 Extending our prior work on NLP Sentiment Scores with our thematic identification framework SmartBuzz, we develop trading signals based on machine identified short-term thematics found in the stock news articles and company earnings result transcripts for ASX 200 stocks.

 We build SmartBuzz factors for various trading focused thematics on topics such as Earnings & Profits, Dividends, and ESG – testing for short-term moves in stock price after the results call and news publication. We use SmartBuzz to track the emergence and persistence of these themes through time, noting an increase in sentiment towards Earnings and Dividends Thematics recently.

 What we find is that these text sources offers unique insights, information frequencies, timeliness, and challenges. We can exploit the text to trade longshort for up to 10 days after publication with an average return spread of up to 1% and more with combinations of signals and filters.

 Identifying certain peculiarities with sentiment scores between both text sources, we find that using just the ‘Pos’ or ‘Neg’ labels works well for News, but for Transcripts we find a distinct bias that needs to be neutralised.

Relevance also proves to be a useful tool to filter some results.

 Using a SmartBuzz factor built around ‘Reporting’ theme terms we tested the 10-day performance around the event dates. We found the original source of the event (i.e. the company transcripts) holds the strongest alpha, while News articles are useful with relevance filters.

 Despite the biased ‘positive outlook’ found in management earnings calls, the SmartBuzz ‘Reporting’ factor is still effective at distinguishing between subsequent outperformers and underperformers as shown in Figure 1.