![Sample our Behavioral Sciences journals, sign in here to start your access, latest two full volumes FREE to you for 14 days]({"type":"image" , "src" : "/sda/110801/TOC-Subject-Sample-2021-Behavioral-Sciences.jpg"})
Abstract
Relying on over 300 million news sentiment scores from a global news source covering a battery of different news topics and their sentiment, we construct a novel measure called top news sentiment based on rolling correlations with equity returns. We find an effect of top news sentiment on stock returns with robust timeframes. These findings add to the sentiment literature with the first dynamically created sentiment variable. We are able to construct profitable trading strategies based on top news sentiment. It therefore pays off for investors to be inattentive to most news topics and their sentiment.
Notes
1 We will address these studies later in detail in the section “Literature Survey.”
2 Frazzini (Citation2006), for instance, shows that the disposition effect among investors induces underreaction to news. Sinha (Citation2016) also/underreaction to news in the US stock market in general. See Nofsinger and Sias (Citation1999) for herding in trading by both institutional and individual investors. This type of herding appears to be related to stock return momentum. According to Hillert et al. (Citation2014), media makes price momentum.
3 The theory of rational inattention offers models which elicit how investors simplify, ignore and summarize available information based on their cognitive abilities. Sims (2003) provides a complete framework for this disciplined behavioral model, describing error-prone behavior.
4 Gigerenzer et al. (Citation1999) show that agents ignore part of the information (consciously or unconsciously) to be more efficient with their time allocation. In that sense, heuristics are rational since they are adapted to the immediate surroundings of the agents and her needs as well as capabilities.
5 The database comprises tick data of all news articles includes their sentiment, headline, and topic codes of each respective article.
6 See https://sircaknowledgebase.force.com/s/article/Thomson-Reuters-News-Analtyics-TRNA with a direct link to the TRNA White Paper, last accessed 4 July 2019.
7 See Appendix A1 for examples of the sort of processing that can be achieved for sentiment.
8 All of the included topics have a potential economic relevance for and/or influence on equity prices. For the relationship between commodity and equity markets, see, for example, Delatte and Lopez (Citation2013) who find a time varying and symmetric dependence between commodity and stock markets most of the time.
9 Among the largest datasets examined, Heston and Sinha’s (2017) study applied a large news sentiment dataset, comprising over 900,000 news stories. Uhl (Citation2018) considered over 8 million news articles for a study on equity index options. However, all of these studies considered static topics.
10 It is noteworthy to point out that we do not filter for relevance of the news because relevance only applies to asset names (i.e. companies), but is irrelevant and misleading when considering topics.
11 This hypothesis draws from earlier research in the sentiment space, which has identified that longer-term news sentiment momentum is predictive of stock returns. See, for instance, Heston and Sinha (Citation2017) or Uhl et al. (Citation2015).
12 These windows were chosen based on economic intuition rather than on a data mining approach.
13 We deliberately choose a longer rolling window frame as with shorter time frames a graphical representation would not really help due to the constant change in correlation regimes.
14 This selection was done based on the highest adjusted R-squared value of the regressions.
15 See Appendix A2 for the computation of the out-of-sample forecasts.
16 One of the reasons why we identify such a high and stable correlation structure could be due to the mechanical way of constructing top news sentiment. However, if none of the topical sentiments were as highly correlated as observed, the conclusions would be different since the observed correlation structure would also be different.
17 See Appendix A3 for further details and figures.
18 We assume 6 bps transaction costs for a 100% transaction in equities. The 2-day lag is realistic in practical terms because we would always calculate the news sentiment on day t with data up to t − 1, so that we could implement the strategy on t + 1 (taking always market closing prices).
19 CMAs help to take into account the rate of change of the data as opposed to the level, eradicating potential level bias in the data.
20 Note that a similar effect can also be achieved with frequency filters, such as the cumulative sum filter, as shown in Uhl et al. (Citation2015), for instance.