Periodic Reporting for period 1 - InvestorCliques (Investor Cliques in Stock Markets)
Reporting period: 2018-05-01 to 2020-04-30
Complex networks theory during the last two decades has gained novel results and profound insights into many complex systems in engineering, social and biological and also financial fields. This project applied cross-disciplinary complex network approaches to investigate investor behaviors and explain their impact on asset price dynamics. Moreover, information diffusion in financial markets is a fundamental question, which was addressed by this research.
This research is not only academically but also practically relevant as financial supervisory bodies and regulators can benefit from the study to monitor investors in stock markets. Sometimes, investors’ behavior can be related to a (local) information leakage. Additionally, practitioners can benefit from understanding better the origins of crises and booms, derived from the actual observations of the financial agents (investors). Our research allows regulators to plan effective policies to contrast and overcome financial crises in stock markets.
The main objectives of every Work Package are given in the following:
1. Network estimation and visualization methods for investor networks
2. Clustering algorithm to detect communities of the investor networks
3. Identify the evolution of investor cliques and track the change in structure of networks
4. Analyze association between network evolution and market movement
5. Develop parallel computing solutions for the existing network models and clustering algorithms using distributed computing platform and big data technology
This research is not only academically but also practically relevant as financial supervisory bodies and regulators can benefit from the study to monitor investors in stock markets. Sometimes, investors’ behavior can be related to a (local) information leakage. Additionally, practitioners can benefit from understanding better the origins of crises and booms, derived from the actual observations of the financial agents (investors). Our research allows regulators to plan effective policies to contrast and overcome financial crises in stock markets.
The main objectives of every Work Package are given in the following:
1. Network estimation and visualization methods for investor networks
2. Clustering algorithm to detect communities of the investor networks
3. Identify the evolution of investor cliques and track the change in structure of networks
4. Analyze association between network evolution and market movement
5. Develop parallel computing solutions for the existing network models and clustering algorithms using distributed computing platform and big data technology
WP1 Network estimation and visualization:
We developed a method that can capture information flow and trading pattern in stock market. We estimated investor stock trading networks based on the Granger Causality by using a vector autoregressive model. The obtained networks are directed, and they can capture the flow of information and trading patterns across a portfolio. Directed trading networks are inferred separately for financial-insurance companies and household investors. In addition, an aggregated network is inferred, whose nodes contain the trading information of different investors based on similarities in their socioeconomic attributes. By focusing on the bow-tie structures and results of multiplex analysis of investor networks, we aimed to take the first step toward understanding the information flows and trading patterns encoded in directed investor networks at the market level. Our working paper for the first deliverable D1.1 WP1 can be found at:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3578039
Furthermore, we developed software tool to visualize and compare investor networks (D4.3):
https://sigma-networks.herokuapp.com/
Users can visualize common nodes and links of two input network. Most recently, we have included an option that can detect maximum common substructures. This can be used for other work packages.
WP2 Network dynamics:
We developed efficient methods for comparing investor networks. We used these methods to compare networks prior to and during the global financial crisis of 2007-2008 and found important empirical results. We observed the herding tendency and high synchronization in trade timing during the crisis, which is supported by the literature. These findings can be used to develop early-warning signals for crises in stock markets (D2.1).
In D2.2 we used this network comparison techniques to capture the topological changes of investor networks over time and study the joint dynamics of investor networks and stock price time series. We found that the network changes are positively correlated with volatility while negatively correlated with the returns. However, their relationships with absolute changes in stock prices are non-linear and positive. This finding can be helpful to study investor trading behaviors in different market conditions (D2.2).
Furthermore, dynamic network aspect was also addressed in D3.1. We proposed a method that can track the evolution of investor clusters and cliques over time. Our working papers for D2.1 D2.2 and D3.1 can be found at:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3575664
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3589316
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3575580
WP3 Parallel computing solutions:
We addressed the technical problem of community and clique detection in investor network. We introduced a pruning technique for investor clique enumeration based on specific economic and financial contexts. It can be incorporated into the central clique solver to reduce the computing time and improve the clique-based community detection method (D3.1). Furthermore, we presented several use cases in the study on investor behavior. Empirical findings proved that the algorithm is useful for tracking the evolution of the community structure of the networks of the household and institutional investors. Our working paper for D3.1 can be found at:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3575580
We developed a method that can capture information flow and trading pattern in stock market. We estimated investor stock trading networks based on the Granger Causality by using a vector autoregressive model. The obtained networks are directed, and they can capture the flow of information and trading patterns across a portfolio. Directed trading networks are inferred separately for financial-insurance companies and household investors. In addition, an aggregated network is inferred, whose nodes contain the trading information of different investors based on similarities in their socioeconomic attributes. By focusing on the bow-tie structures and results of multiplex analysis of investor networks, we aimed to take the first step toward understanding the information flows and trading patterns encoded in directed investor networks at the market level. Our working paper for the first deliverable D1.1 WP1 can be found at:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3578039
Furthermore, we developed software tool to visualize and compare investor networks (D4.3):
https://sigma-networks.herokuapp.com/
Users can visualize common nodes and links of two input network. Most recently, we have included an option that can detect maximum common substructures. This can be used for other work packages.
