MAY 2012 AIRCURRENTS: BLENDING SEVERE THUNDERSTORM MODEL RESULTS WITH LOSS EXPERIENCE DATA A BALANCED APPROACH TO RATEMAKING EDITOR S NOTE: The volatility in year-to-year severe thunderstorm losses means that traditional ratemaking approaches are not sufficient for developing actuarially sound prices. AIR Senior Vice President David Lalonde introduces a balanced ratemaking methodology that properly reflects the tail risk from potential severe thunderstorm events. INTRODUCTION Severe thunderstorms can affect most regions of the mainland United States, but are particularly common in the Southeast, Great Plains, and the Midwest. On an annual aggregate basis, they account for approximately half of all insured catastrophe losses in the United States. Comprising tornadoes, hail, and straightline winds, these systems can impact many states over a period of days to a week, although the most severe damage is typically concentrated in highly localized areas. In 2011, six severe thunderstorm events generated insured losses of more than $1 billion each, with total losses from 24 storms exceeding 26 billion USD. The most damaging events the Tuscaloosa, Alabama, tornado and the Joplin, Missouri, tornado are the two costliest in U.S. history. THE ARTICLE: Puts the 2011 season in context of past and potential losses and discusses how the AIR Severe Thunderstorm Model for the United States can be used in ratemaking to manage the high volatility in insured losses. HIGHLIGHTS: AIR s severe thunderstorm model provides a reliable and stable view of severe risk and avoids shifts in loss costs caused by the volatility in loss experience data. But company claims data is valuable and should not be discounted entirely either. AIR s ratemaking approach integrates both, which properly accounts for tail risk and prevents volatility at granular levels of geography While it was a record-breaking year in terms of insured loss, 2011 should not be considered a clear indication of things to come. A number of conditions that favor storm formation came together in 2011 that, combined with the levels of exposure in the affected areas, contributed to the high losses. Year-to-year variability in severe thunderstorm activity is unpredictable and can be very extreme. This article will put the 2011 season in context of past and potential losses and discuss how the AIR Severe Thunderstorm Model for the United States can be used to manage the high volatility in insured losses, focusing in particular on using model results in ratemaking. Figure 1. Average annual severe thunderstorm loss, based on 1991 2011 PCS losses trended to 2011 dollars.
PUTTING 2011 IN CONTEXT In April of 2011, a confluence of atmospheric factors created a prolonged period of favorable conditions for tornado activity. Between April 22 and 28, a violent severe thunderstorm outbreak the largest in U.S. history affected thirteen states in the South and parts of the Midwest and Northeast. One particularly damaging storm during this outbreak spawned the EF-4 Tuscaloosa-Birmingham tornado, which traveled some 80 miles across Alabama and destroyed several thousands of homes and flattened entire neighborhoods. In all, a total of 550 tornados were reported in the seven-day period and it was the second deadliest severe thunderstorm outbreak in U.S. history after the Tri-State tornado outbreak of 1925. ISO s Property Claim Services (PCS) estimates insured losses of 7.3 billion USD (Catastrophe Serial No. 46), making it the single costliest outbreak in U.S. history. One month later, another massive severe thunderstorm system tore through a wide swath of the country stretching from Lake Superior to central Texas and east through Missouri, Tennessee, Kentucky, Ohio, and to the East Coast. Thousands of buildings were damaged, hundreds more were completely destroyed, and more than a thousand people were injured. Missouri was the most severely hit. An extraordinarily violent tornado rated an EF5 with winds of at least 200 mph touched down just inside the Missouri border and cut across the city of Joplin (population 49,000) before continuing to the east. The tornado was on the ground for nine minutes and left more than 8,000 homes and apartment units and more than 500 commercial properties heavily damaged or destroyed. Almost 1,500 people were injured, and 138 killed. It was the deadliest tornado to strike the United States in more than half a century. Across the 20 states affected by the weeklong outbreak, PCS estimates insured losses of 6.9 billion USD (Catastrophe Serial