رویکرد مدیریت ریسک بازار روزانه بر اساس تجزیه و تحلیل مستندات

The management of financial risk is one of the most challenging tasks of financial institutions. In the last two decades, diverse quantitative models and approaches have been developed and refined to address the impact of volatile markets on business. Whereas existing approaches have intensively utilized structured data such as historical price series, little attention has been paid to unstructured (textual) data, which could be a large source of information in this context. Previous empirical research has shown that certain news stories, such as corporate disclosures, can cause abnormal price behavior subsequent to their publication. On the basis of a data set comprising such news stories as well as intraday stock prices, this paper explores the risk implications of information being newly available to market participants. After showing that such events can significantly drive stock price volatilities, this research aims at identifying among the textual data provided those disclosures that have resulted in most supranormal risk exposures. To this end, four different learners — Naïve Bayes, k-Nearest Neighbour, Neural Network, and Support Vector Machine — have been applied in order to detect patterns in the textual data that could explain increased risk exposure. Two evaluations are presented in order to assess the learning capabilities of the approach in the context of risk management. First, “classic” data mining evaluation metrics are applied and, second, a newly developed simulation-based evaluation method is presented. Evaluation results provide strong evidence that unstructured (textual) data represents a valuable source of information also for financial risk management — a domain in which, in the past, little attention has been paid to unstructured data. With regard to classification performance, it is also shown that there exist significant differences between the applied learning techniques.

Financial modeling of market risks, i.e. the management of losses due to movements in financial market prices, has been a subject of research for the last few decades. Today, traditional financial risk management tools, such as Value-at-Risk and Stress Testing, make use of quantitative data stored in structured databases [37]. While the approaches to analyze such structured data, e.g. historical price series, have been continuously improved (for an example of intraday effects see [25]) , little attention has been paid in this context to the analysis of unstructured qualitative data. Especially when assessing intraday market risk exposures resulting from market events, such as critical corporate disclosures that were not anticipated by market participants, there exists limited quantitative data (at event time) that could be analyzed in such situations. However, the disclosures contain qualitative data representing a potential source of information that is not taken into consideration by traditional risk management tools. Nonetheless, the management of intraday market risk is a challenge for different market participants engaged in frequent trading, such as high frequency traders, floor traders, and market makers [25]. Also, using quantitative intraday data to forecast intraday volatilities is still at its infancy and quite unsuitable to address event risks as described before [39]. Event risk plays an important role, especially in the case of small time horizons when relevant information about a company is newly available to the market. In this context, Campbell et al. [10] have observed that firm-level variance has more than doubled in the last three decades while market and industry variances have remained stable over that period. Therefore, the goal of this paper is to explore how existing risk management approaches can be supported by utilizing unstructured textual data. In contrast to the wide range of publications that focus on quantitative risk management approaches, we will focus on the potentials of qualitative data and the corresponding data analysis methodologies. The remainder of our paper is structured as follows: In Section 2, we provide a literature review on related work in the fields of financial risk management and on text mining in the context of financial forecasting. Furthermore, a theoretical foundation is provided. Then, Section 3 illustrates how and why we have selected our data set. In Section 4, we present a text mining approach to identify significant risk exposures that result from textual information newly available to market participants. In 5 and 6 we provide evaluations of our approach. We present both a “classic” evaluation based on traditional evaluation metrics (Section 5) and a domain-specific evaluation on the basis of a simulation (Section 6). In Section 7, we conclude with a summary and an outlook on further research.

To date, “traditional” risk management tools have largely neglected one of the largest sources of information, i.e. unstructured qualitative data. Therefore, existing risk management approaches are not able to sufficiently capture/predict extreme intraday market movements (at event time) triggered by new information released to the market. However, as the mitigation of intraday market risk is important to many market participants [25], we propose an intraday risk management approach that also makes use of qualitative, unstructured data. The associated task of text analysis has been especially challenging, as we identified a majority of those regulatory-driven corporate disclosures followed by statistically significant abnormal volatility levels. The underlying text mining approach was designed to identify those corporate disclosures that are associated with highest abnormal volatility levels. By means of an empirical study, we show that today's technology is capable of extracting valuable information from corporate disclosures for risk management purposes. Both a “classic” and a newly developed domain-specific simulation-based evaluation confirm the suitability of our approach to identify most critical, i.e. volatility-enhancing, market events. We therefore conclude that intraday market risk exposures can be discovered utilizing text mining techniques. Moreover, we show that more sophisticated classification methods such as kNN, NNet, or SVM perform better than NB. Taking into account both classification results and computational efficiency, SVM turns out to be the method of choice for this particular learning task. Future research will include the application of above intraday text mining approach to an extended data set. It is, for example, of greatest interest whether or not the proposed approach also works in times of market turmoil. Existing research on trend forecasting could provide a basis to further refine our risk management approach by estimating up and downside risk exposure by developing corresponding classifiers.