Two nonlinear machine learning methods, support vector regression (SVR) and neural networks (NN), together with linear regression (LR), were used to forecast the maximum 5-day accumulated precipitation over winter in Canada at lead times of 3, 6, 9 and 12 months. The precipitation data were obtained from 108 Canadian stations for the winters (DJF) of 1950 to 2006. We tested the following predictors: the 5 leading principal components (PCs) of the global tropical sea surface temperatures (SST), the SST index from the Niņo 3.4 region in the tropical Pacific, the 500 hPa geopotential height (Z500) over the Northern Hemisphere, the North Atlantic Oscillation (NAO) index, the Pacific/North American (PNA) teleconnection index and the Pacific Decadal Oscillation (PDO) index. The leave-one-out cross-validation results showed overall improved skill using the SVR model relative to the NN and LR models. For the SVR and NN models at short lead time (3 months) and LR, using the global tropical SST PCs and the Niņo 3.4 SST index as predictors gave the best seasonal extreme precipitation forecasts. For the longer lead times (6, 9 and 12 months), the SVR and NN models were improved when PDO was added as an extra predictor, suggesting that PDO contributes nonlinearly to the Canadian seasonal extreme precipitation.