Improvements in Tropical
Cyclone Track Forecasting
in the Atlantic Basin, 197098
Colin J. McAdie and Miles B. Lawrence
NOAA/NWS/NCEP/Tropical Prediction Center, Miami, Florida
ABSTRACT
Tropical cyclone track forecasts issued by the Tropical Prediction Center/National Hurricane Center for the Atlantic basin have improved over the period 197098. Improvement is shown at 24, 48, and 72 h. Although this improvement can be shown without any preconditioning of the data, the question of accounting for forecast difficulty is addressed,
building upon the work of Neumann. A decrease in the initial position errors over the same period is also shown.
Track forecast errors generated by the Atlantic climatology and persistence (CLIPER) model (run in best-track
mode) are used as a measure of forecast difficulty. Using the annual average CLIPER errors in a regression against the
official forecast errors yields an equation giving an expected error for each year under consideration. The expected error
(representing forecast difficulty) is then subtracted from the observed official errors. The resulting set of differences can
then be examined for long-term trends, difficulty having been accounted for.
Fitting a straight line to these differences (197098) yields the result that official forecast errors have decreased by
an average of 1.0% per year at 24 h, by 1.7% per year at 48 h, and by 1.9% per year at 72 h. A second-order fit, however,
suggests that the rate of improvement has increased during the latter half of the period.
1. Introduction
The Tropical Prediction Center/National Hurricane
Center (NHC)1 issues tropical cyclone track and intensity forecasts and warnings for both the Atlantic and
eastern North Pacific basins.2 As responsibility for the
In 1995, the National Hurricane Center was renamed the Tropical Prediction Center (TPC), one of nine centers within the National Centers for Environmental Prediction, as part of the National
Weather Service restructuring and modernization. At the time of
this writing, only the hurricane forecast unit within TPC retained
the title National Hurricane Center.
By convention, areas of the globe prone to tropical cyclone occurrence are referred to as basins. The Atlantic basin includes
the North Atlantic Ocean, the Caribbean Sea, and the Gulf of
Mexico. In the eastern North Pacific basin, NHCs area of forecast responsibility extends from the Pacific coasts of North and
South America westward to 140W.
Corresponding author address: Mr. Colin J. McAdie, Tropical
Prediction Center, National Hurricane Center, 11691 SW 17th
Street, Miami, FL 33165-2149.
E-mail: colin@nhc.noaa.gov
In final form 20 December 1999.
Bulletin of the American Meteorological Society
eastern North Pacific was transferred to NHC in 1988
from what was then the Eastern Pacific Hurricane
Center in Redwood City, California, this analysis of
long-term trends in track forecast error will be confined to the Atlantic basin.
The costs of warning U.S. coastal residents for
hurricane landfall are significant. Jarrell and DeMaria
(1998) give a rough estimate of $600,000 per mile of
warned coastline and find an average length of warned
coastline over the last decade of about 400 n mi
(741 km). Given the long-term average of 1.7 hurricane landfalls, or near-landfall events, per year along
the U.S. Gulf of Mexico and Atlantic coastlines
(Neumann et al. 1999), the average annual warning
costs for hurricanes are then about $400 million. Pielke
and Landsea (1998) find average annual hurricane
losses (normalized to 1995 economic conditions and
population) of $4.8 billion.
The issue of long-term trends in tropical cyclone
track forecast error was addressed comprehensively by
Neumann (1981). Neumann demonstrated the importance of first removing forecast difficulty from track
errors before determining trends in forecast skill. That
technique is described and used in this paper.
989
The objectives of the present study are to provide
an updated version of Neumanns earlier work, and
to extend the method to include the 48- and 72-h forecasts. This study also includes weighting of trend lines
by number of cases. We do not attempt to remove the
effect of initial position errors on the forecast, as was
done by Neumann. Any improvements in initial positioning are included here as a contribution to the
forecast process. Simplifications to the procedure used
to account for forecast difficulty are also discussed.
