Hybrid Criteria for Nearest Neighbor Selection with Avoidance of Biasing for Long Term Time Series Prediction

Nearest neighbor is pattern matching method for time series prediction in which most recent values of the time series are compared with previous available values and forecasting is achieved by finding the best match pattern (nearest neighbor). Usually Euclidean distance is used to check the similarity of pattern. In this paper two hybrid criteria of pattern matching are being proposed and evaluated for multistep-ahead time series prediction. The first selection criterion is hybrid of “Maximum distance and Cross-Correlation” and second is hybrid of ‘Manhattan distance and Cross-correlation”. Better forecasting has been achieved using these algorithms.