时变参数模型的最优滚动窗宽选择标准及应用

宏观经济领域中存在严重的结构突变性，模型估计量的优劣对估计样本规模是敏感的。本文针对时变参数模型，建立了滚动窗宽选择标准，通过最小化估计量的近似二次损失函数及最大化各子样本估计量间的曼哈顿距离选择窗宽大小，权衡了模型估计量的准确性和时变性两个相悖目标。蒙特卡罗模拟实验表明，本文所提出的方法在各种结构突变情形下均适用，能够应用于线性关系和非线性关系的时变参数模型中，且均具有稳健性。将该方法应用于我国金融网络的结构突变识别过程，显著改善了传统窗宽选择方法的结果。

Optimal Rolling Window Selection for Time-Varying Parameter Model and Its Application

There are serious structural changes in the macroeconomic field. The performance of the model estimators is sensitive to the choice of estimation sample size, while methods to select the window size in rolling time-varying parameter model have received little attention.In this paper, a new approach is developed to select the rolling bandwidth for capturing the time-varying parameter in models with potential breaks. More specifically, the function forms are unknown, which can be set as a single-index semi-parametric model that can capture the linear or nonlinear relationship between variables, also can be extended to the linear or generalized linear regression model where only need to use the corresponding model estimation method.
Our new approach, to balance the accuracy and the time-varying objectives of the model estimators, solves the multi-objective optimization problem that minimizing bootstrap approximation quadratic loss function of model estimators and maximizing the Manhattan distance between the sub-sample estimators. Monte Carlo simulations show that using the window size selected by our procedure can significantly improve upon the performance of the model estimators. And also our method is applicable to all kinds of structural changes and time-varying parameter models of linear and nonlinear relations, not sensitive to the parameters choice in the same data generation process.When applied to capture the structural changes of China's financial network, 30 financial institutions, from 16 October, 2010 to 26 September, 2015, are included. Our results suggest that our procedure can capture the structural changes of the financial system, also significantly improve upon the performance of the financial network model estimators compared to traditional methods which just according to the subjective intention and forecasting performance. Our research and conclusions are helpful for the optimization and application of time-varying parameter model, and have important theoretical value and practical significance.