Partial Least Squares and Penalized Regression in Time Series for Macroeconomic Forecasting

Many studies have applied factor models to reduce the dimension of the subspace spanned by the predictors through factor analysis. This paper revisits partial least squares to investigate the forecasting performance when this reduction is related to the forecast goal. The most well-known method to estimate the common factors, called principal components, is used for comparison. This study revisits three different approaches of partial least squares to investigate whether forecasting accuracy can be improved over this widely used factor forecasting method. In addition, a regularization and variable selection method, called the elastic net, is applied to the same data from the Stock and Watson database, as another method to forecast while the dimension among the predictors is reduced. One static and one dynamic partial least squares approach show good improvements over the principal components method. The elastic net method has relatively good forecast accuracy, but fails to improve the forecast performance of principal components in most cases.