[Previous Article]
[Next Article]
LAD Multiple Linear Regression Forecasts of Atlantic
Tropical Storm Activity for 1996
contributed by William Gray
Colorado State University, Fort Collins, Colorado
A specific version of multiple linear regression is used
by Dr. William Gray of Colorado State University to make forecasts separately
for each of several parameters of tropical storm activity in the Atlantic
Basin (Gray et al. 1992). Least absolute deviation (LAD) multiple linear
regression is used to develop the prediction model for these forecasts.
LAD regression was discussed briefly in the December 1992 issue, and is
described more fully in Gray et al. (1993a) and references therein. In
essence, the calculation of the weights for the predictor variables is
based on least absolute deviation (LAD) of forecasts with respect to observations,
rather than on least squared deviation as in ordinary regression. Forecasts
for each year's storm season are made at three times: in late November
of the preceding calendar year, in early June, and finally in early August
at the beginning of the storm season. This year an additional April forecast
update was issued. Forecast parameters include (1) named storms, (2) named
storm days, (3) hurricanes, (4) hurricane days, (5) intense hurricanes,
(6) intense hurricane days, (7) hurricane destruction potential, and (8)
net tropical cyclone activity. The predictors used, and the skill expected,
at each of the three forecast times were tabulated in the September 1993
issue of this bulletin. However, the prediction system has been revised
several times as opportunities for skill increases emerged. In 1994, for
example, prediction of the state of ENSO to be expected during the fall
storm season was incorporated into the regression equation rather than
being an a posteriori human adjustment (Gray et al. 1994). The following
list identifies the five predictor clusters, their expected influence on
Atlantic tropical storm activity, and their status regarding the 1996 storm
season. More detail on the influences of these predictor groups is found
in Gray et al. (1993a, 1993b).
(1) The QuasiBiennial Oscillation (QBO) at 50 and 30 mb in the northern tropics expected at the onset of the hurricane season: Westerly QBO winds enhance storm activity, while easterly winds suppress it (Gray et al. 1992). A large shear between the two levels is also an storm-inhibiting factor. In late summer 1996 the QBO is expected to be easterly at both 50 and 30 mb, which will suppress hurricane activity. The shear is not expected to be significant.
(2) El Nino/Southern Oscillation (ENSO): Warm eastern equatorial Pacific sea surface temperature (SST), or El Nino, reduces storm activity, while anomalously cool SST enhances it. Going into summer we currently have a weak negative SST anomaly in the NiZo 3 region. Dr. Gray's ENSO prediction scheme calls for a continuation of this state or possibly some further weakening, resulting in near normal to slightly cool conditions during the late summer (about 0 to -0.5 EC anomalies). This would tend to slightly enhance hurricane activity.
(3) African rainfall: Intense hurricane activity is enhanced when the Western Sahel and Gulf of Guinea regions in West Africa have above average precipitation the previous late summer and fall (implying favored chances for the same anomaly sign for the upcoming late summer and fall, which is the actual "predictor"), and is suppressed when that precipitation is below average (Landsea et al. 1993). Rainfall in the Western Sahel in Aug-Sep of 1995 was above average, and Gulf of Guinea rainfall last August through November was also above average. From these observations and from the current rainfall outlook, a non-drought 1996 rainy season is anticipated and thus there will be no tendency for suppressed intense hurricane activity.
(4) West Africa westtoeast surface pressure and temperature gradients: Above average westtoeast surface pressure and (often associated) easttowest surface temperature gradients from February to May are associated with enhanced hurricane activity later that year. For February through May of this year the gradients for both variables were near average, and the consequent influence on tropical storm activity is neutral.
(5) Caribbean basin sea level pressure anomaly (SLPA) and upper tropospheric (12 km) zonal wind anomaly (ZWA): Negative anomalies of either one of these weakly imply enhanced storm activity, while positives weakly associate with reduced activity. For AprilMay 1996 SLPA averaged +1.0 mb, exerting some inhibiting influence on storm activity. However, ZWA was somewhat below its mean despite the normally positive correlation between SLPA and ZWA. Taken together, a neutral to slightly inhibiting influence on 1996 storm activity is indicated.
In a newly developed prediction procedure for early June
forecasts, a set of 11 potential predictors is introduced for each predictand,
and the set 4 to 7 of these that maximize the cross-validated estimate
of sample explained variance is used for each of the 9 predictands on an
individual predictand basis. The use of SST in the NiZo 3.4 region serves
as a substitute for the previously used NiZo 3 SST and the SOI (see Goldenberg
and Shapiro 1995). Table 1 shows which predictors were selected for each
predictand.
Table 1. The predictors chosen out of the potential pool
of 11 for each of the 9 tropical storm predictands.
