贝叶斯潜在阈值TVP-FAVAR模型的构建及其应用

伴随着大数据时代的到来，高维时间序列的建模愈显必要，其应用也越来越广泛，但其中存在的“高维诅咒”导致其参数估计的不确定性并降低变量预测的性能。本项目拟构建贝叶斯潜在阈值时变因子扩展的向量自回归（TVP-FAVAR）模型，利用动态因子模型对变量个数降维，借助潜在阈值模型对参数矩阵稀疏化，运用时变模型反映变量之间的时变性，并使用贝叶斯框架下的MCMC方法对复杂模型进行有效估计，进一步把它应用到GDP和CPI等变量的预测和货币政策的效应分析中。创新之处在于：方法上，本项目拟构建贝叶斯潜在阈值TVP-FAVAR模型，解决“维数灾难”问题；应用上，拟利用该模型提高变量的预测精度，增加货币政策效应分析的有效性。

Modeling high-dimensional time series for widely ranging applications is becoming more and more important in the era of big data. This project aims to develop a Bayesian latent-threshold, time-varying parametric, and factor-augmented vector autoregression (TVP-FAVAR) model for analysis and predictions of various econometric and financial variables, which provides an effective solution to tackle the complications from the “curse of dimensionality” presented in the data. Before being applied to GDP and CPI forecasting and effect analysis of various monetary policies, the proposed model: 1) makes use of dynamic factor modeling for dimension reduction, 2) resorts to latent thresholding approach to obtain a desired sparse variance-covariance matrix, 3) formulates the effect-expressing parameters to be time-varying for capturing the inherent correlation dynamics, and 4) constructs Bayesian MCMC procedures for efficient estimation of the underlying complex models. Significance of this project lies on the methodology --- an innovative Bayesian latent threshold TVP-FAVAR model; and applications --- enhanced prediction precision and improved analysis efficiency.