Prior training with jittered series for time series forecasting

The improvement of forecasting accuracy is always an important and difficult task in many areas. In this study, we propose a method of constructing the jittered series to improve neural network forecasting performance for a short time series. My experiment shows that jittered series has a significant impact on forecasting performance of a neural network especially in a short time series. The noise level of time series has no significant impact on the size of jittered series. The larger the size of training sample is, the less impact of a jittered series will be. The results of the linear simulated data show different from those of the nonlinear simulated data in forecasting performance while training with jittered series. The smaller size of the training sample can improve the forecasting performance 30% higher for the linear simulated data and 50% to 60% higher for the nonlinear simulated data.