A comparison of parameter selection measures for sensor learning from financial news events
The popularization of web platforms promoted a significant increase in the publication of financial news and reports in digital media. In this sense, a multidisciplinary research area called “learning to sense” (or sensor learning) has received attention recently. Unlike traditional machine learning methods, in sensor learning there is an interest in obtaining a time series that indicates the activity of a particular topic over time. A sensor is represented by a set of parameters learned from a historical news events dataset. The sensor generates time series as news events are processed and these time series are used in decision support systems. This paper presents an overview of sensor learning for financial news. We compared six parameter selection measures for sensor learning, with the differential of considering an unsupervised scenario. The general idea is to use the concept of k-recurrent events, i.e, news events that are similar and occur together in different periods of up-trends and down-trends of a financial time series. Thus, if a specific event (extracted from news) occurred at least k times in the past always associated with up-trends, then such news is labeled as positive news. Analogously, it can be labeled as negative. The experimental results from real data provided evidence that the approach investigated in this work is a promising alternative for sensor learning from financial news events, especially in contexts where there are no domain experts or external information to label a training set.