No. 48). The severe thunderstorm losses from 2011 were the highest in recent decades and accounted for some 80% of insured U.S. catastrophe losses for the year. Figure 2 shows the annual occurrence exceedance probability curve generated using the AIR severe thunderstorm model and U.S. insured industry exposures as of December 31, 2010. Occurrence losses from the two individual outbreaks fall within the 5 10% exceedance probability range on this curve. Losses of this magnitude are not extreme tail events in the AIR model, and those managing risk with the AIR model already considered such scenario losses. Occurrence Losses Aggregate Losses AOL 20% 10% 5% 2% 1% 0.4% 0.2% As compared to a peril like hurricanes, where a single event can potentially generate insured losses in excess of 100 billion USD, the risk from a higher frequency peril such as severe thunderstorm is better reflected through an aggregate loss distribution. Figure 3 shows the annual aggregate loss exceedance curve from the AIR model based on U.S. insured industry exposures as of December 31, 2010. On an annual aggregate basis, the cumulative losses from the 2011 season losses were rarer, with an exceedance probability of about 0.5%, as shown in Figure 3. However, the probability is still within the range of industry losses captured by AIR s model. MANAGING LOSS VOLATILITY Because of the myriad climate and weather factors that give rise to severe thunderstorms and the localized pattern of damage, insured losses can be very volatile from year to year, as shown in Figure 4 for the entire U.S. PCS 48: May 20 27 PCS 46: April 22 28 Exceedance Probability Figure 2. Occurrence losses from the two major 2011 outbreaks were not extreme outliers. 2006, 2009: ~$10B 2008, 2010 total: ~$11.5B AAL 20% 10% 5% 2% 1% 0.4% 0.2% Exceedance Probability 2011: ~$26.2B Figure 3. Aggregate losses from the 2011 season have an exceedance probability of ~0.5%. Note that aggregate losses from 2000, 2001, 2002, 2003, 2004, 2005, and 2007 are less than AAL. 2
Billion USD 30 25 20 15 10 5 0 AIR s model results have a number of applications for insurers, including in underwriting, reinsurance structuring, and portfolio management. The model not only provides a holistic view of the potential losses that can be expected in the future, but can offer insight into what components of the portfolio are driving severe thunderstorm risk. The model can also help companies evaluate where accumulations of risk are located, where to grow and retract business, whether to change the portfolio s mix of lines of business, and how to optimize the usage of wind pools. Billions USD 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0 0.5 0.0 Figure 4. U.S. severe thunderstorm losses can be extremely volatile. (Source: PCS) At higher geographic resolutions, losses are even more volatile. To illustrate this point, losses for two individual states not normally considered at high risk of severe thunderstorms are shown in Figure 5. Looking at 20 years of experience up to 2009 for Arizona, one would perhaps be surprised by the large losses in 2010. Similarly in Minnesota, the level of loss in 1998 was not in line with more recent historical experience. As the impact of 1998 diminished in subsequent years, companies relying solely on experience would have been again surprised in 2008. This demonstrates that even 20+ years of historical data is not able to capture the true tail of industry losses, let alone an individual company s risk from severe thunderstorms. Minnesota Arizona Figure 5. Losses are increasingly volatile at higher levels of geographic resolution. Shown here are insured industry losses for Minnesota and Arizona. (Source: PCS) AIR developed the insurance industry s first probabilistic model to help companies proactively manage severe thunderstorm losses. The AIR Severe Thunderstorm Model for the United States, first released in 1987, simulates large atmospheric systems that can spawn hundreds of individual micro-events (tornadoes, hailstorms, and straight-line windstorms). AIR uses smoothing and data augmentation techniques to overcome historical event and claim reporting bias and to provide full spatial coverage of storm potential, including where none have been observed in the past. As will be discussed in more detail in the next section, AIR model results can also be used in developing sound pricing. BLENDING MODEL RESULTS WITH LOSS EXPERIENCE DATA, BOTH CAT AND NON-CAT Because of the high volatility in losses, traditional ratemaking approaches have been modified