Having assessed vendor market loss estimates for recent US hurricanes, it is clear that no catastrophe model is perfect. Given the multiple areas of uncertainty in determining the hazards, exposures and vulnerabilities during major events, vendor loss estimates are always likely to fall short of the precision levels desired by the market. With this being the case for hurricane risks, even greater variability can be expected for other natural hazards such as earthquakes, floods and convective storm outbreaks, as they are comparatively underdeveloped as modelled perils.
Challenged modelled loss accuracy for hurricane events is not an unexpected conclusion to emerge from the study and should not be interpreted as degrading the value of vendor modelling tools. After all, catastrophe models were not designed to predict the costs of individual events in real time and their primary purpose of assisting carriers in understanding and quantifying their risks is undisputed. The results are clear: whilst several carrier insolvencies followed Andrew (and Katrina to a lesser degree), billions of dollars in claims have been paid out post-HIM with no significant capital impairments.
A number of key conclusions emerge from this study, some, such as the importance of understanding differences between estimates, are more clear-cut than others. Having a range of views post-event can actually benefit carriers, brokers and investors as long as the important drivers are clearly communicated, particularly in situations where significant divergences occur. Catastrophe modelling firms can assist the market further here by better communicating the levels of uncertainty contained within each estimate and providing more transparency around the various assumptions that are driving loss estimates. It is likewise incumbent on market participants to review rigorously, or even challenge, some of the more extreme loss estimates released by modelling firms. Scrutinising assumptions that can drive vastly different views of events (such as the physical parameters of the hazard or insurance take-up rates) is recommended best practice in establishing whether loss estimates pass initial tests of credibility.
This study also reveals important trends which can help the market gauge the likely accuracy of modelled loss estimates for hurricane events in the lead up to landfall, or immediately thereafter (see Figure 11). Analysis conducted for this report shows that hurricane losses which are both conventional (i.e. driven by wind/ surge) and contained (i.e. less than USD 10 billion in total) are far more likely to be captured accurately by the vendor catastrophe models.
Figure 11: Template for Gauging Accuracy of Modelled Industry Losses (Source: JLT Re)
Conversely, more complex hurricanes which often bring unforeseen and unmodelled consequences are far more challenging to predict. Not only are such events unlikely to be replicated by the stochastic event sets initially used by catastrophe modelling firms, but any unusual storm characteristics are difficult to simulate even after landfall. The surge that caused the levees to fail in New Orleans after Katrina struck, the unusual trajectory of Superstorm Sandy or the recordbreaking precipitation that inundated Houston with Hurricane Harvey are such examples.
All three of these storms saw vast differences in estimates between AIR and RMS due to un-modelled elements. In each case, un-modelled loss components accounted for a significant proportion (if not the majority) of the total cost. Ultimately, making real-time loss predictions for events that bring new loss phenomena is more art and science. This is reflected by the highly divergent views between the modelling companies.
Over time, however, vendor firms will draw on lessons learned during such events to refine their models and incorporate a whole host of un-modelled perils. Contributions from the academic and engineering communities will also continue to assist them in this area. Hurricaneinduced flood and LAE are likely to be key areas of focus in the near term following HIM. Another theme that is starting to emerge from 2017 claims data, particularly for Irma, is the strong performance of newer construction and roofs in the highest wind zones. One possible implication for future model updates is that the building codes implemented and enforced, post the 2008 financial crisis, have performed even better than current modelling assessments. This may go some way to explaining why AIR’s and RMS’s estimates in 2017 broke the overriding trend of underestimating major losses and raises questions over whether recent model revisions (which were only first tested during HIM) may have overcorrected previous shortcomings.
Technological advancements will also drive future improvements to catastrophe models. Remote sensing on next generation satellites, as well as drone data, are already having an impact by facilitating the capture of more accurate satellite-derived wind speeds, improving post-event damage assessments (augmented by artificial intelligence detection methods) and enhancing the exposure calibration processes for higher resolution IEDs. Hurricane Maria was a stark example of how differing views on industry exposure can lead to disproportionate levels of variance between modelled estimates.
By leveraging increased skill in weather prediction, catastrophe modelling firms will soon be able to move away from pure stochastic trackdriven estimates pre- (and immediately following) landfall and utilise realtime clustering methods to generate probabilistic scenarios. This will greatly enhance the windfield generation process and lead to a more robust statistical assessment of outcomes. Improved weather forecasting will also bring more tangible impacts by enabling further hurricane preparation and loss mitigation measures ahead of landfall.
Even with these advances, modelled loss estimates will no doubt continue to be the subject of scrutiny and scepticism in future years. And yet, their utility looks set to only grow as demand from carriers, brokers and investors for real-time loss information will only increase in today’s datahungry world. With peak hurricane season upon us, JLT Re’s Analytics team is committed to providing differentiated real-time catastrophe reporting through its CATz blog to help clients understand and quantify the uncertainties associated with each event and any published modelled loss estimate. JLT Re’s Cat Model Insight (CMI) function has also been created to assist clients with the model validation process by assessing various model components and identifying potential un-modelled elements of loss.
Download the full report here: