[Previous Article] [Table of Contents] [Next Article]


Early August Forecasts of Atlantic Tropical Storm Activity
for the Balance of the 1996 Season, Using Poisson Models

contributed by James Elsner, Gregor Lehmiller and Todd Kimberlain
Department of Meteorology, Florida State University, Tallahassee, Florida


The models used here have been developed using data through July, and thus apply to forecasts only for the balance of the storm season (August 1 through November 30).

Forecast methodology for basin-wide activity

The two components of hurricane activity that we have studied are seasonal number of hurricanes (H) and seasonal number of intense hurricanes (I). To predict the number of hurricanes we use a Poisson regression with a maximum likelihood criterion to estimate the number of tropical-only hurricanes Ht (Elsner et al. 1996, Hess et al. 1995), to which we add a seasonal average number of baroclinically-influenced hurricanes Hb, adjusted conditionally on the number of tropical-only hurricanes. The conditional adjustment amounts to removing the variance of Hb explained by H t. This is accomplished by using a Poisson model to predict Hb from Ht. Ht and Hb are weakly negatively correlated.

The Poisson regression used for prediction of tropical type hurricanes uses four predictors, as originally suggested by Gray et al. (1993). This prediction model can be expressed as:

H = H t + H b

where Ht = exp (b0 + b1x1 + b2x2 + b3x3 + b4x4)

and where the b's are coefficients on the predictors x.

The four predictors include (1) a 1-month forward extrapolation of the 30 mb zonal winds at 10N (a measure of the QBO phase), (2) Jun-Jul average rainfall anomalies (expressed in standard deviations) for the West Sahel region in Africa, (3) the past Aug-through-Nov average rainfall standardized anomalies in the Gulf of Guinea region, and (4) the Southern Oscillation Index (SOI) averaged over June and July. The direction of the relationships are as highlighted in Gray's predictions: positive correlations with the rainfalls, with the 30 mb zonal wind, as well as with the SOI.

We also use a Poisson regression to estimate the number of intense hurricanes (3 or more on the Saffir/Simpson scale), as detailed in Elsner and Schmertmann (1993):

I = exp (g0 + g1x1 + g2x2 + g3x3 + g4x4)

The predictors used to forecasts intense hurricanes are identical to those used to predict tropical-only hurricanes except that the 50 mb zonal wind is used instead of the SOI, providing a more complete description of the QBO state. This set of predictors is changed slightly from that used in the longer-lead forecast for the 1996 storm season shown in the December 1995 issue of this Bulletin because it was determined, based on statistical tests, that the shear parameter used in the early December forecast model adds nothing significant to the early August model. The December 1993 issue of this Bulletin briefly summarizes the reasoning behind the beneficial use of the Poisson as compared to ordinary multiple linear regression, particularly for relatively infrequently occurring events such as intense hurricanes.

Both the hurricane and intense hurricane models have been skill-evaluated using a hold-one-out cross-validation strategy. The correlation between actual and predicted number of hurricanes is 0.65 with a mean absolute error of 1.51 storms. For the number of intense hurricanes the correlation between actual and predicted is 084, with mean absolute error 0.74 storms.

Forecast methodology for sub-basin activity

In addition to basin-wide activity we are now predicting activity in four sub-basins of the 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. 1996). We use logistic regression to predict hurricane landfalls along the Northeast and Southeast coasts and the presence or absence of intense hurricanes in the Gulf and Caribbean. As with the approach taken for basin-wide activity, we express the sub-basin forecasts in terms of probabilities.

Logistic regression is a statistical model used to predict events in a yes/no framework by estimating coefficients for several predictor variables. Here we use a maximum likelihood technique to obtain the coefficients. A logical regression can be expressed as

prob(yes) = -------------------------------------------------

where the a's are the coefficients on the p predictors x.