A 29-yr period (197098) is examined. While long
enough to show any statistically significant trends, this
period is reasonably uniform in the basic technology
available to the forecaster. It does not, for example,
extend back into the presatellite or precomputer era.
Geostationary satellite data and computer-generated track
forecast model guidance were available through the period, although in increasingly refined form in the latter
part. Global model forecasts became available in 1974.
2. Data
The set of average3 annual initial position errors
and official forecast errors for forecasts issued by NHC
during the period 197098 are given in Table 1. The
origin of the data is as follows. The forecaster first
determines the current (0 h) center position of a tropical cyclone, specified to tenths of degrees latitude and
longitude. This position, two past positions (12 h and
24 h), the initial motion (i.e., forward speed and direction), and the 12 h motion are supplied as input
to the numerical guidance. The suite of track models
(DeMaria et al. 1990; Gross 1999) that compose the
guidance is run on demand, usually as soon as the input parameters have been determined. After the guidance has been run, the set of track forecasts thus
generated are evaluated by the forecaster, along with
the forecast fields generated by the global and regional
models. The forecaster synthesizes this information
and determines the official track forecast (specified
as a set of latitudelongitude points representing the
forecast center positions) for 12, 24, 36, 48, and 72 h.
The process is repeated four times a day at 0000, 0600,
1200, and 1800 UTC.
Verification of these forecasts is provided by the
Atlantic best-track file, also known as HURDAT
Average is here defined as the arithmetic mean and is used in
this sense throughout.
990
(Jarvinen et al. 1984), updated annually. This file contains the final, official estimates of tropical cyclone
position (given in tenths of degrees latitude and longitude) and maximum 1-min sustained winds 10 m
above the surface (given in knots), as determined in
poststorm analysis at NHC. Data that may not have
been available in real time are used to modify and refine the operational track. The revised tracks are then
entered into the best-track file.
In this study, initial position error is defined as the
great-circle distance between each operationally determined initial position and its best-track (verifying)
position. Likewise, forecast error is defined as the
great-circle distance between each official forecast
position at 24, 48, and 72 h and its corresponding besttrack position.
Only forecasts for tropical cyclones having maximum 1-min sustained winds of 34 kt (17.5 m s1; tropical storm strength) or greater have been included. This
requirement applies to both the time of the forecast and
the verification time. Subtropical cases have been included, but extratropical cases have been excluded.4
3. Determination of trends in initial
position error
As noted above, the first step in the forecast process is to determine the initial (0 h) position. Average
annual initial position errors are given in Table 1 and
plotted in Fig. 1. A downward trend is readily apparent. An unweighted linear least squares trend line (not
shown) significant at the 95% level, using the F-test
criterion (Draper and Smith 1966), gives an average
annual improvement of about 2.1% (Table 2). This
estimate does not, however, take into account the fact
that interannual variation in tropical cyclone activity
can cause a substantial change in the number of cases
(Table 1). One method of accounting for this sample
variability is to weight the linear least squares fit by
the number of cases. This weighted line (shown in
Fig. 1) yields an average annual improvement for ini-
The term subtropical refers to a low pressure system that develops over subtropical waters, initially having a nontropical circulation, but with some tropical cyclone cloud structure. Many
evolve into tropical cyclones. The term extratropical refers to a
tropical cyclone that has undergone modification by moving into
a nontropical (baroclinic) environment; characteristic changes
include expansion of wind field, decrease in maximum winds, and
increasing asymmetry of the wind field (Neumann et al. 1999).
Vol. 81, No. 5, May 2000
TABLE 1. NHC average initial positioning (0 h) errors, and average track forecast errors at 24, 48, and 72 h, 197098. Units are
nautical miles (n mi). Average best-track (BT) CLIPER errors are calculated for a homogeneous sample.
