MODULE 02 · INSIDE THE CAT MODEL: THE FOUR COMPONENTS

The Vulnerability Module

📖 ~25 min · Lesson 6 of 15 · 3 micro-checks + quiz · Foundation Level · Free

Why the Same Storm Destroys One Building and Spares the Next

The 2011 Christchurch earthquake (Mw 6.2) struck at shallow depth beneath a densely built city. The peak ground acceleration in the CBD reached 0.7g, strong enough that building collapse was widespread. But two buildings standing 80 metres apart on the same street produced very different outcomes. One, an older five-storey unreinforced masonry structure built in the 1930s, partially collapsed, killing workers inside. The other, a modern reinforced concrete frame completed in 2008, sustained repairable damage but remained standing.

Same earthquake. Same ground motion. Same street. Completely different damage.

The hazard module gave both buildings the same intensity score. The coordinate, the hazard grid cell, the PGA, identical. But the loss estimates the cat model produced for these two buildings were separated by an order of magnitude. That difference came entirely from the vulnerability module.

How does the vulnerability module translate a physical intensity, a number measuring ground shaking or wind speed, into a financial damage ratio? And why does the building matter so much more than the event?

What the Vulnerability Module Does

The vulnerability module answers a single question: given this hazard intensity at this location, what proportion of the building's value is expected to be destroyed? Its output is a damage ratio, a number between 0 and 1, where 0 means no damage and 1 means total loss. Multiply the damage ratio by TIV and you have the ground-up loss for that location in that event.

The vulnerability module sits between the hazard module and the financial module. The hazard module tells it how hard the ground shook or how fast the wind blew. The financial module will use its output to calculate what the insurer actually pays. Everything that determines whether physical intensity becomes a small claim or a catastrophic loss passes through the vulnerability module.

Its operation is conceptually simple even if the underlying science is not. The module looks up a damage function, a mathematical relationship between intensity and damage, for the specific type of building at the location. It applies that function to the hazard intensity and returns a damage ratio. The entire complexity of structural engineering, material science, and building physics is compressed into that relationship.

The Central Insight

The vulnerability module is where the building matters. The hazard module does not know or care what sits at a coordinate, it calculates intensity at a point regardless of what is there. The vulnerability module is what transforms a physical force into a building-specific damage estimate. This is why the quality of exposure data is the primary driver of model output reliability, without accurate knowledge of what the building is, the vulnerability module has nothing meaningful to work with.

Damage Functions - The Relationship Between Intensity and Damage

A damage function is a curve that maps hazard intensity to expected damage ratio for a specific type of building. At low intensities the curve sits near zero, minor shaking, negligible damage. As intensity rises, damage rises with it, accelerating through a zone where structural systems are genuinely stressed, before approaching 1.0 at extreme intensities where total loss becomes likely. The shape of the curve, and where it sits relative to the intensity axis, depends entirely on the type of building being described.

Illustrative Damage Functions - Two Buildings, Same Earthquake Intensity Scale

1.0 0.75 0.5 0.25 0.0

Mean Damage Ratio

Peak Ground Acceleration (increasing →)

PGA ≈ 0.08g · 1-in-10yr

~2% / ~0%

Minor shaking. Masonry sees minor cracking. RC frame essentially undamaged.

PGA ≈ 0.38g · 1-in-100yr

~55% / ~12%

Same intensity. Masonry sustains heavy structural damage. Modern RC frame is repairable. 4× difference in damage ratio.

PGA ≈ 0.70g · 1-in-500yr

~90% / ~42%

Extreme shaking. Masonry approaches total loss. RC frame sustains major structural damage but likely survives.

The gap between the two curves is not a modelling artefact, it reflects real physics. Unreinforced masonry has no ductility: when shaking forces exceed its capacity, it fails suddenly and catastrophically. Reinforced concrete frames are designed to deform gradually under extreme loading, absorbing energy and remaining standing even when severely damaged. The damage function for each building type encodes this structural behaviour as a mathematical curve, calibrated against observed damage from real earthquake events.

Every commercial cat model contains hundreds of damage functions, one for each combination of construction class, occupancy type, year of construction, and sometimes additional modifiers like number of storeys or roof type. The model selects the appropriate function based on the COPE data in the exposure file. When COPE data is missing and the model applies a default, it is selecting a damage function for a hypothetical average building rather than for the building that actually exists.

