讲座题目: Deviance Information Criterion for Model Selection: Theoretical Justification and Applications
主讲嘉宾:曾涛 副教授 (浙江大学经济学院)
讲座时间:2023年4月25日(周二)下午15:00
讲座地点:经济学院407会议室
Abstract:This paper gives a rigorous justification to the Deviance information
criterion (DIC), which has been extensively used for model selection based on MCMC output. It is shown that, when a plug-in predictive distribution is used and under a set of regularity conditions, DIC is an asymptotically unbiased estimator of the expected Kullback-Leibler divergence between the data generating process and the plug-in predictive distribution. High-order expansions to DIC and the effective number of parameters are developed, facilitating investigating the effect of the prior. DIC is used to compare alternative discrete-choice models, alternative GARCH-type models, and alternative copula models in three empirical applications.
摘要:离差信息准则(Deviance information criterion,DIC)已经被广泛应用在基于马尔可夫蒙特卡洛后验结果的模型选择之中。该论文给出了DIC严格的理论证明,指出当使用插入法预测分布的时候,在满足相关正则条件的情况下,DIC是真实数据生成过程与插入法预测分布之间KL距离的大样本无偏估计。论文推导了DIC以及有效参数个数的高阶展开形式,可以用于考察先验信息的影响。实证应用包括离散选择模型、GARCH模型和copula模型。
作者介绍:曾涛,浙江大学经济学院长聘副教授,浙江大学金融研究院研究员,浙江大学资产管理研究中心副主任,浙江大学经济学院数字经济实验室副主任,金融专硕项目主任。研究领域为贝叶斯模型选择、机器学习、量化投资等。代表性成果发表于Journal of Econometrics(4篇)、Journal of Financial Econometrics、《金融研究》等。