Dr Elsner's team generates their multi-year hurricane forecasts using a three-pronged statistical approach. First, the team members manipulate the best-track database to remove any spurious long-terms trends. One such trend is the artificial increase in hurricane numbers during the last century due to improved observational capabilities. The current network of reconnaissance aircraft, satellites and ocean buoys records virtually every storm, whereas earlier storms may have gone undetected if they remained at sea or failed to cross paths with a ship. Detrending the data minimises this potential bias.
The team also normalises the database prior to analysis. It does so by subtracting the long-term mean annual number of hurricanes (5.8) from each year's hurricane tally, and then dividing the sum by the standard deviation (2.5) of the best-track record.
After detrending and normalising the data, the team uses the singular spectrum approach (SSA) and maximum entropy method (MEM) to reveal any cyclic patterns in the frequency of Atlantic hurricanes between 1886 and present. These statistical techniques reveal the 2.5, 5.6, and 7.4-year oscillations discussed in the text.
The team then applies two separate prediction models to each of the underlying oscillations - an autoregressive moving average (ARMA) model, and a model that repeats the SSA procedure starting with a random guess for the forecast. The ARMA model explains 81% of the total variation of the high-frequency (2.5-year) oscillation, 72% of the mid-frequency (5.6-year) oscillation, and 83% of the low-frequency (7.4-year) oscillation.
The five-year forecast is based on the mean of the number of hurricanes predicted by the ARMA and iterative models.
Multi-year hurricane forecasts could provide better information for risk management decisions. David Malmquist describes work now underway.
An experimental five-year hurricane forecast predicts that the next two Atlantic hurricane seasons will be busier than normal, with a return to normal or below normal conditions in 2001 and 2002. The forecast predicts another busy year in 2003.
Climate scientists at Florida State University (FSU) issued the forecast on 1 June, the official start of the Atlantic hurricane season. Development of multi-year hurricane forecasts has been encouraged and funded by the Risk Prediction Initiative (RPI), an academic-business partnership of the Bermuda Biological Station for Research,1 as part of its effort to make climate forecasts more usable to insurers.
Multi-year hurricane forecasts are designed to overcome an important limitation of current seasonal forecasts- their failure to provide sufficient lead time for many risk-management decisions. The earliest seasonal forecasts now appear in early December, six months before the start of the following year's June-November hurricane season, and nine months before its September peak. Yet according to Jeremy Hindle, vice president and underwriter at XL Mid Ocean Re, a six- to nine-month lead is of limited use given that catastrophe insurers have written most of their US windstorm policies by December 1. This temporal disparity helps drive reinsurers' interest in multi-year hurricane forecasts.
A team led by James Elsner, head of the forecast group at FSU bases its forecast on the “best-track” archive compiled by the US National Hurricane Center. This is the official hurricane record for the Atlantic basin. It contains information about the location and intensity of each tropical cyclone at six-hour intervals throughout its lifetime. The best-track database reaches back to 1886 and contains 567 hurricanes.
By statistically examining the underlying patterns in the best-track database (see box for a technical description of their forecast technique), Dr Elsner's team finds that the interplay of three separate climatic oscillations largely explain the observed fluctuations in Atlantic hurricane frequency from year to year and decade to decade. To predict what hurricane activity will be like during the next five years, Dr Elsner's team models the behaviour of each of the oscillations, then uses the model results to project each oscillatory pattern into the future. The team generates the final forecast by recombining the individual predicted oscillations.The statistical techniques that Dr Elsner uses to extract and recombine these oscillations have their roots in the fields of signal processing and wavelet analysis. They resemble the techniques that a sound-studio musician might use to isolate and recombine the sounds of individual instruments into an orchestral movement. The three “instruments” in the Elsner hurricane concerto perform on time scales of 2.5, 5.6, and 7.4 years.
