Machine Learning in International Trade Research - Evaluating the Impact of Trade Agreements

国际贸易对现代经济至关重要，世界各国政府试图通过各种干预措施来影响本国的出口和进口。鉴于通过世界贸易组织（世贸组织）进行的贸易谈判所面临的问题，各国越来越多地转向只涉及一个或少数伙伴的优惠贸易协定。与此同时，注意力已从降低进口关税转向非关税壁垒的作用，如条例和技术标准的差异。因此，现代的优惠贸易协定除了削减关税外，还在服务贸易、竞争政策、公共采购等不同领域包含了大量的条款。如何评估优惠贸易协定及其各项条款对贸易流动的影响，是国际贸易研究的一个关键问题。我们认为，机器学习文献中的方法可以帮助解决这一挑战，并且这些方法通常上级现有方法。我们使用“机器学习”一词来指代用于统计预测的算法，这些算法在可用数据的子集上进行训练，以预测可量化的结果（这里是：贸易流量）。虽然这种算法已经开始应用于经济研究，但它们还没有用于PTA的分析，也没有用于更广泛的国际经济学。首先，机器学习可以帮助评估现有方法对PTA效果的评估是否合适。这种方法通过将协定执行后观察到的贸易流量与表明在没有优惠贸易协定的情况下贸易流量会发生什么的所谓反事实结果进行比较来评价优惠贸易协定。这种反事实的说法总是基于一个特定的统计模型。目前，最常见的模型是所谓的重力方程。估计的影响当然取决于引力方程对反事实贸易流的预测程度。我们将使用机器学习来开发更灵活的预测，我们可以将重力方程的预测能力与之进行比较。机器学习还可以帮助改进PTA评估的现有方法。间接地说，基于重力方程的方法通过使用未参与优惠贸易区的国家之间贸易流量变化的平均值来构建一个反事实。除国际贸易外，在一系列情况下也采用了类似的办法。最近的方法学进展表明，这些方法可以通过应用机器学习来选择比简单平均值更复杂的控制单元组合（此处：未参加PTA的国家）来改进。尽管它们的潜力，这些技术还没有被应用在国际贸易研究，我们建议他们适应这种context.Finally，机器学习可以用来确定个别PTA条款的相对重要性。现有研究面临的主要挑战是，许多优惠贸易协定包含类似的条款，因此难以分别估计它们对贸易流量的影响。因此，研究人员通常以某种方式汇总规定，例如将其合并为大类。这就限制了决策者的相关性，因为他们需要知道是否应在优惠贸易协定中列入某项具体条款。这个问题让人想起机器学习中的“特征选择”问题，其中算法必须决定将许多潜在相关变量中的哪些变量用于预测目的。我们计划使用这些方法的一个子组，可以识别具有最大影响的变量子集（这里：条款），并准确估计其影响。总体而言，拟议的研究将加深我们对PTA如何影响贸易流的理解。这一点，以及我们计划开发的实证技术，将有助于参与PTA设计和评估的研究人员和政策制定者，并最终有助于制定更好、更循证的贸易政策。

International trade is of vital importance for modern economies, and governments around the world try to shape their countries' exports and imports through numerous interventions. Given the problems facing trade negotiations through the World Trade Organization (WTO), countries have increasingly turned to preferential trade agreements (PTAs) involving only one or a small number of partners. At the same time, attention has shifted from reductions of import tariffs to the role of non-tariff barriers such as differences in regulations and technical standards. Accordingly, modern PTAs contain a host of provisions besides tariff reductions, in areas as diverse as services trade, competition policy or public procurement.A key question in international trade research is how to estimate the effects of PTAs and their individual provisions on trade flows. We argue that methods from the machine learning literature can help address this challenge, and that such methods are often superior to existing approaches. We use the term 'machine learning' to refer to algorithms used for statistical prediction that are trained on subsets of the available data to make forecasts of quantifiable outcomes (here: trade flows). While such algorithms have started to be applied in economic research, they have not been used for the analysis of PTAs nor in international economics more generally.First, machine learning can help evaluate the suitability of existing methods for estimating PTA effects. Such methods evaluate PTAs by comparing the trade flows observed after the implementation of an agreement to a so-called counterfactual outcome that shows what would have happened to trade flows in the absence of a PTA. This counterfactual is invariably based on a specific statistical model. Currently, by far the most common model is the so-called gravity equation. The estimated effect does of course depend on how well the gravity equation predicts counterfactual trade flows. We will use machine learning to develop a more flexible forecast to which we can compare the gravity equation's predictive power.Machine learning can also help improve existing methods for PTA evaluation. Implicitly, approaches based on the gravity equation construct a counterfactual by using an average of the changes in trade flows between countries not involved in a PTA. Similar approaches have been applied in a range of contexts besides international trade. Recent methodological advances have shown how these approaches can be improved by applying machine learning to select more complex combinations of control units (here: countries not participating in a PTA) than simple averages. Despite their potential, these techniques have not been applied in international trade research, and we propose to adapt them to this context.Finally, machine learning can be used to determine the relative importance of individual PTA provisions. The key challenge existing research has faced is that many PTAs contain similar provisions, making it difficult to estimate their effect on trade flows separately. Thus, researchers usually aggregate provisions in some way, for example by combining them into broad groups. This limits the relevance to policymakers who need to know if they should include a given individual provision in a PTA. This problem is reminiscent of the issue of 'feature selection' in machine learning where algorithms must decide which of many potentially relevant variables to include for forecasting purposes. We plan to use a subgroup of these methods that allow to identify the subset of variables (here: provisions) with the largest effect and to accurately estimate their impact.Overall, the proposed research will deepen our understanding of how PTAs impact trade flows. This, and the empirical techniques we plan to develop, will help researchers and policymakers involved in the design and evaluation of PTAs and ultimately contribute to a better, more evidence-based trade policy.