Catastrophe Modelling

Hurricane Andrew in 1992 bankrupted eleven insurers and exposed a fundamental gap: the industry had no systematic way to estimate losses from events that had not yet happened. Three decades and billions of dollars of modelling investment later, the gap between modelled and actual losses for secondary perils remains stubbornly wide. Catastrophe modelling has solved the problem it was built for, and now faces the harder problem of perils that physics-based models struggle to capture.

What is catastrophe modelling?

Catastrophe modelling is the quantitative estimation of potential losses from natural and man-made catastrophic events. A cat model simulates thousands or millions of possible events, each with a defined hazard footprint, and calculates the resulting damage to an insurer's portfolio of exposures. The output is an exceedance probability curve: the probability of losses exceeding various thresholds over a given period. This curve drives treaty pricing, capital allocation, and reinsurance purchasing decisions.

The three dominant vendor models, RMS (now Moody's RMS), AIR (Verisk), and CoreLogic, each combine a hazard module (where and how intense events occur), a vulnerability module (how buildings and contents respond to hazard intensity), and a financial module (how insurance policy terms translate physical damage into insured losses). For peak perils like US hurricane and Japanese earthquake, these models are well-calibrated against historical events and widely trusted.

The challenge is secondary perils. Convective storms, wildfire, flooding, and winter storms have caused insured losses exceeding 50 billion dollars per year globally in recent years, often exceeding the modelled estimates. These perils are harder to model because the events are more localised, the vulnerability functions are less well understood, and the hazard is more sensitive to climate variability. The industry's reliance on vendor models calibrated to historical data is most fragile precisely where the hazard is changing fastest.

The landscape

The PRA expects insurers using internal models for Solvency II capital requirements to demonstrate that their catastrophe models reflect current risk, not just historical patterns. Supervisory Statement SS4/17 requires firms to validate their models against emerging experience and to justify their model choices. For perils where the hazard is non-stationary, this means annual recalibration is not sufficient; firms need to demonstrate that their models account for trends in the underlying hazard.

EIOPA's work on natural catastrophe risk in Solvency II has highlighted the challenge of model divergence. Two vendor models applied to the same portfolio can produce materially different loss estimates, particularly for secondary perils and non-peak regions. Regulators are asking firms to understand and justify the differences, which requires deeper engagement with model assumptions than many firms have historically maintained.

The Lloyd's Realistic Disaster Scenario (RDS) framework and the emerging systemic risk scenarios are pushing the market to consider correlated and cascading events that single-peril cat models do not capture. A major earthquake triggering a tsunami, a port closure, and supply chain disruption across multiple lines of business is a scenario that requires multi-peril, multi-line modelling capabilities that vendor models are only beginning to address.

How AI changes this

Machine learning is augmenting physics-based models for secondary perils where traditional approaches underperform. For wildfire, ML models trained on satellite vegetation indices, soil moisture data, and historical fire perimeters improve hazard estimation beyond what fire spread physics models alone achieve. For flood, ML-based surrogate models can approximate computationally expensive hydrodynamic simulations at a fraction of the runtime, enabling real-time exposure monitoring during events.

Real-time exposure monitoring during catastrophe events uses AI to estimate losses as events unfold. Satellite imagery analysis, social media data, and IoT sensor networks feed into models that update loss estimates hourly rather than daily. For reinsurers managing retrocession programmes and capital providers holding cat bonds, faster loss estimation reduces uncertainty and enables quicker capital redeployment.

Parametric trigger calibration benefits from ML's ability to model the relationship between an observable event parameter, wind speed at a weather station, earthquake magnitude at an epicentre, and the actual insured loss. Better calibration of parametric triggers reduces basis risk, the gap between the parametric payout and the actual loss, making parametric products more attractive to buyers and more accurately priced for sellers.

Climate trend integration into forward-looking hazard views is where AI's value is most strategically significant. Rather than adjusting historical catalogues with simple trend factors, ML models can learn the complex relationships between climate variables and hazard intensity, producing event sets that reflect projected conditions rather than past conditions. This is essential for long-tail reserving and multi-year treaty pricing where the relevant time horizon extends decades into the future.

What to know before you start

Do not attempt to replace vendor models with ML models for peak perils. RMS, AIR, and CoreLogic have decades of development, validation, and regulatory acceptance behind them. The opportunity for AI is in augmenting these models for perils and regions where they underperform, and in providing complementary views that challenge vendor model assumptions.

Validation of catastrophe models is uniquely difficult because the events of interest are rare by definition. A model that performs well against historical events may fail on the next event because the next event is, by definition, different from the ones that came before. Backtesting against historical events is necessary but not sufficient. Stress-test your AI-augmented models against plausible scenarios that are outside the historical record.

Data resolution matters enormously. A flood model that operates at one-kilometre resolution misses the street-level elevation differences that determine which properties flood. High-resolution exposure data, geocoded to property level with accurate building characteristics, is the prerequisite for any cat model improvement. If your exposure data is geocoded to postcode centroid, invest in data quality before model sophistication.

Start with a secondary peril where your portfolio has material exposure and the vendor model has known limitations. Flood in the UK, convective storm in continental Europe, or wildfire in the US and Australia are common starting points. Build an ML-augmented view alongside your vendor model, compare the outputs, and use the divergence to drive better-informed underwriting and risk assessment decisions. The goal is a richer understanding of uncertainty, not a single better number.