(PDF) Recent advances in modelling systemic risk using network analysis-ECB

R E C E N T A DVA N C E S I N M O D E L L I N G S Y S T E M I C R I S K U S I N G N E T WO R K A N A LY S I S J anuary 2 010 E U RO P E A N C E N T R A L B A N K  RECENT ADVANCES IN MODELLING SYSTEMIC RISK USING NETWORK ANALYSIS JANUARY 2010 In 2010 all ECB publications feature a motif taken from the €500 banknote. © European Central Bank 2010 Address Kaiserstrasse 29 60311 Frankfurt am Main, Germany Postal address Postfach 16 03 19 60066 Frankfurt am Main Germany Telephone +49 69 1344 0 Website http://www.ecb.europa.eu Fax +49 69 1344 6000 All rights reserved. Reproduction for educational and non-commercial purposes is permitted provided that the source is acknowledged. ISBN 978-92-899-0611-1 (online) CONTENTS PREFACE 4 INTRODUCTION 5 DETAILED SUMMARY OF THE THEMES 9 SESSIONS I Analysis of network topology, recent advances and applications 9 II Interdependencies among institutions, sectors and systems 16 III Interbank credit, markets and liquidity management in large value payment systems 21 IV System-level liquidity effects and networks in early warning models 25 ECB Recent advances in modelling systemic risk using network analysis January 2010 3 PREFACE In October 2009, the ECB hosted a one-day Against this background, and in light of the workshop called “Recent advances in modelling recent institutional reforms concerning the global systemic risk using network analysis”, which and European institutions for macroprudential gathered together experts from central banks and supervision, improving the analytical capacity from international organisations working in the of central banks and international organisations fields of financial stability and payment system entrusted with responsibilities in the areas of analysis/oversight. The aim of the workshop was financial stability and payment system oversight to exchange views and experiences in modelling has become paramount. The presentations and and analysing systemic risk in different kinds of discussions in the workshop that are summarised networks that are relevant to financial stability in this paper provide one contribution to that end. and payment systems. The workshop also aimed They highlight the potential of network theory to improve the awareness of network modelling to enhance the tools for market infrastructure in general, to enhance knowledge about the oversight, counterparty risk management and possibilities and limitations of this field of macroprudential analysis and propose several analysis, and to exchange experiences in the use avenues for future research. and formatting of data, computing techniques and analysis of results that have been obtained Any opinions expressed by the presenters, in various institutions. discussants, or chairpersons of sessions that are quoted in this paper are their own and do not The global financial crisis that erupted in necessarily reflect the views and opinions of August 2007 clearly illustrated the role of their respective institutions. financial linkages as a channel for propagation of shocks. Indeed, the spreading of the financial turmoil from the US sub-prime mortgage market via the securitisation instruments to the banks’ off-balance-sheet vehicles and further to the banks’ balance sheets and to other financial and non-financial sectors exposed unforeseen counterparty linkages and eroded confidence in a way which further amplified the effect of the initial shocks. Research in the area of financial network analysis has shown that modelling the interlinking exposures either between financial institutions, among the sectors of the economy or across entire national financial systems, can assist in detecting important shock transmission mechanisms. Simulation exercises using these networks may then reveal that parts of the systems that might not be considered vulnerable to given adverse scenarios could still be affected due to their close interconnection with agents that are directly confronted with the unforeseen events. Policy recommendations could then be targeted towards structural changes that mitigate the adverse consequences that may emerge in closely intertwined systems in times of crisis. ECB Recent advances in modelling systemic risk using network analysis 4 January 2010 INTRODUCTION INTRODUCTION Introductory remarks by Gertrude Tumpel- And can network methodologies provide us with Gugerell, Member of the Executive Board of a useful tool in this respect? the ECB With these questions in mind, I have structured Ladies and Gentlemen, my introductory remarks into three parts. I will first give a short assessment of the relevance of I would like to welcome you to the workshop systemic risk in the modern financial system. on “Recent advances in modelling systemic Then I will discuss the use of network theories risk using network analysis” here at the ECB. for the analysis of systemic risk. Finally, I will A workshop on systemic risk that provides an briefly refer to network applications to payment analytical focus on the financial sector as a and financial systems. network of financial agents could not come at a more timely moment. SYSTEMIC RISK IN THE MODERN FINANCIAL SYSTEM In 1896 the German sociologist Georg Simmel Systemic risk refers to the possibility that a stated in his book “The Philosophy of Money”: triggering event