Parameter Proliferation in Nowcasting

This paper evaluates three approaches to address parameter proliferation issue in nowcasting: (i) variable selection using adjusted stepwise autoregressive integrated moving average with exogenous variables (AS-ARIMAX); (ii) regularization in machine learning (ML); and (iii) dimensionality reduction via principal component analysis (PCA). Utilizing 166 variables, we estimate our models from 2007Q2 to 2019Q4 using rolling-window regression, while applying these three approaches. We then conduct a pseudo out-of-sample performance comparison of various nowcasting models—including Bridge, MIDAS, U-MIDAS, dynamic factor model (DFM), and machine learning techniques including Ridge Regression, LASSO, and Elastic Net to predict China's annualized real GDP growth rate from 2020Q1 to 2023Q1. Our findings suggest that the LASSO method outperform all other models, but only when guided by economic judgment and sign restrictions in variable selection. Notably, simpler models like Bridge with AS-ARIMAX variable selection yield reliable estimates nearly comparable to those from LASSO, underscoring the importance of effective variable selection in capturing strong signals.