Mixed-Frequency Bayesian Predictive Synthesis for Economic Nowcasting
提出一种新的混频数据动态建模框架，将不同频率数据视为潜在因子并动态综合，用于即时预测美国季度GDP，在点预测和密度预测上均优于现有方法，与专业预测者调查结果相当。
Abstract We develop a novel framework for dynamic modelling of mixed-frequency data using Bayesian predictive synthesis. The proposed framework—unlike other mixed-frequency methods—considers data reported at different frequencies as latent factors, in the form of predictive distributions, which are dynamically synthesized and updated to produce coherent forecast distributions. Time-varying biases and interdependencies between data reported at different frequencies are learnt and effectively mapped onto easily interpretable parameters with associated uncertainty. Furthermore, the proposed framework allows for flexible methodological specifications based on policy goals and utility. A macroeconomic study of nowcasting two decades of quarterly US GDP using monthly macroeconomic and financial indicators is presented. In terms of both point and density forecasts, our proposed method significantly outperforms competing methods throughout the quarter, and is competitive with the aggregate Survey of Professional Forecasters. The study further shows that incorporating information during a quarter, and sequentially updating information throughout, markedly improves the performance, while providing timely insights that are useful for decision-making.