Seeing the full spectrum of climate disaster risk: Why probabilistic catastrophe models matter in a world of escalating extremes

Why probabilistic catastrophe models matter more than ever

From devastating floods in Europe and Asia, to record wildfires in North America and Australia, to increasingly destructive hurricanes and typhoons worldwide, climate-driven disasters are imposing growing human and economic costs. Annual losses from natural catastrophes now regularly reach hundreds of billions of dollars. For businesses, banks, investors, insurers and governments, understanding acute physical climate risks has become a strategic priority.

Yet many climate risk assessments still rely on non-probabilistic approaches that do not simulate the full range of possibilities at the event level and that typically capture only a fraction of the risk. In this article, we explain why probabilistic, or stochastic, catastrophe models – used by the insurance industry for over three decades – provide a more complete and decision-relevant view of natural catastrophe risk.

Key takeaway messages:

• Real resilience requires understanding the full range of possible outcomes. Averages and single events hide risk.
• Understanding risk as a distribution, not a single number, is key to informed decision-making.
• Lessons learnt by the insurance industry, which has relied on probabilistic catastrophe models for over 30 years, can be applied to non-Insurance sectors.
• Probabilistic catastrophe models simulate thousands of plausible events, not just what has happened before.
• They go beyond the historical record to assess events that are plausible, even if never previously observed.
• These models capture both everyday losses that accumulate over time and rare extremes that drive financial distress.Uncertainty is explicitly quantified and made decision-relevant.
• Stochastic models capture the hidden vulnerability of interconnected systems, revealing aggregate risk concentrations and how losses compound across locations. Probabilistic models translate physical hazards into financial impacts across assets and portfolios in a traceable way.
• Loss outputs are validated via benchmarking against observed losses from historical events.
• Stochastic catastrophe models can be climate-conditioned to assess how risk may evolve under different future climate scenarios.

Probabilistic catastrophe models consider full risk distributions

Probabilistic catastrophe models are built using a multidisciplinary approach that brings together experts in climate science, engineering, geospatial analysis, and economics to construct realistic representations of how natural hazards interact with the built environment. These models simulate the physical behavior of perils such as hurricanes or wildfires, the hazards they generate—such as wind, rainfall, heat, or embers—and how those hazards impact different types of properties, from multi-story office buildings and factories to single-family homes.

Non-probabilistic physical risk models typically rely on outputs from global and regional climate models under a limited set of emissions or warming scenarios. Hazard intensities are derived from aggregated climate variables – such as temperature, precipitation, wind, or sea level – often using multi-model ensembles from the Coupled Model Intercomparison Projects (CMIP) coordinated by the IPCC and spatial downscaling techniques to translate coarse climate model outputs to asset-level metrics.

These approaches generally characterize changes in hazard conditions at specific time horizons rather than simulating large numbers of discrete events. As a result, they may smooth variability across models and scenarios and provide limited representation of event-level extremes and tail risk compared with stochastic, event-based catastrophe models that explicitly sample a wide range of plausible hazard realizations and their probabilities.

Other non-probabilistic models may rely on a limited number of deterministic scenarios, for instance, a “100-year flood,” a repeat of a notable historical wildfire, or a single severe hurricane track. While such scenarios are intuitive, they offer only a narrow window into risk. They implicitly assume that the past is a reliable guide to the future, and that a handful of events can capture the full range of possible outcomes.

Importantly, simpler non-probabilistic approaches are not just less complete, they can actively introduce risk by creating a false sense of confidence. A single “precise” scenario can feel reassuring, but precision is not accuracy when the real world contains wide variability and rare extremes. In practice, this can lead to:

• False comfort: decisions anchored on one outcome, while the true range of plausible outcomes is much wider.
• Underestimated tail risk: capital and contingency plans sized for the “average year,” not the stress year that can drive financial distress.
• Missed accumulation effects: resilience plans that fail because they overlook how losses can compound across events and locations.
  

Probabilistic catastrophe models take a fundamentally different approach. By simulating a wide range of plausible events and their probabilities, they move beyond single-scenario thinking and capture the full spectrum of risk, including rare but consequential extremes, events that have never been observed historically but remain physically possible, and how losses can compound across events and locations. In doing so, they address the key shortcomings of simpler approaches—reducing false comfort, making tail risk visible, and supporting decisions that are robust not just in the average year, but in the stress years that matter most.

