Relating Interesting Quantitative Time Series Patterns with Text Events and Text Features

摘要

In many application areas, the key to successful data analysis is the integrated analysis of heterogeneous data. One example is the financial domain, where time-dependent and highly frequent quantitative data (e.g., trading volume and price information) and textual data (e.g., economic and political news reports) need to be considered jointly. Data analysis tools need to support an integrated analysis, which allows studying the relationships between textual news documents and quantitative properties of the stock market price series. In this paper, we describe a workflow and tool that allows a flexible formation of hypotheses about text features and their combinations, which reflect quantitative phenomena observed in stock data. To support such an analysis, we combine the analysis steps of frequent quantitative and text-oriented data using an existing a-priori method. First, based on heuristics we extract interesting intervals and patterns in large time series data. The visual analysis supports the analyst in exploring parameter combinations and their results. The identified time series patterns are then input for the second analysis step, in which all identified intervals of interest are analyzed for frequent patterns co-occurring with financial news. An a-priori method supports the discovery of such sequential temporal patterns. Then, various text features like the degree of sentence nesting, noun phrase complexity, the vocabulary richness, etc. are extracted from the news to obtain meta patterns. Meta patterns are defined by a specific combination of text features which significantly differ from the text features of the remaining news data. Our approach combines a portfolio of visualization and analysis techniques, including time-, cluster- and sequence visualization and analysis functionality. We provide two case studies, showing the effectiveness of our combined quantitative and textual analysis work flow. The workflow can also be generalized to other application domains such as data analysis of smart grids, cyber physical systems or the security of critical infrastructure, where the data consists of a combination of quantitative and textual time series data.