Solent Re is a London-based specialty reinsurer writing global property catastrophe alongside marine, energy, and a growing cyber book. After a major flood event in Central Europe, the claims team reports a final net loss of USD 118M against a modelled loss estimate of USD 71M, a 66% overshoot.
The head of cat modelling is asked to explain the discrepancy. Her analysis identifies three contributing factors. First, a large industrial park on the western edge of the affected area had been geocoded to postcode level, the actual buildings were in a lower-lying area with 40% higher flood inundation depth than the postcode centroid. Second, the model's flood vulnerability functions for Eastern European industrial construction were calibrated on Western European data, a structural mismatch. Third, and largest: two of the affected locations carried business interruption cover that the cat model had not been configured to capture. The BI losses alone accounted for USD 28M of the gap.
"The model was right about what it modelled," she tells the board. "The problem is what it didn't model at all."
Every cat model comes with documentation listing what it covers, which perils, which geographies, which building types. But the boundary of a model is rarely as clear as its documentation suggests. A model described as covering "European flood" may cover riverine flood in France and Germany while treating flash flood, coastal surge, and groundwater flood very differently, or not at all. A model described as covering "US tornado" may use wind speed as its primary intensity metric without fully capturing the damage from large hail that frequently accompanies severe convective storms.
The most dangerous aspect of a model's boundary is that crossing it produces a number, not an error. When a location falls outside the model's calibrated scope, a construction class it has no data for, a geography where its hazard grid is sparse, a secondary peril its vulnerability functions were not built to capture, the model does not stop. It applies a default and continues. The output looks identical to output produced within the model's well-calibrated scope. There is no asterisk, no warning, no indication that the number was produced outside the model's area of reliable knowledge.
A cat model output of USD 71M is the same number whether it was produced with complete, well-calibrated data within the model's core scope, or with sparse data at the edge of the model's domain using default assumptions for everything the data did not provide. The number does not carry its own quality label. The analyst who produced it is the only source of that information, and if the caveats do not travel with the output, the downstream decision-maker cannot know what the number actually represents.
It helps to separate model limitations into three distinct categories. They are different in nature and different in what can be done about them, but all three contribute to the gap between what a model says and what actually happens when a loss occurs.
The irreducible randomness of the physical world. Even a perfect model cannot predict which specific events will occur in a given year, or exactly how strong the ground shaking will be at a specific point from a known earthquake. This uncertainty is captured in the stochastic catalogue, it is the entire reason we run tens of thousands of synthetic events rather than a single deterministic prediction. No amount of better data or better science eliminates it. It is the randomness the EP curve and AAL are designed to quantify.
Uncertainty from incomplete knowledge, things the model does not know but could in principle know better. The choice of Ground Motion Prediction Equation for earthquake, the calibration of flood vulnerability functions to a specific regional building stock, the rate assumptions in the stochastic catalogue for a peril with limited historical data, all of these reflect the limits of current scientific knowledge, not the limits of knowledge in principle. Better data, better calibration, and more research can reduce epistemic uncertainty over time. It is the primary source of vendor model divergence: two vendors making different but defensible scientific choices about the same uncertain physical processes.
Losses from perils, geographies, or coverage types that the model does not attempt to represent at all. This is categorically different from uncertainty, it is absence. If a model does not include a business interruption vulnerability function, the BI loss is not uncertain within the model, it is simply not there. The model produces a number as if BI did not exist. Non-modelled items do not show up as uncertainty ranges or confidence intervals. They show up as a gap between the modelled loss and the actual loss after the event, as happened with Solent Re's USD 28M BI shortfall.
Non-modelled items fall into four broad categories. Each represents a class of real-world loss that standard vendor cat models either do not attempt to capture or capture only partially. Understanding these gaps does not require knowing how to handle them professionally, it requires knowing they exist and that the model's output is silent about them.
Standard cat models focus on physical property damage. Business interruption, the revenue lost while a building is unusable after a cat event, is modelled separately and often not configured in routine runs. Contingent BI, which arises when a supplier or customer's premises is damaged rather than the insured's own, is rarely modelled at all. For commercial and industrial risks, BI can exceed the property damage loss.
A major earthquake can trigger landslides, fires, and tsunami. A tropical cyclone produces storm surge as well as wind. A severe hailstorm accompanies a tornado outbreak. Standard models typically capture the primary peril intensity but may underrepresent the secondary perils that frequently accompany it. Fire-following-earthquake and tsunami are the most commonly cited secondary peril gaps for earthquake models.
Vendor models are calibrated to available data. In markets with limited historical loss records, sparse building stock information, and incomplete hazard data, much of sub-Saharan Africa, parts of Southeast Asia, Central Asia, the models' calibration is thinner and their uncertainty bands wider. Output for these regions carries higher epistemic uncertainty than the same vendor's output for US hurricane or Japan earthquake.
After a major catastrophe, the sudden demand for construction materials and skilled labour drives up repair costs, sometimes by 20–40% above pre-event levels. Standard model outputs are calibrated to normal market conditions. The additional cost of rebuilding in a disaster-affected economy, demand surge, is a systematic source of underestimation for large, geographically concentrated events.
Some risks do not yet have mature cat model frameworks. Cyber catastrophe, a systemic attack affecting many policyholders simultaneously, has no established stochastic catalogue comparable to those for hurricane or earthquake. Pandemic business interruption, climate-change-driven non-linear hazard shifts, and novel liability scenarios share the same challenge: the physical science or loss history needed to build a calibrated probabilistic model does not yet exist or is too short to be reliable.
