10 Correlations, Hierarchies, Networks and Clustering
267
• During a crisis period, is there an increase or decrease of clusters stability? Most
papers find a decrease (e.g. [78, 102]), but at least one [75] (using an alternative
clustering methodology, the p-median problem) advocates for an increase.
• In many papers, researchers observe that Minimum Variance portfolios tend to
select the same assets as Network-based portfolios [108, 116, 119]; In [64], authors
show that in general there are no relations between the two methods, suggesting
any empirical relations observed come from special properties of the financial
correlations. Which ones? This question is left open to these days.
10.8 Conclusion
Correlation networks (and complex networks in general) provide a useful set of quantitative tools for several tasks in finance, e.g. monitoring systemic risk and acting
preemptively on a few institutions to prevent cascading effects; building diversified
portfolios of assets or strategies; designing statistical arbitrage strategies; residualizing returns by removing the main hierarchical risk factors; finding less crowded
cross-sectional risk premia.
However, in our opinion, researchers in this field need to overcome several challenges so that results and techniques exposed in this review can become part of the
standard toolbox of the practitioner. We can observe that there is i) a lack of reproducibility for a couple of studies (hence some contradictory claims); ii) difficulty
to compare methods due to re-implementation biases and lack of open source code;
and iii) no widely accepted common tasks or benchmarks; iv) data is not shared and
most of the time confidential.
To tackle these issues, we recommend researchers to provide code and data (or at
least synthetic datasets). When the underlying data is confidential, the use of Generative Adversarial Networks (GANs) or other generative models could be a solution to
i) share anonymized synthetic data having properties similar to the original dataset;
ii) define a common task and iii) set benchmark results enticing other researchers
to challenge. This approach has hugely benefited the progress of machine learning
as a research field. We think it should help researchers in correlation networks (and
complex networks in general) having more impact as well.
Acknowledgements Many thanks to the researchers which have contributed to improve this review
(in chronological order): David Matesanz, Tiziana Di Matteo, Diego Garlaschelli, Damiano Brigo,
Fabrizio Lillo, Quang Nguyen, Thomas Guhr.
Disclaimer: Views and opinions expressed are those of the authors and do not necessarily represent
official positions of their respective companies.
267
• During a crisis period, is there an increase or decrease of clusters stability? Most
papers find a decrease (e.g. [78, 102]), but at least one [75] (using an alternative
clustering methodology, the p-median problem) advocates for an increase.
• In many papers, researchers observe that Minimum Variance portfolios tend to
select the same assets as Network-based portfolios [108, 116, 119]; In [64], authors
show that in general there are no relations between the two methods, suggesting
any empirical relations observed come from special properties of the financial
correlations. Which ones? This question is left open to these days.
10.8 Conclusion
Correlation networks (and complex networks in general) provide a useful set of quantitative tools for several tasks in finance, e.g. monitoring systemic risk and acting
preemptively on a few institutions to prevent cascading effects; building diversified
portfolios of assets or strategies; designing statistical arbitrage strategies; residualizing returns by removing the main hierarchical risk factors; finding less crowded
cross-sectional risk premia.
However, in our opinion, researchers in this field need to overcome several challenges so that results and techniques exposed in this review can become part of the
standard toolbox of the practitioner. We can observe that there is i) a lack of reproducibility for a couple of studies (hence some contradictory claims); ii) difficulty
to compare methods due to re-implementation biases and lack of open source code;
and iii) no widely accepted common tasks or benchmarks; iv) data is not shared and
most of the time confidential.
To tackle these issues, we recommend researchers to provide code and data (or at
least synthetic datasets). When the underlying data is confidential, the use of Generative Adversarial Networks (GANs) or other generative models could be a solution to
i) share anonymized synthetic data having properties similar to the original dataset;
ii) define a common task and iii) set benchmark results enticing other researchers
to challenge. This approach has hugely benefited the progress of machine learning
as a research field. We think it should help researchers in correlation networks (and
complex networks in general) having more impact as well.
Acknowledgements Many thanks to the researchers which have contributed to improve this review
(in chronological order): David Matesanz, Tiziana Di Matteo, Diego Garlaschelli, Damiano Brigo,
Fabrizio Lillo, Quang Nguyen, Thomas Guhr.
Disclaimer: Views and opinions expressed are those of the authors and do not necessarily represent
official positions of their respective companies.