such as a bank failure or a market “money is the spider that spins society’s web”. disruption could cause widespread disruption With this, Simmel already at the time pointed of the financial system, including significant to the network aspect of money, how financial difficulties in otherwise viable institutions or innovation can transform the economy and markets. Preventing these negative externalities society; and the transformation process as from impairing the functioning of the system changes in the complexity, size and nature of and from spilling over to the real economy is a economic and societal networks. crucial element of the mission of central banks and of supervisory authorities. The recent financial crisis has strikingly illustrated the interconnectedness that In the last two years, the functioning of the characterises the global financial system. global financial system has been challenged by In providing a framework for strengthening an extraordinary sequence of such triggering financial stability, policy makers are currently events. This brought to the fore how complex not only refining the regulatory and institutional and interconnected the financial system had set-up, but also looking for new analytical become and, consequently, how problems in tools that help to better identify, monitor and one part of the system could reach other parts, address sources of systemic risk. Therefore, also very distant ones. I believe network analysis can make a relevant contribution and I am delighted that you have In July and August 2007 the asset-backed come together today to present and discuss new commercial paper (ABCP) market collapsed work in this field. when investors realised that money market mutual funds had invested in paper backed by Let me give you three questions (from the sub-prime assets. Investors became suddenly perspective of a policy maker) which today’s distrustful of all forms of private credit, workshop would ideally shed light on: especially structured products and other complex and opaque instruments, and this What are the key channels and systemically caused the funding for structured investment important players that need special attention? vehicles and special-purpose vehicles to dry up. Difficulties faced by conduits and other How can macro-prudential supervision take the asset-backed programmes in rolling over their interconnectedness into account? short-term funding forced them to look to bank ECB Recent advances in modelling systemic risk using network analysis January 2010 5 sponsors for liquidity (this was the case, for From a micro-prudential perspective, a instance, for IKB and Sachsen LB in Germany) strengthened supervision of individual or to sell assets. A crisis of confidence ensued institutions’ risk-taking incentives is also which gripped money market mutual funds and important. A key element of the risk management the commercial paper market, notwithstanding framework of banks is that they take into their distance from the US housing market. account, in terms of credit and liquidity risks, the exposure they have to particular (potentially Such unstable dynamics, set off by increasing systemically relevant) counterparties. Systemic uncertainty about the size of losses in the risk is, in principle, outside the control of each system and, maybe more importantly, about individual institution. But, by keeping liquidity their exact location, continued in the course of buffers and capital reserves and by limiting large 2008. Then, the collapse of Lehman Brothers exposures and addressing dependencies, banks in September 2008 transformed a pessimistic can contribute to an increase in the resilience of and disoriented mood into full-blown panic the system as a whole. and paralysis.1 THE USE OF NETWORK THEORIES The biggest negative surprise following FOR THE ANALYSIS OF SYSTEMIC RISK Lehman Brothers’ default was its effect on The financial crisis has reminded us how money market funds. When one fund, Reserve important it is to look at the links and Primary, “broke the buck” (that is, the value connections of the financial system. We saw of investors’ money fell below the notional that major disruptions such as failure or a near amount invested), the sector was hit by a wave failure of certain institutions rapidly spilled over of redemptions that fuelled instability in the to the whole financial system. credit markets. Again, banks and companies relying on short-term funding through Therefore, network theory can help us to commercial paper or ABCP (i.e. debt backed analyze the systemic risk of such disruptions by mortgages, credit cards and other consumer (i) by looking at how resilient the system is to loans) could not roll over their debt, except at contagion; and (ii) what the major triggers and overnight maturities. channels of contagion are. The ensuing dynamics in market participants’ An important aspect of the analysis of systemic behaviour clearly illustrate the presence of risk is that an apparently robust system may in knock-on effects, negative externalities, and fact be very fragile. This comes from the fact a coordination failure in the market network. that a high number of interconnections within Each institution responded rationally given the network will serve as shock-amplifiers rather individually available information. However, than as absorbers. each