254
G. Marti et al.
– Authors in [66] notice that there is an “ecology of clusters”: They “can survive
for finite periods of time during which time they may evolve in some identifiable
way before eventually dissipating or dying”.
– In [115], the authors track the merging, splitting, birth, death, contraction, and
growth of the clusters in time.
10.5 Clusters, Hierarchies, and Networks Based on
Alternative Data
In this review, we cited a number of publications computing and studying clusters,
hierarchies and networks based on time series of returns and their correlations, i.e.
readily available and cheap data. Another avenue of research consists of using alternative data to estimate lead-lag or contemporaneous relations between companies.
10.5.1 Based on Textual Data
• In [55], authors suggest that media network based investors’ attention is a powerful predictor of market premium: In brief, the more frequent the stocks are comentioned by media news, the more non-shareholders attention is triggered, and
the higher probability of overvaluation for connected stocks. They create an attention index, by aggregating across all the stocks in the market on monthly basis,
and find that their index can forecast the market premium with a significantly
negative coefficient, a 5.97 and a 5.80% monthly in-sample and out-of-sample
R
2 respectively. In addition, they show that the findings hold when controlling for
alternative attention proxies, news-based predictors, fundamental information predictors and other standard factors. Their indicator consistently provides negative
return forecasts for both time-series and cross-sectional portfolios. As a final test,
they also provide evidence that their indicator captures investor attention by sorting cross-sectional portfolios on news co-occurrence frequencies and by checking
the performance of average correlation of Google search and Bloomberg search
frequencies.
• Authors in [49] propose to derive a company risk indicator from news-sentiment
networks. For their research, they gathered from Seeking Alpha a set of articles,
which contain a body of text and the author’s self-reported sentiment regarding
the targeted entities in the articles. The sentiment in the articles, can be seen as a
reflection of the author’s future expectations. They build co-occurrence networks
from the parsed news articles on a quarterly basis, starting from Q1 2011 until
the end of Q2 2016. These co-occurrences can cover partnerships, joint ventures,
competitors, and suppliers, but can also cover other type of relationships, such
as for example an author listing his or her opinion on current best stock picks or
G. Marti et al.
– Authors in [66] notice that there is an “ecology of clusters”: They “can survive
for finite periods of time during which time they may evolve in some identifiable
way before eventually dissipating or dying”.
– In [115], the authors track the merging, splitting, birth, death, contraction, and
growth of the clusters in time.
10.5 Clusters, Hierarchies, and Networks Based on
Alternative Data
In this review, we cited a number of publications computing and studying clusters,
hierarchies and networks based on time series of returns and their correlations, i.e.
readily available and cheap data. Another avenue of research consists of using alternative data to estimate lead-lag or contemporaneous relations between companies.
10.5.1 Based on Textual Data
• In [55], authors suggest that media network based investors’ attention is a powerful predictor of market premium: In brief, the more frequent the stocks are comentioned by media news, the more non-shareholders attention is triggered, and
the higher probability of overvaluation for connected stocks. They create an attention index, by aggregating across all the stocks in the market on monthly basis,
and find that their index can forecast the market premium with a significantly
negative coefficient, a 5.97 and a 5.80% monthly in-sample and out-of-sample
R
2 respectively. In addition, they show that the findings hold when controlling for
alternative attention proxies, news-based predictors, fundamental information predictors and other standard factors. Their indicator consistently provides negative
return forecasts for both time-series and cross-sectional portfolios. As a final test,
they also provide evidence that their indicator captures investor attention by sorting cross-sectional portfolios on news co-occurrence frequencies and by checking
the performance of average correlation of Google search and Bloomberg search
frequencies.
• Authors in [49] propose to derive a company risk indicator from news-sentiment
networks. For their research, they gathered from Seeking Alpha a set of articles,
which contain a body of text and the author’s self-reported sentiment regarding
the targeted entities in the articles. The sentiment in the articles, can be seen as a
reflection of the author’s future expectations. They build co-occurrence networks
from the parsed news articles on a quarterly basis, starting from Q1 2011 until
the end of Q2 2016. These co-occurrences can cover partnerships, joint ventures,
competitors, and suppliers, but can also cover other type of relationships, such
as for example an author listing his or her opinion on current best stock picks or