03 February 2009

Published in: Cat risk - ILS

Why we favour a statistical approach to hurricane forecasting

The Tropical Meteorology Project (TMP) at Colorado State University headed by Dr. William Gray has been issuing Atlantic basin seasonal hurricane forecasts since 1984. The TMP currently issues seasonal hurricane forecasts at four lead times, in early December, early April, early June and early August. Prior to the mid-1980s, there was no objective methodology available for determining how active the upcoming Atlantic hurricane season was likely to be.

Since we are attempting to forecast small-scale weather events (e.g. hurricanes) months in advance, we use a statistical climate modeling approach for our forecasts. We are constantly working on improving and refining our seasonal forecast schemes. New statistical models were unveiled for the 2008 Atlantic hurricane season and led to forecasts which were very close to observations.

Statistical climate forecasts are based on the premise that those global oceanic and atmospheric conditions which preceded comparatively active or inactive hurricane seasons in the past provide meaningful information about similar trends in future seasons. We build our forecasts using a hindcast technique, whereby we select predictors that explained a considerable amount of the variability in hurricane seasons over the past sixty years. Several of the key features that are investigated in these seasonal hurricane forecasts include current and predicted states of El Niño (anomalously warm eastern and central tropical Pacific sea surface temperatures), tropical and North Atlantic sea surface temperatures and sea level pressures and tropical Atlantic vertical wind shear (the change in wind direction with height). Typically, more active Atlantic hurricane seasons are associated with La Niña (anomalously cool eastern and central tropical Pacific sea surface temperatures), warmer than normal tropical and North Atlantic sea surface temperatures, lower than normal tropical Atlantic sea level pressures and reduced levels of tropical Atlantic vertical wind shear.

Our methodology is to be contrasted with numerical modeling which involves observing atmospheric fields such as winds and pressures and numerically integrating them forward in time using known mathematical formulae. This technique works very well out to about 7-10 days when the atmospheric wind and pressure fields are the dominant factor. However, after this point, forecasts using numerical modeling tend to degrade towards chaos, because one must take into account a plethora of inter-relationships between the atmosphere, the ocean and the land surface.

Interpreting the hurricane forecasts

Seasonal forecasts issued by the TMP since 1984 have shown moderate levels of skill above that specified by a climatological average forecast or a previous 5- or 10-year running mean. Forecasts issued in June and August have shown the highest levels of skill, which is to be expected, since we are then closest to the events (e.g., the upcoming hurricane season) that we are trying to predict. Even though the hurricane season runs from 1 June 1 to 30 November, we feel that issuing a forecast on 1 August is still of considerable use, as approximately 95% of all major hurricanes (storms with one-minute sustained winds of more than 111 mph) form after this date.

Table 1 displays seasonal forecasts issued by the TMP in early August for named storms and hurricanes forming after that date. Observations only include storms that formed after 1 August. Note that these early August forecasts have either exactly verified or forecasted the correct deviation from climatology in 23 of 25 years for named storms and 19 of 25 years for hurricanes. If we predict an above- or below-average season, it tends to be above or below average, even if our exact forecast numbers do not verify.

Table 1

Verification of the authors' early August forecasts of Atlantic named storms (NS) and hurricanes (H) 1984-2008

Year	Predicted NS	Observed NS	Predicted H	Observed H
1984	10	12	7	5
1985	10	9	7	6
1986	7	4	4	3
1987	7	7	4	3
1988	11	12	7	5
1989	9	8	4	7
1990	11	12	6	7
1991	7	7	3	4
1992	8	6	4	4
1993	10	7	6	4
1994	7	6	4	3
1995	16	14	9	10
1996	11	10	7	7
1997	11	3	6	1
1998	10	13	6	10
1999	14	11	9	8
2000	11	14	7	8
2001	12	14	7	9
2002	9	11	4	4
2003	14	12	8	5
2004	13	14	7	9
2005	13	20	8	12
2006	13	7	7	5
2007	13	12	8	6
2008	13	12	7	6
Average	10.8	10.3	6.2	6.0
1984-2008 Correlation		0.62		0.58