Micro-Check · Before You Continue

The Meridian SOV lists the Osaka distribution centre with unknown construction class. The model applies a default damage function for a representative Japanese warehouse. The 1-in-100 year ground motion at that location is 0.35g. Why might the default damage function significantly misstate the loss, and in which direction is the error likely to go if the actual building is a pre-1981 concrete frame?

The error is likely to understate the loss, potentially by a large margin. Japan's 1981 seismic code revision substantially raised design requirements for reinforced concrete buildings, structures built before 1981 were not required to meet the same ductility and strength standards as post-1981 buildings. At 0.35g, a pre-1981 RC frame might carry a damage ratio of 35–45%, while the default representative warehouse function, likely calibrated to a notional post-1990 building, might return 15–20%. The model is applying the wrong damage curve to a building that is structurally weaker than the default assumes. The result looks plausible, there is no error flag, but it systematically underestimates the loss for this location at every intensity level where the age difference matters.

Why COPE Fields Are the Bridge Between Hazard and Loss

The COPE fields in the exposure data are not just labels, they are the inputs that select the damage function. Each letter corresponds to a dimension along which damage functions differ:

COPE Field	What It Tells the Vulnerability Module	Why It Changes the Damage Function
Construction class	The structural material and system, masonry, RC frame, steel, wood	Different materials have fundamentally different relationships between applied force and structural failure. This is the primary damage function selector for earthquake and wind.
Year built	Which building code era the structure was designed to	Building codes have tightened dramatically after major events, Northridge 1994 (California), Hanshin 1995 (Japan), Marmara 1999 (Turkey). A pre-code building and a post-code building of identical material can differ in damage ratio by a factor of 3 or more at design-level ground motion.
Occupancy type	What the building is used for and what is inside it	Occupancy affects both the contents damage function (hospital equipment vs warehouse storage behave differently) and the structural design norms associated with the use type. A hospital and a warehouse of identical construction can have different damage functions because design requirements differ.
Number of storeys	The building's height and approximate natural period of vibration	Taller buildings have longer natural periods and respond more to lower-frequency ground motion components. A 30-storey tower and a 3-storey building experience the same ground motion differently at the structural level, the vulnerability module accounts for this through height-dependent function variants.

This table reveals something important about exposure data quality. A blank field does not just leave a gap in the record, it forces the vulnerability module to guess which damage function to apply. The guess is not random; it uses a representative default for the region and occupancy type. But the right damage function for a specific building can sit anywhere across the range the default spans. The exposure data is the only mechanism by which the model can know the difference.

Micro-Check · Before You Continue

The Meridian portfolio includes a waterfront hotel in Miami, 22 storeys, RC Shear Wall, 2019 build. The cat model assigns a hurricane wind damage function for high-rise post-2000 RC construction. A neighbouring property of identical dimensions, also RC Shear Wall, was built in 1978. Both are exposed to the same hurricane wind gust. Why do they receive different damage functions, and which property is expected to sustain more damage at the same wind speed?

The 1978 property receives a different damage function because it was designed to an older building code, and it will sustain significantly more damage at the same wind speed. Florida's building codes have strengthened considerably since Hurricane Andrew (1992) and again after the 2004–05 hurricane season. A 22-storey building completed in 2019 was designed to meet current Florida Building Code requirements, including specific wind load provisions for high-rise coastal construction. A 1978 building was designed to the standards of that era, lower design wind speeds, less stringent cladding attachment requirements, and less demanding roof-to-wall connection specifications. For a Category 4 hurricane at close range, the 1978 building might carry a damage ratio of 30–40%, while the 2019 building might carry 12–18%, roughly twice the damage for identical geometry, material, and hazard intensity. Year built is often the single largest differentiator between damage functions for wind peril in markets with active building code evolution.

Secondary Uncertainty - Why the Model Gives a Range, Not a Number

Every damage function produces a mean damage ratio, the expected damage given a specific intensity for a given building type. But actual damage does not fall exactly on the mean. Two unreinforced masonry buildings of identical specification, subjected to identical ground motion, will not sustain exactly the same damage. One might have a slightly different foundation condition. The other might have had previous repair work that altered its structural behaviour. Their contents might be arranged differently. The ground motion, even at the same point, varies in its directionality and frequency content in ways the single PGA metric does not fully capture.