for severe thunderstorms. Using the last three to five years of claims experience, as is often done for the non-wind component of rates, is not a suitable approach for formulating an appropriate charge for the peril of severe thunderstorm. Many companies are using the Insurance Services Office (ISO) excess wind procedure (or an internally developed variation), removing wind losses from their historical data and accounting for those losses using a factor derived from the last 20 30 years of experience. There are two main components to these approaches: First, and importantly, determine what losses to remove from the experience data to avoid double counting and second, develop a sound method to add back in the loss potential. The frequency of severe thunderstorms may lead some insurance companies to believe that they have sufficient data to manage and price severe thunderstorm risk without the use of probabilistic tools. However, an excess wind factor is unlikely to capture the tail risk associated with this peril because the method relies only on the company s historical claims data. Most companies are unlikely to have experienced a normal amount of extreme low frequency/ high severity losses within the timeframe of the available data and, further, historical claims data do not reflect a company s current exposure distributions. The Casualty Actuarial Society (CAS) Statement of Principles Regarding Property and Casualty Ratemaking states that A rate provides for all costs associated with the transfer of risk. And thus we are motivated to explore additional ways to incorporate the tail risk. 3
Location ID One approach is to use modeled (catastrophe) average annual losses (AAL) combined with the non-catastrophe loss experience data. Because AIR s modeled losses are in line with the PCS definition of a catastrophe event, it is a straightforward exercise to remove the PCS catastrophe events from a company s loss experience data and to replace these with model results. As illustrated in Figure 6, ZIP Code AAL from non-pcs catastrophe events are combined with modeled AAL by ZIP Code. After performing this exercise at the ZIP Code level, losses can then be aggregated to the state or territory level appropriate for ratemaking or classification studies. Note that the model losses reflect current exposures, but the actual historical loss data must be trended before blending.. State County ZIP Code Ground-Up AAL Gross AAL 1 Kansas Morton 67953 134 127 10 Kansas Stanton 67862 144 136 100 Kansas Morton 67950 157 149 1000 Kansas Stevens 67951 157 148 1001 Kansas Stevens 67951 164 156 1002 Kansas Stevens 67951 122 114 1003 Kansas Stevens 67951 130 122 Year Raw Model Output PCS Cat Number Claims Data ZIP Code Gross Incurred Loss 2006 46 67855 60,500 2006 46 67953 33,265 2006 67951 21,200 2006 67954 11,050 2006 47 67838 32,450 2006 47 67846 54,780 2006 45 67880 14,350 2006 57 67952 18,900 2006 56 67950 65,700 2006 67857 3,600 2006 67 67861 85,650 2005 80 67879 45,600 2005 82 67862 23,025 Modeled AAL by Geography ZIP Code Non-Catastrophe Average Annual Loss ZIP Code Average Loss 67838 27,958 67846 7,166 67851 23,029 67855 50,845 67857 36,483 67860 19,648 67861 8,947 67862 20,467 67870 8,250 67877 8,104 67878 74,209 67879 7,761 67880 117,990 67950 25,145 + 67951 58,449 Average Loss 67855 26,545 67953 8,600 67951 1,650 67954 9,700 67838 4,355 67846 1,247 67880 18,562 67952 10,467 67950 5,754 Figure 6. Combining non-catastrophe claims data with modeled loss costs for actual events as defined by PCS. The approach described above discards the catastrophe component of loss experience data for severe thunderstorms and substitutes modeled losses. But this data is valuable and should not be discounted entirely. Rather than use one or the other, AIR has formulated a balanced approach that uses both. Company excess wind loss experience for catastrophe losses can be used up to a threshold (T), and the modeled loss factor can be used above this threshold (see Figure 7). The estimated loss cost thus is made up of three components: attritional (non-catastrophe) losses, excess wind losses below a threshold, and modeled wind losses above a threshold: This approach uses actual trended loss experience for the last three to five years, excluding excess wind losses; we call this term A. (Exactly which losses are considered attritional loss versus excess wind loss is determined by the company. For the purpose of this example, we will simplify and assume the exclusion of all wind loss.) Next, an excess wind factor is developed