show positive results similar to ours, reflecting improvements in news processing in
recent years.
6.9 Summary and areas for additional research
We have shown that there is exploitable alpha in news. News analytics produce results
superior to naive ‘‘buy on the good news, sell on bad’’ strategies. This is shown in both
event studies and historical portfolio simulation.
6.9.1 Directions for future research. Is this just for quants?
The effect of these stringent filters is to reduce the firehose of news events to a much
smaller flow of significant events. In this work, we have focused on quantitative
machine-driven approaches using those events as signals.
Modern visualization and information extraction tools, relatives of the Event Study
Explorer shown here, deliver this information in a way that is useful for judgment-driven
fundamental investing as well. Protoype systems based on ‘‘sentiment indices’’ calcu-
lated across countries, sectors, capitalization classes, and styles can act as intelligence
amplification (IA) tools for investors of all flavors.
In the quant sphere, this study has only scratched the surface of what can be done with
the RNSE feed. Even restricting one’s attention to ‘‘slow’’ alpha, accumulating over
days to months, the RNSE feed provides an incredibly rich array of features from which
to build detectors for longer term arbitrage opportunities.
Intentionally omitting portfolio construction techniques and ‘‘pre-news’’ analysis, the
following are among the research projects that the authors find most compelling using
RNSE output as source data:
.Sentiment surprise This study tallied days on which extreme sentiment occurred. A
more sophisticated version would be to measure ‘‘surprise’’ extreme sentiment. Re-
turning to the initial investment hypothesis that people take a long time to process
complex and ambiguous new information, ‘‘surprise’’ sentiment changes should in-
duce sufficient cognitive dissonance to justify a hunt for alpha. Looking at the strong
sentiment of a firm that is very different from other firms in its peer group (by sector,
industry, market, or even itself through time) is a natural next step.
.Topic and product codes RNSE aggregates news information from a wide variety of
sources throughout the Thomson/Reuters organization. Furthermore, it provides a
rich array of over 800 different human and machine-generated topic codes. It is a
reasonable hypothesis that people pay more or less attention to more or less popular
products, and have differing reactions to stories of differing product types.
.Machine learning from streams of headlines The headlines exposed by RNSE provide
ample opportunity for supervised or online machine learning over streams of text
using subsequent return, subsequent volatility, or subsequent analyst judgments as
sources of labels.
.Story linkage RNSE output allows one to trace the evolution of a breaking news
story as more information is learned and put out into the news stream. It seems
reasonable to try to build classifiers for news that is self-consistent with respect to
170 News and abnormal returns