Using global sensitivity analysis to understand the drivers of model variability
Catastrophe models inform underwriting, capital management, regulatory reporting and investment decisions. As these models become more sophisticated, understanding why their results change is becoming just as important as understanding the results themselves. Here, Owen Hinks, Catastrophe Model Developer at JBA Risk Management, explains how global sensitivity analysis (GSA) is helping us understand which model assumptions have the greatest influence on model outputs, and why that matters for interpreting model results and supporting better model development and governance.
Flood catastrophe models continue to evolve, incorporating higher-resolution terrain data, improved exposure datasets, updated flood defences and increasingly sophisticated representations of physical processes. These developments improve our ability to represent flood risk, but also make it harder to understand why modelled losses change. For model developers and model users alike, this raises an important question: when flood loss estimates change, which components are responsible?
One method of answering this question is global sensitivity analysis (GSA), which systematically explores how changing multiple inputs to a model influences its outputs. Rather than simply showing that losses change, it helps identify why they change by quantifying the relative influence of different model components.
We applied this approach to JBA’s recently updated Global Flood Model. As described below, the analysis shows that there is no single dominant driver of modelled flood losses. Instead, the relative importance of terrain, climate representation, vulnerability and flood defences varies between locations, illustrating how different components influence model results in different contexts. This work was presented at EGU 2026 and builds on our recently published paper: Towards global sensitivity analysis of large-scale flood loss models.
When many model components change at once
Recent updates to JBA’s Global Flood Model span multiple components, including hazard maps, exposure representation, flood defences and climate scenarios. Individually, these developments aim to improve how the model represents real-world flood risk. Collectively, however, they make it difficult to determine why modelled losses change between model versions.
This study focuses particularly on improvements to terrain representation, while recognising that these interact with other components throughout the modelling chain.
Terrain data and global flood maps
JBA continually updates its global flood maps as improved datasets become available. In this model release, improvements include 30 m resolution flood maps across 17 countries, alongside the continued use of high-resolution 5 m mapping in selected regions. A global grid alignment has also been implemented to ensure consistency between datasets of differing resolution.
One important aspect of these updates is improved terrain representation. The fidelity of water depth estimates is closely linked to both the resolution and accuracy of digital elevation models. Consequently, many countries have transitioned from digital surface models (DSM) to digital terrain models (DTM). These bare-earth DTMs remove buildings and vegetation, providing a clearer representation of overland flow pathways in flood modelling. Figure 1 demonstrates the significant difference in terrain representation with different data types and resolution in Wrocław, Poland.
Understanding interacting model components
Improvements to terrain data are only one part of the story. Recent updates have also improved how exposure is distributed, moving from population-based weighting to building footprints.
This illustrates a broader challenge in catastrophe modelling. Flood catastrophe models combine many interacting and changeable components, including hazard data, exposure distribution, vulnerability functions, defence assumptions and climate signals. Improvements to any one component may increase model realism, but understanding why those changes occur requires looking at the model as a whole.
Using GSA to investigate model behaviour
By applying GSA to JBA’s Global Flood Model (Figure 2), we can compare the relative influence of terrain representation, flood defences, climate assumptions and other model components.
Study design
This study focused on two countries to isolate different aspects of terrain representation. Austria was used to investigate the influence of terrain resolution (30 m vs 5 m), while Poland was used to compare terrain type (DSM vs DTM). The GSA framework used was developed through an InnovateUK Knowledge Transfer Partnership and is described in more detail in our recent research into applying GSA to large-scale flood catastrophe models.
Across both countries, the analysis also varied flood defence representation and the choice of global climate model, representing different plausible patterns of future climate change, allowing the influence of terrain to be assessed alongside other important model components.
By systematically varying these inputs, GSA produces sensitivity indices ranging from 0 (little influence) to 1 (high influence), allowing the relative influence of each model component on flood loss estimates to be compared consistently.
What we found
The analysis showed that there is no single dominant driver of variability in modelled flood losses. Instead, the relative influence of terrain, vulnerability, flood defences and climate assumptions depends on both geography and the modelling choices being considered.