This variability around the mean is what the cat model calls secondary uncertainty. It is distinct from the primary uncertainty about which event occurs and what its intensity will be, that is captured by the stochastic catalogue. Secondary uncertainty is the variability in building response given that the event and its intensity are already fixed. Even if you knew with certainty that a 0.4g earthquake would strike tomorrow, you could not predict with certainty what fraction of any specific building's value would be destroyed. Secondary uncertainty quantifies that remaining unpredictability.

Primary Uncertainty

Which events happen
and how intense they are

Captured by the stochastic catalogue, tens of thousands of possible events, each with a probability. This uncertainty cannot be reduced by collecting more building data. It is irreducible randomness in the physical world.

Aleatory, irreducible

Secondary Uncertainty

How a specific building
responds to a known intensity

Captured by a probability distribution around the mean damage ratio. A building expected to sustain 25% damage might actually sustain anywhere from 5% to 65%, the mean is the best estimate, not the guaranteed outcome. Better building data narrows this range but cannot eliminate it.

Epistemic + aleatory mix

Secondary uncertainty matters most at the tail of the loss distribution. When the cat model calculates the 1-in-100 year loss, it is not just asking "which event has a 1% annual probability?" It is also accounting for the possibility that a moderate-probability event produces an unusually bad outcome because buildings across the affected area happened to sustain damage at the high end of their secondary uncertainty range simultaneously. The 1-in-100 year portfolio loss incorporates both primary event uncertainty and secondary building-response uncertainty.

In practice, secondary uncertainty widens the confidence interval around any single-event loss estimate. If the model says an event produces a mean loss of USD 30M for the Meridian portfolio, the actual loss could plausibly be anywhere from USD 15M to USD 55M depending on how actual building response tracks against the mean damage functions. This is not a modelling failure, it is an honest representation of the genuine unpredictability in how physical forces translate into structural damage. A model that produced a single precise number without uncertainty bounds would be more confident than the underlying science warrants.

Micro-Check · Before You Continue

After a moderate earthquake affecting three Meridian locations in Chile, the actual combined loss is USD 8.4M. The cat model had estimated a mean loss of USD 5.1M for an event of that magnitude and location. The underwriter asks whether the model is "wrong." How do you explain the relationship between the model's mean estimate and the actual outcome?

The model is not wrong, the actual outcome is within the range the model's secondary uncertainty implies. A mean damage ratio is the expected value across many buildings of that type subjected to that intensity many times. In any single event, actual damage will scatter around the mean due to building-specific factors the model cannot observe: minor construction variations, previous damage, soil heterogeneity below the footprint, the specific directional character of the ground motion at that instant. An outcome of USD 8.4M against a mean estimate of USD 5.1M, a 65% exceedance, is not unusual. It means the buildings happened to perform at the adverse end of their secondary uncertainty range for this event. The model's mean estimate remains the best prediction available before the event. The right question is not "was the model wrong this time?" but "does the model's mean estimate remain well-calibrated across many events?", which requires a much larger dataset than three locations in one earthquake to assess.

Common Misconceptions

"The vulnerability module uses the same damage function for all buildings of the same material"

Construction material is the starting point, not the full specification. Year built, occupancy, height, and region all further differentiate the damage function. Two steel-frame buildings, one a 1970s warehouse and one a 2020 hospital, have different functions even though their primary material is the same. The model has hundreds of functions, not a handful.

"A higher damage ratio always means more total loss"

Not without knowing the TIV. A damage ratio of 0.6 applied to a USD 1M building produces a USD 600K loss. The same ratio applied to a USD 20M building produces a USD 12M loss. The vulnerability module produces the ratio; the financial module multiplies it by TIV to produce the loss. The two numbers are meaningless without each other.

"Secondary uncertainty is a sign the model is imprecise"

Secondary uncertainty is a property of the physical world, not the model. Buildings of identical specification genuinely respond differently to identical forces because no two buildings are truly identical and no two ground motion records at adjacent points are truly identical. A model that claimed zero secondary uncertainty would be less accurate, not more, it would be asserting a precision that physical reality does not support.

"The vulnerability module only matters for structural damage"

The vulnerability module covers contents damage and business interruption loss as well as structural damage, though typically through separate damage functions. Contents in a flooded basement sustain different losses from the same inundation depth than the structural walls. A hospital that sustains moderate structural damage may have business interruption losses far in excess of the repair cost because its revenue generation is interrupted while the building is made safe. All of these require separate vulnerability functions.

