Hybrid System with Dynamic Classification for Combining Time Series Forecasts.

In general, time series are composed of linear and nonlinear components. For this reason, hybrid techniques have been widely used for time series forecasting. In these systems, one approach is to decompose the series into its linear and nonlinear components. Generally, different linear and nonlinear techniques are used to individually model the original and residual series, respectively, improving the accuracy of the final forecast. The last step of hybrid models is the combination of linear and nonlinear predictions. For example, the final prediction can be obtained by adding the individual predictions, assuming a linear correlation between them. However, this is not always the best alternative. Another way to combine forecasts is through a nonlinear technique. Defining the best combination function for predictions is still an open problem. Thus, this paper proposes a new approach that incorporates in the system the choice of combining the linear and nonlinear components’ forecast in a linear or nonlinear way. This process is carried out considering a pool of classifiers responsible for predicting, for each time instant, how the combination should be handled. The pool is based on Decision Tree (DT), AdaBoost (AB) and Support Vector Machine (SVM). For linear and non-linear modeling, Auto Regressive Integrated Moving Average (ARIMA) and Support Vector Regression (SVR) techniques are used, respectively. The obtained results show that this alternative reduces the final forecast error and improves the accuracy of the system.