Forecasting GDP with Many Predictors Using Sparse-Group LASSO MIDAS

We conducted an investigation into four econometric models designed to handle mixed-frequency data. Our primary objective is to leverage a vast array of monthly macroeconomic variables to enhance the accuracy of forecasting quarterly Gross Domestic Product (GDP). To achieve this, we compared the following models: (1) The Autoregressive (AR) model, (2) The Mixed Data Sampling (MIDAS) model, which enables the combination of data at different frequencies, (3) The Lasso-MIDAS model, as proposed by [27], aimed at addressing issues related to inconsistent data frequencies and the curse of dimensionality arising from high-dimensional data, (4) the Sparse-group LASSO model, introduced by [3], which accommodates for the mixing processes and recognizes that the financial and the macroeconomic data may have heavier than exponential tails. An empirical analysis of forecasting GDP growth reveals that the sparse-group LASSO model consistently outperforms other models when forecasting four steps ahead, both before and after COVID-19 episodes. For short-term forecasting, both the MIDAS and sparse-group LASSO models exhibit favorable performance compared to alternative approaches. When comparing our findings before and after the COVID-19 episodes, it becomes evident that the MIDAS model significantly outperforms other models when incorporating COVID-19 data. Utilizing high-frequency data without any form of regularization appears to play a substantial role in improving forecasting performance, particularly during abrupt economic downturns. In essence, these two models can serve as an alternative “benchmark” for forecasting when sudden economic fluctuations occur, rendering conventional models like the AR model quickly outdated.