Resumen de Essays on forecasting with partial least squares methods
Julieta Fuentes de Díaz
In this thesis we look at partial least squares (PLS), a technique which constructs a scheme for extracting orthogonal unobserved components based on the covariance between the predictors and the forecasting variable. Nonetheless, PLS methods are based on all variables, that is, it gives weight to all the predictors in the dataset; and then, the weight given to predictors with high predictive power is weakened. Therefore, we introduce into the economic analysis the sparse partial least squares (SPLS) approach, proposed by Chun and Keles (2010) in the context of chemometrics. In order to obtain a sparse solution, that is, to construct the factors based on a reduced number of predictors, this method includes an L1 penalty in the PLS formulation. This type of regularization allows selecting the predictors that contain relevant information and discarding those that have redundant information or negligible effect on the forecasting target variable. We extend the static implementation to a dynamic one, with the aim of considering the macroeconomic time series properties.
