Time-varying model averaging for FAVAR models with smooth structural changes*
提出一种时变模型平均方法（TVMA-FAVAR），通过局部留一交叉验证准则估计最优时变权重，在平滑结构变化下提升经济和金融变量的预测精度，蒙特卡洛模拟和实证均优于现有方法。
Factor-augmented vector autoregressive (FAVAR) models serve as a powerful tool for forecasting economic and financial variables, particularly in environments with a large number of predictors. In this study, we propose a time-varying model-averaging method for FAVAR (TVMA-FAVAR) that accommodates smooth structural changes. To estimate the optimal time-varying combination weights, we develop a local leave-one-out cross-validation (LLOCV) criterion that is asymptotically unbiased for the local mean squared error (LMSE), as it excludes a term irrelevant to the weight vector. We establish the asymptotic optimality of the TVMA-FAVAR method by demonstrating that its LMSE attains an infeasible lower bound. In addition, we derive the convergence rates of the selected weights and the TVMA-FAVAR estimator. Monte Carlo simulations show that the proposed method outperforms existing popular model averaging and selection approaches. An empirical application to forecasting U.S. macroeconomic variables further illustrates the superior predictive performance of the TVMA-FAVAR method.