PREDICTOR |
|||||||||||||||||
NS |
NSD |
H |
HD |
IH |
IHD |
HDP |
NTC |
MPD |
|||||||||
QBO:Zonal wind, 50mb |
x |
x |
x |
x |
x |
||||||||||||
QBO:Zonal wind, 30mb |
x |
x |
|||||||||||||||
QBO:| 30mb - 50mb shear | |
x |
x |
x |
x |
x |
||||||||||||
Jun - Sep Sahel rain, 4-yr avg |
x |
x |
x |
x |
|||||||||||||
Aug-Nov Gulf Guinea rain |
x |
x |
x |
x |
x |
x |
x |
x |
|||||||||
Feb - May E - W P grad, W Sahel |
x |
x |
x |
x |
|||||||||||||
Feb - May E - W T grad, W Sahel |
x |
x |
x |
x |
x |
x |
|||||||||||
Apr - May Carib. SLP |
x |
x |
x |
x |
x |
||||||||||||
Apr - May 200mb Carib. Zonal wind |
x |
x |
x |
x |
|||||||||||||
Apr - May Nino 3.4 SST |
x |
x |
x |
x |
x |
x |
x |
x |
x |
||||||||
(Apr-May) - (Jan-Feb) Nino 3.4 SST |
x |
x |
The LAD multiple regression predictions, made first in
November 1995 and then in early June 1996, for each tropical storm parameter
are shown in Table 2. The mean values based on the 1950-95 period are shown
to the right. The first column shows expected skill for the forecasts in
terms of percent variance explained.
Table 2. Predicted Atlantic tropical storm historical hindcast skills and forecasts for the 1996 season, as of late November 1995 and early June 1996. Skill expected for independent (real-time) forecasts made in June has been estimated using resampling simulations. These are equal to approximately 90% of the hindcasting skill levels obtained using cross-validation with a partially dependent sample.
Forecasts |
||||||||
Expected Skill |
June |
|||||||
1950-95 |
November |
1996 |
||||||
Atlantic Tropical |
(% Variance |
1995 |
----Fcsts.---- |
1950-95 | ||||
Cyclone Parameter |
Explained) |
Fcst. |
Objective |
Final |
Mean |
|||
Named Storms |
0.48 |
8 |
8.0 |
10 |
9.3 |
|||
Named storm days |
0.59 |
40 |
24.9 |
45 |
46.6 |
|||
Hurricanes |
0.51 |
5 |
4.4 |
6 |
5.8 |
|||
Hurricane days |
0.60 |
20 |
14.8 |
20 |
23.9 |
|||
Intense hurricanes |
0.57 |
2 |
2.0 |
2 |
2.3 |
|||
Intense hurricane days |
0.53 |
5 |
4.1 |
5 |
4.7 |
|||
Hurricane destruction potential |
0.59 |
50 |
57.6 |
60 |
71.2 |
|||
Net trop. cyc. activity (% avg) |
0.66 |
85% |
66.3% |
95% |
100.0% |
|||
Maximum potential destruction |
0.59 |
41.5 |
60 |
66.0 |
Approximately near normal tropical storm activity is
predicted for 1996, as it was in the late November forecast. The March
update (not shown in Table 2) indicated generally slightly higher storm
activity than that shown in the present forecast, due in part to the expectation
of a continuation of the cool tropical Pacific conditions that had not
yet weakened substantially by early April.
The best analog years to 1996 are 1954, 1979 and
1989--especially 1989.
There is a significant difference between the objective
and final forecasts for some of the 1996 storm activity predictands, in
the direction of upward adjustments. This is the case because the Gray
team believes that the positive April-May Caribbean SLP anomaly is unrepresentative
of conditions to be expected by the time late summer arrives. Additionally,
all three of the best analog years listed above were underpredicted. Still
other factors that are not taken into account in the objective equations
are: late spring northern tropical Atlantic SST, salinity and SST in the
north Atlantic (speeding up of the thermohaline circulation), and a slightly
downwardly revised recent outlook for NiZo 3.4 SST for this season. The
predictands that use the Caribbean SLP as a predictor (Table 1) have been
adjusted upward by the greatest amounts relative to the objective forecast.
Goldenberg, S.B. and L.J. Shapiro, 1995: A new look at
the relationships between El NiZo, West African rainfall, and north Atlantic
tropical cyclone activity. Preprints, 21st Conference on Hurricanes
and Tropical Meteorology. American Meteorological Society, April 24-28,
Miami, Florida, 585-587.
Gray, W.M., C.W. Landsea, P.W. Mielke, and K.J. Berry,
1992: Predicting Atlantic seasonal hurricane activity 611 months in advance.
Wea. Forecasting, 7, 440455.
Gray, W.M., C.W. Landsea, P. Mielke and K. Berry, 1993a:
Predicting Atlantic basin seasonal tropical cyclone activity by 1 August.
Wea. Forecasting, 8, 7386.
Gray, W.M., J.D. Sheaffer, P.W. Mielke, K.J. Berry and
J.A. Knaff, 1994: Predicting ENSO 914 months in advance. Proceedings of
the 18th Annual Climate Diagnostics Workshop, Boulder, Colorado, November
15, 1993, 390-393.
Landsea, C.W., W.M. Gray, P.W. Mielke and K.J. Berry,
1993: Predictability of seasonal Sahelian rainfall by 1 December of the
previous year and 1 June of the current year. Preprints, 20th Conference
on Hurricane and Tropical Meteorology, AMS, San Antonio, Texas, 473476.