The problem with any data-driven model, in any realm of finance or life, is that it’s backward looking. Data is a record of the past.
This is nothing new, and insurance actuaries know how to make data work, by aggregating a lot of it, over time. Time and history even out life’s vagaries. Outliers always revert to the mean: had investment banks’ real-estate modeling included the 1920s and 1930s, their subprime mortgage businesses in 2008 might have turned out a little safer.
But what if the historical record deteriorates? Human nature doesn’t change, and financial markets reflect human psychology. But nature’s nature is changing, and nature doesn’t care what humans think or do.
This is a problem for the reinsurance industry and the rest of the insurance value chain when it comes to covering physical catastrophic risk, such as damage from floods, fires, and typhoons. (Or earthquakes and volcanoes, but their patterns aren’t impacted by climate change.)
This is not just for direct protection, but for modeling insurance-linked securities (ILS), aka catastrophe ‘cat’ bonds, where capital markets are utilized to mobilize other resources to provide cover.
“Cat-bond models need to address and struggle with climate change,” said Mohit Mehra, vice president for analytics at Peak Re in Hong Kong. “It’s having an impact on underwriting, pricing, and portfolio management.”
Fortunately, artificial intelligence is going to bring more reliability the models the industry uses to assess risk, particularly in the field of cat bonds. But AI algorithms rely on data, and that data needs to be detailed enough to enable the kind of cross-pollination that makes predictive analytics useful.
And in Asia that type of granular data is rare and hard to come by. This is why the cat-bond market in Asia is tiny. As a result, when natural disasters strike, there’s no insurance or capital market to absorb the financial blow. It’s left to taxpayers, businesses and individuals to pay.
What is the data that is so vital to supporting a cat-bonds market for Asia? First we’ll look at the rising costs to Asian countries. Then we’ll have Mehra describe how the models work. Finally we’ll ask how the right combination of private-sector development in AI and regulation can bring about real improvements.
Rising toll
The number of ‘peril losses’ is increasing along with the value of damage. This is partly because the value of underlying property has grown – a good problem to have, as it reflects economic prosperity and population growth. But it’s also because weather- and climate-related events are happening a lot more frequently.
North America represents around 74 percent of the global $50 billion ILS market, according to the OECD. Europe is 19 percent, Japan 14 percent. Beyond Japan, Asia isn’t much of a participant – 3 percent out of Australia, and the remainder of the region’s cat-bond issuance is an asterisk.
But Asia, with its huge population and dynamic economy, is sustaining heavy losses from weather events. In 2023, a typhoon ripped through China, killing 81 people in the Beijing and Tianjin areas. It caused $32.2 billion in economic losses, according to Mehra, but only $1.4 billion of that was insured. The difference is called the protection gap, and this one is big. Mehra says the protection gap against flooding in China is even bigger.
Last year, while mega-hurricanes like Milton made the US once again the point of greatest insured losses, Asia Pacific was close behind, accounting for 40 percent of losses, according to Christian Aid, an NGO. Those losses are based on what’s insured, so in poorer areas of the world, the devastation is far greater. Indian heat waves, typhoons hitting the Philippines, floods in Pakistan and Indonesia, Australian fires – this region is seeing an uptick in peril events.
How the models work
Cat models quantify in monetary terms of the impact of a natural disaster, by asking three questions: where’s been hit, by what, and how (if) it is insured.
Mehra spoke recently at the Insurtech Insights conference, where he described how cat-bond models work. It’s helpful to know this to understand the data that is required to make them work well.
Models rely on three inputs to answer this.
First is a hazard model, which is like a menu of events – earthquakes, floods, typhoons – with ‘intensity parameters’ (such as wind speed or an earthquake’s ‘peak ground acceleration’). This is the where.
Second is a vulnerability model, a catalogue to describe the buildings in the impact zone. The hazard model’s type of disaster and its intensity is then compared to this catalogue to generate a ‘damage ratio’. This is the what.
These models needs to be specific, down to each building’s level of vulnerability. This is based on its design, the number of stories, its age, its occupancy, and the quality of its construction. These factors are then mapped against the hazard model.
A four-story brick building from 1970 will have a higher damage ratio than a modern concrete-and-steel building whose construction followed a strict safety design.
Third is a financial model that takes the damage ratio and uses a variety of statistical methods to determine the insurance at risk, and the loss in dollar terms. This is the how much.
When a earthquake hits Beijing, say, the quake’s peak ground acceleration times a building’s damage ratio provides a calculated value of the property and its loss. Overlay that on top of the fine print from an insurance contract (including any deductibles and policy limits), and the insurer determines the gross loss.