According to Dr Elsner, the high-frequency, 2.5-year oscillation reflects the well-known link between hurricane activity and the quasi-biennial oscillation, or QBO. This link was first identified by hurricane forecaster Bill Gray of Colorado State University. The QBO refers to a predictable shift in the upper-level easterly winds that encircle earth's equator. The speed of these winds varies on a time scale of about two years. When the upper-level winds blow strongly, they tend to outpace the tradewinds below, which are what carry hurricanes that form in the eastern tropical Atlantic westward toward the Caribbean islands. The difference in speed between surface and upper-level winds creates a situation in which the column of rising air within a developing hurricane is likely to be sheared off as it moves upward. This vertical shear disrupts hurricane formation. When the upper-level winds are weak, the vertical shear is reduced, and hurricanes are more likely to form and intensify.
The middle-frequency, 5.6-year oscillation is very likely tied to El Niño, which recurs about every 4 to 6 years and has been observed to reduce hurricane activity in the Atlantic. The warm ocean water that invades the eastern Pacific during an El Niño, fuels the growth of thunderstorms. As air inside these storms rises from its heat source in the water, it begins to cool and spread laterally, producing high-level westerly winds that blow across the Caribbean and tropical North Atlantic. Dr Gray hypothesises that these winds can shear the top off any hurricane they may encounter. During El Niño's opposite - La Niña - the eastern Pacific is abnormally cold, the high-level westerly winds diminish and hurricane formation is enhanced. The probability that two or more hurricanes will strike the US during an El Niño is 28%. The probability of two or more US landfalls during La Niña is 66%.
The origin of the low-frequency, 7.4-year oscillation is less clear, according to Dr Elsner. He believes that it may be associated with fluctuations in sea surface temperatures in the North Atlantic. Many climate scientists see a link between warmer surface waters in the North Atlantic and increased numbers of hurricanes.The precision of these multi-year forecasts is still very much in question. Dr Elsner says that they most likely will “result in only marginal skill above climatology.” In other words, someone could predict future activity nearly as well simply by relying on the long-term climatic average. In the Atlantic, yearly averages for hurricane activity are 9.3 named storms, 5.8 hurricanes and 2.2 intense hurricanes. Intense hurricanes are those greater than category 3 on the Saffir-Simpson hurricane intensity scale, with winds exceeding 111 mph. Dr Elsner's forecast calls for eight hurricanes in 1999 and 2000, five hurricanes in 2001, four in 2002, and nine in 2003.
But a marginal increase in skill may still be of value to the insurance industry, where small shifts in probability can translate into millions of dollars in premium revenue or insured loss. Insurers can use long-term forecasts to adjust the landfall probabilities generated by catastrophe risk models. These models calculate mean probabilities by extrapolating from the historical record, and most models do not account for short-term fluctuations around those mean values.
In addition to its efforts to promote long-term hurricane forecasts, the RPI also works to make real-time and single-season hurricane forecasts more valuable to insurers. Jeremy Hindle notes that the most useful forecasts attempt to predict the likelihood that an intense tropical cyclone will come ashore at a certain time and place. Intense hurricanes generate more than 80% of US hurricane damage, even though they comprise only 20% of the storms that make landfall. According to Mr Hindle, insurers and reinsurers need to be able to increase their confidence in the probability of intense landfalling hurricanes to benefit fully from forecasts.
Dr Elsner and Colorado State University's Bill Gray have used RPI funding to issue the first-ever seasonal landfall forecasts, which predict several months in advance the likelihood of hurricane landfall along specific segments of the US and Caribbean coastlines. RPI-funding has also encouraged Isaac Ginis of the University of Rhode Island to help optimise the computer model he uses to predict the intensity of an existing storm at landfall.
Dr. David L. Malmquist is assistant research scientist, Bermuda Biological Station for Research (BBSR). Phone: +441 297 1880 x255; fax +441 297 2890: e-mail: email@example.com. Web page: www.bbsr.edu/rpi.
The RPI is sponsored by ACE USA Inc., American International Group Inc./IPC Reinsurance Company, Ltd, American Re, Employers Reinsurance Corporation, General Reinsurance Corporation, Renaissance Reinsurance Ltd, State Farm, USAA, Willis, X.L. Mid Ocean Reinsurance Company, Ltd, and Zurich Group/Centre Solutions (Bermuda) Ltd.
1. (see Global Reinsurance, 1997, v.6,pp119-121.)