Agent Computing and AI to Achieve the 2030 Agenda: New Methods to Infer Policy Priorities from Open Fiscal Data and Sustainable Development Indicators

我们如何在2030年前实现可持续发展目标？这是大多数国际论坛上最常出现的问题，也是世界各国政府如何制定优先政策的核心因素。然而，我们如何知道这些优先事项是否有助于实现可持续发展目标？发展中国家会重蹈采纳千年发展目标（被许多学者视为失败的议程）的覆辙吗？如何改进政府制定政策重点的方式？该项目将利用有关公共支出和发展指标的新数据、尖端机器学习技术和最先进的计算模拟方法来解决这些问题。在此过程中，它将对政府如何优先考虑政策问题产生深刻的见解，并重新绘制指导循证决策的问题和方法的格局。该项目由三个支柱组成：(1)将公共支出数据与可持续发展目标联系起来；(2)确定易受直接政策干预影响的发展指标；(3)对政策优先排序过程进行建模，以评估发展战略。第一个支柱建立在越来越多的财政数据公开运动的基础上。这个想法是通过深度学习将公共支出计划分类到可持续发展目标中。为了训练这个分类器，我将使用一个来自墨西哥的新数据集（世界上独一无二的），其中政府专家为4000个支出计划分配了可持续发展目标标签。这和Netflix用来分类电影的技术是一样的。由于聘请电影专家对每部电影进行分类在经济上是不可行的，该公司使用专家分类样本，并利用描述情节的文本来训练算法并预测标签。以类似的方式，我将利用描述墨西哥支出计划的文本，以便为未分类数据分配可持续发展目标标签。第二个支柱包括确定“工具性”发展指标。这些指标容易受到获得专门资源的具体政策的干预。例如，一项疫苗接种运动（有资源的政策）旨在改变麻疹发病率指标（指标）。有趣的是，有几个发展指标不是工具性的。例如，人均GDP是各种因素的综合衡量，没有一个政府有具体的政策直接干预它。因此，确定工具指标是了解和评价政策优先事项的关键，因为政府只将资源分配给具有政策工具的发展问题。我建议在政策制定者和专家之间进行一项在线调查，他们将被要求从随机样本中确定有用指标。在开发计划署和GIFT的支持下，这项调查将向世界各地的开发计划署工作人员和政府官员进行。通过这次调查，我希望能分类出大约100到150个发展指标。第三个支柱建立在1和2的基础上，以校准策略优先级的代理计算模型。我之前开发了一个类似的模型，并在各种出版物中验证了它，例如，通过估计政策优先级、政策弹性、政策一致性、事前政策评估和法治的有效性。我的模型的一个显著特征是，政策优先级（以跨发展指标的资源分配形式）从适应性政策制定过程中内生产生，该过程考虑了可持续发展目标之间复杂的相互联系网络（可持续发展文献中的一个中心主题）。因此，可以根据工具性指标来确定这些优先事项，并且可以对模型进行校准，以符合支柱1中估计的经验支出模式。目前还没有工具可以做到这一点。

How can we reach the Sustainable Development Goals (SDGs) by 2030? This is the most recurrent question in most international forums, and a central factor in how governments are formulating policy priorities around the world. However, how can we know if those priorities are conducive to the SDGs? Will developing countries repeat the same mistakes from adopting the Millennium Development Goals (regarded by many scholars as a failed agenda)? How can we improve the way in which governments formulate policy priorities? This project will harness novel data on public expenditure and development indicators, cutting-edge machine learning techniques, and state-of-the-art computational simulation methods to tackle these questions. In doing so, it will produce profound insights into how governments prioritise policy issues, and redraw the landscape of questions and methods that guide evidence-based policymaking.The project is structured in three pillars: (1) linking public expenditure data to SDGs, (2) identifying development indicators that are susceptible of direct policy interventions, and (3) modelling the process of policy prioritisation to assess development strategies. The first pillar builds on the growing movement of open fiscal data. The idea is to classify public expenditure programmes into the SDGs through deep learning. To train this classifier, I will employ a novel dataset from Mexico (unique in the world), in which government experts have assigned SDG labels to 4,000 expenditure programmes. This is the same technology used by Netflix to classify movies. Since hiring movie experts to categorise every movie is economically unfeasible, the company uses a sample of expert classifications, and exploits the texts describing the plots to train an algorithm and predict labels. In a similar way, I will exploit the texts describing Mexican expenditure programmes in order to assign SDG labels to unclassified data.The second pillar consists of identifying development indicators that are 'instrumental'. These are indicators susceptible of being intervened by specific policies that receive dedicated resources. For example, a vaccination campaign (a policy with resources) is designed to transform the indicator of incidence of measles (the indicator). Interestingly, there exist several development indicators that are not instrumental. For example, GDP per capita is a composite measure of various factors and no government has a specific policy to directly intervene on it. Thus, identifying instrumental indicators is key to understand and evaluate policy priorities, as governments only allocate resources to those development issues with policy instruments. I propose conducting an online survey across policymakers and experts who will be asked to identify instrumental indicators from a random sample. With the support of the UNDP and GIFT, this survey will be administered to UNDP functionaries and government officials around the world. Through this survey, I expect to classify approximately 100 to 150 development indicators.The third pillar builds on 1 and 2 in order to calibrate an agent-computing model of policy prioritisation. I have previously developed a similar model and validated it throughout various publications, for example, by estimating policy priorities, policy resilience, policy coherence, ex-ante policy evaluation, and the effectiveness of the rule of law. A distinctive feature of my model is that policy priorities (in the form of resource allocations across development indicators) emerge endogenously from an adaptive policymaking process that takes into account the complex network of interlinkages between SDGs (a central topic in the sustainability literature). Thus, these priorities can be defined over instrumental indicators, and the model can be calibrated to match the empirical expenditure patterns estimated in pillar 1. There is currently no tool that can achieve this.