Model averaging based on leave-subject-out cross-validation for vector autoregressions

• Published in: Journal of econometrics Vol. 209; no. 1; pp. 35 - 60
• Main Authors: Liao, Jun, Zong, Xianpeng, Zhang, Xinyu, Zou, Guohua
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.03.2019; Elsevier Sequoia S.A
• Subjects: Algebra; Asymptotic optimality; Averaging; Consistency; Cost analysis; econometric models; Econometrics; economic analysis; economic evaluation; economic theory; Leave-subject-out cross-validation; Model averaging; prediction; Regression analysis; Simulation; Validity; Vector autoregressions; C53; C52; Model averaging; Consistency; Asymptotic optimality; Vector autoregressions; Leave-subject-out cross-validation
• ISSN: 0304-4076; 1872-6895
• DOI: 10.1016/j.jeconom.2018.10.007

The vector autoregressive (VAR) model is a useful tool for economic evaluation and prediction. This paper develops a leave-subject-out cross-validation model averaging (LsoMA) method to average predictions from VAR models. The approximate unbiasedness of LsoMA and its asymptotic optimality in terms of obtaining the lowest possible quadratic errors are established. The rate of the LsoMA based weights converging to the optimal weights minimizing the expected quadratic errors is also derived. Simulation experiments show that our method is generally more efficient than the other frequently used model selection and averaging methods. Two empirical applications further illustrate that the proposed method is promising.