Generating Loss Estimates in Real Time: An Inside Look at the ALERT™ Process—and What Sets It Apart
Oct 25, 2011
Editor's Note: AIR Vice President of Consulting and Client Services Rob Newbold talks about AIR's ALERT™ (AIR Loss Estimates in Real Time) service—its complexities and uncertainties—and how it differs from other services available in the market.
AIR has provided estimates of insured losses from natural catastrophes in real time since Hurricane Hugo in 1989. Initially, the service was limited to hurricanes and earthquakes affecting the United States, and the communications were sent by agonizingly slow fax.
In August 1998, AIR launched the ALERT™ (AIR Loss Estimates in Real Time) website. The breadth of coverage was extended to all modeled perils and regions, and the new means of delivery ensured that companies had access to the information much more quickly.
Today, AIR's ALERT service is in higher demand than ever before, not least because of the reliability of the loss estimates. ALERT keeps companies well informed at critical times, allowing them to communicate effectively within their organizations and to set expectations for investors. AIR's real-time loss estimates can help determine if reinsurance is adequate, and manage reserves. ALERT's highly localized information allows insurers to understand where to suspend or continue writing business, and how to most effectively deploy claims resources. Timely information is also extremely valuable in light of increasingly common financial instruments for hedging against catastrophe losses.
Each ALERT posting begins with a publicly available summary (which is also delivered by email to NewsALERT subscribers) that compiles the known hazard and damage information along with an analysis of the event by AIR scientists and engineers. For events covered by AIR models that cause significant insured losses, the posting also includes an estimated range of insured losses for the industry as a whole. Companies that license AIR software can download event sets to estimate the losses to their own portfolios.
This article explores the ALERT process in the context of three recent events: Hurricane Irene, which in August served as a dramatic reminder of hurricane risk in the U.S. Northeast; the March 11th Tohoku earthquake; and the outbreak of severe thunderstorms that ravaged large swaths of the central and eastern United States in late May.
Tropical Cyclones: Hurricane Irene's Impact on the U.S. East Coast in August, 2011
AIR has posted loss estimates for more than 30 U.S.-landfalling tropical cyclones since the ALERT website was launched in 1998. Practice, however, does not make the process less resource intensive. The estimation of losses for a hurricane in real time requires three primary storm parameters—storm track, central pressure and the radius of maximum winds (Rmax).
The National Hurricane Center (NHC) provides information on the current and forecast track and intensity—along with the uncertainty in those forecasts—based on an ensemble of statistical and dynamical models. Rmax, which is not provided by the NHC, is estimated by AIR researchers using other meteorological sources, including vortex reports on the storm structure and radar images.
Prior to landfall, the range of loss estimates issued on ALERT is wide, reflecting the sensitivity of losses to landfall location and intensity, which are still quite uncertain. As the storm approaches land, the uncertainty in storm parameters reduces and the range of losses narrows. Once the hurricane has made landfall, AIR researchers use any available instrumentally-recorded surface wind speeds to estimate the efficiency with which winds aloft (gradient level) are being transferred to the surface—a parameter in the AIR U.S. hurricane model called the "gradient wind reduction factor."
In the case of Hurricane Irene, storm surge was of particular concern in light of the forecast landfall location just left of low-lying Lower Manhattan. AIR's storm surge module uses as input the primary meteorological characteristics of the storm, as well as bathymetry, elevation and tide height information. Lending uncertainty to the loss estimates is the extent to which storm surge damage will ultimately be paid as wind losses.
Hurricane Irene made first landfall in North Carolina on August 27 and a second landfall in New Jersey on August 28. On August 29, AIR posted a final loss estimate for Irene's impact in the United States of between USD 3 billion and USD 6 billion. On September 20, PCS issued a preliminary loss estimate of $3.65 billion—an estimate that may well rise with subsequent resurveys.
Earthquakes: The Tohoku Earthquake of March 11, 2011
In the case of earthquakes, key information is available from worldwide seismological agencies within minutes of the event, including preliminary estimates of magnitude, epicentral location and depth. Estimates of slip distribution along the rupture plane follow—and these typically vary from reporting agency to reporting agency.
However, while highly sensitive instruments inform these estimates, they are still only estimates, and may change—sometimes significantly—over time. Estimates of losses can be highly sensitive to such uncertainties. In the case of the Tohoku earthquake, for example, uncertainties in the southern extent of the rupture plane—towards the high concentrations of exposure in Tokyo and Chiba prefectures—had major implications for modeled losses.
Although AIR researchers can make rough estimates of losses using the "source parameters" of magnitude, depth, and rupture length and direction, more refined estimates can only be made if actual ground motion recordings are available. After the Tohoku earthquake, Japan's national seismic network remained offline for nearly a week. Once back online, AIR used the recorded ground motion data to constrain the simulated ground motion, and integrated that with the best available estimate of slip distribution to achieve more reliable estimates of shake damage.
