A novel grey prediction model for seasonal time series

Considering the weakness in the discrete grey seasonal model, a new grey seasonal model is put forward by introducing a time trends item. Moreover, some properties of this proposed model are deduced, such as the unbiased feature, to provide more information to perceive this model. Subsequently, four time series concerning the quarterly and monthly electricity and petroleum consumptions that have various features of the upward, downward, and wave tendencies from China, America, Japan, and Germany, are adopted to verify the availability and generality of this new model. Experimental results from these four case studies demonstrate that, on the one hand, the proposed method can strikingly improve the simulating and forecasting performance compared with the conventional discrete grey seasonal model, indicating this new model is capable of describing seasonal time series with different tendencies. On the other hand, this new technology is validated to have superior forecasting ability over a range of grey models, econometric models, and machine learning methods. Finally, the impact of sample size on the precision for the new model is further discussed, and results suggest that the modeling sample length should be at least four times the number of cycles in a seasonal sequence in order to ensure the satisfied and stable forecasting accuracy.