using the ratio of wind T losses limited to threshold (T) to non-wind losses, which we will call B. Finally, using the severe thunderstorm model, the average annual loss excess of T can be determined, which we will call C. The blended loss cost can then be calculated as: A(1 + B) + C This blending approach properly accounts for changes in demographics and building stock over time, and companies can perform sensitivity analyses by setting and then adjusting the loss threshold. Loss Cost = Non Cat + (Experience based Cat Load < T) + (Model Cat Load >T) 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Figure 7. Actual loss experience (left) and modeled loss potential (right). T Sound pricing requires rates that are sufficient to cover tail scenarios and that can represent the variability of risk at an appropriate level of geographic granularity. AIR s catastrophe models are especially well suited for this purpose because they provide a reliable and stable view of severe thunderstorm risk and avoid shifts in loss costs caused by the volatility in loss experience data. 4
CASE STUDY A hypothetical insurer wishes to test the sensitivity of loss costs to different thresholds between loss data and modeled losses when blending model output and claims data for ratemaking. They first determine their annual aggregate exceedance probability curves for both modeled loss and trended actual losses, as shown in Figure 8. Next, the insurer picks losses that correspond to several points on the modeled annual aggregate loss exceedance probability curve. The choice of which points to pick can depend largely on the variability in the loss data and the insurer s confidence in its loss history. Finally, loss costs are calculated according to actuarial formulas with the different thresholds chosen. Table 1 shows the different loss costs obtained from using actual losses only, modeled losses only, and blended results with the different thresholds shown in Figure 8. These loss costs can then be used as the input into traditional ratemaking formulae and the impact of using different thresholds on the overall rate indication can be assessed. Table 1. Loss costs vary depending on ratemaking methodology used Millions USD 350 300 250 200 150 100 50 0 Trended Actual Annual Aggregate Cat Losses by Year 5% EP Modeled Loss Point 10% EP Modeled Loss Point 20% EP Modeled Loss Point SCENARIO EARNED HOUSE YEARS (000 S) NON-CAT AAL ($M) A: NON-CAT LOSS COST (AAL/EHY) B: EXPERIENCE- BASED CATAS- TROPHE LOAD (SUM {WIND LOSS<T/TOTAL LOSS})/20 C: MOD- ELED CA- TASTROPHE LOSS COST (AAL/EHY) A(1 + B) + C TOTAL LOSS COST Actual Cat Losses Only 330 150 454.55 0.366626 0.00 621.20 Excl. Losses > 5% EP Point (T=$150M) Excl. Losses > 10% EP Point (T=$82M) 330 150 454.55 0.284259 38.43 622.18 330 150 454.55 0.232670 52.94 613.25 Figure 8. Modeled loss exceedance probability points are superimposed on company loss data. Excl. Losses > 20% EP Point (T=$44M) 330 150 454.55 0.183478 68.88 606.83 Modeled Losses Only 330 150 454.55 0.0 136.66 591.21 CLOSING THOUGHTS Managing the risk from low frequency, high impact events like severe thunderstorms should be a part of every insurer s risk management strategy. The AIR Severe Thunderstorm Model for the United States helps insurers minimize accumulations of risk, strategically price their policies, and determine effective reinsurance and risk transfer strategies. In the realm of ratemaking, using model output in conjunction with historical loss data allows insurers to develop sound pricing that properly reflects the entire range of potential severe thunderstorm loss experience, especially the tail risk that is often missed by looking solely at claims data. 5
ABOUT AIR WORLDWIDE AIR Worldwide (AIR) is the scientific leader and most respected provider of risk modeling software and consulting services. AIR founded the catastrophe modeling industry in 1987 and today models the risk from natural catastrophes and terrorism in more than 90 countries. More than 400 insurance, reinsurance, financial, corporate, and government clients rely on AIR software and services for catastrophe risk management, insurance-linked securities, detailed site-specific wind and seismic engineering analyses, and agricultural risk management. AIR is a member of the Verisk Insurance Solutions group at Verisk Analytics (Nasdaq:VRSK) and is headquartered in Boston with additional offices in North America, Europe, and Asia. For more information, please visit www. air-worldwide.com. 2012 AIR WORLDWIDE. ALL RIGHTS RESERVED. 6