[Previous Article] [Table
of Contents] [Next Article]
Experimental Canonical Correlation Analysis (CCA) Forecasts
of Canadian Temperature and Precipitation
-- Jan-Feb-Mar 1997
Contributed by Amir Shabbar1 and Anthony Barnston2
1Climate Research Branch, Atmospheric Environment
Service, Downsview, Ontario, Canada
amir.shabbar@ec.gc.ca
2Climate Prediction Center, NOAA, Camp Springs,
Maryland
In the last three issues of this Bulletin, forecasts of Canadian
temperature and precipitation using the multivariate statistical technique
of canonical correlation analysis (CCA) were presented. For Canada, we
have developed the predictive relationships between evolving large scale
patterns of quasi-global sea surface temperature, Northern Hemisphere 500
mb circulation, and the subsequent Canadian surface temperature and precipitation.
In this issue we present the forecasts for Jan-Feb-Mar 1997 using the predictor
fields through August 1996. These forecasts are made with lead times of
7 months, where lead time is defined as the time between the end of the
latest predictor season and the end of the predictand season. Cross-validation
skill scores in historical data indicate that CCA-based seasonal forecasts
are most skillful for winter predictions. Further detail about Canadian
CCA-based seasonal climate prediction is found in Shabbar (1996a, 1996b)
and Shabbar and Barnston (1996).
During the Oct-Nov-Dec period, historically expected forecast skill is
generally nonsignificant across most of Canada. Exceptions are found in
southern British Columbia, where modest skill is demonstrated and above
normal temperatures are predicted for Oct-Nov-Dec 1996, and southern Ontario
where modest skill is also found for below normal precipitation this fall.
With the weakening of the cold episode in the equatorial Pacific, the Oct-Nov-Dec
forecast is not expected to be appreciably influenced by the ENSO condition
this fall even in those regions where ENSO is a source of predictive skill.
Figure 1 shows the CCA-based temperature forecast for the 3 month period
of Jan-Feb-Mar 1997 expressed as a standardized anomaly. The temperatures
are forecast to be above normal from the west coast to the St. Lawrence
Valley. Below normal temperatures are expected throughout most of the east
coast and northern Canada. Table 1 shows the value of the standard deviation
in C at selected stations. The mean skill over all 51 stations is given
in the caption beneath each forecast map. The field significance is also
shown, reflecting the probability of randomly obtaining overall map skill
equal to or higher than that which actually occurred. Field significance
is evaluated using a Monte Carlo procedure in which the forecast versus
observation correspondences are shuffled randomly 1000 times. The field
of cross-validated historical skill (correlation) for the forecast time
period is shown in Figure 2. This forecast has a mean national score of
0.39 and a "perfect" field significance of 0.000. In fact, Jan-Feb-Mar
is the best period for temperature forecasts in Canada. The skill of forecasts
for Jan-Feb-Mar deteriorates only very slightly as the lead time of the
prediction is increased out to a year (see Fig. 11-1(a) in the September
1995 issue of Experimental Long Lead Forecast Bulletin). Modest but usable
skill is found throughout southern Canada south of the 60 degree latitude.
Skill is particularly high (>0.6) over southwestern Ontario.
Figure 3 shows the CCA-based precipitation forecast for Jan-Feb-Mar 1996
expressed as a standardized anomaly. Table 1 shows the value of the standard
deviation (mm) at selected stations. Cross-validated historical skill (correlation)
for this time period is shown in Figure 4. The forecast has modest expected
skill, with a mean national score of 0.18. However, the field significance
is 0.000, the strongest value possible. Local skill is highest along the
upper Great Lakes and in southeastern British Columbia. Precipitation is
expected to be below normal in southern British Columbia and over southern
Ontario and southern Quebec. With lower expected skill, an excess in winter
precipitation is forecast for the Yukon, the Mackenzie Valley and along
the east coast of Canada.
Most of the statistical and dynamical models are predicting further weakening
of the cool conditions in the equatorial Pacific Ocean over the fall season
and a possible return toward a weak warm ENSO episode in late winter and
spring of 1997. The Jan-Feb-Mar temperature forecast is somewhat reminiscent
of a warm episode; however, the standardized anomalies are weak. Factors
other than ENSO may also be influencing the precipitation forecast for
the Jan-Feb-Mar period.
Table 1. Standard deviation of temperature (Temp) and precipitation
(Prcp) for the 3 month period January through March at selected Canadian
stations.
|
Station
|
Temp (C)
|
Prcp (mm)
|
|
|
|
|
Whitehorse
|
5.7
|
8.6
|
|
Fort Smith
|
4.2
|
9.1
|
|
Innujjuak
|
3.4
|
7.4
|
|
Eureka
|
3.5
|
2.0
|
|
Vancouver
|
1.6
|
51.9
|
|
Edmonton
|
4.5
|
10.8
|
|
Regina
|
3.9
|
9.3
|
|
Winnipeg
|
3.4
|
11.9
|
|
Churchill
|
3.1
|
10.1
|
|
Moosonee
|
3.1
|
18.6
|
|
Toronto
|
2.3
|
20.7
|
|
Quebec City
|
2.6
|
35.8
|
|
Halifax
|
2.0
|
56.7
|
|
St. John's
|
2.5
|
55.0
|
References
Shabbar, A., 1996a: Seasonal prediction of Canadian surface temperature
and precipitation by canonical correlation analysis. Proceedings of
the 20th Annual Climate Diagnostics Workshop, Seattle, Washington,
October 23-27, 1995, 421-424.
Shabbar, A., 1996b: Seasonal forecast of Canadian surface temperature by
canonical correlation analysis. 13th Conference on Probability and Statistics
in the Atmospheric Sciences. American Meteorological Society, San Francisco,
California, February 21-23, 339-342.
Shabbar, A. and A.G. Barnston, 1996: Skill of seasonal climate forecasts
in Canada using canonical correlation analysis. Mon. Wea. Rev., 124,
2370-2385.
Figures
Fig. 1. CCA-based temperature forecast for the 3 month mean period
of Jan-Feb-Mar 1997. Forecasts are represented as standardized anomalies.
Fig. 2. Geographical distribution of cross- validated historical skill
for the forecast shown in Fig. 1, calculated as a temporal correlation
coefficient between forecasts and observations. Areas having forecast skill
of 0.30 or higher are considered to have utility. The mean score over 51
stations is 0.39. Field significance is 0.000.
Fig. 3. As in Fig. 1 (CCA anomaly forecast), except for Jan-Feb-Mar
1997 precipitation.
Fig. 4: As in Fig. 2 (geographic distribution of correlation skill)
except for the precipitation forecast shown in Fig. 3. The mean score over
69 stations is 0.18. Field significance is 0.000 (see text).
[Previous Article]
[Table of Contents] [Next
Article]