Key Terms

Vulnerability Module

The second cat model component. Takes a hazard intensity and building description, applies the appropriate damage function, and returns a mean damage ratio. Sits between the hazard module and the financial module in the calculation chain.

Damage Function

A mathematical curve mapping hazard intensity to expected damage ratio for a specific building type. Different functions exist for each combination of construction class, year built, occupancy, and other COPE modifiers. The model selects the appropriate function from the exposure data.

Damage Ratio

The output of the vulnerability module, the proportion of TIV expected to be destroyed given a specific hazard intensity. Ranges from 0 (no damage) to 1.0 (total loss). Multiplied by TIV to produce the ground-up loss estimate.

Mean Damage Ratio

The expected damage ratio from the damage function at a given intensity, the best single estimate of the proportion of TIV destroyed. Actual damage will vary around this mean due to secondary uncertainty. The mean is an expectation, not a guarantee.

Secondary Uncertainty

The variability in building response given that the event intensity is already fixed. Even knowing that 0.4g ground motion will occur at a location, damage can range widely due to building-specific factors not captured in the damage function. Represented as a probability distribution around the mean damage ratio.

Ground-Up Loss

The total physical damage loss before any insurance policy conditions, deductibles, limits, coverage restrictions, are applied. Damage ratio × TIV = ground-up loss. The financial module then applies policy conditions to produce the insured loss.

Knowledge Check - Lesson 2.3

Five questions · 4 of 5 correct (80%) to pass

1. The hazard module assigns a PGA of 0.42g to two adjacent Meridian locations in Istanbul, a 1975 unreinforced masonry office and a 2012 RC shear wall tower. The vulnerability module returns a damage ratio of 0.68 for the masonry building and 0.14 for the RC tower. The TIV of each is USD 12M. What are the respective ground-up losses, and what does this illustrate about the vulnerability module's role in the calculation chain?

ABoth buildings receive the same ground-up loss of USD 5.04M, because the hazard module assigned identical intensity, the vulnerability module can only differentiate losses when buildings are at different locations with different hazard scores.

BThe masonry building produces a ground-up loss of USD 8.16M (0.68 × USD 12M) and the RC tower produces USD 1.68M (0.14 × USD 12M), a nearly fivefold difference from identical hazard. This illustrates precisely the vulnerability module's role: it is the mechanism by which the same physical event produces completely different financial outcomes depending on what kind of building is at the exposed location. The hazard module treats both buildings identically. The vulnerability module is where building characteristics become financial differences.

CThe ground-up loss for both buildings is USD 8.16M, because the policy limit applies at the TIV level and both buildings have the same TIV. The damage ratio differentiates the damage but not the insured loss once the financial module is applied.

DThe calculation cannot be completed without knowing the secondary uncertainty range for each building type, the mean damage ratios of 0.68 and 0.14 are point estimates that need to be expressed as ranges before a ground-up loss figure can be produced.

2. A cat model's damage function for wood-frame residential construction in coastal Florida produces a mean damage ratio of 0.45 at 130 mph 3-second gust wind speed. In practice, after a hurricane producing exactly 130 mph gusts across a neighbourhood of 500 identical wood-frame homes, the observed damage ratios range from 0.12 to 0.91. Is the model wrong?

AYes, a well-calibrated model should produce mean damage ratios that match observed outcomes within ±10% for a sample of 500 buildings. A range of 0.12 to 0.91 indicates the damage function is poorly calibrated for this building type and wind speed.

BNo, the wide range is exactly what secondary uncertainty predicts. The 500 homes are not truly identical: they have different roof-to-wall connection details, different ages within the same code era, different maintenance histories, slightly different local wind exposure from trees and neighbouring buildings, and different orientations relative to the storm track. The mean damage ratio of 0.45 is a prediction of the expected value across all those variable conditions, not a prediction that every building suffers exactly 45% damage. The model should be assessed on whether the mean of the observed distribution approximates 0.45, not on whether individual outcomes cluster tightly around it.

CThe question cannot be answered without knowing whether the 130 mph figure represents the sustained wind speed or the 3-second gust, the two metrics produce very different damage functions and a mismatch between the metric used in the model and the metric observed in the field would explain the wide range.

DYes, secondary uncertainty should narrow, not widen, as the sample size increases. For 500 buildings, the law of large numbers should produce observed damage ratios clustered much more tightly around the mean. A range of 0.12 to 0.91 at this sample size indicates systematic model error rather than natural variability.

