Crisis Econometrics and High Dimensional Nonstationary Regression

金融史表明，资产价格泡沫、危机和恐慌是间歇性的，但却是全球金融市场的长期特征。2007-2009年全球金融危机和当前主权债务危机的经验强调了这些历史教训，并见证了金融不稳定在实体经济活动和就业方面的代价。有关资本要求的新监管举措和《巴塞尔协议III》（Basel III）国际金融协议强调，有必要加强对银行和金融市场的监管，以帮助避免通常伴随着资产价格泡沫的“过度信贷创造”。在这种市场监督中，一个重要的实际问题涉及评估现有证据中什么是“过度”。除非央行经济学家和监管者能够评估投机泡沫是否存在，否则他们无法消除投机泡沫。这个难题最近的一个突出例子发生在20世纪90年代互联网泡沫的早期阶段，在不到十年的时间里，创造和摧毁了8万亿美元的股东财富。1996年，艾伦•格林斯潘（Alan Greenspan）发表了著名的“非理性繁荣”（irrational exuberance）言论，表达了对可能出现的资产价格通胀的担忧，但没有任何计量经济学证据支持，也没有对市场升级或随后的崩盘产生影响。同样，21世纪初，经济学家罗伯特•希勒（Robert Shiller）对房价过度上涨的多次警告，未能让政策制定者警惕正在形成的房地产投机泡沫。计量经济学工作可以通过量化与基本面相关的市场过度行为，帮助监测金融市场。首席研究员（PI）的研究提供了金融繁荣和市场压力的预警系统。该指数的一些方法已被中央银行的研究和监测小组采用，以加强对金融资产、房地产和大宗商品价格市场的监测。这项工作涉及计量经济学分析，重点关注个人金融时间序列的揭示属性和数据中存在的繁荣。目前的NSF项目扩展了检测技术的工作，以提供可预测的日期算法，帮助监管机构在存在金融传染和危机随时间和不同市场串联风险的市场监测活动中。基于证据的预警诊断作为警报机制对市场参与者和监管机构都很有用。但是，要在金融监督和监管工作中发挥作用，计量经济学预警系统需要可靠地揭示市场中的通胀上升——低误检率以避免不必要的政策措施，高正检率以确保适当的政策实施。气泡现象的非线性结构通常会降低测试机制的判别能力。这些能量的减少使计量经济年代测定的尝试复杂化，并增加了对不受这个问题影响的新方法的需求。当前项目的挑战之一是开发和测试这些方法，以便它们可以在积极的政策环境中使用。第二个挑战在于将这些方法应用于不断膨胀的主权债务和信用违约互换市场，尤其是在欧盟外围国家。实证应用计划使用信用违约互换利差来分析欧洲债务危机，并通过金融体系和实体经济跟踪现象的迁移。除了危机计量经济学的主要研究分支外，该项目还将开发高维时间序列系统的自动化方法，允许变量之间的共同运动和联系以及数据的非平稳性。最近电子数据可用性的大量改进为统计分析提供了新的机会，包括在实际工作中使用非常高维的数据集。现代计量经济学实践现在经常遇到变量的维度可能超过样本量的系统。在这种情况下，稀疏统计方法可以用于解决自由度问题，但该方法目前仅用于具有平稳回归量的线性系统。金融和宏观经济学中的计量经济学工作通常涉及非平稳和可能共同移动的数据序列。这些特征引入了具有挑战性的非线性和识别问题，必须在使用稀疏估计技术时加以解决。这个项目的第二个分支执行这些扩展。项目中的一些相关研究将考虑大尺寸动态面板，其中序列的未知子集具有共同特征或参数，如自回归单位根。新方法将为回归中的子集分类提供一种数据确定的方法，其中在面板中某些个体的回归特征中存在一些共性。这种类型的计量经济学分类涵盖了许多感兴趣的经验例子，例如在全球和区域经济增长分析中出现的趋同俱乐部。对这些面板分类装置的研究将大大扩展现有收缩方法的范围及其在经济学中的潜在实证应用。

Financial history shows that asset price bubbles, crises and panics are intermittent but perennial characteristics of world financial markets. Experience from the Global Financial Crisis (GFC) over 2007-2009 and the present sovereign debt crisis underscores these historical lessons and bears witness to the cost of financial instability in terms of real economic activity and employment. New regulatory initiatives on capital requirements and the Basel III international financial accord have emphasized the need for improved surveillance of banks and financial markets to help avoid the "excessive credit creation" that typically accompanies asset price bubbles. An important practical issue in such market surveillance involves the assessment of what is "excessive" in the available evidence. Central bank economists and regulators cannot work to offset a speculative bubble unless they are able to assess whether one exists.A prominent recent example of this conundrum occurred during the early phase of the 1990s Internet Bubble which created and destroyed $8 trillion in shareholder wealth in less than a decade. In 1996 Alan Greenspan famously spoke of "irrational exuberance", expressing concern about possible asset price inflation but with no supporting econometric evidence and no effect on market escalation or the subsequent crash. Similarly in the early 2000s, repeated warnings of excessive rises in house prices by the economist Robert Shiller failed to alert policy makers of the emerging speculative bubble in housing.Econometric work can assist this complex exercise of monitoring financial markets by quantifying market excesses in relation to fundamentals. Research by the Principal Investigator (PI) has provided an early warning alert system of financial exuberance and market stress. Some of the PI's methods have been adopted by central bank research and surveillance teams to enhance the monitoring of