Combining Text Mining and Multivariate Time Series Modelling

文本集被认为是应用经济分析的宝贵信息来源。在获取大量文件方面的最新发展，科学摘要、文章、新闻、社交媒体信息或不同机构的声明，以及为从文本中提取信息而开发的方法，增加了对这类数据的兴趣。然而，关于这些方法的性能的知识，特别是当与通常的计量经济学方法相结合时，仍然相当有限。因此，TEXTMOD项目的目标是促进方法的发展，并提高对如何将文本挖掘获得的信息纳入计量经济模型的理解。因此，重点是多变量时间序列模型。这些指标是使用模型构建的，这些模型试图在没有人为干预的情况下从大量文件中确定相关主题。基于文本的时间序列的一个例子是一个主题趋势，它描述了给定主题（例如与通货膨胀有关的主题）的重要性如何随时间变化，它可能在经济研究中引起兴趣，并可以为经典的真实的经济指标添加信息内容。虽然在过去几年中提出了大量方法来确定专题及其长期趋势，但几乎没有证据表明这些程序的统计特性、其相对性能及其与更传统的建模方法的相互作用。因此，该项目的一个中心目标是研究参数设置的敏感性，对文本样本变化的鲁棒性以及与这些算法相关的不确定性。在该项目中，将提出更多的方法来比较不同样本的专题建模结果或不同方法的结果。在另一个重要步骤中，将考虑推导主题趋势的不同方法，并最终考虑将它们纳入时间序列模型的后果，例如，广泛使用的向量自回归模型，将进行研究。将特别强调对结果的适当解释、对使用基于文本的数据得出的更多见解的评价以及对将通过联合置信带获得的估计不确定性的严格测量。这些方法将被应用于研究真实的经济指标和趋势之间的关系，发现在经济学科学语料库从波兰和德国的主题。

Collections of texts are considered as a valuable source of information for applied economic analysis. Recent developments in the access to large sets of documents, e.g., scientific abstracts, articles, news items, social media messages or statements of different institutions, and in the methods developed for extracting information from texts increase the interest in this type of data. However, the knowledge about the performance of these methods, in particular when combined with the usual econometric methods is still rather limited. Therefore, the objective of the TEXTMOD project is to contribute to the development of methods and to improve the understanding of how the information obtained from text mining can be incorporated in econometric models. Thereby, the focus is on multivariate time series models. The indicators are constructed using models which try to identify relevant themes in large collections of documents without human intervention. An example of text-based time series, which can be of interest in economic research and can add information content to classical real economic indicators, is a topic trend describing how the importance of a given topic (e.g. related to inflation) changed over time. While a substantial number of methods have been proposed over the last few years for identifying topics and their trends over time, there is little evidence on the statistical properties of these procedures, their relative performance and their interaction with more traditional modelling approaches. Consequently, a central aim of the project is to investigate sensitivity to parameter settings, robustness to variations of the textual sample and uncertainty associated with these algorithms. In the project, additional methods for comparing the results of topic modelling across samples or resulting from different methods will be proposed. In a further important step, different methods for deriving trends in topics will be considered and finally the consequences of including them in time series models, e.g., the widely used vector autoregressive model, will be studied. Special emphasis will be put on the appropriate interpretation of results, evaluation of additional insights from using text-based data and rigorous measurement of the estimation uncertainty which will be captured by means of joint confidence bands. The methods will be applied to study the relationships between real economic indicators and trends in topics found for scientific corpora in economics from Poland and Germany.