Machine Learning in Macroeconomic Modeling

该奖项将用于研究将机器学习算法纳入宏观经济模型是否能带来更好的理解现代经济的方式。该项目将首先假设个人消费者、工人和公司可以被建模为根据特定类型的机器学习算法(助推算法)做出决策。该项目试图了解机器学习算法在什么条件下可以模仿理性个体的决策过程。当这些算法在各种机构和信息假设下使用时，该团队还将分析可能的长期经济结果。因此，该项目可能开发出一种有价值的新技术，用于在整个经济系统的背景下对个人决策进行建模。它还可以帮助我们理解，越来越多地使用机器学习方法来帮助甚至取代人类的决策，可能会如何影响未来的经济结果。因此，该项目可以帮助指导提高美国经济竞争力的努力。研究团队将利用机器学习算法的核心组件之一，称为Boosting算法，从一系列基本的、可能不准确的预测规则集合中构建高度准确的预测规则。经济模型中通常的方法是假设代理人(个人、公司等)通常被赋予错误指定的型号。在这种情况下，个人或公司的决策过程通常依赖于简单但很适合的预测规则，这可能不同于真实的数据生成过程。该团队的目标是了解一名拥有有缺陷但适合的模型的代理是否可以表现得好像她知道真正的数据生成过程。该团队计划分两步实现这一目标。在项目的第一部分，团队将研究错误指定的模型下的学习动态。作为一个例子，它研究了一类新的学习模型，在这种模型中，代理必须学习增长率，而不是兴趣变量的水平。在许多宏观经济模型中，调查的主要焦点是增长率(例如通货膨胀率)，而不是变量的水平(例如价格)。假设代理人通过递归学习过程而不是理性预期来学习增长率，可能会导致更好地解释重要的宏观经济动态，如反复出现的恶性通货膨胀和股价波动。研究计划的第二步调查特定机器学习算法的动态，有两个研究目标：[1]如果一个代理被赋予错误指定的模型，决策者如何测试和建立新的模型以改进预测，以及[2]构建新模型的过程的渐近性质，特别是该代理能否在长期内模仿理性的代理。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This award funds research that will examine whether incorporating machine learning algorithms into macroeconomic models can result in better ways to understand modern economies. The project will begin by assuming that individual consumers, workers, and firms can be modeled as making their decisions according to a particular type of machine learning algorithm (a boosting algorithm). The project seeks to understand the conditions under which a machine learning algorithm can emulate the decision-making process of a rational individual. The team will also analyze likely long run economic outcomes when these algorithms are used under various institutional and informational assumptions. The project, therefore, may develop a valuable new technique for modeling individual decisions in the context of an entire economic system. It could also help us understand how future economic outcomes may be affected by the increased use of machine learning methods to aid or even substitute for human decision making. The project could therefore help guide efforts to improve the competitiveness of the US economy.The research team will exploit one of the central components of the machine learning algorithm, called the boosting algorithm, to build a highly accurate forecasting rule from a collection of rudimentary and possibly inaccurate forecasting rules. The usual approach in economic models is to assume that the agents (individuals, firms, etc.) are typically endowed with misspecified models. When this is the case, an individual or firm's decision-making process typically relies on simple, yet well fit, forecasting rules, which can differ from the true data generating process. The team aim to understand whether an agent endowed with flawed but well fit models can behave as if she knows the true data generating process. The team plans to pursue this objective in two steps. In the first part of the project, the team will examine learning dynamics under misspecified models. As an example, it examines a new class of learning models in which the agent has to learn the growth rate instead of the level of a variable of interest. In many macroeconomic models, the growth rate (e.g., inflation rate) rather than the level of a variable (e.g., price) is the main focus of the investigation. Assuming that the agent learns the growth rate through a recursive learning process rather than rational expectations may result in better explanations of important macroeconomic dynamics, such as recurrent hyperinflation and stock price volatility. The second step in the research plan investigates the dynamics of a specific machine learning algorithm with two research objectives: [1] if an agent is endowed with misspecified models, how the decision maker can test and build a new model to improve the forecast, and [2] what are the asymptotic properties of the processes of constructing new models, in particular whether the agent can emulate the rational agent in the long run.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.