ABSTRACT
Recent research has found the language sentiment in financial news to be a substantial driver of prices in financial markets, though there are two diametrically opposed interpretations for this: either markets perceive news sentiment as fundamental information (thus leading to changes in the valuation of assets) or news sentiment conveys a noise signal (thus contributing to the stochastic component of prices). The opposite roles are resolved in the context of crude oil prices by decomposing price movements into two components referring to fundamental and noise trading. Contrary to theoretical arguments in prior literature, we find empirical results supporting both interpretations.
Disclosure statement
No potential conflict of interest was reported by the authors.
Notes
1 In some cases, the different relationships could theoretically be attributed to the different datasets under study. For instance, Tetlock (Citation2007) measures the effect of the general Abreast of the Market column in the Wall Street Journal on U. S. stock market returns, while Tetlock, Saar-Tsechansky, and Macskassy (Citation2008) specifically narrow the analysis to unexpected earnings of individual stocks. However, if it was true hypothetically that the different settings (stock returns vs. unexpected earnings) were responsible for the opposing interpretations (noise vs. fundamental information), then we would expect that newswires only convey fundamental information based on their nature. However, we later observe mixed findings in the sense that newswires drive both fundamental and noise trading, which is the reason why we build upon a rather broad motivation that presents both interpretations independent of the content.
2 We explicitly refrain from applying the Kalman decomposition to news sentiment. At first glance, one might be tempted to think that it separates the underlying trend of news sentiment from its news surprise component, but this assumption does not hold true when carefully reflecting the nature of newswires. Each disseminated news story is supposed to convey a novel set of information and thus encodes only the delta with regard to the known body of knowledge at that point in time. Hence, the news sentiment variable at time should be unrelated to its lags. This rules the assumption that news sentiment follows an inherent process invalid; however, it presents one of the prerequisites of the Kalman decomposition. Instead, we must regard news sentiment as a random walk, which is further supported by an analysis where we – despite the aforementioned theoretical arguments – split the news sentiment into a discrete-time process and a stochastic component, only finding that both components become orthogonal to each other and that the process component reveals no relationship with any of the price dynamics, as sentiment variable is largely absorbed by the stochastic component.
3 The Kalman filter is related to the somewhat similar but different concept of the Kalman smoother, which provides a post-processing algorithm that can be applied to an already-known time series. More specifically, the Kalman smoother estimates the values at time step based on the full dataset
. It thereby incorporates observations that occurred after the current time step
and encodes their knowledge in the estimated fundamental price. In our analysis, we apply a Kalman filter but not a Kalman smoother, as we want to avoid retrospective corrections to our predictions by the empirical observations. We thereby eliminate a potential look-ahead bias and thus study the effect of news sentiment on the estimated price components, rather than the post-observation-corrected values from the Kalman smoother.
4 We have also repeated our analysis with the non-standardized sentiment variable and observe results comparable to those of standardized sentiment.
5 Reuters Newswire Services. URL: http://agency.reuters.com/en/products-services/products/newswires.html.