关键词: 大维变量选择, 混频因子模型, 现时预测, 国内生产总值

Abstract: The large shock to China’ s macroeconomy caused by Covid-19 crisis has imposed tremendous challenges on nowcasting China’s Gross Domestic Product ( GDP ) accurately. In this paper, from the perspective of the Covid-19 crisis, we combine Model Confidence Set and U-MIDAS to propose a new method to select variables in the setting of mixed frequency data. This method allows us to select variables with stable predictive performance when used to forecast GDP via exploiting variables’ heterogeneous response to the shock. Then, TFA-MIDAS models with different specifications are built with those stable variables and we compare these models with other types of (T)FA-MIDAS models in the aspect of accuracy when used to forecast GDP before and after the outbreak of the epidemic. Firstly, when macroeconomy ran smoothly before the epidemic, two-factor MIDAS models with variables covering a wide range of macroeconomy have excellent predictive power; single-factor U-MIDAS models with stable variables also perform well. Secondly, when the economy continued to recover from the shock, the two-factor U-MIDAS model with some certain stable variables can continuously nowcast GDP in a much more accurate way by capturing the core dynamics of GDP. Thirdly, setting long lag orders for the predictive variables when the economic situation is stable can improve the forecast performance; while during the recovery period, shorter lag orders are preferred in order to avoid the negative impact of those outliers that occurred in the crisis on forecasting. The method proposed in this paper aims to provide some enlightenment to macroeconomic forecasting during recession and recovery period of huge exogenous shocks.

Key words: Variable Selection, Factor-MIDAS, Nowcasting, GDP