MODEL SELECTION AND FORECASTING ABILITY OF THEORY-CONSTRAINED FOOD DEMAND SYSTEMS

Out-of-sample forecasting of annual U.S. per capita food consumption, applying data from 1923 to 1992, is used as a basis for model selectio n among the absolute price Rotterdam model, a first-differenced linear approximate almost ideal demand system (FDLA/ALIDS) model, and a firs t-differenced double-log demand system. Conditional-on-price consumpti on forecasts derived from elasticities are determined to be superior t o direct statistical model forecasts. Models with consumer theory impo sed through parametric restrictions provide better forecasts than mode ls with little theory-imposition. For these data, a double-log demand system is a superior forecaster to the Rotterdam model, which is super ior to the FDLA/ALIDS model.