Topic-based summarization to objectively analyze Central Bank statements and market sentiment

The Riksbank (Sweden’s Central Bank) releases the minutes of the Monetary Policy Meetings every two months. SEB Bank, as most of the players in the market, analyzes and creates reports about the opinion of each board member of Riksbank, to know how market changes. The goal of this work was to create automatic tools to help the Research Team of SEB. Two text mining techniques have been used: Sentiment analysis, to uncover the position (Hawkish or Dovish) of each Board Member, and Summarization, to analyze the most important statements in the minutes. In particular, a human-based topic summarization algorithm is used to summarize beliefs for each board member on different topics. Moreover, an automatic topic summarization algorithm based on Latent Semantic Analysis is proposed. The topics retrieved by the Topic Model used are similar to the ones proposed by humans. Summary evaluation has been based on human judgment: if enough sentences are retrieved, most of them are relevant, but some key points could lack. Author-based and time-dependent topic models could be a good improvement of this work.