Comparison of Probabilistic Statistical Forecast and Trend Adjustment Methods for North American Seasonal Temperatures

Abstract

he three multivariate statistical methods of canonical correlation analysis, maximum covariance analysis, and redundancy analysis are compared with respect to their probabilistic accuracy for seasonal forecasts of gridded North American temperatures. Derivation of forecast error covariance matrices for the methods allows a probabilistic formulation for the forecasts, assuming Gaussian predictive distributions. The three methods perform similarly with respect to probabilistic forecast accuracy as reflected by the ranked probability score, although maximum covariance analysis may be preferred because of its slightly better forecast skill and calibration. In each case the forecast accuracy for North American seasonal temperatures compares favorably to results from previously published studies. In addition, two alternative approaches are compared for alleviating the cold biases in the forecasts that derive from ongoing climate warming. Adding lagging 15-yr means to forecast temperature anomalies improved forecast accuracy and reduced the cold bias in the forecasts, relative to using the more conventional lagging 30-yr mean.