Flexible and robust mixture models for the identification of structural shocks in financial time series

向量自回归模型的同期结构效应识别是多元时间序列分析中的一个重要问题，因为它是体现模型经济含量的结构效应。一个重要的例子是对金融市场冲击的传导和(可能)传染效应的分析。我们可以衡量资产之间的相关性，并观察到，在牛市和熊市期间，这些相关性往往有很大不同。然而，风险经理和政策制定者通常需要更多信息。也就是说，他们需要知道特定市场的冲击向其他市场传递的强度和方向，以及在繁荣和危机时期，冲击传递的模式是否相同。这些信息不能直接从关联结构中读出。本研究旨在扩展和发展新的方法，以帮助识别结构性冲击，特别是在金融数据中。众所周知，可以利用金融时间序列的典型分布特性来达到这一目的。特别是大多数金融数据的条件异方差以及显著的尖峰现象，即经验分布通常比高斯分布具有更厚的尾部和更高的峰值。这些特征可能已经足以识别结构效应。本研究项目中要开发和研究的模型是体制转换模型的一部分。这些模型在经验金融学中相当受欢迎，因为它们对所提取的制度具有良好的拟合性和经济解释力。目前，采用并扩展了现有的通过制度转换效应进行识别的方法，以实现与所研究数据的相关性质的最佳匹配。因此，构建的模型融合了厚尾创新、独立成分和马尔可夫切换GARCH效应，其中后一类模型已被证明与以更高频率衡量的金融回报特别接近，例如每日或每周。随后，通过将它们应用于金融经济学中的一系列相关问题，如冲击的传递、外汇市场的价格形成以及大宗商品市场的投机效应，说明了这些模型的有效性。

Identification of the contemporaneous structural effects in vector autoregressive models is an important issue in the analysis of multivariate time series since it is the structural effects which incarnate the economic content of a model. An important example is the analysis of transmission and (possibly) contagion effects of shocks in financial markets. We can measure the correlations between assets and observe that often these correlations are substantially different in bull and bear market periods. However, risk managers and policy makers typically need more information. Namely, they need to know how strong and in which directions shocks in specific markets are transmitted to other markets, and whether the pattern of shock transmission is the same in boom and crisis periods. This information cannot be directly read off the correlations structure.This research project aims at extending and newly developing methods which can help to identify structural shock particularly in financial data. It is known that typical distributional properties of financial time series can be exploited to reach this goal. These are, in particular, the conditional heteroskedasticity as well as the pronounced leptokurtosis of most financial data, i.e., the fact that the empirical distribution typically has thicker tails and higher peaks than the Gaussian distribution. These features may already be sufficient to identify the structural effects.The models to be developed and investigated in this research project are members of the class of regime-switching models. These models are rather popular in empirical finance due to their good fit and economic interpretability of the extracted regimes. Currently existing approaches to identification via regime-switching effects are adopted and extended in order to achieve an optimal fit to the pertinent properties of the data under study. Thus models are constructed incorporating thick-tailed innovations, independent components, and Markov-switching GARCH effects, where the latter class of models has been proven to deliver a particularly close fit to financial returns measured at higher frequencies such as daily or weekly. Subsequently, the usefulness of the models is illustrated by applying them to a set of relevant problems in financial economics, such as transmission of shocks, price formation in foreign exchange markets, and the effects of speculation in commodity markets.