Effect of different detrending approaches on computational intelligence models of time series

This paper analyzes the impact of different detrending approaches on the performance of a variety of computational intelligence (CI) models. Three approaches are compared: Linear, nonlinear detrending (based on empirical mode decomposition) and first-differencing. Five representative CI methods are evaluated: Dynamic evolving neural-fuzzy inference system (DENFIS), Gaussian process (GP), multilayer perceptron (MLP), optimally-pruned extreme learning machine (OP-ELM) and Support Vector Machines (SVM). Four major conclusions are drawn from experiments performed on six time series benchmarks: 1) qualitatively, the effect of detrending is remarkably uniform for all the CI methods considered, 2) extraction of the overall trend does not improve performance in general 3) the EMD-based method provides better performance than linear detrending (while the difference is negligible in most cases, it is noticeable in some cases), and 4) first-differencing, while effective in some cases, can be counterproductive for series showing common patterns.