The gathering, organized by the Willis Research Network (WRN), heard how a perfect storm and its accompanying storm surge could wreak residential losses of between $36 billion and $140 billion.

The wide range of loss estimates reflects 89 track variations and the sensitivity of using differing assumptions in calculating damages. Some disaster scenarios envisaged storm surges, coinciding with a high tide 10ft higher than Manhattan’s sea walls.

Rowan Douglas, chairman of WRN, acknowledges that the wide range of possible loss outcomes demonstrates how much variance still exists in models—but he says that model technology is improving all the time.

“The insurance industry has always used historical data to form the basis of the event sets that drive insurance industry loss models,” he told lloyds.com. “But today’s climate is more dynamic than previously and it means that we are now at the limits of the value we can extract from historical data alone.”

But, as the Princeton conference delegates heard, climate scientists are developing physically based models of the planet that simulate the world’s oceans and atmosphere. They’re called General Circulation Models or, sometimes, Global Climate Models.

“We’re entering a new era whereby we can begin to assess the regional impacts of a changing climate,” he says. “At the Princeton seminar we discussed how this emerging science will inform future cat models the insurance industry uses in its catastrophe risk management.”

Representatives from the leading model shops AIR, Eqecat and RMS explained to the conference how they were actively incorporating the latest scientific thinking into their products.

“We’re often asked by clients what is the best model to use,” Rowan Douglas says. “But there is no single ‘best’ model. Our philosophy is to take an ensemble approach by obtaining more than one modelled output.”

While it may take many months before the full extent of insurance claims are known, the (re)insurance industry, analysts and regulators are keen to know asap how badly each company has been impacted. To fill the information void, the big three catastrophe modelling companies scramble to publish real-time loss estimates based on their proprietary model technology and their view of industry exposure. 

But why? Why do they take the enormous downside risk of getting it wrong when there is so much uncertainty pre-landfall and in the first few days following a hurricane event? There is a grave danger of misinforming clients as well as giving cynics license to engage in a bit of model-bashing if they miss the mark.

The answer: they have to. 

First, Real-time loss estimation has been sold as a product feature, targeting both buyers and sellers of Live Cat protection.  Also, supporting certain Insurance Linked Security (ILS) transactions commits model vendors to developing event windfield footprints, damage and loss estimatesto synthetic industry data.

So how well do they do? Naturally all will tell you that they do better than the others! AIR Worldwide feel sufficiently pleased with their performance in Ike that they have released a report comparing their model estimates with the latest figures from PCS. 

Personally, I think Ike demonstrates precisely why model companies should resist the rush to publish early loss estimates.

First, damage to offshore energy interests was unusually high considering the Saffir Simpson cat 2 classification that Ike had, not to mention the enormous scale of the windfield. We have learned from Katrina and Rita that it also takes a while to fully appreciate the extent of claims for damage to subsea interests.

Second, none of the ‘hurricane’ models extend far enough inland to generate meaningful losses to Ohio in the way that Ike did (over $1bn and counting).

Unlike some of the experience of model usage in the banking world, the insurance industry  has benefitted enormously from disciplined adoption of catastrophe modelling technology to support portfolio and capital management. Unfortunately real-time loss estimation is not where the models add real value.