A NONLINEAR EXTENSION OF THE NBER MODEL FOR SHORT-RUN FORECASTING OF BUSINESS CYCLES

To avoid the pitfalls of the widely used NBER model, in this paper we have adopted neural networks to forecast business cycles. We find that our model has overcome some of the main deficiencies of the classical leading indicators model: first, the model was able to correctly forecast all reference points in in-sample and out-of-sample data; second, the model can forecast the future value of reference series; and third, the model has a constant forecast horizon. Sensitivity analysis suggests there are some nonlinear relationships between the reference variable and selected leading indicators. This explains why we were able to improve the forecasting performance of the original model. Copyright 2005 Economic Society of South Africa.