0h
24 h
48 h
72 h
Year
Official
Cases
Official
BT clip
Cases
Official
BT clip
Cases
Official
BT clip
1970
22.0
46
84.3
63.2
34
185.9
234.2
13
254.0
525.9
1971
22.6
203
110.7
66.7
183
242.1
183.3
137
381.9
330.4
118
1972
24.0
66
142.4
112.4
57
390.5
370.0
38
687.9
699.7
25
1973
24.3
98
116.8
93.0
84
246.3
245.6
54
363.2
374.8
29
1974
18.2
103
97.2
64.0
89
206.6
151.6
64
348.3
223.8
42
1975
15.5
143
117.0
84.6
121
256.9
210.3
92
401.9
316.4
68
1976
20.3
159
128.2
95.3
142
287.8
265.4
112
433.0
416.2
85
1977
11.4
40
133.0
108.5
30
331.1
324.2
14
484.7
574.0
1978
18.5
123
144.3
99.4
101
323.6
260.1
59
422.8
357.1
33
1979
17.8
156
89.5
63.1
138
160.2
149.7
98
238.6
244.0
83
1980
15.6
209
128.7
95.0
188
273.3
226.4
140
404.8
360.1
109
1981
19.7
211
126.0
83.3
190
248.4
222.0
146
422.9
394.4
106
1982
16.9
55
131.2
90.8
44
244.7
218.3
28
269.3
267.3
20
1983
10.2
45
85.3
71.0
33
197.1
219.1
17
439.6
443.5
1984
22.2
178
132.1
90.9
156
266.7
262.7
121
390.1
416.4
88
1985
16.5
175
109.9
95.3
149
221.7
246.0
104
330.1
350.2
67
1986
20.0
76
107.5
73.7
66
237.6
191.9
42
383.6
322.6
27
1987
19.0
131
108.6
82.5
119
229.7
242.7
95
349.1
418.9
67
1988
12.2
152
69.7
63.8
133
143.1
160.6
109
230.9
267.9
90
1989
17.5
239
95.7
68.4
215
192.9
182.9
166
283.5
287.4
129
1990
15.0
235
101.3
69.5
205
194.8
188.5
156
303.7
313.5
113
1991
15.0
69
113.6
100.1
55
192.0
243.8
31
296.8
391.2
17
1992
10.3
138
83.0
71.0
122
166.8
221.3
98
278.4
423.2
75
1993
14.0
103
102.4
88.4
80
177.9
235.0
58
240.7
348.6
41
1994
12.9
97
102.8
115.2
83
209.8
286.1
62
341.8
382.6
50
1995
12.2
444
87.1
72.3
402
159.4
191.7
335
233.3
292.4
272
1996
10.2
290
72.0
64.7
260
128.2
180.4
217
189.9
317.8
183
1997
11.8
91
85.5
72.9
74
150.2
182.0
51
229.3
325.1
38
1998
12.5
317
84.1
73.6
284
144.9
191.1
231
201.8
285.7
191
Bulletin of the American Meteorological Society
Cases
991
tial position errors of 2.2%. Although the percentage
improvement is about the same as that taken from the
unweighted line, the percentage of variance explained
has increased from 50% to 62% (Table 2).
Increased skill in determining the initial position
is clearly a contributing factor in lowering forecast
error, primarily because successive positions determine the estimate of initial motion (speed and direction). This was shown by Neumann (1981), who
sought to remove the effect of initial position errors
from the forecasts, and in doing so found an average
error reduction in those forecasts of 6%, 2%, and 1%
at 24, 48, and 72 h, respectively. In other words, a perfect initial position would result, on average, in an
improvement of 6% in the 24-h forecast, but the benefit damps out quickly beyond 24 h.
Several possible explanations for the improvement
in initial positioning have been investigated. First,
however, the possibility must be considered that factors independent of skill are responsible for the decrease in errors or at least bias the true trend. For
example, well-defined, intense hurricanes with small
eyes are easier to locate than ill-defined tropical storms
(Mayfield et al. 1988). If a climatological increase in
average intensity were present, the initial position errors shown in Fig. 1 might be attributable to that alone
and not to any increase in skill. An examination of the
best-track file, however, shows no such trend in average intensity. Another factor to be considered is the
availability of aircraft reconnaissance, which is a key
factor in center location (Gray et al. 1991). Data provided by the Air Force Reserve 53d Weather Reconnaissance Squadron were examined to determine if the
number of center fixes (normalized by storm duration)
TABLE 2. Percentage improvement per year, taken from trend
lines. Variance explained given in parentheses.
