Financial Time Series Summarization

The use of Machine Learning techniques for financial data analysis has become increasingly popular. The problem of explaining financial data with the help of Machine Learning and Natural Language Processing techniques has become crucial for companies that need to support business decisions and to understand machine learning results. This thesis proposes a method for summarizing financial time series based on textual protoforms, which consist of a set of informative summaries used to explain the performance of a certain stock over time. The method aims at synthesizing the key information conveyed by both historical prices and the economic stock indicators. The analyzed financial time series are represented into a unified vector space by using a popular embedding algorithm to leverage time series similarities in the generation of the output summaries. The resulting summaries can explain the performance of different stocks in different time periods, compare multiple stocks with respect to a performance measure or compare the performances of different sectors. In addition, several indicators are generated about the quality and the informativeness of the summaries. The indicators allow comparison between different summaries and gives the final user informations about the confidence of the provided statements.