Novel multi-scale predictor for complex time series

The task of complex time series predicting is hard to be accomplished with only one single predicting model. In this paper, a complex time series is decomposed into a series of intrinsic mode functions and a residue signal. Then a RBF network is constructed for an intrinsic mode function or the residual signal. Finally output of every predicting model is integrated into one output with equal weighted. As the sifting process of EMD is an approximate frequency dividing process, the relationship between the logarithm of optimal parameter of every RBF network and its corresponding Intrinsic Mode Function is also approximate linear. This relationship can be utilized to alleviate the computing burden of the model selecting with cross validation method. Experimental results showed that the proposed method outperformed the single RBF network in the task of predicting complex time series.