It is October 2017. Hurricane Maria made landfall in Puerto Rico three weeks ago. The Category 4 storm destroyed an estimated 80% of the island's power grid, knocked out water service for more than a million people, and killed, close to 3,000 people. It will take eleven months for electricity to be fully restored.
The economic losses from Maria are later estimated at USD 90 billion. The insured losses, the amount the insurance industry actually pays out, are approximately USD 31 billion. That means USD 59 billion of economic destruction simply has no insurance coverage. It falls on households, small businesses, and the government, on people who, in many cases, have spent decades paying insurance premiums on their homes and businesses.
Two questions will shape this entire course. First: why did the insured loss fall so far short of the economic loss? Second: how did insurance companies, and their reinsurers, estimate in advance how much Maria might cost them, so that they had set aside enough money to pay the claims?
The insurance industry handles loss every day. Burst pipes. Shop thefts. A kitchen fire. These events happen constantly, at a rate the industry has measured for centuries. They are independent of each other, the burst pipe in Manchester has nothing to do with the shop theft in Bristol, and their frequency and severity are predictable enough that an insurer can set premiums today, confident that the aggregate claims next year will fall within a manageable range. This is the everyday business of insurance. It is called attritional loss.
A catastrophe is different in kind, not just in scale. When Hurricane Maria struck Puerto Rico, it did not cause damage to one building, or ten, or a hundred. It caused damage to hundreds of thousands of buildings simultaneously, across the entire island, within a matter of hours. The key word is simultaneously. Simultaneous loss across a large number of policyholders at the same moment, driven by a single event, is what defines a catastrophe for insurance purposes.
This simultaneity destroys the statistical independence that makes everyday insurance work. When losses are independent, they average out. An insurer with a million policyholders can be fairly confident that if 5% of them suffer a loss in a given year, it will be roughly 50,000 claims. The law of large numbers works in the insurer's favour. But when a hurricane hits, the loss correlation jumps to near-100% within the storm track. The law of large numbers no longer applies. All the claims arrive at once.
Insurance works by pooling independent risks. Catastrophes break that independence. When the risks in your pool all move together, because the same hurricane, earthquake, or flood has hit them all simultaneously, you need a fundamentally different approach to estimating how bad things might get. That different approach is catastrophe modelling.
Not all large losses are catastrophes. A single large industrial fire might cost USD 200 million but it affects one policyholder, it is an attritional large loss, not a catastrophe. A catastrophe has three characteristics that together distinguish it from any other type of insurance loss:
A single physical event causes damage to many insured properties simultaneously. The correlation is driven by the hazard footprint - the geographic area over which the event produces damaging levels of wind, ground motion, flooding, or fire intensity. Every property within the footprint that is exposed to the hazard suffers some probability of damage at the same moment. The closer to the centre of the event, the higher the damage probability. This spatial correlation is the defining financial characteristic of a catastrophe.
The spatial correlation means that geographic concentration of insured exposure is the primary driver of cat loss. An insurer with 10,000 policyholders uniformly distributed across the United States faces very different hurricane cat risk from one with 10,000 policyholders all in coastal Florida, even if the total premium income and TIV are identical. The Florida-concentrated portfolio has a dramatically higher probability of a single event affecting a large fraction of the book simultaneously. This is why cat models are built around geographic exposure accumulation - it is the mechanism through which correlation becomes financial loss.
Catastrophes are rare enough that historical experience is a poor guide to future loss. A major hurricane makes direct landfall on a specific coastal city perhaps once every several decades. An earthquake of magnitude 7+ strikes a major urban area once every generation. This low frequency means that in any given insurer's historical claims experience, major cat events may appear only a handful of times, or not at all. The statistical approach that works well for attritional loss, estimate the average annual loss from historical data and add a margin, systematically underestimates catastrophe risk because the history does not contain enough events to capture the true distribution of potential outcomes.
This is why catastrophe modelling exists. Rather than relying on the limited historical record of actual events, cat models construct a large synthetic catalogue of potential events, hundreds of thousands of simulated storms, earthquakes, or floods, calibrated to the physical science of each hazard. This catalogue provides a much more complete sampling of the full distribution of possible losses, including events more severe than anything that has occurred in the modern historical record.