Lessons already learned by the insurance industry

Before probabilistic catastrophe models became mainstream in the Insurance industry in the late 1980s and 1990s, catastrophe risk assessment relied largely on the actuarial approach. This approach extrapolated risk from historical insurance claims and past loss experience, assuming that the future would broadly resemble the past. It worked reasonably well for frequent, moderate losses, but it struggled to capture rare, extreme events and their devastating financial consequences [1].

Then, a succession of extreme disasters culminated in events such as Hurricane Andrew in 1992. Andrew and other major catastrophes produced losses far beyond the historical record – many times greater than the insurance premiums collected – and contributed to the financial collapse of several insurers and reinsurers. These shocks made clear that backward-looking actuarial methods were insufficient, and accelerated the shift in the insurance industry toward more sophisticated, data-driven, probabilistic modeling approaches, leading to the widespread adoption of stochastic catastrophe models capable of capturing rare but financially dominant risks. These models capture the whole range of physically plausible events, returning financial losses and their probabilities.

Modern probabilistic catastrophe model frameworks increasingly combine probabilistic foundations with data-driven, machine-learning techniques to improve components of the models.

Going beyond the historical record

History alone is not enough, especially for rare but severe events.

Extreme floods, major wildfires, or catastrophic hurricanes may occur only once in many decades or even centuries. The available historical record is simply too short to capture them all, and too short to provide a statistically representative sample. Hundreds of years of observations might be needed, but are rarely available, as reliable historical records only exist for approximately the last 6 decades. Probabilistic models address this by explicitly modeling plausible events with non-zero probability, even if they have never been observed in recorded history.

This matters because:

• A past loss is just one possible outcome
• No loss in a historical event does not imply zero risk
• An event like the next devastating event may have never happened previously in history.
  

Why the “probabilistic” or “stochastic” element is so powerful

The defining strength of probabilistic models lies in their stochastic event sets. The goal of these event sets is not to predict specific future disasters, but to generate enough “points” on the risk curve that its essential shape and statistics are captured. Individual simulated events have limited value on their own; it is their collective behavior that matters.

The stochastic catastrophe modeling framework

Probabilistic (or stochastic) catastrophe models take a broader, more realistic view. Rather than asking “What happened before?” they ask “What could plausibly happen?”

At a high level, these models follow a structured framework:

• Stochastic event generation. Thousands to hundreds of thousands of simulated events for floods, wildfires or storms, are generated using Monte Carlo probabilistic techniques across a range of connected potential probability distributions. These events are not forecasts. Instead, they are a statistically representative sample of the full range of plausible outcomes, consistent with climate conditions and physical constraints.
• Hazard modeling. Each event is translated into local hazard intensity, e.g., flood depth meters, wind speed, or wildfire burn probability. This step ensures that risk is assessed at a granular, asset-level scale rather than through coarse regional averages.
• Vulnerability modeling. This translates hazard intensity into physical damage. Different assets, e.g. residential, commercial or industrial buildings, respond differently to the same hazard intensity. A meter of floodwater does not have the same impact on a warehouse as it does on a data center or a retail property. Vulnerability functions capture these differences by estimating the fraction of asset value damaged at each hazard intensity. A broad catalogue of vulnerability functions is typically available for different specific building characteristics. These functions are derived empirically, based on engineering simulations and expertise, damage data collected by event-response teams on the ground after events, and loss data from large volumes of insurance claims.
• Financial loss modeling. Physical damage is translated into monetary impacts considering different types of damage, e.g. building structural damage, contents damage and revenue loss from business interruption. The result is a coherent estimate of financial loss for each simulated event, which can then be aggregated across years, portfolios, regions, or sectors.

The result is not a single number, but a full loss distribution that includes high-frequency, low-severity events, as well as low-frequency, high-severity events, for each acute peril.

Because probabilistic models are built from explicit components – events, hazards, vulnerability, and losses – their assumptions can be traced, tested and examined, rather than hidden inside opaque black-box predictions.

