Nonlinear Forecasting Using a Large Number of Predictors

Abstract

This paper aims to introduce a nonlinear model to forecast macroeconomic time series using a large number of predictors. The technique used to summarize the predictors in a small number of variables is Principal Component Analysis (PC), while the method used to capture nonlinearity is artificial neural network, specifically Feedforward Neural Network (FNN). Commonly in principal component regression forecasts are made using linear models. However linear techniques are often misspecified providing only a poor approximation to the best possible forecast. In an effort to address this issue, the FNN-PC technique is proposed. To determine the practical usefulness of the model, several pseudo forecasting exercises on 8 series of the United States economy, grouped in real and nominal categories, are conducted. This method was used to construct the forecasts at 1-, 3-, 6-, and 12-month horizons for monthly US economic variables using 131 predictors. The empirical study shows that FNN-PC has good ability to predict the variables under study in the period before the start of the "Great Moderation", namely 1984. After 1984, FNN-PC has the same accuracy in forecasting with respect to the benchmark.

Keywords: artificial neural networks, Bayesian regularization, factor model, forecasting, principal components analysis

JEL Classification: C13, C33, C45, C53