Although few investors would argue that news has value, the big questions are, how much and for how long? This study explores the extent to which media articles predict stock returns and whether news moves stock prices over the long term or only in the short term. It also seeks to discover whether positive or negative news is more predictive of returns.

Previous studies are contradictory, probably because they used different methodologies. In particular, they struggled with the enormous number of news sources now available, leading them to focus on a narrow range of sources and limited time frames.

For logistical reasons, earlier studies limited themselves to analysis of a small subsection of news flow, such as a specific newspaper column. The authors of this study use a dataset of about 900,000 news stories from the Thomson Reuters news archive, sometimes with multiple stories about a particular company.

And whereas other studies typically followed stock prices for a day or two after the news, this study tracks returns for a full quarter to determine whether the effects of news flow are fleeting or long lasting. In particular, they examine whether a week’s worth of news on a single company has greater predictive power for the shares than a single day’s coverage.

A sentiment measure, which assesses investors’ reactions to news, was developed by the authors by applying a neural network to a dataset of news stories from 2003 to 2010. This neural network, called Thomson Reuters NewScope Data, breaks sentences down into their constituent parts—nouns, adjectives, verbs, adverbs, and intensifiers. This approach enables identification of parts of speech that convey tone, the intensity of tone, and positive and negative sentiment.

News substantially affects stock prices: On the day of a news story, average excess returns are 1.99%. (Excess returns are the difference between positive-news portfolios and negative-news portfolios.) Although the short-term return is high, this and previous studies show that price spikes are usually quickly reversed.

A more relevant finding, then, is that the neural network produced returns of 0.17% on the day after the news and 0.04% the following day. The authors believe these second- and third-day returns represent permanent value. After that, excess returns essentially disappear for the rest of the quarter, implying that the news is quickly priced in.

So, returns can be predicted for just a day or two when the news flow lasts for a day. But when news flow lasts for a week, there is a dramatic increase in how long the effect on prices lasts. In the week following the announcement, the excess return is some 3.75%, and over the subsequent 13 weeks, average excess returns are more than 2%.

The neural network is also useful in revealing the effects of positive, negative, and neutral news. Companies that are the subject of neutral news outperform those that are not written about at all. This effect would seem to negate the saying that “no news is good news.”

Deep textual analysis via a neural network appears to detect news that is persistently underincorporated in stock prices. Given that stock prices predominantly react to this news only at the time of the next earnings announcement, this finding suggests that the best trading opportunities exist just before corporate announcements.

For investors with the capabilities and mandate to short stocks, a premium can be captured through buying the stocks of companies in the 24–48 hours following publication of positive stories about them and shorting those with negative news flow in that length of time.

The opportunity persists beyond the initial 24–48 hours for the stocks of companies that are the subject of news flow for a number of days. Companies with good news over a one-week period outperform companies that are the subject of a week’s worth of bad news. Portfolios formed on this basis earned excess returns for up to 13 weeks in the period studied.