Predictors for these models include those used for basin-wide activity, with the addition of several others. For example, vertical shear is computed from July averaged winds at Cape Hatteras and Miami using 00 and 12 UTC rawinsonde observations. The shear is the Euclidean distance between the 700 and 200 mb wind components:

[ (u700-u200)2 + (v700-v200)2 ]. Additionally, we use sea level pressures (SLPs) averaged over four locations along the U.S. East Coast including Miami, Charleston, Cape Hatteras and Boston, and sea surface temperature anomalies averaged from April through June in the North Atlantic and the tropical Atlantic and Caribbean (Gray et al. 1996) and the zonal wind anomaly from upper tropospheric (12 km) zonal winds averaged over June and July (Gray et al. 1996).

Prediction error for the logistic models expressed as a cross-validated accuracy ratio ranges from 81.4% (compared with a climatology of 58.1%) for the U.S. Southeast coastal hurricane landfall model, to 78.3% (climatology of 52.2%) for the Gulf of Mexico intense hurricane model, to 82.6% (climatology of 52.2) for the Caribbean intense hurricane model. Forecast accuracy for the U.S. Northeast coastal landfall model is not significantly better than climatology.



Predictions for the 1996 Atlantic Hurricane Season

Some of the predictor data for the statistical models was obtained from Dr. Chris Landsea on August 7. The predictor variables and regression coefficients for the number of tropical-only hurricanes and the number of intense hurricanes, estimated from the 1950-95 data, are shown in Table 1 and the resulting forecast probabilities are shown in Tables 2 and 3.

The forecasts indicate a near average 1996 hurricane season with a better than 70% chance of observing less than 2 intense hurricanes in addition to Bertha (which reached intense hurricane strength in July). Moreover, there is a better than 65% chance of observing between 3 and 6 hurricanes from August onward (i.e. in addition to the two tropical cyclones that reached hurricane strength before the beginning of August). The models indicate a 40% chance of observing at least one more (following July) intense Caribbean hurricane this year, and reflect only a slim likelihood of an East Coast hurricane landfall in addition to Bertha. Despite the expectation of a near average season, Table 3 indicates that the forecast probabilities for East Coast hurricane landfall are lower than the climatological probabilities.



Table 1. Model specifications for total Atlantic hurri-cane activity. See text for more detail about the nature and timing of the predictors.

Coeff: Coeff:
1996 Tropical Intense
Predictor Term Value Hurric. Hurric.
constant ----- 1.036 1.117
50mb zonal wind -20 m/s ----- 0.021
30mb zonal wind -33 m/s 0.014 0.013
western Sahel rainfall -0.60 sd 0.369 0.410
Gulf of Guinea rainfall 0.10 sd 0.670 0.428
SOI 1.30 sd 0.163 -----

Table 2. Atlantic hurricane activity probabilities for 1996. H refers to hurricanes, I intense hurricanes. The expected value of H for the balance of the 1996 season is 4.5, and the expected value of I is 1.1.

Number
0 1 2 3 4 5 6 7 8 9+
H 0.011 0.050 0.112 0.169 0.190 0.171 0.128 0.082 0.046 <0.040
I 0.344 0.367 0.196 0.070 0.019 0.004 0.001 0.000 0.000 0.000



Table 3. Probabilities of a hurricane landfall along the Southeast and Northeast coasts and the probabilities of the presence or absence of an intense hurricane in the Gulf and Caribbean for 1996. H denotes hurricanes, I intense hurricanes. The asterisk indicates that the model has statistically significant forecast skill, given the present sample size of events.

Climatological
Sub-basin Probability Probability
H SE Coast 0.088 0.488
H NE Coast 0.010 0.152
I Gulf 0.130 0.478
I Caribbean 0.400 0.478



Acknowledgments: Partial support for this work came from the Risk Prediction Initiative (RPI) of the Bermuda Biological Station for Research (BBSR).

References

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

Elsner, J.B., G.S. Lehmiller and T.B. Kimberlain, 1996: Objective classification of Atlantic basin hurricanes. J. Climate, 9, accepted.

Gray, W.M., C.W. Landsea, P.W. Mielke, and K.J. Berry, 1993: Predicting Atlantic seasonal tropical cyclone activity by 1 August. Wea. Forecasting, 8, 73-86.

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

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



[Previous Article] [Table of Contents] [Next Article]