Forecasting GNP with noisy data: a case study

Article type: Research Article

Authors: Howrey, E. Philip^{a; *}

Affiliations: [a] Research Seminar in Quantitative Economics, Department of Economics, The University of Michigan, Ann Arbor, MI 48109, USA

Correspondence: [*] The author is Professor of Economics at The University of Michigan. He is grateful to Saul Hymans and Janet Wolfe for helpful comments on this research.

Abstract: Forecasts of real GNP are based on preliminary estimates of National Income and Product Accounts data which will subsequently be revised. The purpose of this paper is to estimate the decrease in forecast accuracy which results from the use of preliminary rather than revised data. It is found that forecasts of the level of real GNP are very sensitive to data revision, but forecasts of the rate of growth of real GNP are relatively impervious to data revision. Over the period 1986–1991, the period during which real GNP was reported in constant 1982 dollars, the variance of forecast errors of the level of real GNP was four times what it would have been if the finally revised data had been available at the time the forecasts were made. The error variance of forecasts of the rate of growth of real GNP was only about 5 percent greater using preliminary data. These empirical results for the 1986–1991 period are consistent with theoretically determined error bounds for forecasts from trend stationary and difference stationary models with autoregressive disturbances.

Keywords: Prediction error, measurement error, trend stationary model, difference stationary model, autoregressive forecasts

DOI: 10.3233/JEM-1996-22302

Journal: Journal of Economic and Social Measurement, vol. 22, no. 3, pp. 181-200, 1996