Forecast
Unweighted
Weighted by
no. of cases
Adjusted and
weighted
0h
24 h
48 h
72 h
2.1% (0.50)
1.2% (0.27)
2.1% (0.43)
2.2% (0.38)
2.2% (0.62)
1.5% (0.45)
2.3% (0.60)
2.5% (0.62)
N/A
1.0% (0.80)
1.7% (0.88)
1.9% (0.87)
had changed with time. No evidence of such a trend
was found.
Geographic location is another possible influence.
More tropical cyclone occurrences in the western (data
dense) portion of the basin over a span of years might
result in lower initial position errors. Nevertheless, the
annual average initial longitude shows no identifiable
trend.
Having ruled out these possible spurious effects in
trend, the strongest contributing factor to lower initial
position errors appears to be the forecasters enhanced
ability to acquire, view, and manipulate satellite data.
Notable in this regard is the acquisition of the University of Wisconsin Man-computer Interactive Data
Access System (McIDAS) in the early 1980s, and the
installation of the visible and infrared spin scan radiometer atmospheric sounder Data Utilization Center
at NHC in 1989 (Sheets 1990).
We conclude then, based upon these data, that there
has been an improvement of about 2% per year in the
initial position errors over the 29-yr period.
4. Determination of trends in track
forecast error
FIG. 1. Average annual NHC initial positioning errors, 1970
98. Data appear in Table 1. Trend line is weighted by number of
cases. Units are nautical miles.
992
a. Unadjusted trend lines
As in the case of initial position errors above, the
most straightforward approach to finding a trend in
track forecast errors is to employ a linear least squares
fit of the average errors against time. The unadjusted
average annual track forecast errors (Table 1) with
trend lines (weighted by number of cases) are shown
in Figs. 2a (24 h), 2c (48 h), and 2e (72 h). The trend
lines have negative slopes, indicating improvement,
for all three forecast periods.
Improvement found is 1.5% per year at 24 h, 2.3%
per year at 48 h, and 2.5% per year at 72 h (Table 2).
(For comparison, corresponding rates of improvement
Vol. 81, No. 5, May 2000
taken from unweighted trend lines are 1.2%, 2.1%, and
2.2% at 24, 48, and 72 h, respectively.) Although the
weighted trend lines are significant at the 95% level,
using the F-test criterion, note that considerable variability remains about these lines (Figs. 2a, 2c, and 2e).
A reduction of variance of 45%, 60%, and 62% at 24,
48, and 72 h, respectively (Table 2), indicates that time
has accounted for some of the year-to-year variation
in average error but leaves a portion unexplained.
Weighting by number of cases, as was true with initial position error, does not change the rate of improvement very much but does obtain a greater reduction
of variance.
tities into a stepwise multiple regression against official errors demonstrated that neither latitude nor the u
component of motion provided any significant additional reduction of variance beyond that provided by
CLIPER error alone. Thus it emerged as the best single
predictor of official forecast error or, stated in another way, as a predictor of forecast difficulty.
In the Atlantic, tropical cyclones traveling westward, generally south of 20N, are propelled by a persistent deep easterly flow and tend to have small
CLIPER track errors. As systems become caught up
in the westerlies (recurve), forward speed increases and
tracks become more difficult to predict; CLIPER errors are large. CLIPER errors may thus be used to
characterize areas of typical tropical cyclone motion
within a basin or, indeed, to characterize an entire basin (Pike and Neumann 1987). When the CLIPER
model is run retroactively on archived tracks, its errors become a measure of forecast difficulty. This is
the central contribution of Neumanns technique. The
b. Use of CLIPER to account for forecast difficulty
In the computation of the trends in Figs. 2a, 2c, and
2e forecast difficulty has not been taken into account.
