Verification of our Forecasts for the1998 North Atlantic Hurricane Season and Predictions for the 1999 Season



contributed by James B. Elsner



Department of Meteorology Florida State University Tallahassee, FL 32306-3034



1. Forecast methodologies



The two components of hurricane activity that we have studied are the seasonal number of hurricanes (H) and the seasonal number of intense (or major) hurricanes (IH). To predict the number of hurricanes we use an ordinary least squares (OLS) linear regression to estimate the number of tropical-only hurricanes () to which we add a seasonal average number of baroclinically-enhanced hurricanes (). The model can be expressed as

where



and where the 's are coefficients on the predictors as described in Hess et al., (1995). The five predictors include a 10-month forward extrapolation of the 50 mb and 30 mb zonal winds, the magnitude of the vertical shear of these winds, and the average rainfall anomalies (expressed in standard deviations) from the Gulf of Guinea and Sahel regions of west Africa. These are the predictors originally suggested by Gray et al. (1992).

To predict the number of intense hurricanes we use a nonlinear Poisson regression with maximum likelihood criterion. The model can be expressed as

where the 's are the coefficients on the predictors and exp() is the exponential function. Details of this model are given in Elsner and Schmertmann (1993). The five predictors used in forecasting intense hurricanes are identical to those used in predicting tropical-only hurricanes.

In addition to basin-wide activity we also predict activity in four sub-basins of the North Atlantic including the Caribbean Sea, the Gulf of Mexico, the southeast U.S. coast (Cape Hatteras south to Key West) and the northeast U.S. coast (Cape Hatteras north to the Canadian border) (Lehmiller et al., 1997). We use logistic regression to predict hurricane landfalls along the southeast coast and the presence or absence of intense hurricanes in the Gulf and Caribbean. As with the approach taken for basin-wide intense hurricane activity, we express the sub-basin forecasts in terms of probabilities.

Logistic regression is a statistical model used to predict yes/no events by estimating coefficients for several predictor variables. "Yes" indicates the occurrence of the event. Here we use a maximum likelihood technique to obtain the coefficients. A logistic regression can be expressed as







where Pr is the probability of occurrence and the 's are the coefficients on the p predictors. The only model with significant long-lead skill above climatology is the model for Caribbean hurricanes (HC). The Caribbean hurricane model uses only the two rainfall parameters (Gulf of Guinea and Sahel) as predictors.



2. Verification of our early December forecasts for the 1998 hurricane season

Using the above statistical models on data received by early December last year, we estimated that hurricane activity over the North Atlantic would be below normal for 1998, including 3 hurricanes and 1 intense (or major) hurricane (3 on the Saffir/Simpson scale). This was a serious underestimation of the actual activity. There were 10 hurricanes (14 tropical storms) during 1998 of which 3 reached major hurricane intensity. Both total hurricane activity and intense hurricane activity were well above the long-term averages (1886-1996) of 5.0 and 1.9, respectively. The average number of hurricanes over the shorter, more reliable period of 1950-1996 is 5.9. The early December Poisson model estimated a 7% chance of exactly 3 intense hurricane for 1998. We anticipated less than a 20% chance of a Caribbean hurricane during 1998. Both hurricane Georges and Mitch were intense Caribbean hurricanes.

Table 1 lists the hurricanes during 1998. None of the published seasonal forecasts predicted this much activity though NOAA's Climate Prediction Center issued a "forecast of opportunity" that indicated the season would be above average. Five of the 10 hurricanes were tropical only. Karl and Nicole were the two baroclinically-initiated hurricanes of the 1998 season. Over the period 1950-1996 an average of 1.5 baroclinically-initiated hurricanes occurred per season. In some respects the active 1998 season is analogous to the active 1969 season. In 1969, the season's first hurricane occurred during the second week of August and the season's last hurricane dissipated during the last week of November. Five of the 12 hurricanes in 1969 were tropical only.

The 1998 season reversed the activity of the previous season when only 3 hurricanes formed. The relative inactivity during the early part of the 1998 season was exceptional, but not unanticipated as El Niño conditions were expected to inhibit early-season activity. The 1998 season can be described as a late-starting season. The first hurricane of the season, formed after August 15th; in fact, tropical depression number two did not become tropical storm Bonnie until August 19th. Other late-starting seasons include 1924, 1948, and 1974, among others. A survey of the historical records shows that 39% of the hurricane seasons between 1886 and 1996 with at least one hurricane began after the middle of August. Most recently the inactive period from 1991 through 1994 featured all late-starting seasons. The frequency distribution of hurricanes in seasons that begin late suggested that 1998 would be below average (see Table 2). The distribution is symmetric about 4 hurricanes. The expected number of hurricanes in late-starting seasons is 4 (the average is 3.7). A maximum of 7 hurricanes occurred in the two late-starting seasons of 1949 and 1981. The late-starting 1998 season with its 10 hurricanes is unprecedented in the historical record.

