Performance evaluation of a residual-based approach and clustering for time series with application to demand forecasting

Demand forecasting is a critical issue in predicting customer demand and improving the corresponding management plans; this is commonly done through the analysis of the so-called time series. However, it is possible that adequacy of a forecasting approach is violated, due to model misspecification or disturbances of the phenomenon that are not completely captured, leading to biased or inefficient predictions. In this thesis, a residual-based approach is evaluated to improve and correct forecasts, obtained previously with other models by using their remaining residuals. Such approach involves the training of different models on such series of residuals, combining the obtained predictions and evaluating their performance in terms of accuracy. Furthermore, some clustering approaches are used to improve the knowledge about both the original time series and those obtained from the residuals, in such a way to combine the outputs of such technique in the process of evaluation for the proposed approach.