The chance occurrence of a group of relatively easy
forecasts toward the end of the period might create the
erroneous impression of an increase in skill or exaggerate the slope of the trend line. Following
Neumann (1981), the track forecast
errors generated by the Atlantic climatology and persistence (CLIPER)
model (Neumann 1972; Neumann and
Pelissier 1981) are taken as a measure
of forecast difficulty. CLIPER is a
simple track prediction model derived
by regression against a set of previous
tracks with future position as the dependent variable. In the version of the
Atlantic CLIPER model used in this
study, 240 tropical cyclone tracks occurring between 1930 and 1970 provide 3156 dependent cases (Neumann
1972).
The use of the CLIPER model errors as a measure of forecast difficulty
has a quantitative basis. Neumann
(1981) considered a number of possibilities, including initial positioning
error, latitude, longitude, translational
speed, u and v component of motion,
CLIPER errors, and others. Screening
for correlation with official forecast
error, the three most highly correlated
were CLIPER error, latitude, and u
component of motion. Because of the
FIG. 2. Average NHC official track forecast errors over the period 197098 for
nature of its derivation, however, the Atlantic basin. Average errors are shown unadjusted for difficulty at (a) 24, (c)
CLIPER contains information about 48, and (e) 72 h, and adjusted for difficulty using the best-track CLIPER errors at (b)
the other two. Inserting all three quan- 24, (d) 48, and (f) 72 h. Vertical axis units are nautical miles.
Bulletin of the American Meteorological Society
993
advantage is that years during which a preponderance
of forecasts occur south of 20N can be fairly weighed
against those years containing more erratic tracks; further, increased skill in predicting those erratic tracks,
if present, will become evident by comparison with the
CLIPER bench mark.
In this work, CLIPER-model errors have been generated using best-track data rather than operational
input. Although either will serve the purpose outlined
above, the use of the best-track CLIPER has an advantage in that it is completely objective and reproducible, whereas the operational CLIPER is subject to the
skill of a particular forecaster in determining initial
motion.
For each official forecast (197098), a best-track
CLIPER error was generated, forming a homogeneous
sample. Yearly averages of these best-track CLIPER
errors are given in Table 1. In spite of the simplicity
of the CLIPER model, note in Table 1 that the average best-track CLIPER errors are lower than the average official errors at 24 h, in all years except 1994.
c. Longitude as an additional measure of forecast
difficulty
Although CLIPER error is the single best predictor of forecast difficulty, chosen objectively in favor
of latitude and u component of motion, Neumann
(1981) found that longitude provided some additional
reduction of variance (5.6% at 24 h)5 when used in
conjunction with CLIPER. Longitude acted to decrease expected forecast difficulty as one traveled
westward in the basin. Presumably, this reflected the
increased data density in the western part of the basin, resulting in a better analysis of the steering flow
and better model initialization.
In the present work the reduction of variance due
to longitude is much smaller (about 0.9% at 24 h). At
48 h the reduction of variance is about 0.1%, and at
72 h is negligible. For comparison with Neumanns
work, these numbers were computed without accounting for number of cases. When the number of cases is
accounted for, the reduction of variance due to longitude essentially vanishes at 24 and 48 h, and is about
1.7% at 72 h. We have thus chosen to exclude longitude in this work and use CLIPER as the sole predictor of forecast difficulty. It is possible that advances
Neumann used the period 195480. Since 48- and 72-h forecasts
did not begin until 1961 and 1964, respectively, they were not
available for comparison over the same interval at that time.