rational response had repercussions for the whole system. Another key aspect of the analysis is that within the network of the financial system, The impact of systemic risk depends very there are players with only a few connections, much on the collective behaviour of financial but also players that are highly connected. institutions and their interconnectedness, as well Obviously, such networks are extremely as on the interaction between financial markets vulnerable if those highly connected players and the macroeconomy. Systemic stability is are disrupted. In fact, when a shock hits the a public good. The recognition of this public system, the number of affected participants can good property underpins the recent emphasis on a macro-prudential approach to regulation and 1 G. Tett (2009), “Markets 12 months after Lehman collapse”, supervision. Financial Times, 9 September. ECB Recent advances in modelling systemic risk using network analysis 6 January 2010 INTRODUCTION be especially low, but the shock may still and systemic connections in many different propagate system-wide. Payment systems, for segments of the financial markets, ranging from instance, are networks with such a property.2 money markets to networks of credit default swaps (CDSs), and from large-value payment Clearly, large and highly connected financial systems to cross-sector exposures in the euro institutions are systemically important. This area financial system. has important implications for macro-prudential surveillance, and hence for financial stability. We see that this research gives important insights Network analysis is crucial for the identification into the various amplification mechanisms of such systemically important institutions and in the global web of financial connections. markets which are critical players in the web Such amplification very much depends on a of exposures. Monoline insurance providers number of factors, such as the size of aggregate and AIG provided an example of such critical macroeconomic shocks, asset price volatility, institutions; key custodian banks or large liquidity risk and financial leverage. Moreover, correspondent banks play a similar role. network analysis can be used to simulate the effect of credit and funding shocks on banking Let me add to this, that a particular institution and financial stability by taking into account – might not only be critical to the functioning beyond the direct balance sheet exposures – also of financial markets or market infrastructures the impact of contingent claims and credit risk because other institutions are financially transfer techniques. exposed to it, but also because other market participants rely on the continued provision I am glad that the workshop brings together a of its services. For us as policy makers this is wide variety of applications. It demonstrates two a crucial point, as the impact of a failure of a key points: first, network analysis is advancing given market player also hinges on the ability as a common tool for assessing dynamics of the financial infrastructure to support its within the various parts of the financial sector resolution and to facilitate the orderly unwinding (from payment systems to interbank balance of positions. So let me now turn to the specific sheet exposures); and second, it reveals that application of network theory to payment and a truly systemic perspective needs to combine financial systems. the focus on various parts of the financial sector with an analysis of the interlinkages among NETWORK ANALYSIS APPLICATIONS TO PAYMENT them, ideally including the interaction with the AND FINANCIAL SYSTEMS real economy. This is, of course, an ambitious Research in network theory has received objective that calls for further research. relatively little attention in economics until the last decade. Therefore, I am delighted to see that CONCLUSIONS this literature is growing and today’s workshop Let me conclude. The recent financial crisis has clearly illustrates its growing importance. underscored the need for policy makers and regulators worldwide to track systemic linkages. The papers from today’s program highlight how direct and indirect interlinkages and contagion dynamics among financial institutions, as well as 2 See M. Pröpper et al. (2008), “Towards a network description of interbank payment flows”, DNB Working Paper No. 177, for among institutions, markets and infrastructures, an analysis of Dutch payment flows; C. Puhr and S. W. Schmitz can be significantly influenced by three (2009), “Structure and stability in payment networks – a panel data analysis of ARTIS simulations” in H. Leinonen (ed.), important network characteristics: First, the Simulation analyses and stress testing of payment networks, degree of connectivity, second, the degree of Bank of Finland, for the Austrian large-value payment system; concentration and third, the size of exposures. and K. Soramäki et al. (2007), “The topology of interbank payment flows”, Physica A, Vol. 379, pp. 317-333, for an We see from the papers that network analysis analysis of Fedwire, the large-value payment system operated by can help to better understand the interlinkages the Federal Reserve. ECB Recent advances in modelling systemic risk using network analysis January 2010 7 Network analysis offers a very relevant significant reduction in counterparty risk, hence