DHB: Financial Markets as an Empirical Laboratory to Study an Evolving Ecology of Human Decision Making

该项目利用来自金融市场的丰富数据集来研究人类决策的进化生态。社会主体的分散行动如何导致全球社会秩序，这种秩序又如何随着时间的推移而变化？在市场的背景下进行这项研究有两个主要原因：第一，市场在组织人类行为方面发挥着重要作用，而且为了它们自己而有趣。其次，这些丰富的金融数据集提供了一个独特的机会来寻找决策策略集合中的模式，这使得研究它们的相互作用并在一段时间的演变跨度中跟踪它们成为可能。这个项目研究的数据来自五个不同的证券交易所，具有不同的交易和价格形成规则，使得研究机构在决定行为中的作用成为可能。它们包括数百亿条详细记录，使人们能够以高度的统计精度和时间分辨率识别和区分人类行为模式。大多数数据集包括代理商的行为(买入、卖出或取消订单)以及价格和交易量。这使得研究代理人行为与市场反应之间的相互作用成为可能。其中许多银行都有标识，将每项操作与其机构和交易账户联系起来。这使得对战略行为的异质性进行研究和分类成为可能。这些数据集涵盖了长达11年的时间跨度，这使得在比我们预期战略变化的时间尺度更长的时间尺度上研究战略的演变和相互作用成为可能。调查人员还拥有公共新闻公告的记录，这使得研究信息到达和特工反应之间的关系成为可能。他们把对这些数据进行建模的方法称为经验行为建模。通过观察与兴趣现象直接相关的背景下的行为，并表征其关键规则，他们构建了简约的代理模型，其组成部分是以经验为基础的。然后，他们使用这些模型对感兴趣的经济现象做出准确的定量预测。通过在我们可以直接观察异类代理行为的环境中构建这样的模型，我们相信我们可以制作出既能对市场功能提供实用见解，又能对人类决策策略的性质和演变产生更深层次见解的模型。金融策略经历了选择、继承和创新，因此为研究文化进化提供了一个理想的(尽管承认受到限制的)环境。虽然详细的机制与生物学中的机制完全不同，但研究人员认为，这种类比仍然很强。通过与生物学家的密切合作，研究人员应该能够确定这些想法是否对社会科学具有定量预测价值。如果它们在金融市场上没有用处，因为金融市场的策略选择很强，而利他主义等复杂因素很少，它们就不太可能在其他地方有用。这个项目使用了一种新的方法，它综合了行为经济学、基于代理的建模、进化生物学、生态学和统计物理学的想法。这些方法的成功的令人信服的演示可能会对整个社会科学中的代理建模产生广泛的影响。该项目的其他影响包括教育和指导本科生、研究生和博士后研究人员。

This project takes advantage of rich data sets from financial markets to study an evolving ecology of human decision making. How do the decentralized actions of social agents result in global social order, and how does this change through time? There are two central reasons for doing this study in the context of markets: First, markets play an important role in organizing human behavior and are interesting for their own sake. Second, the richness of these financial data sets provides a unique opportunity to look for patterns in sets of decision making strategies, making it possible to study their interactions and track them across evolutionary spans of time.The data this project studies are from five different stock exchanges, with a diversity of rules for trading and price formation, making it possible to study the role of institutions in determining behavior. They include tens of billions of detailed records, making it possible to identify and differentiate patterns of human behavior with a high degree of statistical precision and temporal resolution. Most of the data sets include the actions of agents (orders to buy, sell, or cancel) as well as prices and trading volume. This makes it possible to study the interaction between agent behavior and market response. Many of them have identifiers labeling each action with its institution and trading account. This makes it possible to study and classify the heterogeneity of strategic behavior. The data sets range over long spans of time, as long as eleven years, making it possible to study the evolution and interaction of strategies on time scales that are longer than those on which we expect strategies to change. The investigators also have a record of public news announcements, making it possible to study the relationship between information arrival and agent response. They call their approach to modeling these data empirical behavioral modeling. By observing behavior in a context that is directly related to the phenomenon of interest, and characterizing its key regularities, they construct parsimonious agent models whose components are empirically grounded. They then use these models to make accurate quantitative predictions of the economic phenomena of interest. By constructing such models in a setting where we can directly observe heterogeneous agent behavior, we believe we can make models that both give practical insights into the functioning of markets and yield deeper insights into the nature and evolution of human decision making strategies. Financial strategies undergo selection, inheritance and innovation, and thus provide an ideal (though admittedly restricted) setting in which to study cultural evolution. While the detailed mechanisms are all quite different from those in biology, the investigators believe the analogy nonetheless remains strong. By working in close collaboration with biologists the investigators should be able to determine whether these ideas have quantitative predictive value for social science. If they are not useful in financial markets, where the selection of strategies is strong and complicating factors such as altruism are minimal, they are unlikely to be useful elsewhere.Broader impacts. This project uses a new approach that synthesizes ideas from behavioral economics, agent-based modeling, evolutionary biology, ecology, and statistical physics. A convincing demonstration of the success of these methods could have a broad impact on agent modeling throughout social science. Other impacts of this project include educating and mentoring undergraduates, graduate students and postdoctoral researchers.