In the immediate aftermath of a major catastrophe, the cost of repairing any individual building rises sharply. Skilled labour, materials, and equipment are in acute short supply across an entire region simultaneously. Contractors who charge USD 15,000 to replace a roof in normal times charge USD 25,000 or more in the months following a major hurricane. This is demand surge, and it means that the financial loss from a catastrophe is higher than what you would estimate simply by multiplying the number of damaged buildings by a normal repair cost.
Claims complexity is a related amplifier. Wind versus flood coverage disputes (as seen after Katrina and Sandy), business interruption valuation disagreements, the interaction between primary and reinsurance coverage layers, and the sheer volume of simultaneous claims straining adjusters and legal systems all contribute to the final settled cost being higher and slower than the initial estimate. Both demand surge and claims complexity are explicitly modelled in modern cat models, though imperfectly.
Return to the Hurricane Maria question: why did USD 59 billion of a USD 90 billion loss go uninsured? The protection gap, the difference between total economic losses and insured losses from a catastrophe, is one of the most important and persistent features of the catastrophe risk landscape. Understanding it is foundational to understanding why cat modelling exists as a discipline and what it is trying to achieve.
The protection gap has multiple causes, and they vary significantly by geography and peril. In Puerto Rico's case, the dominant driver was low insurance penetration: many households and small businesses simply did not have property insurance at all, or had policies with very low limits relative to the value of their assets. But the gap manifests differently in different contexts:
The protection gap is not merely an equity concern, it is directly relevant to how cat models are used. An insurer pricing a reinsurance treaty for a Turkish earthquake portfolio needs to understand that the insured TIV represents only a fraction of the total economic exposure in the affected area, and that the correlation between insured and total economic loss may not be constant across different event severities. A cat model calibrated to insured loss data will reflect the decisions already made about what to insure, which means it systematically excludes the parts of the risk landscape that insurance has not yet reached.
The term catastrophe covers a wide range of physical hazards. The insurance industry has traditionally focused on natural perils, events driven by geophysical or meteorological forces. In recent years it has increasingly engaged with man-made and technological catastrophes as well. The Event Library on this platform covers landmark events across the peril spectrum.
Ground motion, liquefaction, and building collapse. Cat models failed to account for the scale of building code non-compliance.
Wind, surge, and rainfall flooding. Harvey's inland flooding was almost entirely uninsured. Maria's island geography compounded the gap.
Flash flooding in small catchments not covered by large-river cat models. Germany's voluntary flood system produced a large uninsured fraction.
Urban conflagration, fire spreading structure to structure. Largest insured wildfire in history. California's insurance market crisis amplified the impact.
A Russian state-deployed cyberweapon that escaped its target. Triggered years of war exclusion litigation. The founding event of cyber cat modelling.
1,400 deaths in a wealthy region. No heat cat model exists. The ultimate example of a peril with a near-100% protection gap.
Each of the events shown above, and more, has a full analytical case study in the Event Library, examining what happened, why the losses were what they were, and what the cat models got wrong. The Event Library is free to access and provides the real-world anchoring for every concept in this course. You do not need to read them now, they will be referenced at relevant points throughout the courses.
By the late 1980s, the insurance industry's approach to catastrophe risk was based almost entirely on historical experience. Underwriters used tables of historical storm tracks, historical earthquake records, and historical flood extents to estimate how often a major loss might occur and how large it might be. This worked tolerably well when the worst events the market had experienced were not dramatically worse than average years.
Hurricane Andrew changed everything. When Andrew made landfall in South Florida on 24 August 1992, it caused USD 27 billion in insured losses, more than twice the largest previous hurricane loss in history, and more than the industry had collected in premiums from Florida homeowners over the preceding decade. The historical data had not contained an event of this severity. The market had been pricing as if Andrew could not happen. Eleven insurers went insolvent. The entire Florida property insurance market faced a capital crisis.
The response to Andrew was the birth of modern catastrophe modelling. Within months of the storm, three firms, AIR Worldwide, RMS (Risk Management Solutions), and EQECAT, were developing probabilistic models that did not rely solely on the limited historical record. Instead, they constructed synthetic catalogues of thousands of possible hurricane seasons, calibrated to the physical science of tropical cyclone behaviour, and ran entire insurer portfolios through each simulated event. For the first time, underwriters could estimate the probability of an event worse than anything in the historical record, and price their risks and hold their capital accordingly.
