pricing of information risk, hence the companies that are not covered in the news have
superior performance. These studies, however, are more concerned with the intensity of
news coverage than news sentiment per se. The work by Paul Tetlock (Tetlock, 2007) is
probably the closest to our own. He used a quantitative content analysis program,
known as General Inquirer, to analyze correlations between sentimental values of words
(as determined by theHarvard IV-4 Psychosocial Dictionary) in theWall Street Journal
‘‘Abreast of the Market’’ column and the performance of the Dow Jones Industrial
Average. As might be expected, media pessimism is shown to predict lower returns on
the Dow, at least on the day the pessimistic news is published. Unfortunately, General
Inquirer has some major limitations. For example, it takes individual words as inputs
and not combinations of words. More intelligent software is needed to study such a
multifaceted notion as sentiment. The ideal solution would be a machine that simulates
the complex emotional reactions of investors to news stories about companies they
owned or might like to own. Counting words in aWall Street Journalcolumn is a good
start, but quite a bit more must be done in order to navigate the thorny labyrinth of
investor psychology.
The above observations seem to imply that technological developments along the lines
of artificial intelligence might be a`propos. This is precisely the approach of Dow Jones
News Analytics (DJNA). Recently, with an eye towards making news sentiment
quantifiable, Dow Jones, one of the largest news agencies in the world, formed a
partnership with RavenPack International, a company specializing in algorithmic text
analysis, machine-learning, quantum information theory, and the like. DJNA is the
fruition of this collaboration. The proprietary RavenPack software takes financial news
as inputs and assigns scores using various linguistic classification techniques, thus
quantifying sentiment according to story type or the machine-simulated reaction of
professional analysts, specifically with respect to the impact of such information on
equity prices or trading volume. The DJNA platform assigns these quantitative scores
to each news story and delivers the data to customers within milliseconds of
publication.
High-frequency trading funds have an obvious interest in products such as DJNA.
The ability to instantaneously assimilate the sentiment around a company and trade
before everyone else trades sounds like a good way to make money. But what about
long-only money managers? It seems rude to leave them out of the great news sentiment
harvest feast. How might a humble money manager, say with assets under management
somewhere below $1bn, use DJNA to add value in a concentrated portfolio of stocks?
This is the question we attempt to answer.
Our event study proceeds as follows. First, news sentiment relative to companies must
be measured in a meaningful way. The metric ofnet news sentimentis defined as our
primary measurement tool. We then attempt to discern sentiment-related events in the
life of a company that might signal a temporary misvaluation of its stock. In particular,
the event of asentiment reversalis identified via the net news sentiment metric. The idea
is that an extensive period of negative news coverage around a company causes down-
ward pressure on its stock price, and when the negativity has sufficiently subsided there
is value to be unlocked on the long side. Universes of sentiment reversals are constructed
and their performance is computed. Finally, highlighting our principal theme of real-
world investing, Monte Carlo simulations are implemented for concentrated portfolios
of sentiment reversals.
232 News and abnormal returns