Three of the 10 hurricanes (Bonnie, Earl, and Georges) made landfall in the United States. Although no major hurricane made landfall in the United States during 1998, it was the most active season for U.S. landfalls since 1989 when Chantal, Hugo, and Jerry all crossed the U.S. coastline. Georges made two landfalls, one over the Florida Keys and the other over Mississippi. Our early August sub-basin forecasts also proved to be less accurate than expected based on hindcast evaluations (see Table 3). In particular we were wrong about the hurricane activity in the Caribbean Sea. The largest contributing factor for the forecast errors is likely the poor indication of rainfall over western Africa. The early indications were for a dry August and September. Dry conditions in western Africa are linked to fewer tropical-only hurricanes and a reduced likelihood of activity over the Caribbean Sea. Rainfall amounts across western Africa during August were above average. According to Ricardo Correa-Torres, the percent of normal rainfall in August was much above the values observed in June and July.

The unanticipated heavy rainfall associated with vigorous tropical waves contributed to the unexpected hurricane activity.



3. Increasing volatility in hurricane activity

The 1998 season continues the trend of increased volatility in hurricane activity since the early 1970s. The much above normal 1998 North Atlantic hurricane season follows a season with below average activity. The 1997 season followed the very active 1995 and 1996 seasons. The 10 hurricanes of 1998 makes the 1990s the first decade featuring at least 3 years with 9 or more hurricanes. More significantly perhaps is the fact that no year during the 1990s featured average, or near average, numbers (5, 6, or 7). The decade of the 1910s had no average years, but only 1916 was above the average the rest were below. The large interannual variability during the 1990s is apparent by considering successive 7-year trimmed standard deviations. The standard deviation is computed over the interval after removing the years having the highest and lowest numbers of hurricanes in the interval. Figure 1 shows the time series of running standard deviations since 1886. Values are plotted on the last year of the 7-year interval. Variability has been increasing over the past several decades. In particular the 1990s have seen unprecedented levels of interannual variability. The last period ending with 1998 has the largest trimmed standard deviation in the record. Although it has been speculated by the CSU group that we are entering a new era of above-average hurricane activity, there is little doubt that interannual variability is on the rise and will possibly continue for the next several years. This contrasts to the active late 1940s and 1950s when interannual variability was considerably lower. Over the 9-year period 1947-1955 no years were below average (<5). Results are similar if the coefficient of variation is used instead of the trimmed standard deviation. A similar tendency is noted in the interannual variations of major hurricanes. If the assumption is made that the distribution of hurricane activity is Poisson, then an increase in the variance is tantamount to an increase in the expected level of activity as suggested by the CSU group. The poorer performance of the seasonal forecast models of late is likely linked to large interannual variability and not to pre-millennial distractions about the end of the world.



4. Past performance

Although our group at FSU has been issuing seasonal hurricane forecasts since 1993, the earliest forecasts were for the frequency of intense hurricane activity only. Table 4 shows the long-lead (early December) forecast probabilities from the Poisson model over the past 6 years. The bold numbers indicate the verification. Forecasts were excellent for the 1993 and 1997 seasons, good for the 1994 and 1995 seasons, fair for the 1998 season, and poor for the 1996 season. We started issuing sub-basin forecasts during 1996. Each forecast is for 4 categories of sub-basin activity. Over the past 3 years, we have been correct on 8 of the 12 (66.7%) forecasts. This compares to 50% for forecasts based strictly on climatological norms.



5. Predictions for the 1999 hurricane season

Does the active late-starting 1998 season portend above-normal activity during 1999? Historically the answer is no. The 6 years following active (6 hurricanes) late-starting seasons combine for an average 5.8 hurricanes. The active 1950 season had 11 hurricanes following the late-starting 1949 season, while the inactive 1982 season had 2 hurricanes following the active late-starting 1981 season (Elsner and Kara 1999).

Here we describe our predictions for the 1999 North Atlantic hurricane season. Regression coefficients for Caribbean hurricane activity and the number of intense hurricanes are estimated from data over the period 1950-1996. The coefficients for the tropical-only hurricanes are estimated from the data over the period 1950-1997. These coefficients along with the new predictor values for 1999 are given in Table 5.

The hurricane model calls for 4.57 tropical-only hurricanes next season. To this we add the seasonal average number of baroclinically-enhanced hurricanes (=2.90) and round the sum (7.47) to the nearest whole number to obtain a forecast of 7 hurricanes for 1999. A forecast of intense hurricane activity based on the above regression coefficients is presented in the form of estimated probabilities for each possible number of intense hurricanes.

The 1999 probabilities are substantially different from 1998 probabilities indicating an increased likelihood of intense hurricane activity during 1999. The expected number of intense hurricanes is 3.56. More specifically, the Poisson model estimates that there is a 69% chance of observing three or more intense hurricanes during 1999. The logistic model for occurrence of a hurricane in the Caribbean indicates a 82% chance of observing at least one hurricane during 1999 in this part of the North Atlantic basin. This is above the long-term average of nearly 50%. Most dynamical and statistical models are predicting a cold ENSO during the early summer of 1999, afterwhich a return of El Niño conditions is possible, which would limit the amount of hurricane activity during the later part of the season. No subjective adjustments are made to these forecasts, however. Forecasts will be updated prior to the start of the the 1999 hurricane season.

