Network Structure Identification in Large-Scale Financial Systems

Project: Research

View graph of relations

Description

Advances in technology and globalization have started to reshape the trading and investment ecosystem. Financial assets of various kind are created to serve the individual needs of investors. The creation of ever new assets makes the global financial markets increasingly complex which calls for continuous efforts to investigate the linkages within the markets, namely the network structure. Newly created assets pose a problem for network analysis due to the short historical data series. An asset class strongly exposed to this problem are cryptocurrenices. Thus methods are required which are capable to handle situations of limited data availability and simultaneous high dimensionality of the system, such that one is able to gain insights about the linkages of new assets to the financial markets.In the area of large-scale network analysis, the investigation of the lead-lag time effect is often pursued with Vector AutoRegressive (VAR) models. To tackle the overparametrization problems in VAR models and to identify the network structure, frequently regularization techniques are utilized and algorithms are proposed which are tailored for the estimator under consideration. Depending on the complexity of the method and algorithm, the estimation time of the VAR model with network structure can be very slow. Bayesian methods also suffer from longer estimation time due to the underlying MCMC algorithms. Most methods additionally rely on a second step of model selection, because the suitable number of lags is usually not determined by the estimator.This project proposes an approach for the network structure identification in financial markets involving new assets, which does not rely on regularization techniques or Bayesian methods, instead utilizes linear programming and adaptive estimation techniques. In comparison to regularization techniques, this relaxes the problem of the optimal choice of the tuning parameters. The estimator also chooses the number of lags, which eliminates the need for a second step of model selection. The proposed network estimator is highly flexible and suitable for the analysis of a variety of financial networks, in particular the ones involving new assets. It is computationally efficient, allows for parallel computation and does not rely on a tailored optimization algorithm. Instead recent and future advances in linear programming can be used to continuously enhance the run-time of the estimator. In the project the theoretical properties of the estimator will be investigated and further tested in a horse-race against other approaches on synthetic and real financial datasets. Finally the approach will be made available to researchers and practitioners via a freely available software package, such that a computationally efficient tool for the investigation of the ever growing number of financial assets & networks is available to the wider community. 

Detail(s)

Project number9048225
Grant typeECS
StatusFinished
Effective start/end date1/09/2110/06/22