Principles of the framework called time series forecasting automation are presented. It is required in processing massive temporal data sets and creating completely user-oriented forecasting software where manual data analysis and a user's decision-making is either impractical or undesirable. Its distinct features are local extrapolation models, their active training, criterion of model performance assessment used in adding new examples to the model training set and in deciding on which one of a group of competing models consistent with the common training set performs best. A generalized algorithm for local model tuning on massive data series that can be run without human intervention is presented.