10 Correlations, Hierarchies, Networks and Clustering
255
current worst stock picks. They show that the highest quarterly risk value outputted
by their risk model is correlated to a higher chance of stock price decline up to 70
days after a quarterly risk measurement.
• Authors of [125, 126] suggest to study the co-occurrence network of European
banks mentioned in financial discussions to better understand systemic risk in the
financial system. Unlike conventional approaches that estimate interdependence
based either on non-publicly disclosed information (such as interbank asset and
liability exposures) or publicly available co-movements in market data which only
measure their dependence indirectly (co-movements may be driven by other factors) and may not be forward-looking (co-movements are estimated on past market data), their text-to-network process can be applied to publicly available data
(such as discussions in financial forums) which may be forward-looking. However, authors document some of their methodology limits: “the loose definition
of co-mention context, as an entire post, lessens reliability of relations extracted,
while yielding more relations. Provided enough source data, the context could be
narrowed down, to increase the likelihood that a relation pair is actually meaningfully related.” They find that centrality in the network can be explained by a set of
standard variables including size (measures of total assets and total deposits).
• In [133], authors aim at predicting the probability that an edge is inserted or
removed in the (correlation-based) financial network estimated on past returns. To
do so, they use both the past past structure of this financial network and also a social
network estimated from the correlations of Twitter sentiment time series provided
by PsychSignal.com on the same set of stocks. They find that considering both
networks considerably improves the predictability of future financial networks
over a naive benchmark of persistent structure. They also find that the so-called
social networks are harder to predict and much less persistent than the financial
ones.
• Following [46, 61] builds similarity networks of US listed firms from text-based
analysis of the product descriptions included in the Securities Exchange Commission (SEC) filings (Form 10-K). Authors use topological features of these networks
as input to machine learning models to predict the future success or failure of the
firms. When these features are combined with typical financial data such as the
book-to-market ratio, measures of leverage, profitability, default risk of the firms,
momentum and liquidity of their stocks, accuracy of the predictions are significantly higher.
In the next three subsections, we list a couple of papers exploring interactions of
different kinds: supply chain, payments, business partnerships, financial contracts,
and mutual ownership.
255
current worst stock picks. They show that the highest quarterly risk value outputted
by their risk model is correlated to a higher chance of stock price decline up to 70
days after a quarterly risk measurement.
• Authors of [125, 126] suggest to study the co-occurrence network of European
banks mentioned in financial discussions to better understand systemic risk in the
financial system. Unlike conventional approaches that estimate interdependence
based either on non-publicly disclosed information (such as interbank asset and
liability exposures) or publicly available co-movements in market data which only
measure their dependence indirectly (co-movements may be driven by other factors) and may not be forward-looking (co-movements are estimated on past market data), their text-to-network process can be applied to publicly available data
(such as discussions in financial forums) which may be forward-looking. However, authors document some of their methodology limits: “the loose definition
of co-mention context, as an entire post, lessens reliability of relations extracted,
while yielding more relations. Provided enough source data, the context could be
narrowed down, to increase the likelihood that a relation pair is actually meaningfully related.” They find that centrality in the network can be explained by a set of
standard variables including size (measures of total assets and total deposits).
• In [133], authors aim at predicting the probability that an edge is inserted or
removed in the (correlation-based) financial network estimated on past returns. To
do so, they use both the past past structure of this financial network and also a social
network estimated from the correlations of Twitter sentiment time series provided
by PsychSignal.com on the same set of stocks. They find that considering both
networks considerably improves the predictability of future financial networks
over a naive benchmark of persistent structure. They also find that the so-called
social networks are harder to predict and much less persistent than the financial
ones.
• Following [46, 61] builds similarity networks of US listed firms from text-based
analysis of the product descriptions included in the Securities Exchange Commission (SEC) filings (Form 10-K). Authors use topological features of these networks
as input to machine learning models to predict the future success or failure of the
firms. When these features are combined with typical financial data such as the
book-to-market ratio, measures of leverage, profitability, default risk of the firms,
momentum and liquidity of their stocks, accuracy of the predictions are significantly higher.
In the next three subsections, we list a couple of papers exploring interactions of
different kinds: supply chain, payments, business partnerships, financial contracts,
and mutual ownership.