An optimized time series combined forecasting method based on neural networks

In the field of time series forecasting, combining forecasts from multiple models significantly improves the forecasting precision as well as often produces better forecasts than each constituent model. The linear method is the main method in the current literatures for it is simpler and more efficient, and usually gives good results. However, the selection of the basic unit models and the determination of combination weights always bring difficulties to combined forecasting model. In addition, there is usually more time consuming in the combining model. To address these problems, this paper proposes an optimized time series combined forecasting method based on neural networks. The principle of the proposed method adheres to two primary aspects. (a) A multi-threaded grid search method is proposed to quickly determine the number of hidden layer nodes in a neural network for a scenario. (b) A method based on sample dynamic partitioning to explore the weight generation mode is proposed to determine the combined forecasting. Empirical results from six real-world time series show that the superiority of our approach in forecasting accuracies.