A large quantity of training data is often required in order to have enough representatives examples to ensure good performances of any learning-based forecasting method. Yet, time series composed of too many data can also be a problem. It would quite possibly take a long time to generate adequate solutions. In this presentation, a methodology based on a stratification of time series based on some clustering procedures has been developed for a prior selection of training examples. The principle for deterministically constructing folds with unsupervised stratification consists in assigning similar instances to the different folds instead of using a random sampling approach. The results show that with a small number of training examples, obtained through stratification of data, we can improve the performance and stability of models such as artificial neural network and support vector regression, while training at much lower cost. We illustrate the properties of the methodology in forecasting one-step ahead hourly Ontario energy prices (HOEP).