Earlier this year, analysis by the US climate change non-profit Berkeley Earth and by the EU’s space Earth-observation programme Copernicus concluded that 2024 was the warmest year since records began in 1850 – and the first to exceed 1.5°C above pre-industrial levels (Figure 1). Moreover, recent studies such as the IFoA’s Climate Scorpion report have shown that climate sensitivity may be greater than we think, meaning we could experience more warming than expected. 

Climate change might affect perils such as hurricane risk, which could increase as the temperature of the North Atlantic Ocean rises. Do our current models underestimate this, and how can we develop better techniques to assess it and put insurers on the front foot?

It is important to understand how sensitive the Earth is to greenhouse gases. This is measured by equilibrium climate sensitivity (ECS) – defined as the expected warming if greenhouse gas levels double from pre-industrial times. The Intergovernmental Panel on Climate Change’s (IPCC) Sixth Assessment Report gave a likely range of 2.5°C to 4°C, with a best estimate of 3°C. However, there is an 18% chance that ECS could exceed 4.5°C, and some scientists have suggested that it could be as high as 4.8°C due to uncertainties over factors such as ocean carbon uptake, aerosol cooling, tipping points and cloud feedback. 

We may be in for more warming than expected if ECS is higher than the IPCC’s 3°C best estimate. How can we incorporate this forward-looking uncertainty into hurricane risk?

The first hurricane risk analytics companies emerged just before Hurricane Andrew hit Florida in 1992, which caused $15bn in insured losses and 11 insurer bankruptcies. The industry’s perception of potential losses was drastically altered, triggering a risk management and pricing overhaul. 

As temperatures rise and extreme weather events become more frequent and severe, we are approaching another pivotal moment – evidenced by the recent bankruptcies of many Florida insurers, and several insurers’ withdrawal from California due the wildfire risk.

When Hurricane Beryl started in the Caribbean in July 2024, it became the earliest-forming category 5 Atlantic hurricane since records began around a century ago. This was attributed to the exceptionally high sea surface temperature (SST), which is a significant driver of hurricane risk – warmer seas provide more energy for hurricane formation and intensification. Beryl sped up very quickly and limited people’s opportunity to react.

Current natural catastrophe (nat cat) models are typically based on historical data, extrapolated to predict future events. However, as the warming climate changes the world beyond historical observations, these models face challenges. 

Many hurricanes form in the Main Development Region (MDR), an area in the North Atlantic stretching from Africa to the Caribbean, which is now warmer than it has ever been. According to Copernicus, the annual average extrapolar ocean SST reached a record high of 20.87°C in 2024 – that’s 0.51°C above the 1991-2020 average. How can we understand the MDR’s current hurricane risk given unprecedented SSTs, possibly breaching the 29°C temperature limit for the first time? Are there any options other than expert judgment? 

One possibility is to consider other hurricane regions such as the Gulf of Florida, which regularly experiences SSTs not yet seen in the MDR. By relating the physical properties of hurricane behaviour in the Gulf to those of the MDR, we can translate the known physics of hurricanes in the Gulf at higher SSTs to those of the MDR.

Nat cat modeller and data provider Reask has pioneered a methodology that combines machine learning on hurricane formation characteristics – including factors such as SSTs, steering flow, wind shear and mean sea level pressure – with advanced stochastic simulations of hurricane formation and tracks. The model simulates millions of hurricanes and provides a climate-informed, forward-looking risk assessment, rather than calibrating to landfall risk distributions.

We are approaching another pivotal moment – evidenced by the recent bankruptcies of Florida insurers

The model learns by overlaying historical climate patterns with historical hurricane event characteristics, creating a robust database on which to train machine learning algorithms that link climate patterns to hurricane characteristics such as global frequency, trajectory and intensity. The historical record is then used to validate the model by checking that it reproduces hurricane risk from observed climates, rather than calibrating to historical landfall risk distributions. 

This framework could help to quantify two types of uncertainty: the fact that the climate we have observed is just one realisation of multiple climates that could have occurred; and the fact that any one fixed climate could produce many different hurricane outcomes. 

Reask’s model gives the industry the tools to force hurricane risk models with different climates, to assess how climate change affects the entire risk curve, including the tail. 

Figure 2 shows how climate change affects landfall risk along the US coastline, with the coastline colour demonstrating how SSTs drive landfall risk. Model sensitivities can be run using different climate futures – for example, today’s climate versus hurricane activity based on the climate of a +2°C warming world, incorporating future SST, wind shear and steering projections.

Climate change raises severity in a range of perils but assessment of risk is complicated by lack of historical data and unprecedented conditions. Backwards-looking risk assessment based on trend analysis may miss non-linear shifts in the risk distribution. 

Reask’s methodology provides a more accurate simulation of these risks, using machine learning and AI to incorporate a stronger climate signal into the risk assessment. For example, it shows that hurricane risk could double if SSTs in the MDR hit 29°C. 

Could this methodology be adapted for other climate-driven perils to facilitate better risk management, pricing and assessment? 

Other uses could include better predictions for when insurance companies may start to increase prices or pull back from covering certain perils, so that policymakers and the insurance industry can engage constructively on this issue.

These forward-looking techniques may help us to avoid future surprises and put the industry on the front foot, allowing it to recognise non-linear shifts in risk distributions that are driven by faster-than-expected warming, rather than labelling events ‘black swans’. 

Such techniques might also help to address the criticisms levelled at climate change risk management, improving our understanding of how climate could affect risks, helping society to imagine what effects may be coming and prepare for them, and improving how we communicate the changing risk environment.

is a climate risk and thematic manager at M&G

Browse the latest IFoA climate change report