994
in global modeling, a technology not available during
the first part of Neumanns original dataset, have
damped this longitudinal signal.
d. Adjustment procedure
The sole measure of forecast difficulty (CLIPER)
is now applied to the average official errors, prior to
fitting them against time. This step is accomplished
through regression, exploiting the correlation between
the average official track forecast errors and the average CLIPER errors. The regression, using average
official errors as the dependent variable, is weighted
by the number of cases composing the respective annual averages. As mentioned above, weighting was
considered necessary because the number of cases per
year varies widely. With only one predictor of difficulty (CLIPER), the result of the regression simply
assumes the form
y = a + bx,
where y is the expected (or predicted) official error
for any particular year under consideration, and x is
the annual average best-track CLIPER error for that
year. In this procedure, an expected official error is calculated and then subtracted from the annual average
observed error. The difference (y y ) represents forecast skill, difficulty having been accounted for. The set
of adjusted official errors is finally obtained by adding the mean error for the entire period (197098) to
each of these annual differences. Since the objective
is to identify any trends in forecast skill, the final step
is to fit these adjusted official errors against time
(Figs. 2b, 2d, and 2f). Fitting a straight line to these
data, adjusted official track forecast errors over the
period 197098 are shown to have decreased by an
average of 1.0% per year at 24 h, by 1.7% per year at
48 h, and by 1.9% per year at 72 h (rounded to the
nearest 0.1%). These lines are significant at the 95%
level, using F-test criteria.
e. Use of a second-order fit
Inspection of the adjusted errors in Figs. 2b, 2d, and
2f suggests that the rate of improvement is greater during the latter part of the period. The first half shows a
relatively flat trend and perhaps even some deterioration in skill at 24 h. This change in rate of improvement can be shown with a second-order fit against
time, plotted in Figs. 3ac. The curves are found by
using time (year number) and the square of time as the
independent variables. The curves are significant at the
Vol. 81, No. 5, May 2000
95% level, again using F-test criteria. Using values
from the curves, the rates of improvement for the final 5-yr period (199498), for example, are about 2.1%
per year at 24 h, 3.1% per year at 48 h, and 3.5% per
year at 72 h.
Choice of fit focuses attention on different timescales. The linear fit is perhaps the more conservative
of the two, giving a longer-term perspective. If forecast errors should remain constant or rise slightly over
the next several years, the linear fit will give about the
same result; the second-order fit will react more
quickly.
It is interesting to compare this to Neumanns result for the period 195480, in which he found most
of the improvement in the 24-h forecast during the
earlier part, with essentially no improvement during
the 1970s. There was concern at the time that a plateau in skill had been reached, attributed in part to
a degradation in the initial analysis over the midAtlantic. It is now evident that this plateau was overcome with a substantial increase in skill during the
ensuing 18-yr period.
5. Discussion
Tropical cyclone track forecasting, in its present
state, is heavily dependent upon numerical guidance.
This guidance is provided both by the global models
directly and by the suite of track forecast models using the global model forecast fields as input. There has
been a steady increase in the skill of the global models over the last several decades (Shuman 1989;
Kalnay et al. 1990; Caplan et al. 1997). In addition,
increasingly skillful tropical cyclone prediction models have been introduced. Near the midpoint of the
period under consideration, the statistical/dynamical
NHC83 model was introduced (Neumann 1988) and
later revised as NHC90 (Neumann and McAdie 1991).
The Beta and Advection Model (Marks 1990) became
routinely available in 1990. More recently, a tropical
cyclone prediction model developed at the NOAA/
ERL/Geophysical Fluid Dynamics Laboratory
(Kurihara et al. 1995) became operational in 1995.
This model has performed exceptionally well, with
average 72-h track errors during the 1998 hurricane
season of 223 n mi (Gross 1999). It is our opinion that
improvement such as this in the numerical guidance
has been largely responsible for the decrease shown
in the official tropical cyclone track forecast errors.
This linkage was also suggested by Sheets (1990).
Bulletin of the American Meteorological Society
FIG. 3. As in Fig. 2, except that adjusted errors are shown with
second-order fit. Vertical axis units are nautical miles.
Given this improvement in the numerical guidance,
and the steady decrease in official track forecast errors,
the question arises as to how long this trend might
continue. An estimate of the amount of improvement still
possible in the numerical guidance is provided in
a discussion by Neumann (OFCM 1997) using a statistical/dynamical model in the North Atlantic basin.
995
He finds that average errors of 53, 107, and 145 n mi
at 24, 48, and 72 h, respectively, should be attainable.