Acknowledgments: Predictor values were obtained from the CSU group at http://tropical.atmos.colostate.edu/forecasts/index.html. Our research began as an academic exercise to compare statistical models being used at CSU with our newly-developed algorithms. We acknowledge the encouragement given to our group by the folks at the Risk Prediction Initiative (RPI) to persue and continue operational seasonal forecasting. Special thanks goes to Richard Murnane and Anthony Knap in this regard. We thank Eric Williford, Ricardo Correa-Torres, Brian Mackey, Todd Kimberlain, Andy Devanas, and Jon Rizzo for their valuable insights into the vagaries of North Atlantic hurricane climate. Partial support for this work came from the RPI of the Bermuda Biological Station for Research and from the National Science Foundation.



References

Elsner, J. B., and C. P. Schmertmann, 1993: Improving extended-range seasonal predictions of intense Atlantic hurricane activity. Weather and Forecasting, 8, 345-351.

Elsner, J. B., and A. B. Kara, 1999: Hurricanes of the North Atlantic: Climate and Society, Oxford.

Gray, W. M., C. W. Landsea, P. W. Mielke Jr., and K. J. Berry, 1992: Predicting Atlantic seasonal hurricane activity 6-11 months in advance. Weather and Forecasting, 7, 440-455.

Hess, J. C., J. B. Elsner, and N. E. LaSeur, 1995: Improving seasonal hurricane predictions for the Atlantic Basin. Weather and Forecasting,10, 425--432.

Lehmiller, G. S., T. B. Kimberlain, and J. B. Elsner, 1997: Seasonal prediction models for North Altantic basin hurricane location. Mon. Wea. Rev., 125, 1780--1791.





Table 1: 1998 North Atlantic hurricane season. Type refers to tropical-only (TO), baroclinically-enhanced (BE), and baroclinically-initiated (BI). Date refers to the day the storm was categorized as a tropical storm. Winds refer to highest maximum sustained one-minute average speed in knots. SLP refers to the minimum sea-level pressure in millibars. The asterisk indicates the storms that were hurricanes on September 26, 1998.
Category Type Name Date Wind(kts) SLP(mb)
IH-3 TO Bonnie Aug. 19-30 100 954
H-2 TO Danielle Aug. 24-Sep. 3 90 965
H-2 BE Earl Aug. 31- Sep. 3 85 986
IH-4 TO Georges* Sep. 15-29 130 939
H-1 BE Ivan* Sep. 20-27 80 975
H-2 TO Jeanne1* Sep. 21-Oct 90 970
H-1 BI Karl* Sep. 24-27 90 970
H-1 BE Lisa Oct. 5-9 65 987
IH-5 TO Mitch Oct. 22-Nov. 1 155 906
H-1 BI Nicole Nov. 24-Dec. 1 85 979





Table 2: Frequency of late-starting North Atlantic hurricane seasons. Values refer to the number of years as a function of the number of hurricanes. A late-starting season is defined as a season in which the first hurricane occurs after August 15th.
Number of H 0 1 2 3 4 5 6 7
Frequency 2 4 7 8 11 6 4 2



Table 3: Verification of the 1998 location forecasts. "No" means no hurricaneof a particular category occurred or was forecast.
Category Prediction Actual
IHG No No
IHC No Yes
HC No Yes
HSE Yes Yes



Table 4: Track record of the FSU intense hurricane model. Values are in percent. Bold number indicate the actual number of intense hurricanes. Past performance does not guarantee future success.
IH 1993 1994 1995 1996 1997 1998
0 16.8 15.9 4.7 25.9 16.0 35.5
1 29.9 29.2 14.4 35.0 29.4 36.8
2 26.7 26.9 22.0 23.6 26.9 19.0
3 15.9 16.5 22.4 10.7 16.4 6.6
4 7.1 7.6 17.1 3.6 7.5 1.7
5+ 3.3 3.7 19.4 1.2 3.8 0.4



Table 5: Predictor values and model coefficients for 1999 North Atlantic hurricane activity.
HC HT IH
i. Predictor i i i xi
0. Constant 1.123 4.283 1.250 ---
1. Aug-Nov 98 Gulf of Guinea Rainfall 3.440 2.783 0.614 -0.32 sd
2. Aug-Sep 98 West Sahel Rainfall 1.201 0.752 0.304 +0.23 sd
3. 50 mb Zonal Wind at 10N fcst Sep 99 -- 0.018 0.034 0 m s-1
4. 30 mb Zonal Wind at 10N fsct Sep 99 -- -0.015 -0.015 -1 m s-1
5. Wind shear at 10N fcst Sep 99 -- -0.129 -0.039 +1 m s-1



Table 6: Forecasts of intense hurricane activity for the 1999 North Atlantic hurricane season.
No. of Intense Hurricanes (IH) 0 1 2 3 4 5+
1998 Forecast Probabilities 0.355 0.368 0.190 0.066 0.017 0.004
1999 Forecast Probabilities 0.028 0.101 0.180 0.214 0.190 0.287