Date of Graduation
My dissertation concentrates on comparing the forecasting performance of three types of multivariate models: a structural model, a transfer function model in time domain, and two vector autoregressive models (VAR). These models are specified and estimated to reflect the market behavior of important variables in the U.S. copper market. The forecasting performance of these models is compared in the context of a commodity modeling exercise. The purpose is two fold, to assess the relative forecasting performance of the models, and to determine the role and the necessity of a priori information in forecasting economic time series. Only the standard commodity model is based completely on economic theory. For the transfer function model the inputs and the outputs of the system are determined on a priori grounds and the dynamic structure of the relationships are extracted from the data. Univariate ARIMA and vector autoregressive models are not based on any particular notion of causality. In the Bayesian vector autoregressive model (BVAR) we impose weak a priori restrictions on the parameters but the restrictions have no foundation in economic theory. A structural model consisting of four equations and one identity is specified, i.e., demand, supply, price, foreign trade, and an identity for inventories. For the transfer function model, the inputs and outputs are differentiated based on the economic theory while time series techniques are used to identify the transfer functions relating them. The variables used in the structural model are employed to identify the VAR models. The unrestricted VAR is identified with no a priori restrictions imposed on it while certain loose restrictions are imposed on the BVAR model parameters. The estimated models are simulated to generate one, four, eight and twelve-period-ahead unconditional forecasts for supply, demand, price, and net foreign trade. Results are compared to assess the relative forecasting performance of the models. Overall results indicate that the structural model performs the best for the one-period-ahead forecast while the BVAR model forecasting performance is the best for other forecasting horizons. Furthermore, they indicate that employing some form of a priori information is essential and indispensable. These results are however subject to certain qualifications. (Abstract shortened with permission of author.).
Moallem, Masoud, "Structural and multivariate time-series models: Comparative application in commodity markets." (1988). Graduate Theses, Dissertations, and Problem Reports. 9436.