一种基于LASSO的多变量混频GARCH模型设计与优化算法研究  

Research on a Design for A Multivariate GARCH-MIDAS Model Based on LASSO and Its Optimization Algorithm

作　　者：方彤 苏治 Fang Tong;Su Zhi(School of Economics,Shandong University;School of Statistics and Mathematics,Central University of Finance and Economics)

机构地区：[1]山东大学经济学院 [2]中央财经大学统计与数学学院

出　　处：《数量经济技术经济研究》2021年第12期146-163,共18页Journal of Quantitative & Technological Economics

基　　金：国家社科基金重大项目(15ZDC024);国家自然科学基金面上项目(71871234);山东省自然科学基金青年项目(ZR2020QG034)的资助。

摘　　要：研究目标:在多变量混频GARCH模型中实现低频变量选择。研究方法:构建基于线性波动率预测模型的混频GARCH模型长期波动成分,在对数似然函数中引入LASSO惩罚项,使用多种优化算法实现惩罚似然函数估计并对算法效率进行模拟和评估。研究发现:无权重函数约束的多变量混频GARCH模型在变量选择上较好克服了变量非标准化、参数跳跃和权重参数不可识别问题;单纯形法在较小样本量和较少变量数量时提升了变量选择能力和精度,是BFGS的最优替代算法。研究创新:构建并扩展了多变量混频GARCH模型变量选择,给出性质证明和数值模拟,评估并得出模型估计的算法应用条件。研究价值:给出新的多变量混频GARCH模型变量选择设定,为研究金融市场波动预测性提供新思路。Research Objectives:We realized variable selection in GARCH-MIDAS models.Research Methods:We employ a linear predictive specification in the long-term volatility component of a GARCH-MIDAS model,and include the LASSO penalty function in the natural logarithm of likelihood function.We then use several optimization algorithms to estimate the model,and simulate and evaluate the efficiency of algorithms.Research Findings:We find that the multivariate GARCH-MIDAS model with no weighting function restrictions overcomes the non-standardization of variables,coefficients jumps and non-identification issue of weighting parameters;Nelder-Mead algorithm improves the ability and accuracy of variable selection when the number of observations and variables is small,which makes it a best substitutional algorithm.Research Innovations:We extend the GARCH-MIDAS model with variable selection,provide properties proves and simulations,and evaluate the conditions of algorithms for estimations.Research Value:We provide a new design for multivariate GARCH-MIDAS model and provide new perspectives for financial volatility predictions.

关 键 词：混频GARCH 变量选择 权重函数 LASSO 优化算法 

分 类 号：F064.1[经济管理—政治经济学]