Catastrophe Model Performance for Hurricane Florence: Early Insights

13 September 2018

In July this year, JLT Re published a Viewpoint report titled Catastrophe models: In the eye of the storm. For those who have not yet read it, the paper examines the precision of modelled market loss estimates for significant hurricanes since 2004 and provides a unique perspective in assessing the modelling companies’ real-time loss estimation process. A copy of the full report can be accessed by clicking here.

One of the key reasons for carrying out the study was to gauge the performance of loss estimates during the lifespan of hurricane events and assess whether any trends or lessons could be gleaned for future reference. With Hurricane Florence expected to make landfall in the next 24 hours, this piece provides an early perspective of how catastrophe models are likely to perform for this unique and potentially destructive storm.


Understanding the processes and timelines that catastrophe modelling firms work towards when releasing market loss estimates for hurricane events is crucial to this analysis (see Figure 1 for an illustration). As Hurricane Florence has yet to make landfall, the modelling companies are currently attempting to replicate the characteristics of the storm by selecting tracks from thousands of stochastic event sets that best resemble Florence’s forecasted path and intensity. The number of stochastic tracks will dwindle significantly today and tomorrow as more clarity around the storm’s location, forward speed and windfield size becomes available. 

Figure 1: Timelines for US Hurricane Loss Estimation (Source: JLT Re)

hurricane florence


Replicating Florence’s characteristics will be a real challenge for the modelling firms. The storm has surprised forecasters repeatedly on its approach to the US East Coast. As those who follow JLT Re’s CATz blog will all be too aware, Florence was only a couple of days ago expected to make landfall in North Carolina as a category 4 hurricane, a scenario that would have brought winds of up to 140 miles per hour to communities close to the landfall point. Our study reveals that vendor models have historically performed well for events where wind is the main loss component. Specifically, our findings showed that conventional hurricane events that do not assume super-cat characteristics are captured adequately by catastrophe models, and this has been reflected by a number of loss estimates provided for such events: e.g. hurricanes Charley (FL, 2004), Gustav (TX, 2008) and Irene
(NC & NJ, 2011).


But forecasts for Florence have changed significantly in the last 24 hours. Fortunately, Florence has weakened during this time and is now expected to make landfall close to Wilmington as a category 2 hurricane. But that is where the good news ends. There is still considerable uncertainty associated with Florence’s forecasted track and intensity as it interacts with land but one scenario could see the storm move in a south-westerly direction after making landfall in NC, running parallel to the Carolinas’ coastline and impacting a far wider area than originally anticipated. This could see a wide swathe of the Carolinas’ coastline being subjected to strong hurricane force winds and a huge storm surge over a sustained period as the storm lingers. Perhaps even more worryingly, up to 40 inches of rainfall could fall over parts of the Carolinas over the next few days.

This creates real challenges for the modelling firms and does not bode well for the accuracy of modelled loss estimates that will be released in the coming days and weeks. One key takeaway to emerge from our Viewpoint study was that modelled loss accuracy for hurricanes suffers when events are both complex and costly. There are two key reasons for this:

1. Complex hurricanes 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.
2. These types of events often bring un-modelled consequences that cause losses to spiral. Catastrophe models have not performed particularly well in forecasting losses for such events. Katrina (flooding of New Orleans), Ike (inland damage) and Sandy (surge/flood event in NJ and NY) are three examples of such storms, and Figure 2 shows how both AIR and RMS performed in predicting their losses relative to the ultimate insured costs.

Figure 2: Model Performance for Katrina, Ike and Sandy (Source: JLT Re, AIR, RMS, Munich Re)

hurricane florence


Early indications show the stochastic event sets being used by the modelling firms for Florence do not come close to replicating the storm’s forecasted track. At this point, any published loss estimates should be treated with a strong degree of scepticism given they come with considerable uncertainty and, in the case of the main commercial modellers, are based upon unrepresentative stochastic tracks. Some estimates will also exclude water-related loss components that could ultimately far exceed costs associated with wind damage: namely flood and surge. Estimates for Hurricane Harvey last year showed the different approaches modelling firms can take to events such as these.

In our Viewpoint report, we called on the catastrophe modelling firms to better communicate the levels of uncertainty contained within each estimate. We echo that call today: an open dialogue will lead to greater levels of transparency and increase market confidence in the post-event loss estimation process.


Natural Catastrophe Insights

Read more

Natural Disaster Resources

Read more

CAT Activity Zone

Read more