In some jurisdictions, post-event legal and social dynamics inflate insured losses well above the physical damage the model estimates. Assignment of benefits disputes, class action litigation, expanded policy interpretations by courts, and regulatory intervention can all cause actual paid claims to significantly exceed what the cat model calculated from pure physical damage assessment. US litigation-driven loss amplification has been a material source of model underperformance in certain property markets.
When two vendors produce materially different loss estimates for the same portfolio, it is tempting to conclude one of them is wrong. Usually, neither is wrong, they are making different but defensible choices within areas of genuine scientific uncertainty. Understanding why divergence occurs is the first step to using multi-vendor analysis productively.
Divergence typically grows with return period. At AAL, dominated by moderate, frequent events where both vendors have reasonable calibration data, the gap is 18%. At 1-in-100 the gap is 48%. At 1-in-250 it reaches 71%. This pattern is normal and expected: both vendors are well-calibrated for the body of the distribution where historical data is abundant, and diverge at the extreme tail where data is sparse and model assumptions dominate.
The gap between vendors is not noise to be averaged away, it is information. Where vendors agree closely, the science is relatively settled and the output is more reliable. Where they diverge significantly, the divergence marks the boundary of current scientific knowledge on that peril and geography. A practitioner who understands why the vendors diverge for European flood at the 1-in-250 year level understands something important about the limits of what current cat models can reliably say about that risk.
Beyond the technical limits of what models can measure lies a structural reality: most economic catastrophe loss is not insured. Cat models estimate insured losses, the losses falling within the coverage terms of the policies in the portfolio. The total economic loss from a major event is typically far larger. The gap between total economic loss and insured loss is the protection gap, and it is one of the defining challenges in global catastrophe risk management.
The insured fraction varies enormously by country and peril. In the US, it may exceed 60% for hurricane wind. In many developing markets, it may be under 5% for earthquake. The cat model only sees the insured portion, the same physical event produces very different model output depending on where it strikes.
The protection gap matters to practitioners for three reasons. First, it explains why a model can be technically correct, accurately estimating insured losses, while massively understating the total economic impact of an event. The model is not wrong; it is answering a narrower question than the physical scale of the disaster suggests. Second, it explains why insurance penetration matters as a context for model output, a portfolio covering 5% of the economic exposure in a region is very differently positioned from one covering 40%, even if both produce similar model outputs for the covered losses. Third, it identifies where future market growth will occur: the protection gap is the unfilled demand for insurance coverage, and closing it is the business case for many emerging market initiatives.
The model produces a number for every event in its catalogue regardless of whether the coverage in the policy actually responds to that peril. A wind-only policy on a flood-prone property will produce a cat model output, for the wind component, that looks complete, while the flood exposure the model does not cover goes entirely unrepresented. The output covers what the policy covers and what the model models. Both limits apply simultaneously.
Vendor divergence reflects genuine scientific uncertainty, not error on one side. Both vendors are typically making defensible choices within the limits of available data and science. The divergence is most informative when treated as a measure of that uncertainty rather than as evidence that one model needs correcting.
In Solent Re's Central European flood loss, the non-modelled BI component alone accounted for USD 28M of a USD 47M gap, more than half the total discrepancy. For commercial and industrial risks with high revenue intensity, BI can exceed property damage. For events in litigation-prone markets, loss amplification can add tens of percent to the modelled loss. Non-modelled items are systematic, directional, and sometimes larger than the modelled loss itself.
Practitioners working in emerging markets, structuring parametric products, pricing climate-linked covers, or assessing sovereign risk programmes directly encounter the protection gap as a core professional challenge. It is also the context within which every insured loss estimate must be understood, a USD 22M insured loss in a market with 8% penetration represents a very different physical event from a USD 22M insured loss in a 60% penetration market.
The irreducible randomness of the physical world, which events occur and at what intensity. Quantified by the stochastic catalogue and the EP curve. Cannot be reduced by better data or science. The fundamental reason cat modelling uses probability distributions rather than single predictions.
Uncertainty from incomplete scientific knowledge, things the model does not know but could in principle know better. The primary source of vendor model divergence. Can be reduced over time with better data, better calibration, and more research.
Losses from perils, geographies, or coverage types the model does not attempt to represent. Categorically different from uncertainty, not uncertain within the model, simply absent. Common examples: business interruption, secondary perils, demand surge, loss amplification.
The difference between total economic loss and insured loss from a catastrophe event. The portion of physical damage that is not covered by any insurance policy. Cat models capture only the insured portion. In many developing markets, the protection gap exceeds 90% of total economic loss.
Post-event inflation of construction and repair costs driven by sudden excess demand for labour and materials in the affected area. Typically 20–40% above pre-event market rates for major concentrated events. A systematic non-modelled item that causes actual losses to exceed standard model output.
The inflation of insured losses beyond physical damage through legal and social dynamics, litigation, assignment of benefits disputes, expanded judicial interpretations of policy language. A significant non-modelled source of underperformance in certain property markets.
The version or release date of a vendor cat model. Models are updated periodically to incorporate new science, post-event recalibration, and new data. Two models of different vintages will diverge even if they use similar methodologies, because their underlying calibration data differs.
Five questions · 4 of 5 correct (80%) to pass