Leslie et al. (1998) used a baroclinic model to obtain inherent lower bound average errors of 52, 83, and 121 n mi
at 24, 48, and 72 h, respectively, in the Atlantic.
Given either of the above estimates, and taking the
final 5-yr average (weighted by number of cases) of
the official NHC track forecast errors given in Table 1
as a measure of current skill in the Atlantic basin (84,
151, and 221 n mi, at 24, 48, and 72, h, respectively)
it would appear that there is room for further improvement. Looking ahead, it is entirely possible that other
plateaus in skill such as that experienced during the
decade of the 1970s will be reached, and it should
therefore not be assumed that the rate of improvement
shown over the last decades will apply to the future.
6. Conclusions
A small, but statistically significant, long-term
downward trend is evident both in the initial position
errors and in the official tropical cyclone track forecast errors, in the Atlantic basin, over the period 1970
98. Forecast errors are found to decrease at all forecast
intervals examined, specifically at 24, 48, and 72 h.
(Official NHC tropical cyclone forecasts do not currently extend beyond 72 h.)
During this period, tropical cyclone initial positioning errors have decreased by an average of 2.2% per
year, and official track forecast errors have decreased
by an average of 1.0% per year at 24 h, by 1.7% per
year at 48 h, and by 1.9% per year at 72 h, when determined by linear trend lines, using the best-track
CLIPER as a measure of forecast difficulty, and
weighting by number of cases. These trend lines are
significant at the 95% level.
While accounting for difficulty does (sometimes
markedly) change year-to-year comparisons, it does
not change the sense of the trend, nor does it significantly change the magnitude of the trend. Removal of
difficulty does, however, result in greater reduction of
variance.
A second-order fit gives a shorter-term improvement over the last five years (199498) of 2.1% per
year at 24 h, 3.1% per year at 48 h, and 3.5% per year
at 72 h.
Acknowledgments. The authors acknowledge the assistance of
Jim Gross of the Tropical Prediction Center in making available the initial positions and official forecasts through the ATCF
996
(Automated Tropical Cyclone Forecast) database. We would also
like to thank Joan David for her expert assistance with the graphics. Major Douglas Lipscombe of the Air Force Reserve 53d
Weather Reconnaissance Squadron deserves our thanks for his
time and effort in providing the center-fix database used in the
analysis of initial position errors. We acknowledge Bob Burpee,
former director of the Tropical Prediction Center, for his encouragement in publishing this work. Finally, we wish to thank Charlie
Neumann for the many helpful discussions that we have had over
the years on the use of multiple regression as it applies to tropical
cyclone forecasting.
References
Caplan, P., J. Derber, W. Gemmill, S.-Y. Hong, H.-L. Pan, and
D. Parrish, 1997: Changes to the 1995 NCEP Operational
Medium Range Model Analysis-Forecast System. Wea. Forecasting, 12, 581594.
DeMaria, M., M. B. Lawrence, and J. T. Kroll, 1990: An error
analysis of Atlantic tropical cyclone track guidance models.
Wea. Forecasting, 5, 4761.
Draper, N. R., and H. Smith, 1966: Applied Regression Analysis.
John Wiley and Sons, 407 pp.
Gray, W. M., C. J. Neumann, and T. L. Tsui, 1991: Assessment
of the role of aircraft reconnaissance on tropical cyclone analysis and forecasting. Bull. Amer. Meteor. Soc., 72, 18671883.
Gross, J. M., 1999: 1998 National Hurricane Center forecast verification. Minutes, 53d Interdepartmental Hurricane Conference, Rockville, MD, Office of the Federal Coordinator for
Meteorological Services and Supporting Research, B24B63.
Jarrell, J. D., and M. DeMaria, 1999: An examination of strategies to reduce the size of hurricane warning areas. Preprints,
23d Conf. on Hurricanes and Tropical Meteorology, Dallas,
TX, Amer. Meteor. Soc., 5052.