Terrain resolution is a primary driver of loss variability in Austria
In Austria, terrain resolution is the dominant influence on average annual loss (AAL) and the second most influential factor for the 200-year aggregate exceedance probability (AEP) (Figure 3). At higher return periods, where flood depths are greater, changes in flood depth are less likely to vary the damage ratio, which may explain the reduced sensitivity of 200-year AEP to terrain resolution.
Using higher-resolution (5 m) terrain data produces more refined flood extents and depths, reducing both the magnitude and variability of modelled losses compared with 30 m data. Figure 3c demonstrates this effect, with both the median 200-year AEP loss and the spread across the model runs decreasing.
Terrain data type has a secondary influence in Poland
In Poland, terrain type (DSM versus DTM) has a much smaller influence on modelled losses than terrain resolution had in Austria and is less influential than several other model components (Figure 4). This reflects the fact that flood extents are broadly similar between the two representations, while differences in elevation modelling primarily affect water depth.
Although not as influential as other parameters in the GSA, improved terrain data type does vary flood loss estimates (Figure 4c). DTM-based modelling improves the representation of flow pathways, leading to improved depth estimates and, consequently, differences in damage ratios. The influence of terrain representation therefore depends on the modelling context.
Importance of vulnerability
Across both countries, vulnerability functions are consistently among the most influential model components. This is unsurprising: vulnerability functions sit close to the loss calculation, translating hazard intensity into damage. As a result, relatively small changes in vulnerability assumptions can produce comparatively large changes in modelled losses.
This also highlights the importance of developing hazard and vulnerability components together. Improvements in one part of the model are unlikely to deliver their full benefit unless they remain consistent with the others.
Sensitivity varies spatially
The Austria case study also illustrates that model sensitivity varies geographically rather than remaining constant across a country (Figure 5).
Mountainous areas exhibit greater sensitivity to terrain resolution due to their more complex topography. In contrast, lower-relief and more urban areas, where higher-resolution terrain data provides the greatest refinement of flood extents and depths, corresponding improvements in vulnerability representation are also needed to reduce uncertainty in modelled losses.
In highly defended areas, such as Vienna, flood defence representation can become the dominant driver of variability in modelled losses, reflecting the influence of the city’s defence designed to a 100-year standard of protection.
This spatial variation reinforces the need to assess model performance at appropriate geographic scales.
Model components interact and must be balanced
Taken together, these findings highlight an important principle: model improvements cannot be assessed in isolation. Improvements to terrain representation, exposure, vulnerability, or flood defences all influence model behaviour, but their effects depend on how the model represents the whole system. Understanding those interactions is therefore essential when interpreting changes in modelled losses and deciding where future model development effort should be focussed.
These results describe how the current model responds to different assumptions. They complement, rather than replace, other forms of model evaluation based on observations, benchmark models and physical understanding.
Implications for financial decision-making
Flood risk information is increasingly used beyond the insurance and reinsurance sector, supporting decisions in banking, investment and climate-related reporting. Organisations need to understand not only the level of risk indicated by a catastrophe model, but also how sensitive that assessment is to different assumptions and data sources.
Global sensitivity analysis provides an additional line of evidence when evaluating catastrophe models. It complements observations, benchmarking and physical understanding by helping explain how models respond to different assumptions. In turn, this helps identify which model components are therefore likely to have the greatest influence on model outputs used in risk assessments.
For insurers and reinsurers, these insights can support pricing, underwriting and portfolio management by improving understanding of how different model components influence loss estimates. For banks, investors and organisations undertaking climate-risk or ESG reporting, this understanding provides greater transparency around model outputs and help identify where improvements to data or model representation are most likely to influence the risk assessments built on those outputs.
Final thought
As catastrophe models continue to evolve, understanding how they behave is becoming just as important as improving their performance.
Global sensitivity analysis helps build that understanding by showing how different model components influence results and how those influences vary between locations and applications. Used alongside observations, benchmarking and physical understanding, it provides another valuable line of evidence for developing, evaluating and applying catastrophe models with confidence.
How JBA could help you
JBA provides flood hazard data and catastrophe models for every country in the world under both present-day and future climate scenarios.
Our flexible and consistent modelling approach enables you to tailor model data, parameters and assumptions that help you understand flood risk at property, portfolio and regional scales.
If you would like to learn more, get in touch.