The Intelligent Quarterly from the publishers of The Insurance Insider

Spring 2012

Search archive

Advanced search

"Catastrophe models are like TomToms; they give you a good idea of where you're heading but you still need to look out the window to get there," says Aspen Re's head pricing actuary Paul Gates.

So are catastrophe models just that: useful tools, but potential distractions from what is on the road ahead? Or is there a problem with the underlying assumptions that insurers the world over have put their faith in?

Since their debut in 1987, catastrophe models have become the primary risk management tool for those operating in the P&C sector. But in the past six years there have been calls to end overreliance on models, in favour of a more broad-based approach to managing catastrophic risk.

Eqecat president Bill Keogh said the catastrophe model's greatest strength is its ability to help "to set rational expectations about risk". In other words: expectations of a company's exposure and its probable maximum loss (PML) in a 1-in-100 year event. But what can users rationally expect of a model's capabilities?

Not 'accuracy' - which in Keogh's opinion has no place in modelling discourse. Those expecting exact loss estimates have misunderstood that models exist to simplify a very complex reality.

The best way to think about a model is as "a platform for understanding uncertainty in catastrophic risk" says Keogh. They are probabilistic, not deterministic and like any tool have limitations.

One of the greatest limitations is models' reliance on historical record, which is very sparse for low frequency, high intensity hazards. This year's Japanese earthquake and tsunami will provide a rich source of data for the future. Until Tohoku, Japan had never seen a magnitude 9.0 quake, so there were no mega-quakes in the model. Seismologists studied the country for years, but none predicted waves of 38.9 metres, travelling 10km inland.

Similarly, the Christchurch quake occurred on an unknown fault. Scientists had mapped out many scenarios for a huge quake in Wellington - the expected location of a large New Zealand quake - but not for Christchurch. Following the quake, the fault has been identified and can be factored into the model, but there are so many other unknowns that will only come to light when disaster strikes.

"[Catastrophe models] are a tool to use in pricing that tells you whether it is high or low risk," says one senior property underwriter. "We know there are biases in the models, so you don't take the number as gospel, but it's an indicator".

While model output can be volatile, the general consensus seems to be that if you can afford to buy a model licence, it would be remiss not to. "It's a tool - it's far superior to having no tool," says Kirk Bitu, senior vice president of reinsurance at BMS.

Sleepless nights
At a recent Eqecat conference an industry panel was asked what kept them awake at night in relation to catastrophe models. Omissions, biases and the quality of exposure data topped the list.

Jeff Tennis of Lockton's catastrophe analytics team agrees that data quality is a problem, but says the wider issue of how data is utilised and understood is what really needs to be addressed. "Since Katrina, data quality has improved dramatically within the US, but there is still room for improvement," he says.

"Insurance companies and their intermediaries should understand the reasonability and impact of data assumptions, which directly impact modelling output. Are 5,000 square foot homes common in the North East? How sensitive is the portfolio to storm surge?"

A clear understanding of modelling output is also vital he says. "Do the consumers of catastrophe modelling output really understand what a 1-in-100 year PML figure represents?"

Blowing in the wind
With the Atlantic hurricane season upon us, the record of hurricane models is under the spotlight once more. The predictions of forecasters at AIR, RMS and Eqecat came very close to the actual number of storms in 2007 and 2008. They underestimated the number of storms in 2009 and 2010, but were still in the ballpark. When it comes to predicting landfalling storms, however, their record has been patchy at best. Predicting where storms will make landfall is difficult because short-term weather conditions often determine the location.

Insured loss estimates for landfalling storms have often been far off the mark also. In the past decade, actual incurred losses from US hurricane events have frequently deviated from modelled losses by a factor of two or greater.

For a primary insurer the accuracy of the models isn't a huge concern, according to Bitu. "Models are just a tool to manage the book, whether they are right or wrong isn't as important as the fact that you transfer the risk," he says.

However, there is still the possibility of an insurer being dramatically under-reinsured due to an anomalous model output. The recent RMS update is evidence of this. Since the 2004-05 season the industry has felt that RMS was too low in its loss estimates, says Jeff Tennis, and has adjusted model output to account for it.

Therefore, those who treated the numbers in previous RMS models as gospel will be punished by the model update. AM Best took up this theme in a recent memo, explaining how model updates are incorporated into ratings. The rating agency emphasised the importance of recent data in the analysis of catastrophe risk and its resulting impact on risk-adjusted capitalisation. In cases where models have been updated AM Best expects companies to incorporate revisions in model output as soon as practicable.

When asked whether downgrades were inevitable in the aftermath of model updates, the firm stressed that it considers a company's overall approach to risk management.

The firm said in a recent webinar that although some "marginal companies" may have experienced downgrades, there have been no ratings changes due to model changes alone.

Advantages and limitations
While he believes that many of Eqecat's clients are starting to use models appropriately, Keogh thinks that long periods without a big disaster lead to a resurgence of bad habits. "The longer we go without a huge loss event people think models don't only predict but control risk," he says. "Models should never surprise us, but events will always surprise us."

However, both are optimistic that the industry is moving towards a better understanding of models and their capabilities. While catastrophe models still influence the business insurers choose to write and how much reinsurance they choose to buy, they are seen for the most part as a portfolio optimisation tool and are good for identifying whether a portfolio is sufficiently diversified or shows evidence of strong correlation. They can also be applied to assessing the portfolios of parties considering a merger, to determine the correlation between the two entities.

Where models still fear to tread is in the area of secondary perils. Tsunami risk, levee failure, contingent business interruption claims and the nuclear perils are still not included in most models.

Likewise, models cannot capture every change in every niche market. There will be specialist lines where the underwriters themselves are aware of building code enforcements or regulatory updates that are not in the models. Models are simply not attuned to the subtle nuances of specialist business.

Models are a platform for understanding uncertainty. Many in the P&C industry recognise that the old way of approaching them, with a fixation on PMLs, needs to evolve. Some however remain wedded to the false sense of security that models can provide.

How catastrophe models are constructed