Jarvinen, B. R., C. J. Neumann, and M. A. S. Davis, 1984: A tropical cyclone data tape for the North Atlantic basin, 18861983:
Contents, limitations, and uses. NOAA Tech. Memo. NWS
NHC 22, 21 pp. [Available from Environmental Science Information Center, Environmental Data Service, NOAA, U.S.
Department of Commerce, 3300 Whitehaven St. NW, Washington, DC 20235.]
Kalnay, E., M. Kanamitsu, and W. E. Baker, 1990: Global numerical weather prediction at the National Meteorological Center.
Bull. Amer. Meteor. Soc., 71, 14101428.
Kurihara, Y., M. A. Bender, R. E. Tuleya, and R. J. Ross, 1995:
Improvement in the GFDL Hurricane Prediction System. Mon.
Wea. Rev., 123, 27912801.
Leslie, L. M., R. F. Abbey Jr., and G. J. Holland, 1998: Tropical
cyclone track predictability. Meteor. Atmos. Phys., 65, 223231.
Marks, D. G., 1990: The beta and advection model for hurricane
track forecasting. NOAA Tech. Memo. NWS NMC 70, 89 pp.
[Available from Environmental Science Information Center,
Environmental Data Service, NOAA, U.S. Department of
Commerce, 3300 Whitehaven St. NW, Washington, DC
20235.]
Mayfield, M., C. J. McAdie, and A. C. Pike, 1988: A preliminary evaluation of the dispersion of tropical cyclone position
and intensity estimates determined from satellite imagery.
Vol. 81, No. 5, May 2000
Tropical Cyclone Studies, Office of the Federal Coordinator
for Meteorological Services and Supporting Research, 2-1217.
Neumann, C. J., 1972: An alternate to the Hurran tropical cyclone
forecasting system. NOAA Tech. Memo. NWS SR 62, 24 pp.
[Available from Environmental Science Information Center,
Environmental Data Service, NOAA, U.S. Department of Commerce, 3300 Whitehaven St. NW, Washington, DC 20235.]
, 1981: Trends in forecasting the tracks of Atlantic tropical
cyclones. Bull. Amer. Meteor. Soc., 62, 14731485.
, 1988: The National Hurricane Center NHC83 Model.
NOAA Tech. Memo. NWS NHC 41, 44 pp. [Available from
Environmental Science Information Center, Environmental
Data Service, NOAA, U.S. Department of Commerce, 3300
Whitehaven St. NW, Washington, DC 20235.]
, and J. M. Pelissier, 1981: Models for the prediction of tropical cyclone motion over the North Atlantic: An operational
evaluation. Mon. Wea. Rev., 109, 552538.
, and C. J. McAdie, 1991: A revised National Hurricane Center NHC83 Model (NHC90). NOAA Tech. Memo. NWS NHC
44, 35 pp. [Available from Environmental Science Informa-
Bulletin of the American Meteorological Society
tion Center, Environmental Data Service, NOAA, U.S. Department of Commerce, 3300 Whitehaven St. NW, Washington,
DC 20235.]
, B. R. Jarvinen, C. J. McAdie, and G. R. Hammer, 1999:
Tropical Cyclones of the North Atlantic Ocean, 18711988.
Historical Climatology Series, Vol. 6-2, National Oceanic and
Atmospheric Administration, 206 pp.
OFCM, 1997: National Plan for Tropical Cyclone Research and
Reconnaissance (19972002). National Oceanic and Atmospheric Administration, 91 pp.
Pielke, R. A., and C. W. Landsea, 1998: Normalized hurricane
damages in the United States, 192595. Wea. Forcasting, 13,
621631.
Pike, A. C., and C. J. Neumann, 1987: The variation of track forecast difficulty among tropical cyclone basins. Wea. Forecasting, 2, 237241.
Sheets, R. C., 1990: The National Hurricane CenterPast, Present
and Future. Wea. Forecasting, 5, 185232.
Shuman, F. G., 1989: History of numerical weather prediction at
the National Meteorological Center. Wea. Forecasting, 4, 286
296.
997