It is important that the insurance industry appreciate that these seasonal forecasts are based on statistical schemes which, owing to their intrinsically probabilistic nature, will fail in some years. Moreover, these forecasts do not specifically predict where within the Atlantic basin these storms will strike. The probability of landfall for any one location along the coast is very low and reflects the fact that, in any one season, most U.S. coastal areas will not feel the effects of a hurricane no matter how active the individual season is. However, it must also be emphasized that a low landfall probability does not insure that hurricanes will not come ashore. Our project, in partnership with the GeoGraphics Laboratory at Bridgewater State College, has recently developed the United States Landfalling Probability Project which provides probabilities of landfall based on historical data and adjusted based on the premise that more active seasons tend to have more landfalls. This data is currently available online at http://www.e-transit.org/hurricane. The landfall website is adjusted with each seasonal forecast update.

What's behind the long-term trends in Atlantic hurricane activity?

There is currently a vigorous debate as to whether storm activity in the Atlantic is getting worse due to human-induced climate change. Our project has done extensive study into this question and finds that, at this point, there are no detectable trends in levels of activity in the Atlantic. Although Atlantic hurricane seasons have been much more active since the mid 1990s, similar active periods are also documented in historical data maintained by the National Hurricane Center from 1926 to1969 and during the late 19th century. The Atlantic basin has more pronounced variability on multi-decadal timescales than does any other tropical cyclone basin (such as the Northeast Pacific and the Northwest Pacific). This variability is likely driven by natural ocean circulation changes that drive changes in Atlantic sea surface temperature, sea level pressure and vertical wind shear patterns in such a way as to make Atlantic seasons more or less active.

As one goes back further in time, our ability to detect and monitor tropical cyclones degrades. Prior to the mid-1960s, there was no satellite data, while prior to the mid-1940s, no aircraft reconnaissance was conducted. However, we find that the United States coastline has been quite densely populated since around 1900, and therefore, we use the number of United States landfalls to investigate potential changes in levels of hurricane activity. For the insurance industry, tropical cyclones are typically only of consequence if they make United States landfall. Table 2 displays United States landfalls from 1900 to 1949 along with the most recent 50-year period (1959-2008). Even though global temperatures have risen during this time period, the number of US landfalls has actually gone down slightly.

Table 2

U.S. landfalling tropical cyclones sorted by landfall intensity during two 50-year periods

Years	Named Storms	Hurricanes	Intense Hurricanes (Cat 3-4-5)	Global Temperature Increase
1900-1949 (50 years)	189	101	39	+0.4C
1959-2008 (50 years)	167	85	33

This decrease in landfall is especially dramatic for the United States East Coast and Florida Peninsula. Figure 1 displays landfalling major hurricanes during the 43-year period from 1923 to 1965 along with the most recent 43-year period of 1966 to 2008. Only seven hurricanes made landfall over the most recent period compared with 24 during the earlier period.

Figure 1

Contrast of tracks of East Coast and Florida Peninsula major landfalling hurricanes (intense hurricanes – IH) during the 43-year period of 1923-1965 versus the most recent 43-year period of 1966-2008

Seasonal hurricane forecasts issued by the TMP generate a considerable amount of interest in the insurance industry and among the general public. This is probably because there is inherent curiosity about how active the upcoming season is likely to be. Using historical data, there is significant hindcast skill (using the past to predict the future) available for predicting the upcoming season. However, one must realize that these are statistical forecasts which will fail in some years. We find that we learn a lot from our forecast errors. Our end-of-the-season verifications give much information on explaining what the factors were that dictated the number and frequency of storms. Some of these factors may not have been considered in our forecasts for that particular year, and we often add new predictors in a quantitative or qualitative manner based on our end-of-the-season verifications. We will continue to revise and, we hope, improve these seasonal forecasts in future years.

Dr Philip Klotzbach is Research Scientist and Dr William Gray Professor Emeritus in the Department of Atmospheric Science, Colorado State University

Back to top