A0984
Title: Optimal predictor and transformation selection for macroeconomic forecasting using variable importance in random forests
Authors: Maurizio Daniele - ETH Zürich, KOF Swiss Economic Institute (Switzerland) [presenting]
Philipp Kronenberg - ETH Zurich (Switzerland)
Tim Reinicke - ETH Zurich (Switzerland)
Abstract: A novel recursive group variable importance measure is proposed in random forests (RF) to select the most relevant indicators for predicting key macroeconomic variables. In contrast to existing RF-based importance measures, the method enhances the modeling flexibility by accounting for the time series structure in economic data. In an out-of-sample forecasting experiment using a large dimensional macroeconomic dataset based on the FRED-MD database, significant improvements are illustrated in forecasting US inflation, when employing our RF-based selection approach for extracting the optimal predictors and data transformations compared to existing selection methods relying on conventional regularization techniques, e.g., the lasso and elastic net. Moreover, the findings reveal that optimal variable transformations uniformly enhance the predictive accuracy of various modeling approaches, including regularization methods, (dynamic) factor models, neural networks, and random forests. The observed forecasting improvements highlight the importance of considering alternative transformations beyond the conventional choices recommended in the FRED-MD dataset. Furthermore, theoretical insights on the RF-based selection criterion are provided in an additive model framework.