This is the fundamental service cat modelling provides: it extends the information available beyond the limits of recorded history. Instead of asking "what happened in the past?", cat models ask "what could happen, and how likely is each outcome?" That shift, from historical frequency to modelled probability, is what makes cat modelling possible and necessary as a professional discipline.
In practice, the insurance industry uses a threshold definition of catastrophe for data collection and reporting purposes. The two most widely used thresholds:
These thresholds are administrative, not physical. They determine which events get tracked in loss databases used to calibrate cat models. An event just below the threshold affects the same underwriting decisions as one just above it, the designation does not change the risk. Cat modellers use these databases as one input among many; the thresholds matter most for understanding what is and is not in the calibration dataset.
In your day-to-day work as a cat modelling analyst, you will encounter the word "catastrophe" used in several distinct senses: the physical event (the hurricane or earthquake), the insurance industry event designation (PCS cat, Munich Re sigma cat), the model output (a simulated event in a stochastic catalogue), and the capital concept (a loss large enough to threaten solvency). These are related but not identical. Being precise about which sense is intended in any given context, a submission, a model run, a reinsurance discussion, is a professional habit worth developing from the beginning.
Scale alone does not define a catastrophe. A single large industrial fire causing USD 500M of damage to one insured is not a catastrophe, it affects one policyholder. A hailstorm causing USD 50M of aggregate losses across 30,000 homes is a catastrophe because of the simultaneous, correlated nature of the loss. The defining characteristic is correlation, not magnitude.
Sometimes true, but often incomplete. In the US after Hurricane Katrina, many policyholders had homeowners insurance, but their policies explicitly excluded flood, which was the dominant loss mechanism. They had insurance, but not the right insurance for the event that occurred. Coverage scope matters as much as penetration rate.
Cat models do not predict specific events. They estimate the probability distribution of possible losses, how likely different loss levels are, not which event will occur next year. A model output of "1-in-100 year loss = USD 2 billion" means there is a 1% probability of exceeding that loss in any given year. It says nothing about whether or when that event will occur.
The most dangerous assumption in catastrophe risk. Hurricane Andrew hadn't happened before it happened. The Tōhoku earthquake in 2011 exceeded all previously modelled maximum magnitudes for the Japan subduction zone. The Pacific Northwest Heat Dome of 2021 reached temperatures that climate scientists had called "virtually impossible." Absence of historical precedent is not absence of risk.
Historically true, but increasingly not. The NotPetya cyberattack (2017) and CrowdStrike outage (2024) demonstrated that man-made events can produce correlated losses at catastrophe scale, affecting thousands of policyholders simultaneously through network connectivity rather than geographic proximity. Cyber cat modelling is now a rapidly growing field using the same underlying framework as natural peril modelling.
An event causing simultaneous, correlated losses across many policyholders, driven by a single physical or technological trigger. Distinct from attritional loss by the correlation structure of the claims, not merely by their total magnitude.
Everyday insurance losses, independent, predictable in aggregate, stable year-to-year. The foundation of standard insurance pricing. Catastrophe loss is the exception to attritional loss behaviour.
Losses that are not independent, when one claim occurs, others in the same geographic area or connected network are also more likely to occur simultaneously. Correlation is the defining financial characteristic of catastrophe risk.
The difference between total economic losses from a catastrophe and the insured losses paid by the insurance industry. Driven by lack of coverage, coverage scope exclusions, or absence of viable insurance products for the peril concerned.
The increase in repair and replacement costs in the immediate aftermath of a major catastrophe, driven by simultaneous excess demand for skilled labour, materials, and equipment across an entire affected region.
A large synthetic set of simulated events, typically tens of thousands, constructed to represent the full distribution of possible catastrophes, including events more severe than anything in the modern historical record. The core input to a cat model.
The probability that a loss will exceed a given threshold in any given year. The primary output metric of a cat model. A 1% exceedance probability corresponds to a 1-in-100-year return period. Covered in depth in Lesson 3.1.
The geographic area over which a catastrophe event produces damaging conditions, the wind field of a hurricane, the ground motion pattern of an earthquake, the inundation extent of a flood. The spatial structure of the footprint drives the correlation of losses.
Five questions · 4 of 5 correct (80%) to pass