Validated loss outputs

Stochastic catastrophe models are validated by comparing modeled losses against observed losses from historical events – using insurance claims, disaster loss databases, and post‑event field assessments – to ensure simulated financial impacts are consistent with real‑world outcomes.

Capturing both the everyday and the extreme

For catastrophe risk, averages can be misleading. Two locations can have the same expected mean annual loss, yet one faces frequent small losses while the other is exposed to rare events capable of causing catastrophic damage. Probabilistic models reveal these differences.

A key advantage of probabilistic models is their ability to support “what-if” analysis across the full risk spectrum. High-probability, low-loss events, such as minor floods or small wildfires, may appear inconsequential in isolation, but their cumulative impact can be material over time. At the other end of the spectrum, rare, low-probability events can generate losses large enough to threaten the viability of firms, portfolios, or even entire financial systems.

By modeling the full loss distribution, probabilistic catastrophe models enable users to quantify average losses, volatility, and tail risk in a consistent framework. This is essential for robust risk management, capital planning, insurance structuring, and climate stress testing.

Quantifying uncertainty, not hiding it

A critical advantage of probabilistic catastrophe models is that they explicitly quantify uncertainty.

Uncertainty is not a weakness of catastrophe modeling – it is a fact of nature. Natural hazards are inherently uncertain:

• We cannot know exactly when or where the next major event will occur
• Damage outcomes vary widely, even for similar hazard intensities.

Probabilistic models embrace this reality and treat uncertainty the way weather forecasts do: by showing ranges, likelihoods, and worst-case possibilities, rather than pretending the future can be reduced to a single outcome. By simulating many possible futures, they allow uncertainty to be measured, communicated, and managed – rather than masked by a single scenario or point estimate. This makes uncertainty a decision input, not a blind spot.

The hidden vulnerability of interconnected systems: How disasters can impact multiple locations at once and concentrate losses across portfolios

Natural disasters such as floods, hurricanes, and wildfires do not affect assets independently: a single event can impact many locations at once, and extreme years often involve clusters of losses across regions. A key strength of stochastic catastrophe models is their ability to capture compounding connected risks by accounting for spatial correlations between events and losses across locations.

By explicitly modeling how events unfold across space, stochastic models can aggregate losses from multiple sites into a coherent portfolio‑level view of risk. This allows organizations to move beyond assessing locations in isolation and instead identify risk concentrations, understand how losses can compound in extreme years, and evaluate joint impacts across assets, regions, or business units. Such aggregate views are essential for stress testing, capital planning, and resilience strategies at the corporate or portfolio level, where it is often the combined effect of correlated losses – rather than any single location – that drives material financial outcomes. 

Climate-conditioned stochastic catastrophe models

Probabilistic catastrophe models can also be made forward‑looking by conditioning them on different future climate states. Rather than relying only on the past, these models adjust key drivers of risk – such as event frequency, intensity, and spatial patterns – using insights from climate science to reflect a number of possible future scenarios.

In practice, this means the same stochastic modeling framework is retained, but its underlying variables are systematically modified to represent plausible future climates. New event sets are then generated that are consistent with those conditions, producing future loss distributions that can be compared with today’s risk metrics. This approach allows decision makers to assess how physical climate risk may evolve over time while preserving the probabilistic structure needed to understand uncertainty, extremes, and accumulation effects.

A decision-useful foundation for climate risk assessment

In a world of escalating climate extremes, risk cannot be reduced to partial views or averages that may wash out extremes. Probabilistic catastrophe models provide a more complete, forward‑looking, and decision‑relevant view of physical climate risk: one that captures accumulation, extremes, and uncertainty in a single, coherent framework.

As climate risks are both systematic and deeply uncertain, understanding risk as a distribution rather than a single number is not just a technical refinement. It is a prerequisite for informed, resilient decision-making.

With an evolving climate, urbanization, and asset concentration amplifying potential losses, understanding the full range of plausible outcomes, not just the most likely one, is becoming essential for resilient decision‑making.

Probabilistic models are designed to support real‑world decisions, not just analysis. By showing how losses build up across frequent small events and rare extreme ones, they help decision‑makers answer practical questions: how much capital is enough, where risks concentrate, and how resilience investments change outcomes – not just what the “average” loss might be.