跨国增长实证研究的模型不确定性问题：机器学习的视角

Model Uncertainty of Cross-country Growth Empirics: Machine Learning Perspective

过去30年间跨国经济增长实证研究领域提出了近150个增长决定因素,而全球200余个国家(地区)的样本限制意味着在总结跨国增长经验时必须考虑模型不确定性问题。有别于该领域经典文献所使用的传统计量方法,本文探索了新近的机器学习方法对该问题的分析所可能有的贡献。本文从小样本、变量排序、非线性特征三个角度说明具有一定特征的机器学习方法较传统计量方法可以更有效地处理模型不确定性问题。利用标准的跨国经济增长数据集,本文考察了10种常见机器学习方法的应用表现,并与3种传统计量方法作了比较。结果显示,套袋法与随机森林法及两者的拓展均能在小样本条件下对经济增长决定因素进行有效排序,灵活捕捉数据的非线性特征,让模型不确定性问题化繁为简,得出更为清晰、稳健的结论。本文旨在说明,机器学习方法的应用有助于跨国增长经验事实的归纳与理解,对于补充传统计量方法的局限与不足具有一定的潜力。

In the past three decades,around 150 factors have been proposed in studies on economic growth.Due to the limited sample size,model uncertainty has become one of the major concerns on analyzing determinants of economic growth.Different from classic growth literature which rely on conventional econometric techniques,we explore and discuss potential contribution of machine learning on this topic.From the perspective of small sample size,variable ranking,and nonlinearity,we demonstrate that specific machine learning techniques are capable of dealing with model uncertainty.We contrast 10 machine learning techniques with 3 conventional econometric methods using standard data set in the literature.The results show that bootstrap aggregation tree and random forest are more robust on capturing the nonlinearity in the data,hence,are more effective on ranking variables in small sample size.Our study implies that the machine learning techniques can be useful alternatives on studying economic growth.