Type-1 fuzzy forecasting functions with elastic net regularization

Abstract

Fuzzy functions have recently been used for forecasting problems. The main concepts behind a fuzzy functions are to cluster the inputs using a fuzzy clustering method and to include the obtained membership grades and their non-linear transformations as new variables in the input matrix. Then, multiple linear regression models are solved for different clusters. However, adding related variables to the input matrix leads to the multicollinearity problem. Thus, the main contribution of the proposed method is to employ an elastic net in fuzzy functions to overcome the aforementioned problem. Two regularization terms occur in an elastic net that come from the ridge and the lasso regression. These regularization terms are optimized using the nested cross-validation approach to overcome the multicollinearity problem in the fuzzy functions method. Twelve practical time-series datasets are analyzed to evaluate the performance of the proposed fuzzy functions. The outstanding performance of the proposed method has been verified in terms of root mean squared errors and mean absolute percentage errors for the selected datasets.