Topic modeling of investment style news.

Abstract

Smart beta exchange-traded funds (ETFs) are increasingly popular investment products among institutional investors. These ETFs can be categorized into different styles depending on the systematic risk factors to which they provide exposure. Hence, the question arises whether certain topics within the news coverage of specific styles influence the investment decision and thereby fund flows towards respective smart beta ETFs. This thesis focuses on partially answering this question by identifying the major topics in investment style news and their importance measured by their frequency of occurrence.

Based on a review of topic models, which are machine learning methods to discover topics in large collections of documents, latent Dirichlet allocation (LDA) is selected to identify the topics in investment style news. Moreover, the most extensive literature survey of LDA in finance (to the best of our knowledge) is compiled in order to optimally apply this method.

Subsequently, the major topics in a unique corpus, which has never before been investigated by topic models (to the best of our knowledge), are identified by LDA. This corpus consists of 1720 articles related to small-cap investing from 9 magazines targeting institutional investors.

The 5 major topics are "equity market (economy)", "analyst research, trading and banking", "retirement planing", "indexes, ETFs and performance" and "fund management and fund launches". These topics either persist, disappear or specialize when the number of topics to identify is increased. Dominant topics of individual magazines correspond to those proposed by the corpus specialist and the short descriptions of the magazines. The dominant topic over time is "fund management and fund launches", which follows a seasonal trend characterized by lower coverage at the end of the year and higher coverage in January, thus suggesting that changes of fund management and fund launches preferentially occur at the beginning of the year.

Since the topic proportions of each article are identified, the correlation between the importance of topics over time and corresponding fund flows can be studied in future research.