EAGER: III: CIFRAM: Dynamic Identification and Interpretation of Emerging Systemic Risks Using Textual Analysis

这个项目将使用语言工具来确定金融危机，如2008年雷曼兄弟破产后的危机，如何以及为什么形成和扩大规模。通过研究通过处理提交给美国证券交易委员会(SEC)的10-K文件中的大量口头数据而收集的文本信息，主要调查人员将使用计算机科学家开发的技术来评估口头主题，并将它们与市场数据联系起来，以评估未来是否正在形成危机。所采用的技术能够识别实际风险，使监管机构和市场参与者能够在重大事态发展之前做出适当反应。在互联网上免费提供数据和计算机代码将降低未来研究人员研究这些问题的成本。主要研究人员还将在会议上介绍研究成果，提交研究成果供出版，并将与研究生合作并对其进行培训。这些材料将在课堂上教授给未来的商业领袖，MBA学生和本科生可以在课堂上公开讨论结果及其影响。这项工作还将提交给监管机构参加的会议，分享如何在潜在危机造成广泛破坏之前将其用于管理的见解。主要调查人员将使用计算语言学的方法，包括潜在狄里克莱特分配(LDA)和文件相似性分析，以确定一组在金融公司、对金融业有敞口的非金融公司，然后是经济中的所有公司中常见的口头主题。然后，调查人员将使用集群和网络方法来评估和分类经济中公司之间的商业联系，并检查它们如何随着时间的推移而演变。然后，将产生的公司相关性网络与不同时间段的市场数据进行比较，以了解股票价格在相邻时期，特别是导致重大危机的时期，如何以及为什么会出现不同的变化。语言因素将是可解释的，因此这项技术将提供一个完全自动化的描述，为什么公司在不同的时间段以不同的方式到来。这种方法将是可复制的，不会受到研究人员的偏见，使数据能够向研究人员提供有关影响市场的最突出问题的信息，即使研究人员事先不熟悉特定系统性风险事件的真正驱动因素。一旦理解了共同变动的文本驱动因素，这些因素就可以用来反向测试动态主题结构在其他系统性事件期间是如何演变的。如果成功，这项研究可以为未来潜在的危机创建一个早期预警系统，并通过在危机发生之前解决危机的驱动因素，作为一种风险管理工具，从而降低解决危机的成本。有关更多信息，请参阅项目网站：http://scholar.rhsmith.umd.edu/khanley/nsf-grant?destination=node/1084

This project will employ linguistic tools to determine how and why financial crises, such as the 2008 crisis following the Lehman Brothers bankruptcy, form and grow in magnitude. By examining textual information gleaned by processing large volumes of verbal data from 10-K filings to the Securities and Exchange Commission, the principal investigators will use techniques developed by computer scientists to assess verbal themes and link them to market data to assess whether future crises are forming. The techniques employed enable the identification of actual risks allowing regulators and market participants the ability to respond appropriately in advance of a major development. The free provision of data and computer code on the internet will lower the cost for future researchers to also examine these issues. The principal investigators will also present the research at conferences, submit the work for publication, and will work with and train graduate students. The material will be taught to future business leaders in the classroom, where MBA students and undergraduate students can openly discuss the results and their implications. The work will also be submitted to conferences attended by regulators to share insights on how they can be used to manage potential crises before they can cause extensive damage.The principal investigators will use methods from computational linguistics, including Latent Dirichlet Allocation (LDA) and document similarity analysis, to identify a set of verbal topics that are common among financial firms, non-financial firms with exposure to the finance industry, and then all firms in the economy. The investigators will then use clustering and network methods to assess and categorize the business links among firms in the economy and to examine how they evolve over time. The resulting firm-relatedness network will then be compared to market data during various time intervals to understand how and why stock prices comove differently in neighboring periods, especially periods leading up to major crises. The verbal factors will be interpretable, and hence this technique will provide a fully automated description of why firms comove in different ways in different time periods. This method will be replicable and not subjected to researcher prejudice, allowing the data to inform researchers regarding the most salient issues affecting markets, even if the researcher is ex-ante unfamiliar with the true drivers of a specific systemic risk event. Once the textual drivers of comovement are understood, these factors can be used to back-test how the dynamic topic structure evolves during other systemic events. If successful, this research could create an early warning system for potential future crises and serve as a risk management tool by addressing the drivers of crisis before they occur, thereby reducing the cost of resolution.For further information see the project web site: http://scholar.rhsmith.umd.edu/khanley/nsf-grant?destination=node/1084