Forecasting Macroeconomic Time Series Using Sparse Random Forest Models

This paper analyzes the forecasting benefits of dimension reducing of predictors prior to estimating random forest (RF) model for macroeconomic time series in high-dimensional data. For that, we propose sparse RF (sRF) models based on a two-step procedure where we consider three classes of variable selection techniques with hard- and soft-thresholding methods as well as screening approaches in the first step. We evaluate the forecasting performance of the sRF models to the RF model, using a large set of macroeconomic and financial variables. The forecasting results show that the sRF models associated with the Elastic-Net method and the sure independence screening procedure based on distance correlation deliver forecasting improvements over the RF model for Inflation, Nonrevolving credit acceleration, Housing start growth and Spread or at least provide equivalent forecasting for Industrial production growth