Unearthing uncertainty: Using risk models to make uncertainty visible

Practitioners in catastrophe risk know the risk landscape has become more fundamentally complex. Today, loss outcomes are not only shaped by hazard intensity but also by urban density, exposure growth, infrastructure dependency, valuation uncertainty, and post-event dynamics that continue to evolve long after an event has passed.

In this dynamic risk environment, catastrophe models play a critical role. But when modeled losses diverge from observed outcomes, it is natural to ask questions and to examine the model in detail, to understand whether it got it wrong.

It is understood that models are not predictors of precise outcomes; models offer frameworks for understanding and managing uncertainty. When divergence occurs, it provides a valuable opportunity to focus on how uncertainty flows from models into decisions, where that process can break down, and allow time for constructive reframing. Time to ask what we can learn, and how models can be used more effectively in the future?

 

Learning from events is not the same as model failure

Over the past two decades, real catastrophic events have repeatedly expanded the industry’s understanding of loss. Storm surge reshaped hurricane risk after Hurricane Katrina in 2005.

Severe convective storms emerged as a tail‑risk peril following the April 2011 outbreaks in the U.S.  In 2017, Hurricane Harvey demonstrated that inland flooding could dominate losses from tropical cyclones.

More recently, wildfires such as Tubbs, Camp, Lahaina, and Los Angeles in 2025 highlighted the importance of urban conflagration and system-level loss dynamics.

Each event represents an opportunity for genuine learning and a chance to build and enhance a risk model. As new loss mechanisms became visible, models have evolved. Expecting models to anticipate every future pathway before it appears in reality is neither realistic nor useful. What matters is how insights are incorporated, and whether users understand the boundaries of what models are designed to represent at any point in time.

 

Where does the perceived ‘model miss’ originate from?

In practice, much of what is labeled as ‘model miss’ does not emerge from hazard science, but from modeling workflows.

Several contributors recur consistently:

• Hazard simplification: Deterministic scores can indicate where an event may arrive, but will often struggle to represent how losses propagate through dense urban systems. In wildfire, for example, a large share of destroyed structures may lie outside mapped high-risk zones because damage is driven by ember transport and structure-to-structure spread rather than vegetation alone.
• Exposure quality and valuation: Undervalued exposure directly leads to understated modeled loss. Indexing insured values to the Consumer Price Index (CPI) rather than construction cost indices can introduce systematic bias, while confusing market value with replacement cost can further distort loss estimates. Rather than a modeling error, it is a user input issue.
• Vulnerability variability: Two adjacent buildings exposed to a similar hazard can experience very different damage. That variability is real and irreducible. Models can represent it probabilistically, but they cannot eliminate it.
• Sampling and convergence: Insufficient stochastic sampling introduces noise that can appear as instability or mispricing. Two users running the same model with different convergence settings can arrive at materially different outcomes.

In our experience, model miss tends to surface when simplifications, data gaps, and uncertainty are not made explicit to your decision makers.

 

Why better data reshapes outcomes even when averages don’t move

One of the most counterintuitive insights from catastrophe modeling is that improving exposure data may leave portfolio-level averages largely unchanged, while materially redistributing loss at the location level.

This matters. When exposure quality improves, mitigated properties are more clearly distinguished, pricing signals sharpen, and mutuality is preserved. When exposure quality is poor, risk differentiation blurs, and ‘model miss’ can become a convenient fall-guy for outcomes driven by information gaps. Better data does not just improve accuracy; it improves fairness and governance.

 

Uncertainty does not end at the event boundary

Another source of surprise is the evolution of insured losses after the event itself. Post‑event dynamics, labor constraints, supply‑chain disruptions, regulatory responses, litigation, social inflation, and recovery programs can materially increase losses beyond initial physical damage. These forces often sit partially or entirely outside the modeled view, yet they are central to financial outcomes.

Treating modeled loss as a point estimate rather than a distribution with drivers makes these additional dynamics feel like a failure of estimating, rather than the continuation of uncertainty through time.

 

A more constructive way forward

For those involved with catastrophe risk, the goal is not to eliminate uncertainty, but to make it visible.

Organizations that manage catastrophe risk most effectively tend to share several practices:

• Clarity on what models cover, and what they do not
• Disciplined attention to exposure quality and valuation
• Sufficient sampling to achieve convergence
• Understanding of variability rather than reliance on averages
• Explicit consideration of post-event dynamics in decision‑making

When these elements are present, divergence between modeled and observed loss becomes explainable and manageable.

 

Closing thought

Practitioners know that catastrophe models are not crystal balls; it would be a perfect world if they were. It is understood that their value lies in understanding uncertainty, not removing it.

When assumptions are transparent, inputs are well understood, and uncertainty is treated as a feature rather than a flaw, models remain the most powerful tools the industry has for navigating an increasingly complex risk landscape. Decisions improve when modeling simplifications are understood and governed.

 

Find out more about Moody's RMS catastrophe models here.