As catastrophe modelling continues to increase in sophistication, it becomes ever more vital for catastrophe reinsurers. Robert Muir-Wood explains what RMS has been up to of late.

With the soft market for global reinsurance set to deepen, catastrophe reinsurers find themselves in increasingly difficult times. Like trawlers in an expanding fishing fleet, plying an ocean depleted in the richest stock, reinsurers must harvest risk in the knowledge that little of it guarantees a good return. More than ever at such times catastrophe modelling technology becomes vital in the search for opportunities in the global risk market: revealing where the market prices risk higher than the underlying technical premium, making it possible to test the fit of each new piece of potential business against the reinsurer's own: seeing where it may be possible to add additional premium without exposing the reserves. With dwindling premium volumes, the spotlight also falls on those retrocessional covers that may formerly have seemed relatively inexpensive: rather than transfer extreme losses, why not manage the portfolio better through diversification (always being careful to calibrate any potential loss correlation) so as to reduce, or even avoid completely, the need for outward reinsurance?

RMS's global models now cover the major perils in countries that reflect more than 90% of worldwide property premiums. By the end of the year there will be models available to be licensed covering 50 territories. However it is the first division of countries that will always dominate reinsurers' portfolios: the US, followed by the UK, Japan, Australia, with New Zealand, France, Germany, Netherlands, Italy, Canada, Puerto Rico also featuring in most reinsurers' top 10. Having complete confidence in how to price risk and analyse portfolio loss in these key territories lies at the heart of global catastrophe reinsurance.

RMS first introduced models for many of these territories in 1993-4, working with the local insurance market. However, in the last two years all of these models have undergone significant upgrades, involving major investments in new research and model calibration. The first of these new upgrade releases was the Australian Cyclone model in 1996, to provide an industry standard for rating risk, in Queensland. In 1997 in alliance with Opus (formerly Works Consultancy - the leaders in earthquake risk assessment in Australasia), RMS released an upgraded earthquake model for both Australia and New Zealand. The most work of all, some 14 man-years of effort, went into revising and re-releasing the US hurricane model in 1997. Upgrades to the US earthquake models have followed the latest seismotectonic and vulnerability research findings: for the west of the country in 1997 and for the central US in 1998. Shortly to be released is a major upgrade of the Japan typhoon model, incorporating high resolution directional windfield modelling and thorough calibration against detailed Japanese insurance loss experience, which was used in the recent successful Pacific Re securitisation. Through its recently announced joint venture with the leading Japanese geotechnical engineering company, OYO, that has the most comprehensive databases on active faults and soils mapping in Japan, the RMS Japan earthquake model is currently being upgraded to be re-released as the industry standard for both the local Japanese market and for international reinsurance.

Perhaps the most exciting development in 1998 has been the introduction of basin-wide modelling across the whole Caribbean. For earthquake this has involved developing a sophisticated earthquake model with around 100 seismic sources, employing the latest seismotectonic information, including fault-trenching studies, on the complex braided Caribbean plate boundary. For hurricane the conventional approach to modelling, that involves setting off hurricanes at a gate close to a coastline, cannot work across a widely scattered net of islands, and instead RMS researchers have developed a completely new procedure based on some of the latest research in turbulent flow, that generates a stochastic population of hurricanes that have complete life-histories, tracks and evolving pressures, from the first formation of the eye to the final decay, maybe a week later, after landfall in the US. While every track is new, the whole population sustains the basin-wide characteristics of hurricane behaviour of the past century. Combined with the high resolution directional windfield models, also used in the Japan typhoon project, the Caribbean hurricanes are linked with those of the US to allow reinsurers, for the first time, to gain a basin-wide overview of the clash potential of their combined Caribbean and US portfolios, showing exactly how Texas risk combines with Jamaica and North Carolina with Puerto Rico and Barbados. As time is tracked in each simulation it is also possible to explore the application of different hours clauses.

Currently the European windstorm model is also being upgraded to provide high resolution (full postcode) windrisk microzonation, to take onboard information on building stock held by CARtograph (merged into RMS earlier this year) and also, for the first time, linking the wind losses with those from the storm surges that can accompany them: initially for the greatest concentration of risk in the east coast UK. Again this will be a first for catastrophe reinsurers: the combined windstorm - storm surge model is being released in time for end 1998 renewals.

