Genetic Algorithm Approach for Estimation of Parameters of Vector Autoregressive Models under Heteroscedasticity.

Abstract

Abstract : Forecasting is one of the core focuses of statisticians working in agricultural research. Obtaining timely as well as accurate forecasts under all possible circumstances is the need of the hour. Most of the forecasting techniques make one or the other assumptions limiting their applications. Vector Autoregression is one such widely used multivariate forecasting technique where homoscedasticity of errors is assumed for estimation of parameters by ordinary least square (OLS) method. This study proposes genetic algorithm (GA), a heuristic search algorithm, which does not make any such assumptions for estimating the parameters under such situation. The developed methodology is empirically validated using simulated bivariate vector autoregressive model of order 1 under heteroscedasticity. The relative error of parameter estimates and Mean Absolute Percentage Error have shown that GA performs better than OLS estimation under heteroscedasticity. The proposed methodology is also tested under homoscedasticity using bivariate data of fish landings. The results indicated that both GA and OLS are equally efficient in estimating the parameters.