The principal Discriminant Method (PDM) of prediction employs a novel combination of principal component analysis and statistical discriminant analysis. Discriminant analysis is based on the construction of discrete category subsets of predictor values in a multidimensional predictor space. A category subset contains those predictor values which give rise a predictand (or observation) in that particular category. A new predictor value is then assigned to a particular category (i.e., a forecast is made) through the use of probability distribution functions which have been fitted to the category subsets. The PDM uses principal component analysis to define the multidimensional probability distribution functions associated with the category subsets. Because of its underlying discriminant nature the PDM is also applicable to problems in data classification. The PDM is applied to prediction problems using both artificial and actual data sets. When applied to artificial data the PDM shows forecast skills which are comparable to those of standard forecast techniques, such as linear regression and classical discriminant analysis. When applied to actual data in a forecast of the 1982--1983 El Ni¿o, the PDM performed poorly. However, in forecasting winter air temperatures over North America, the PDM proved superior to the other forecast techniques, after suitably filtering or smoothing the raw data in order to improve the signal-to-noise ratio. It is expected that the PDM will show its greatest advantage over other forecast techniques when the relation between predictors and predictand is nonlinear. ¿ American Geophysical Union 1988