WP2 Network dynamics:
We developed efficient methods for comparing investor networks. We used these methods to compare networks prior to and during the global financial crisis of 2007-2008 and found important empirical results. We observed the herding tendency and high synchronization in trade timing during the crisis, which is supported by the literature. These findings can be used to develop early-warning signals for crises in stock markets (D2.1).
In D2.2 we used this network comparison techniques to capture the topological changes of investor networks over time and study the joint dynamics of investor networks and stock price time series. We found that the network changes are positively correlated with volatility while negatively correlated with the returns. However, their relationships with absolute changes in stock prices are non-linear and positive. This finding can be helpful to study investor trading behaviors in different market conditions (D2.2).
Furthermore, dynamic network aspect was also addressed in D3.1. We proposed a method that can track the evolution of investor clusters and cliques over time. Our working papers for D2.1 D2.2 and D3.1 can be found at:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3575664
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3589316
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3575580
WP3 Parallel computing solutions:
We addressed the technical problem of community and clique detection in investor network. We introduced a pruning technique for investor clique enumeration based on specific economic and financial contexts. It can be incorporated into the central clique solver to reduce the computing time and improve the clique-based community detection method (D3.1). Furthermore, we presented several use cases in the study on investor behavior. Empirical findings proved that the algorithm is useful for tracking the evolution of the community structure of the networks of the household and institutional investors. Our working paper for D3.1 can be found at:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3575580
Existing works do not facilitate the identification of causalities in investors’ collective behaviors. We contributed to state of the art a new method for inferring investor networks based on the linear Granger Causality under a vector autoregressive (VAR) model, which can capture the trading similarities and information flows with a time lag in an investor network. This research provides new empirical results on the flows of information among investors.
Second, investor networks are often inferred for one specific security. We contributed to state of the art a multilayer network technique that can perform a market-wide analysis across multiple securities (see D1.1 the technical report for details).
Third, the current literature has barely addressed how the structure of investor networks reflects market conditions, which we did in this work package with an extensive data set. We contributed to the current literature on investor networks a method for comparing investor networks accurately. Investor networks have different features, and methods for comparing them have never been addressed before. Many investor nodes enter and leave the system before and during the crisis, making a topological comparison a potentially better methodological choice to reveal similar trading patterns. Investor networks between pre-crisis and crisis periods have very different structures. Moreover, the herding tendency, high synchronization in trade timing, and the emergence of investor hubs are observed during the crisis (see D2.1 the technical report for details).
Fourth, we contributed to the state of the art a method to capture the correlation between topological changes and stock price fluctuation. The most related works studied associations between the dynamically changing investor stock trading networks and stock price dynamics. In these works, the correlation relationships are unclear, or more robust statistical tests are needed to conclude such relationships. Some models can address this issue but do not fit to the low-frequency data that we use in this work (see D2.2 the technical report for details).
Lastly, we improved the algorithm for clique detection and analyzed the investor cliques. Furthermore, we contribute a method to track changes of investor clusters and cliques over time (see D3.1 the technical report for details).
Second, investor networks are often inferred for one specific security. We contributed to state of the art a multilayer network technique that can perform a market-wide analysis across multiple securities (see D1.1 the technical report for details).
Third, the current literature has barely addressed how the structure of investor networks reflects market conditions, which we did in this work package with an extensive data set. We contributed to the current literature on investor networks a method for comparing investor networks accurately. Investor networks have different features, and methods for comparing them have never been addressed before. Many investor nodes enter and leave the system before and during the crisis, making a topological comparison a potentially better methodological choice to reveal similar trading patterns. Investor networks between pre-crisis and crisis periods have very different structures. Moreover, the herding tendency, high synchronization in trade timing, and the emergence of investor hubs are observed during the crisis (see D2.1 the technical report for details).
Fourth, we contributed to the state of the art a method to capture the correlation between topological changes and stock price fluctuation. The most related works studied associations between the dynamically changing investor stock trading networks and stock price dynamics. In these works, the correlation relationships are unclear, or more robust statistical tests are needed to conclude such relationships. Some models can address this issue but do not fit to the low-frequency data that we use in this work (see D2.2 the technical report for details).
Lastly, we improved the algorithm for clique detection and analyzed the investor cliques. Furthermore, we contribute a method to track changes of investor clusters and cliques over time (see D3.1 the technical report for details).