A0420
Title: Multivariate adaptive learning forecasting
Authors: Dimitrios Thomakos - National and Kapodistrian University of Athens (Greece)
Foteini Kyriazi - Agricultural University of Athens (Greece) [presenting]
John Guerard - McKinley Capital Management (United States)
Abstract: A new method is presented for forecasting multivariate time series, either in simultaneous equations or panel form, that utilizes adaptive learning on past forecast errors for effecting root mean-squared error reductions to any input forecast that one wishes to utilize. The method is the multivariate extension of univariate adaptive learning and presents a number of significant advantages over the univariate approach, as most multivariate methods do when dealing with related time series. The multivariate version of adaptive learning can be used in a number of different settings and can be used to extract useful information for forecasting from different forms of covariation among multiple time series. We explore in detail the theoretical foundations of the method, and the computational requirements and present a number of simulation and empirical examples that illustrate both the efficacy and the competitiveness of the method compared to a number of well-known time series benchmarks.
