth's surface directly above. For large earthquakes, rupture propagates along a fault segment extending tens to hundreds of km from the hypocenter.
Ground Motion Prediction Equation. Mathematical relationship estimating PGA or spectral acceleration as a function of magnitude, distance, fault mechanism, and site conditions. The primary source of epistemic uncertainty in earthquake hazard modelling and a major driver of vendor model divergence.
Average shear wave velocity in the top 30 metres of the soil column. The primary proxy for local site amplification in earthquake hazard models. Low Vs30 values (soft soils) produce higher amplification; high values (rock) produce less amplification relative to bedrock input.
Empirical relationship between earthquake frequency and magnitude: the logarithm of earthquake occurrence rate decreases linearly with magnitude. Calibrated to each fault system and seismogenic zone to produce the annual rates assigned to events in the stochastic catalogue.
Loss of shear strength in saturated sandy soils during earthquake shaking, causing the soil to behave temporarily like a fluid. Produces settlement, tilting, and lateral spreading of structures. Common in river deltas, coastal reclaimed land, and areas with high water tables. Often a non-modelled item in standard commercial earthquake releases.
Secondary peril in which fires ignited by ruptured gas mains and downed electrical lines spread through areas where water pressure has failed. Historically significant in Japan and California. Partially modelled in some vendor earthquake releases; the calibration carries wide uncertainty.
The study of past earthquakes recorded in geological evidence β offset geological features, fault scarps, liquefaction deposits, and sedimentary disruption. Extends the earthquake record beyond the instrumental period, providing recurrence interval data for rare large events.
Ground motion intensity at a specific structural period, as opposed to PGA which is the peak value across all periods. SA at the building's natural period is more relevant than PGA for taller structures whose response is dominated by lower-frequency motion. Used in damage functions for multi-storey buildings.