Skill of Seasonal Climate Forecasts in Canada Using Canonical Correlation Analysis

View More View Less
  • 1 Atmospheric Environment Service, Environment Canada, Downsview, Ontario, Canada
  • | 2 Climate Prediction Center, National Centers for Environmental Prediction, NWS/NOAA, Washington, D.C.
© Get Permissions
Full access

Abstract

An empirical system for forecasting 3-month mean surface temperature T and total precipitation P for Canada—canonical correlation analysis (CCA)—has been developed using the 1956–90 data period. The levels and sources of predictive skill have been estimated for all seasons at lead times of up to one year, using a cross-validation design. The predictor fields are quasi-global sea surface temperature (SST), Northern Hemisphere 500-mb geopotential height, and for T forecasts prior values of T itself. Four consecutive 3-month predictor periods are used to detect evolving as well as steady-state conditions in the predictor fields.

While forecast skills are modest for much of the year, winter and spring skills for T forecasts at a 3-month lead time are both highly statistically field significant and good enough to be beneficial to appropriate users. These forecasts average a 0.3–0.4 correlation skill nationwide and greater than 0.6 in the southeastern prairies. Forecast skill for P averages a lower but still statistically field significant 0.2 in winter with local maxima of greater than 0.5 along parts of southern Canada. A weak secondary seasonal maximum in T forecast skill is found in summer. CCA forecasts generally outperform persistence forecasts, and their skill declines only slowly as lead time is increased. Thus, useful forecasts can be made for certain seasons/regions of Canada several seasons in advance.

The CCA diagnostics indicate that the El Nin˜o/Southern Oscillation (ENSO) plays a dominant role in Canadian T anomalies in winter and spring, and P anomalies in winter. Warm SO (El Nin˜o) episodes tend to force positive winter and spring T anomalies in much of western and southern Canada, and suppressed P in roughly similar portions of the country. Below normal T tends to occur in northeastern Canada, and above normal P in the southeastern Northwest Territory, during warm SO episodes. Because of the linearity of CCA, opposite responses are implied for cold SO episodes. Another important skill source. for Canadian winter forecasts is associated with a long-term trend in global SST. Between the 1950s and the 1990s the high (low) latitude SST has tended to cool (warm). The Canadian winter T response has been a cooling from northern Quebec to northeastern Canada and warming in northwest Canada, while a trend toward greater (lighter) P in the northern (southern) prairies is noted. Knowledge of such trends can greatly aid in forecasting anomalies that are defined using normals for a period centered in the past.

In conclusion, statistically based long-lead forecasts of surface climate are shown to deliver useful skin in Canada. This approach also provides a skill benchmark against which the skill of dynamical models can be compared as they enter the forecasting arena.

Abstract

An empirical system for forecasting 3-month mean surface temperature T and total precipitation P for Canada—canonical correlation analysis (CCA)—has been developed using the 1956–90 data period. The levels and sources of predictive skill have been estimated for all seasons at lead times of up to one year, using a cross-validation design. The predictor fields are quasi-global sea surface temperature (SST), Northern Hemisphere 500-mb geopotential height, and for T forecasts prior values of T itself. Four consecutive 3-month predictor periods are used to detect evolving as well as steady-state conditions in the predictor fields.

While forecast skills are modest for much of the year, winter and spring skills for T forecasts at a 3-month lead time are both highly statistically field significant and good enough to be beneficial to appropriate users. These forecasts average a 0.3–0.4 correlation skill nationwide and greater than 0.6 in the southeastern prairies. Forecast skill for P averages a lower but still statistically field significant 0.2 in winter with local maxima of greater than 0.5 along parts of southern Canada. A weak secondary seasonal maximum in T forecast skill is found in summer. CCA forecasts generally outperform persistence forecasts, and their skill declines only slowly as lead time is increased. Thus, useful forecasts can be made for certain seasons/regions of Canada several seasons in advance.

The CCA diagnostics indicate that the El Nin˜o/Southern Oscillation (ENSO) plays a dominant role in Canadian T anomalies in winter and spring, and P anomalies in winter. Warm SO (El Nin˜o) episodes tend to force positive winter and spring T anomalies in much of western and southern Canada, and suppressed P in roughly similar portions of the country. Below normal T tends to occur in northeastern Canada, and above normal P in the southeastern Northwest Territory, during warm SO episodes. Because of the linearity of CCA, opposite responses are implied for cold SO episodes. Another important skill source. for Canadian winter forecasts is associated with a long-term trend in global SST. Between the 1950s and the 1990s the high (low) latitude SST has tended to cool (warm). The Canadian winter T response has been a cooling from northern Quebec to northeastern Canada and warming in northwest Canada, while a trend toward greater (lighter) P in the northern (southern) prairies is noted. Knowledge of such trends can greatly aid in forecasting anomalies that are defined using normals for a period centered in the past.

In conclusion, statistically based long-lead forecasts of surface climate are shown to deliver useful skin in Canada. This approach also provides a skill benchmark against which the skill of dynamical models can be compared as they enter the forecasting arena.

Save