Econ Papers Daily / 期刊论文
出口企业生存的动态与决定因素：一种机器学习方法
Dynamics and determinants of exporter survival: A machine learning approach
摘要
Exporting is precarious. In Australia, one-third of exporters in a given year are entrants; however, around 60 percent of these entrants exit within their first year. These firms start small, but those that survive expand rapidly and drive the country’s total trade in the subsequent years. One explanation in the literature is that exporters are hit with an ex-post shock upon entry; hence, some survive, and others do not. The paper finds that the answer is more nuanced. Using Australian administrative data, the study: i) applies conventional econometric and machine learning techniques, along with Shapley values, to predict and understand exporter survival; ii) provides a systematic comparison of the predictive performance of ML and conventional approaches. The study documents two key findings. First, exporter survival is shaped by firms’ ex-ante heterogeneity, such as export portfolio, size, and prior export experience, as well as strategic complementarities through compatriot-firm networks. Second, ML models outperform conventional approaches in out-of-sample prediction and deliver more balanced reductions in misclassification errors and their associated costs. The results shed light on the determinants of exporter survival and highlight the value of ML for solving prediction problems such as targeting of export promotion activities, which tend to be distributed ad hoc.
中文摘要
出口具有不稳定性。在澳大利亚，某一年度内三分之一的出口商为新进入者；然而，其中约60%的新进入者在第一年内即退出。这些企业起步规模较小，但存活下来的企业会迅速扩张，并在随后年份推动该国总贸易增长。文献中的一种解释是，出口商在进入时会遭受事后冲击，因此部分企业得以存活，另一些则退出。本文发现，答案更为微妙。该研究利用澳大利亚行政数据：i）运用传统计量经济学与机器学习技术，并结合Shapley值，以预测和理解出口商存活；ii）对机器学习方法与传统方法的预测表现进行系统比较。研究记录了两项关键发现。第一，出口商存活受企业事前异质性（如出口组合、规模及先前出口经验）以及通过同国企业网络形成的战略互补性所塑造。第二，机器学习模型在样本外预测中优于传统方法，并在误分类误差及其相关成本方面实现更为均衡的削减。研究结果揭示了出口商存活的决定因素，并凸显了机器学习在解决诸如出口促进活动 targeting 等预测问题上的价值，而此类活动往往以临时性方式分配。