tool for addressing this challenge. Its focus addressing some of the negative externalities on interconnectedness and on systemically that stem from the over-the-counter network important market players makes it especially that has formed over the years.4 relevant for the assessment of the fragility or resilience of the financial system as a whole. Interlinkages within the financial system By applying network theories we can benefit are nothing fundamentally new. However, from the important progress made in other business strategies developed by financial sciences to monitor and assess systemic risks, institutions over the last 20 years and financial direct and indirect linkages, vulnerabilities and innovations have made the system much more contagion. This is because networks allow us to interconnected, complex and opaque than it was look beyond the immediate “point of impact” of in the past. a shock and, hence, also to the spillovers likely to arise from interlinkages in the system. Thus, I believe that policy makers and regulators of network analysis can undoubtedly provide today will be judged in the future on the basis useful guidance for the analysis of systemic risk of the regulatory measures and analytical tools and can be a key tool for the future analysis of they have applied to address the root causes such risk. of the crisis. A key challenge is to transcend a purely national or sector-specific perspective For us, such analysis will be of crucial and to take an approach that matches the global importance. As you know a European Systemic nature of financial networks. A key prerequisite Risk Board will be established with the mandate for network analysis as a surveillance tool to map financial risks and their concentration at remains, however, the availability of relevant the system level for the macro-prudential data. This holds true especially on a cross-border supervision of systemic stability. The mandates basis, but also at bank level. Going forward, of other supranational institutions and fora, such regulators and overseers should continue to as the IMF and the Financial Stability Board, develop ways to systematically collect and also refer to network aspects of the financial analyse data. The crisis has clearly demonstrated system that have become apparent during the that data confidentiality must not stand in the current crisis and that should be taken into way of improvements in systemic risk analysis account in order to obtain new measures of and assessment by policy makers. financial fragility.3 Once more, I welcome you to this workshop Also for the specific field of market and I wish you productive and enriching infrastructures the relevance of network effects discussions on this very relevant topic. is being taken into account. The market for credit default swaps (CDS) has clearly revealed its systemic importance, as the default of one major counterparty has put the whole system under severe strain. Therefore, I welcome very much that central counterparties for credit default swaps have been established to address first, the high degree of interconnectivity between CDS markets and credit and cash securities markets, 3 See IMF (2009), “Global Financial Stability Report”, Chapter II second, the high leverage embedded in these on Assessing the Systemic Implications of Financial Linkages, financial instruments, and third, the significant April, and E. Nier et al. (2007), “Network models and financial concentration of related risks in a small group of stability”, Journal of Economic Dynamics and Control, Vol. 31, pp. 2033-2060. major market players. Effective implementation 4 See also ECB (2009), “OTC derivatives and post-trading of central clearing of derivatives enables a infrastructures”, September. ECB Recent advances in modelling systemic risk using network analysis 8 January 2010 DETAILED SUMMARY OF THE THEMES DETAILED SUMMARY OF THE THEMES SESSION I – ANALYSIS OF NETWORK TOPOLOGY, are broadly defined as collections of nodes RECENT ADVANCES AND APPLICATIONS (banks) and links (in the form of credit and financial relationships). The links that exist The first session of the workshop, chaired by between the nodes affect the attributes of the Ignazio Angeloni, provided an overview of the nodes (for example, banks’ balance sheets are techniques and the methodologies of network affected by existing links with other banks), analysis and of recent applications aiming to and the structure of the links affects the model and better understand the performance of the system as a whole. There are interconnectedness of financial and payment a number of common properties shared by many systems. The first presentation, made by large and complex networks that are of particular Kimmo Soramäki,5 provided the audience with interest for policy makers today, as they allow a general introduction to the topic, as well as for a better understanding of recent financial with concrete applications, illustrating the network dynamics. These are as follows: potential of this tool for policy purposes. The title – “Is network theory the best hope for • The “robust yet fragile” property of scale- regulating systemic risk?” – refers to the recent free networks, i.e. of systems where the argument made by some policy makers and probability of finding a node with a high economists that network topology could degree (high number of links) is very low, represent a new and key tool for taking into while the probability of a node having a few account contagion and systemic risk.6 connections is very high. This property refers The second paper, presented by Sheri Markose, to the robustness of a connected network in provided an in-depth empirical mapping of the the case of random removal of a node (given financial network created by credit default swap the high frequency of low-degree nodes), (CDS) obligations among US banks, and versus its fragility in the case of a targeted between banks and non-regulated entities attack directed against one of the few highly (monoline insurers and hedge funds) involved connected vertices (which could represent, as protection buyers and protection sellers. for instance, a financial hub). The long-term aim of this research is to establish fully digital and database-driven network • The “strength of weak ties”, which refers mappings of key financial sectors for systemic to the relative importance – in terms of risk modelling and assessment. availability/dissemination of information – of weak versus strong connections in shaping INTRODUCTION TO THE TOPICS the topology of the network. Kimmo Soramäki organised his presentation around three policy questions: • “Homophily”, i.e. the concept that certain attributes tend to set up clusters of nodes. 1. How can we measure the systemic importance of a bank? • The “small world phenomenon”, by which the number of links covering the distance 2. Can regulators promote a safer financial between any two nodes tends to be relatively system by affecting its topology? low (or network paths are short). This might have interesting implications for episodes of 3. Is it possible to devise early warning indicators from real-time data? 5 Kimmo Soramäki has recently created a website www. financialnetworkanalysis.com, which aims to gather research in this relatively unexplored field of financial economics. Soramäki provided a brief overview of the 6 See, for instance, IMF (2009), “Global Financial Stability general findings of network theory that make Report”, Chapter II on Assessing the Systemic Implications of Financial Linkages, and A. G. Haldane (2009), “Rethinking the straightforward the potential for its application financial network”, speech delivered at the Financial Student to the analysis of financial networks. Networks Association, Amsterdam, in April. ECB Recent advances in modelling systemic risk using network analysis January 2010 9 contagion in many real world small networks, In applying these findings to financial networks since the number of affected nodes above one needs to consider the process taking place which epidemics propagate system-wide is in the network and behaviour of the nodes in the especially low (and it can be zero).7 particular field of application.8 Chart 1 Types of networks A crucial characteristic of the structure of network processes is their centrality (i.e., in a broad sense, the relevance of the position of a node in the network).9 Centrality might give an Complete network insight into which nodes should be considered of “systemic importance”. However, Soramäki also made clear the limits of available centrality measures, since, although able to capture the type of flow-processes in the network, they do not currently capture any complex behaviour by the vertices, i.e. the drivers behind each node’s choice to set up certain links and the magnitude of the links that are set up. The lack of behavioural aspects is perhaps the main criticism addressed to network analysis by many economists today. The resulting Random network mechanical representation of how the structure is created and evolves over time cannot fully capture feedback loops and endogenous responses which are, however, at the core of financial networks’ developments. Indeed, the current crisis has shown how network processes can change in a sudden and unpredictable fashion. “Agent-based modelling” (relying on algorithms and simulations) is one recent approach devised to tackle this shortcoming.10 7 M. Bech, W. E. Beyeler, R. J. Glass and K. Soramäki in “Network Scale-free network topology and payment system resilience”, BoF Simulation Seminar, 23 August 2006, provide evidence on how the “small world” property might affect interbank payment flows after the occurrence of a payment outage at a large bank. They show that scale-free, long-tailed networks display the highest rate of liquidity absorption after such a shock (the rate of absorption being the rapidity with which a certain amount of liquidity is absorbed by/sent to the distressed bank). 8 S. Borgatti (2005), “Centrality and network flow”, Social Networks. 