Hemant Shah, founder of RMS in 1988, formerly responsible for model development, and now global marketing and strategy, has witnessed the sea-change in the sophistication of catastrophe modelling. "First generation models were assembled from pre-existing science: for example for US hurricane all models had been based on work contained in the 1980s National Weather Service reports NWS 23 and NWS 38, that explored the risk of extreme storm-surges. This research was never intended for catastrophe modelling. Taking the pieces from different research papers simply does not work because the original authors of the work always make specific assumptions that in combination lead to inconsistencies." Of the catastrophe modelling companies RMS is the only one to have the resources (including a 250 person software, GIS and digitisation facility in India) to commit to significant new internal research. This work is undertaken by more than 100 scientists and engineers working across the time-zones in California, UK and India (sometimes effectively operating around the clock, moving data and results from office to office). "Most of the rival models on the market still remain based on off-the-shelf science. Our research budget remains larger than our competitors' revenues."

In RMS catastrophe modelling is now recognised to be a whole new science, demanding the development of a suite of new techniques and solutions. In the 1980s researchers focused on the question of the hazard and risk at a single location, perhaps a new skyscraper or a nuclear power plant. However, the single site risk methodology cannot simply be transplanted to the portfolio, or "risk correlation", problem - the question of the probability of a widely dispersed set of sites being affected in the same event. Simple correlation matrix approaches to the problem oversimplify its true complexity, and have the major defect that individual events cannot be reconstructed at the end of an analysis. Event specific probabilistic modelling, pioneered by RMS, and still at the core of all its models, makes it possible to find exactly which class of windfields or earthquakes are driving the 200 year return period losses, and even more importantly, to add the losses modelled for new reinsurance treaties to the losses already accumulated for the pre-existing portfolio. Without it, for each new piece of business, one has to go back and re-run the analysis all over again.

Some modelling companies have proposed huge (500,000 or greater) numbers of simulations in a model, but that is a clumsy shotgun approach to the problem. Large event sets become extremely cumbersome to calibrate or to use (requiring some fall-back on an over-simplified correlation matrix that has to abandon the core event specific model) and are unnecessary - it is much smarter to use the minimum number (5,000-50,000 for most applications) of events that sustains the veracity and stability of the results for all potential users, something that is achieved by very detailed convergence testing, exploring how results change as the number of simulations is expanded and contracted. Stratified sampling of all key parameters means that we can evenly populate multi-parameter space with the greatest diversity of events, including deep in the tail of the key distributions. In the Caribbean model spatial stratified sampling is used to ensure an adequate population of events to be able to solve two or even three region clash (loss correlation) reinsurance portfolio problems.

As important as an understanding of the hazard, is knowledge of what is actually insured and how it will perform in a flood, windstorm or earthquake. Parallel research efforts are undertaken on the range of different building types, their values, and their vulnerability. Even in an island the size of Puerto Rico, the building stock is broken down into several different building inventories, where various insured classes of property (for example, multi-family residential, light industrial) are composed of varying proportions of separate building types which in turn vary geographically within the territory. Wherever there have been significant major recent losses, it is important to get deep into recent insurance experience, working with original portfolios and losses, generally at an individual risk basis. Even minor loss events highlight what is likely to happen in a major catastrophe, but few insurers have the time to get close to understanding all the implications. Working with insurers, from our analyses of actual loss experience we have been able to show how modest changes in underwriting can make dramatic reductions in the level of loss come the next catastrophe. Mapping loss at high resolution (2x2km in Puerto Rico), also helps confirm the new windfield models and shows just how dramatic is the fall-off in underlying average annualised loss as one passes from the coast inland. For the same building type, on an island in Japan or the Caribbean, technical premium can fall by as much as a factor of 10 in the first 10 kilometres.

Calibration and validation can extend over many months, becoming the mantras of the model development teams: finding new empirical ways to test individual parameters and then see how results compare when parameters are combined in a model. Sometimes model results defy market expectation, when there will be particular challenges to demonstrate exactly why risk behaves in this way. We are always on the lookout for new forms of calibration - for example in a number of European countries there are continuous quantitative records of the volume of forest windfall in major windstorms going back more than a century, that when normalised for forestry production provide a proxy for industry insurance loss.

