New Algorithms for Multi-Step Ahead Forecasting

History tells that every human being desire to foresee, comprehend and ultimately explore the future. Multi-step ahead forecasting is a challenging research area due to propagation of forecasting errors with the increase of forecasting steps. Two interesting architectures based on nearest neighbor method are proposed. Importance of selection criteria in nearest neighbor search plays an important role in multi-step ahead forecasting. Effect of up-sampling of time series and change of effective embedding dimension on the forecasting errors is studied in detail. Effect of five interpolation schemes for up-sampling and comparison of three distance metrics for nearest neighbor search on forecasting performance is also included. A hybrid selection criterion of nearest neighbor with avoidance of biasing is found to be very effective in multi-step ahead forecasting. In the end, predictability analysis of proposed algorithms on ten benchmark time series highlight the effectiveness of the forecasting algorithms in the scenarios of series collected from different kinds of dynamic systems. This book is based on the PhD work of Mr. Rahat Abbas.