Hybrid System for Time Series using Iterative Residual Forecasting Models

Real world time series generally are composed of linear and nonlinear patterns. Several hybrid systems of the literature combine linear and nonlinear models using the residuals modeling with the objective to map separately such patterns. However, problems as misspecification, underfitting, and overfitting in the training process can degenerate the performance of the models and, consequently, of the whole system. This work proposes an Iterative Residual Forecasting (IRF) method with the objective to improve the accuracy of the models that compose hybrid systems from successive corrections using the residual modeling. So, a set of models can be generated to forecast each pattern (linear, and nonlinear), where the posterior model corrects the last one through of the modeling of their residuals. The experimental evaluation with four well known time series of the literature indicates that the IRF method achieves promising results, showing that can be a useful tool to enhance the accuracy of the hybrid systems.