A major focus of the new generation of model upgrades has been the significance given to historical event reconstruction: recreating all the windfield or earthquake intensity footprints and then finding the loss from each to today's industry portfolio and values. In Japan all the landfalling typhoons have been reconstructed back to the early part of the 20th century, for Portugal all the earthquakes back to 1700 and in Western Europe all major windstorms (around 200 of them) since 1868. If our model projects that the 100 year loss event for France is x2 or x3 the 1990A loss it is important that we can actually find a loss approaching this magnitude over the past 130 years! (Insurers and reinsurers are often unsure how to check the results of one model over another, so it is always worth asking the question as to which event actually had losses comparable to those modelled at 50 or 100 year return periods.) The RMS historical reconstruction are made available to the model user so that they too can see how their business would have performed as they time-step through history.

As we review catastrophic losses through history, in many regions we find evidence of decadal variations in activity, in which corresponding average annualised losses have persisted two to five time higher than the long term average for several years. These phases of higher activity and higher losses include parts of the mid 20th Century for US hurricanes, the late 1950s and early 1960s in Japanese typhoons and the 1980s and first three years of the 1990s for European windstorm. Reacting to one of these episodes, as it happens, provides particular challenges to reinsurers (and knowing when it has ended, major opportunities!).

In RMS's new upgraded international catastrophe models, with their core focus on the underlying science and engineering, in many regions now the central uncertainty in results is becoming less the model than the data on the portfolio that is entered into it. Catastrophes are a bit like shouting "Fire!" outside a student dormitory and seeing who comes out: loss emerges out of complex international reinsurer's portfolios in all kinds of unexpected places - revealing the existence of underlying risks of which there may have been no knowledge. "The next challenge", says Mr Shah, "is undoubtedly data: in the first rounds of securitisation the rating agencies consistently focused on the model but now they are increasingly sinking their teeth into the problems of inadequate data. We recognise that our business is increasingly in consultancy - helping insurers and reinsurers manage their risk, through managing their data: showing them how catastrophe modelling can help facilitate profitability." RMS now has a whole series of techniques of data supplementation, and risk auditing, including fire-drills that can test how a complete global portfolio will behave in a real catastrophic loss scenario: preferably one that, like the real world, occurs without warning. We can now model retrocessional treaties, and how their losses will correlate with all the other lines of business in a reinsurer's portfolio.

Through many self-financed initiatives as well as externally funded private projects, there is a lot going on below the surface in catastrophe research at RMS, much of which may not appear on the open market for a year of more: including work in weather, and weather derivatives and, looking at how changes in the climate state translate into changes in the probability of loss. We are working with many of the key players in risk securitisation in which the veracity of the model is fundamental. We recognise that innovations in catastrophe modelling (like the new clash models for Europe and the Caribbean) can help drive innovations in the risk transfer market. Using its international modelling capability RMS has become involved in projects to securitise risk all around the globe.

Our objective is to see catastrophe and weather risk become fully convertible international currencies: we already provide a consistent index of earthquake and wind risk (in terms of average annualised loss and losses at different return periods) for a number of major cities round the world, and we aim to provide consistent maps of worldwide wind-related risk. In the 21st century our goal is everyone, not only in the insurance industry, but even the original insurance purchaser (even school-children!), should perceive risk in this common standard, as something tangible, as real as long-term rainfall: then we will have helped transform the way in which the risk market operates.

Robert Muir-Wood has been with RMS, Inc. since February 1966 as vice president and now technical director of RMS Europe. He is responsible for research and risk analysis, leading a number of international projects on catastrophe model development for RMS worldwide. He has led the significant new research effort to develop and implement European windstorm modelling and he has run the project to develop new basin-wide hurricane and earthquake modelling capability for the Caribbean. Before joining RMS, Dr Muir-Wood was in charge of international model development at EQECAT.

Dr Muir-Wood has an MA and a PhD from the University of Cambridge. He has published around 40 scientific papers and has written more than 100 articles and reviews. He is the author of six books and lectures on natural catastrophe loss modelling for Lloyd's.