A significant driver of loss from the Tohoku quake, however, was the devastating tsunami—currently a peril that is not modeled by AIR. To estimate property losses from the tsunami, AIR researchers used high-resolution wave height and elevation data to create an inundation footprint. This was compared to known flooded regions collected from aerial photography and satellite imagery. Further validation was undertaken using the Princeton Ocean Model (POM)—a numerical grid point model used for a wide variety of oceanographic applications worldwide.
Estimated tsunami losses were then "backed out" of the estimated shake and fire-following losses to avoid double counting. AIR's final estimate issued on March 24 was between JPY 1.5 trillion and JPY 2.5 trillion. The latest estimate from Japan's Financial Services Agency for losses to residential, nonlife and mutual lines is between JPY 2.1 trillion and JPY 2.2 trillion (excluding government recoveries).
Severe Thunderstorms: The Outbreak of May 20-27, 2011
The first, and generally the most time-consuming, part of the ALERT process is the data collection on event parameters. Comprehensive data collection is the foundation of robust loss estimates. Severe thunderstorms outbreaks are the most challenging in this regard. Catastrophic severe thunderstorm events typically occur over the span of several days and are comprised of often hundreds of individual tornadoes, hailstorms and straight-line wind storms. These are the "microevents" that make up the larger "macroevent." While the microevents are highly localized, the macroevent can impact multiple states. According to ISO's Property Claim Services® (PCS®), the severe thunderstorm outbreak of May 20-27 of this year impacted 20 states from Texas to Pennsylvania.
To generate a real-time estimate of U.S. severe thunderstorm losses, AIR researchers collect available information from the National Weather Service's Storm Prediction Center (SPC). At this stage, however, the raw data contains many duplicate reports since different witnesses commonly report the same microevent. Indeed the SPC data remains highly dynamic for weeks; a "final" determination of the actual make-up of the macroevent can take months. In addition, while the AIR U.S. severe thunderstorm model requires input parameters that include starting location, length, width and direction of each microevent's track, as well as intensity variation within the footprint, many of these parameters are missing from SPC data. Moreover, due to the localized nature of these events and to the fact that tornadoes will likely destroy any anemometers present, wind speeds are rarely recorded instrumentally.
To overcome these challenges, AIR researchers have developed a tool to compile available storm data, parse out specific intensity information when present, and discount repetitive reports. Guided by results, AIR simulates the still-missing parameters by using distributions of event characteristics that are built into the AIR model.
The process is labor and time intensive. Just over a week after the end of the outbreak, AIR posted an estimate of what insured losses would ultimately be from 187 modeled tornadoes, 777 modeled hailstorms and 397 modeled straight-line wind storms: USD 4 billion to USD 7 billion. This compares well with a preliminary report from PCS of USD 4.9 billion issued on June 14, which was updated in a subsequent resurvey to USD 6.5 billion.
Generating reliable estimates of insured losses from actual catastrophes in real time is a highly complex process at AIR—one that involves multiple departments across the company, from Research to Client and Consulting Services to Quality Assurance to Corporate Communications. Although AIR has an effective procedure for identifying events from the models' stochastic catalogs with "similar" characteristics—and making these available could be accomplished with relative speed—AIR believes that these are of limited use, and may even present a misleading view of the actual risk. Every natural catastrophe is unique, and even small differences in event parameters may result in large differences in losses. The real value of ALERT is the losses and event sets generated through hours and sometimes days of data collection and cleaning, and careful analysis.
Uncertainty in the losses can never be eliminated, and is the reason why AIR issues a range of loss estimates rather than a single number. In the early aftermath of an event or (in the case of hurricanes) prior to landfall, the uncertainty in the reported and forecast parameters leads to a trade-off between immediacy and accuracy. An example is the magnitude 8.8 earthquake that struck Chile in February of last year. Immediately following the event, three different credible geological agencies, including the United States Geological Survey (USGS), the local Chilean seismological agency, UCSC, and the German Research Centre for Geosciences (the GFZ) all released different estimates of the magnitude of the earthquake. While the difference between a magnitude 8.3 and an 8.8 may not sound significant, it in fact translates to a more than five-fold increase in the total amount of energy released. Early estimates of the spatial extent of the rupture also varied considerably, with some estimates putting the rupture plane much closer to Santiago than others. That kind of uncertainty can have enormous implications for losses and the only way to capture it is to produce a range of loss estimates. While the range may narrow over time as more information comes in, there will always be a range.
Uncertainty is inherent not only in the actual parameters of the event, but also in every component of the models—and indeed in the exposure data, whether at the industry or company level, used to generated the losses. Uncertainty should not be discounted, but should be recognized, reported and understood.
Despite the uncertainties, AIR has established a record over the last three decades of producing the most reliable real-time estimates of catastrophe losses available. But of course the real purpose of catastrophe models is to prepare for potential large losses before they occur. When AIR provides loss estimates immediately after an actual catastrophe, the model is being used in a very limited way. Nonetheless, such estimates are useful—and not only to model users. AIR looks at each real time event as a test of the models, and an opportunity to learn and improve on what we are doing.