Time Series Forecasting Using the Generalized NGBM (1,1)

Abstract

Sequence data are sometimes rare, non-linear and non-normal, the models form grey theory are just suitable for this kind of scenario. An generalized version of originally GM(1,1) is described that takes power exponent, smoothing factor, initial condition and residual modification into account. In order to alleviate the tediousness of manual parameter selection and the problem of over- fitting in training stage, we then conduct the parameter optimization and parameter screening using genetic algorithm (GA) and 2k factorial design, respectively. The above model does not deviate from the idea of simplicity in grey theory. The experiments suggest that the high- precision of proposed method is able to improve the effectiveness of prediction.