9 “Centrality” may be measured by the number of links that terminate upon a node (in degree), by the distance from other vertices (closeness), or by the existing connections to central nodes. A measure of centrality particularly suitable for financial networks is the betweenness centrality of a node, defined as the number of shortest paths that pass through the node. 10 An “agent-based model” (ABM) is a computational model for simulating the actions and interactions of autonomous individual agents with a view to assessing their effects on the system as a Source: whole. A key concept in an ABM is that simple decision-making rules can generate complex behaviour at the system level. ECB Recent advances in modelling systemic risk using network analysis 10 January 2010 DETAILED SUMMARY OF THE THEMES For instance, Soramäki discussed a model of a the likelihood of larger exposure concentrations real time gross settlement (RTGS) payment relative to a “star” format network (i.e. relative system with 15 banks introducing behavioural to the limiting case where clearers are rules for each bank’s decision about (i) the share all direct). of payments it has “queued” at any moment, and (ii) the size of net exposure it wants to have Finally, as regards the scope to devise early towards a single counterparty in relation to the warning indicators from real data, Soramäki total value of sent payments. Running raised the possibility of central banks constructing simulations on the basis of these rules, such indicators – e.g. increased riskiness or the authors study how the centrality of a failing worsened liquidity conditions of banks – bank (removed for the whole day from the by using the same kind of network techniques network) correlates with additional liquidity used by credit card companies on customers’ demand from the whole system. The more payment behaviour to detect card frauds. For non-linearities the system exhibits due to bank instance, payment data could be used to detect behaviour or liquidity constraints, the weaker is features in the timing of payments sent to and the correlation of the failure impact with the from the bank, in net outflows across different centrality measures. systems, in the bank’s money market activity or in the volume of cash withdrawals/deposits made Concerning the possibility for policy makers to by the public, factors which are rather common promote safer topologies, Soramäki referred to across available examples of failed banks. CLS, the world’s largest settlement system for foreign exchange trades, as an example where THE CDS NETWORK the financial links of an institution are severely The paper by Sheri Markose, Simone Giansante, restricted for the purpose of financial safety. Mateusz Gatkowski and Ali Rais Shaghaghi, CLS is not allowed to have any non-FX “Too interconnected to fail: financial networks settlement related links to the financial of CDS and other credit enhancement obligations infrastructure. Another example is the of US banks”, responds in part to Soramäki’s introduction by regulators of sectoral barriers to agenda of key policy questions. The authors banking, such as those introduced by the Glass- apply agent-based modelling to a financial Steagall Act in 1933. A recent case in point is network and use simulation results to devise an the introduction of central counterparty clearing operational measure of systemic risk. The focus for CDSs. Soramäki expanded on this last on CDSs stems from the “unique, endemic and example by outlining research that he has done pernicious role” that these instruments had in the on the topology of the network that develops current crisis. The authors argue that incentives around the central counterparty (CCP).11 provided by the credit risk transfer (CRT) scheme This work studies the impact of different included in the Basel II accord could have network structures – determined by the extent of contributed to the rapid expansion of this market. tiering (i.e. the number of banks that participate One potential consequence of banks’ ability to directly in the CCP) and the concentration of reduce regulatory capital requirements by using clients across first tier (direct) clearers – on the CRT techniques has been the growing popularity maximum exposure of the CCP. The results of synthetic securitisations, with the consequent show that the higher the level of tiering dispersion of products and risks worldwide in (i.e. the lower the number of members clearing complex chains of insurance and reinsurance directly in the CCP and the higher the number against credit default risk (see Chart 2). of indirect participants) and the higher the level According to the authors, the large amounts of clients’ concentration, the lower the CCP’s outstanding and the relatively high concentration maximum expected exposure. However, high tiering (for a given concentration) makes 11 M. Galbiati and K. Soramäki (2009), Central counterparties and CCP’s exposures more dispersed and increases the topology of clearing networks, forthcoming. ECB Recent advances in modelling systemic risk using network analysis January 2010 11 Chart 2 The CDS chain structure and bear raids 1) A “LENDS” to Reference Entity A Reference (Bond Issuer) Entity or Reference Assets in CDO Tranches Premium in bps Default Default Protection Protection for Seller, C CDS Buyer, B “INSURER” Payment X in case of default: X = 100 (1-R) Now 3rd party D receives B sells CDS to D insurance when A defaults; Recovery rate, R, is the ratio but B still owns A’s Bonds! of the value of the bond Party D has incentive to short issued by reference entity A’s stocks to trigger failure: immediately after default Bear Raid to the face value of the bond Source: S. Markose (Workshop presentation, 2009). 