Forecasting GDP Growth Using Artificial Neural Networks

Financial and monetary variables have long been known to contain useful leading information regarding economic activity. In this paper, the authors wish to determine whether the forecasting performance of such variables can be improved using neural network models. The main findings are that, at the 1-quarter forecasting horizon, neural networks yield no significant forecast improvements. At the 4-quarter horizon, however, the improved forecast accuracy is statistically significant. The root mean squared forecast errors of the best neural network models are about 15 to 19 per cent lower than their linear model counterparts. The improved forecast accuracy may be capturing more fundamental non-linearities between financial variables and real output growth at the longer horizon.