The curves move
As a final step when looking head, an actuary lays this scale of severity against one of frequency, and this creates a curve on a graph that represents the probability of any events that impact a policy, says Mehra.
If only it were that simple to calculate a risk.
“Sometimes the curve changes shape,” Mehra said. “There is an increasing chance of a typhoon or flood.” And while the chance of an earthquake remains low, prosperity and population mean the severity of such an event is increasing.
How are cat models used in practice? They generate an Average Annual Loss, or AAL, which is the premium charged to cover the risk from a peril over time. This is the insurer’s starting point to providing a price.
Another factor is the Return Period Loss, which is similar to a variable-at-risk model used by asset managers to define their risk appetite. It estimates the probability of a single, major event exceeding a certain amount: there’s a one-in-a-hundred chance in any given year over the 100-year life of a policy that an event causes more than $75 million of loss.
Insurers rely on Return Period Loss to determine how much risk they think they can handle on a given policy, and how much they’d pass on to a reinsurer. Balancing the appetite (the insurer’s capacity) for risk is also called ‘accumulation management’.
These also need to be disclosed so regulators can monitor the financial stability of insurance companies.
An underwriter must therefore work with pricing inputs from the AAL and capacity limits from the RPL, and also consider regulations and the insurer’s financial status, that is, its ability to pay out claims while maintaining a viable business.
More data, please
From this quick introduction to cat modelling, it is obvious that an insurer with larger data sets with lots of nitty-gritty detail can make better decisions on what protection they can offer, at what price, and how much capacity their reinsurers have.
More data will mean they can make bigger and more diverse markets in protection. When data is scant or vague, insurers are forced to assume the worst. This means less capacity, delivered at higher prices.
This is even more urgent with climate change, because those severity and frequency curves are on the march. But AI is good at using all sorts of traditional and newfangled data (such as satellite images) to predict damage ratios, pricing, and capacity.
But it’s also clear that the amount and granularity of data that is required regarding property is significant. In many emerging markets, the basics such as who owns what land, and what is the boundary of that title, isn’t formalized. Rule of law can be sketchy. Never mind details about every single building.
Nonetheless, authorities around Asia can do more to require more specific data be disclosed, and not just insurance authorities or stock exchanges (for listed companies), but ministries overseeing health and safety, agriculture, construction, small businesses, tax, and other areas.
APIs ASAP
Uunderwriters can do more, too. They are still using manual tools to crunch their numbers, says Mehra. “Today, we have to wait for analysis runs and extracting results. There’s a lot of touchpoints before you have useful information.”
He says APIs are starting to streamline the process, connecting inputs and models among carriers, insurance brokers, specialist modeling vendors, and reinsurers. And as the data inputs become more specific, the models can do a better job of matching the most relevant information to a given peril.
This will not only improve pricing, but also help insurers build a portfolio of cat exposures. They’ll be better able to examine a possible policy and determine whether it diversifies their exposure or doubles down on existing risks – or even if it would breach the insurer’s risk-management limits.
Insurers are only part of this development: regulators also need to find ways to make data more available, and more specific. “A lot of data in Asia is aggregate,” Mehra said. “We talk about making models more granular, but if the only data for a peril like a flood is aggregate, what do you do? You can only work with high-level assumptions, and anyone pricing the risk must assume the worst.”
The World Bank agrees: it says the lack of robust modelling of risks limits the ability of emerging markets to offer cat bond natural disaster coverage, even to governments. To date, the only Asian sovereign that has issued a cat bond is the Philippines.
The promise of AI
But AI can speed up developing new models and make them more reliable. It doesn’t need to rely on traditional sources. If those traditional sources were easy to handle, there’d already be more data available.
But AI is good at integrating unstructured data such as satellite images, real-time weather updates, and ground detection devices with historical records. It can also do this much faster, so that damage and claims can be assessed closer to real time as events unfurl.
AI is also able to identify patterns and anomalies, which means it can use correlating information to make guesses about construction materials and methods that today a human analyst could not calculate.
Finally, AI can also validate data – ensuring data is consistent, for example – before it’s inputted into risk models. That validation will also make for better models.
Overall, then, AI should have a hugely positive impact on the insurance industry’s ability to provide cover against natural disasters, despite new uncertainties because of climate change.
AI can allow the industry to rely less on backward-looking data and more on forward-looking predictive analysis. In theory this should be especially useful in markets where traditional data is hard to come by. But it will require investment throughout the insurance value chain, and a more aggressive push by governments to make both traditional and new types of data easier to access.