financial asset, real estate, and commodity price markets. This work has involved econometric analysis that focuses on the revealed properties of individual financial time series and the presence of exuberance in the data. The current NSF project extends that work on detection techniques to provide anticipative dating algorithms that can help regulators in market-monitoring activity where there is risk of financial contagion and crisis concatenation over time and across different markets. Evidence-based warning diagnostics are useful as alert mechanisms for market participants as well as for regulators. But to be effective in financial surveillance and regulatory work, an econometric warning alert system needs to be reliable in revealing inflationary upturns in the market -- with a low false detection rate to avoid unnecessary policy measures and a high positive detection rate to assure appropriate policy implementation. The nonlinear structure of bubble phenomena typically diminishes the discriminatory power of test mechanisms. These power reductions complicate attempts at econometric dating and enhance the need for new approaches that do not suffer from this problem. One of the challenges of the current project is to develop and test such methods so that they may be used in an active policy environment. A second challenge lies in the application of these methods to the ballooning sovereign debt and credit default swap market, especially in the European Union periphery. Empirical applications are planned to analyze the European debt crisis using credit default swap spreads and to track migration of the phenomena through the financial system and real economies.In addition to the main branch of research on crisis econometrics the project will develop automated methods for systems of high dimensional time series allowing for co-movement and linkages among the variables as well as nonstationarity in the data. Massive recent improvements in the availability of electronic data offer new opportunities for statistical analysis including the use of very high dimensional datasets in practical work. Modern econometric practice now frequently encounters systems where the dimensionality of the variables may exceed the sample size. In such cases sparse statistical methods can be useful in resolving degrees of freedom problems, but that methodology is presently developed only for linear systems with stationary regressors. Econometric work in finance and macroeconomics typically involves nonstationary and potentially co-moving data series. These features introduce challenging nonlinearities and identification issues that must be addressed in the use of sparse estimation techniques. The second branch of this project pursues those extensions. Some related research in the project will consider large dimensional dynamic panels where an unknown subset of series have a common feature or parameter such as an autoregressive unit root. The new methodology will provide a data-determined approach to subset classification in regression, where there is some commonality in the regression characteristics across certain individuals in the panel. This type of econometric classification covers many empirical examples of interest such as convergence clubs that arise in global and regional economic growth analysis. Research on these panel classification devices will substantially extend the range of existing shrinkage methods and their potential empirical applications in economics.