eThesis - Kauppakorkeakoulun sähköiset gradut

» hse_ethesis_13318.pdf   koko: 3 MB (2218459)

PURPOSE OF THE STUDY Interest towards media analytics has increased significantly by both practitioners and academia alike. The hot topic is whether or not qualitative texts contain information relevant to stock financials, and if they do, whether the impact can be used to earn abnormal returns. In order to answer this, we study the impact media factors have on financial metrics in a novel specification that combines all the major media factors in a holistic media model. To transform qualitative texts information into a "sentiment score", we develop a new methodology to estimate sentiment more accurately than currently prevailing methods.

DATA AND METHODOLOGY Our study focuses on the S&P 100 constituents between the time period of 2006 and 2011. As a source of qualitative texts, we use major news publications and earnings announcements retrieved from LexisNexis -database using a web scraper program developed for the purpose of this study. We retrieve the financials data for our study using Thomson Reuters Datastream -database.

In order to estimate investor sentiment, we employ both the customary word count, as well as our novel Linearized Phrase-Structure -methodology. For word count, we test the Harvard Psychological -dictionary and a finance-specific dictionary by Loughran and McDonald (2011). As our data is panel in nature, we analyze the correlations in our error terms in line with Petersen (2009), first without clustering and then clustering by firm and by time. We find time-effect in our error terms, and therefore employ a Fama-Macbeth (1973) methodology with clustering done in quarters. To mitigate a methodological choice driving our results, we run our specifications with a multitude of alternative specifications.

RESULTS We find that Linearized Phrase-Structure (LPS) outperforms the predominant naïve word count methodology. Also, we find that if employing word counts, researchers should employ context dependent dictionaries, such as Loughran and McDonald's (2011). In terms of our main variables, we find that the existing media factors are not mutually exclusive, and impact financial metrics in chorus. Alas, we do not find statistically significant relationship between sentiment and abnormal returns. However, we find a relationship between aggregate market news volume and abnormal returns, and also between sentiment and abnormal volatility. We infer that our findings support limited attention -theory, and provide evidence against market efficiency.