YOU CAN FIND A SHORT EXCERPT FROM THE ARTICLE HERE:
We have developed a method for detecting anomalies on multidimensional time series data and combined this with clusters from news articles to automatically identify potential risks for companies. To do this, we collect data such as the number of articles, the number of positive and negative articles and the share value. These data sets are then combined using a data merge, resulting in a single time series data set. We then used this data set to perform anomaly detection in order to find relevant outliers. In parallel, we found news clusters with a series of news articles related to a real-world event. For each of these clusters, we extracted relevant keywords and compared them with a predefined list. The keywords in this predefined list are then linked to a specific risk so that we can find possible risks within the clusters. Finally, we combined the results of the two previous steps and overlaid the timestamps of the anomalies with the time span of the cluster. If a cluster containing a risk now overlaps with the timestamp of an anomaly, we conclude that the risks can be considered “real” risks. In the future, we plan to expand the approach to include a trend analysis in order to determine the probability of risks occurring. In addition, the system is being expanded to include a warning mechanism that enables potential risks to be identified at an early stage and gives the risk owner the opportunity to take countermeasures.