Time series prediction based on non-parametric regression and wavelet-fractal

In this paper, a short-term time series prediction method is proposed. The method is based on the fundamental character of chaotic time series. By introducing the concept of series fractal time-varying dimension, a new standard of distance between two series is presented. With the wavelet transform, we search for the top k most nearest series in the history data set at different resolution ratio and use their neighbor series for prediction. The final result of prediction is obtained by summing up the individual results on each scale. Finally, we validate the approach on the prediction of real-time traffic data.