3. Halcyon's underwriter is assessing a new account, a portfolio of 30 solar farms across Texas and Oklahoma. She asks you to run the cat model for hail risk. You produce results, but note that the damage functions in the model were developed and calibrated for residential and commercial buildings. Why does this matter for how you present the hail loss estimates?

AIt does not matter significantly, hail damage is primarily driven by stone size and kinetic energy, which affects all surfaces similarly. Damage functions calibrated to roofing materials on conventional buildings translate reasonably well to solar panels, which are exposed horizontal surfaces in the same way as flat roofs.

BIt matters because the model is applying damage functions that describe how hail damage roof tiles, cladding, and windows, not how it damages photovoltaic panels. The failure mode for solar panels under hail is different: panel cracking and electrical system damage at hail sizes that would cause relatively minor cosmetic damage to a conventional roof. A damage function calibrated to building damage will not correctly represent the intensity threshold at which solar panels fail, the steepness of the damage curve in the critical size range, or the total loss threshold. The hail loss estimates for this account carry an additional layer of model uncertainty beyond normal secondary uncertainty, the damage function itself may be the wrong shape for this asset class, and that must be stated clearly when presenting results.

CIt matters only if the solar farms are in areas with high hail frequency. In lower-hazard areas, the damage function calibration is less important because the intensity levels reached are below the range where calibration differences become significant.

DThe model should not be run at all for this account, if the damage functions are not calibrated to solar panels, any output would be misleading and should not be presented. The underwriter should be told the account cannot be modelled until vendor-specific solar damage functions are available.

4. Two Meridian warehouse locations are identical in every respect, same construction class, same year built, same TIV, same occupancy, but one is in Los Angeles and the other is in Auckland, New Zealand. The cat model assigns them different damage functions. Why?

AThe model assigns different functions because Los Angeles has higher earthquake hazard than Auckland, the damage function scales with the local hazard level, producing steeper curves in higher-hazard regions to reflect the greater energy released by local seismic sources.

BThe model assigns different functions because building codes and construction practices differ between countries, even for the same nominal construction class in the same nominal year. A steel moment frame built in Los Angeles in 2005 was designed to US seismic code requirements; the same description in Auckland was designed to New Zealand's standards. The structural detailing, connection specifications, and design forces differ between the two regimes, and the vulnerability functions are calibrated to observed damage from local events in each region, encoding those regional construction differences. The hazard intensity is what changes with location; the damage function encodes how local buildings respond to that intensity.

CThe model assigns the same damage function, construction class and year built fully specify the damage function regardless of geography. The different location affects only the hazard module output, not the vulnerability function selected.

DThe model assigns different functions because the soil conditions in Los Angeles and Auckland differ, and soil type is the primary vulnerability modifier that changes the damage function between geographies.

5. After running the Meridian portfolio, the cat model reports a mean 1-in-100 year loss of USD 38M. The underwriter asks: "So in a 1-in-100 year event, we'll lose USD 38M?" What is the precise correction you need to make to her interpretation?

ANo correction needed, the underwriter's interpretation is correct. The 1-in-100 year loss means that in an event with 1% annual probability, the expected portfolio loss is USD 38M. This is both a statement about event probability and about the loss amount in that event.

BThe correction is about timing: USD 38M is the loss in a 1-in-100 year event, but because secondary uncertainty means damage varies around the mean, the actual loss could be slightly above or below USD 38M even if a 1-in-100 year event occurs. The underwriter should think of the figure as USD 38M ± 10–15%.

CTwo corrections. First, the return period correction from Lesson 2.1: a 1-in-100 year loss does not mean there will be a loss in exactly 100 years, it means any given year has a 1% probability of exceeding USD 38M, which could happen next year, twice in a decade, or not for 200 years. Second, the secondary uncertainty correction from this lesson: USD 38M is the mean estimate. It incorporates the average building response across the damage functions, but actual loss in a real event depends on how each building performs relative to its mean. Even if a precisely 1-in-100 year event occurs, the actual loss could plausibly range from USD 20M to USD 60M depending on realised secondary uncertainty across all 150 Meridian locations. The model gives you the best single number; it does not guarantee that number.

DThe correction is that USD 38M is the occurrence loss, the loss from the single worst event in a year with 1% probability. The aggregate loss in a year that includes a 1-in-100 event would be higher once smaller events are added. The underwriter should be working with the aggregate EP curve rather than the occurrence EP curve for portfolio management purposes.