Author:

Keywords:

macro-econometrics, model selection, penalized regression, variable ranking

Abstract:

Least Angle Regression(LARS)is a variable selection method with proven performance for cross-sectional data. In this paper, it is extended to time series forecasting with many predictors. The new method builds parsimonious forecast models,taking the time series dynamics into account. It is a exible method that allows for ranking the different predictors according to their predictive content. The time series LARS shows good forecast performance, as illustrated in a simulation study and two real data applications, where it is compared with the standard LARS algorithm and forecasting using diffusion indices.