An Application of Similarity Search in Streaming Time Series under DTW: Online Forecasting

Time-series forecasting has had an incessant attraction to many researchers on time-series data mining. In the paper, we introduce an efficient online forecasting method based on similarity search in streaming time series under Dynamic Time Warping (DTW). The proposed method takes the newly incoming time-series subsequence, then finds k nearest neighbor subsequences and makes predictions based on the manner that these best matches evolved in the past. Prior to the similarity search, these subsequences have been extracted from the streaming time series by a novel segmentation technique using major extrema in time series. Experimental results show that for trend and seasonal streaming time series, the proposed method can bring out short-term forecasts with high prediction accuracy and remarkable time efficiency. Furthermore, if the streaming time series has some linear feature and no trend, another version of the online forecasting method, which hybridizes the aforementioned proposed method with simple exponential smoothing, can improve the prediction accuracy.