In this paper, a dynamic simultaneous-equation model is developed, based on the flexible accelerator theory of investment, for describing and forecasting U.S. government investments. The dynamic model consists of five behavioural equations and eight identities, is estimated from quarterly data during the period 1963 through 1983 by the full-information maximum-likelihood method, and is tested for its validity and forecasting ability. The direction and magnitude of the impacts of shifts in the policy and non-policy exogenous variables are identified and discussed. In particular, the model's forecasting performance is assessed against the regression model originally proposed by Lin and an optimal forecasting model. It is concluded that the dynamic model improves the reliability and accuracy of forecasting.