1) CDS on single and multiple reference entities. of risks to a few dominant players has brought to critical hub could have brought down part of the the fore the “too interconnected to fail” paradigm. whole CDS market, with a consequent impact To avoid such problems in the future, Markose on the whole financial system. et al. suggest setting up stress testing exercises for new financial instruments and propose such The analysis reveals the presence of “super- stress tests for the US CDS network. spreaders” in the CDS network, i.e. large protection sellers who are highly “central” in the The authors reconstruct their network using market in terms of clustering and connectivity CDS linkages among the top 25 US banks and measures, and whose capital bases – although the external non-bank insurers. Market shares comfortably fulfilling regulatory requirements – are taken as a proxy for actual bilateral could be considered low when account is taken exposures. Specific “small world” network of the system-wide capital loss they may impose properties and the loss impact suffered by each if they are assumed to fail in these simulation participant (in terms of core capital loss due to exercises. The same experiment is then simulated the failure of a major player) from its activity in on a random (i.e. not hub-clustered) network. the CDS market are computed.12 This allows the It is worth noting that, although the random authors to investigate the robustness of a graph has no nodes which are highly connected topology in which the top five US banks and also has lower network concentration or accounted for 98% of the reported CDS gross clustering, the consequences of its unravelling notional value at the end of 2008, while the non- when hit by a shock may be more severe bank entities had the highest clustering of links with the top four banks. 12 The authors conduct two experiments. In both, they use a 20% reduction of core capital as a threshold to identify bank According to the results of agent-based failures induced by the default of a triggering bank. The first test considers only the loss of CDS cover due to the failed simulations presented in the paper, bailouts bank suspending its guarantees as a counterparty. In the second of institutions with very large numbers of experiment, the triggering bank is itself a CDS reference entity, links – notwithstanding their possible technical which activates obligations from other CDS market participants. Furthermore, loss of cover owing to the triggering bank’s default insolvency – could not be averted: the on exposures to special purpose vehicles and owing to other simulations show how a credit event at one such credit enhancements is considered. ECB Recent advances in modelling systemic risk using network analysis 12 January 2010 DETAILED SUMMARY OF THE THEMES (22 banks out of 25 fail rather than only five that a CCP should hold and the use of network as in the previous case). At the same time, the indicators to make members’ contributions to the dynamics that bring down the system develops CCP’s capital or clearing fund proportional to over several consecutive rounds (following the their potential systemic impact. The authors also demise of the triggering bank), and not just mention the possibility of changing the existing after the first one. This might have important requirements on initial and on variation margins implications for regulators and central banks that market participants are required to post and aiming at promoting a safer financial system, and hold as part of their risk controls. suggests the need to be cautious in promoting any form of “ideal” network topology. MAIN COMMENTS AND DISCUSSION Commenting on both papers, Johannes Markose and her co-authors emphasise the Lindner, discussant for this first session, agreed need to incorporate institutional rules and with the presenters on the crucial importance behavioural aspects to obtain an adequate of agent-based modelling for understanding modelling of systemic risk and financial financial networks and especially their complex contagion. In particular, they discuss how agent- dynamics under distressed conditions. He said based models can address the failure of other that this would be decisive in further economic tools to take into account systemic strengthening of financial network analysis and risk, heterogeneity in agents’ strategies, and would allow its establishment as an additional interconnectedness of relationships, that make analytic tool for policy makers and regulators. the system prone to non-linear and extreme non-Gaussian dynamics when hit by a shock. While recognising the potential of this new instrument, the discussant also pointed out As regards regulatory solutions to the implicit the importance of a clearer categorisation “too big to fail” insurance enjoyed by large of which shocks and crisis dynamics could market players, the authors argue in favour of a be best understood using network analysis price/tax to be imposed/levied on super- rather than other tools. For instance, network spreaders – possibly identified on the basis of statistics computed after past failures of a the proposed “systemic risk ratio” (SRR) – to market participant could be more useful inputs reflect the negative externalities imposed by to early warning indicators designed to predict these market participants on the whole system.13 the impact of sudden idiosyncratic shocks on More generally, a price on the operations of a financial entity than variables which capture “systemically important” players is regarded as the build-up of macro imbalances over time. an adequate measure to provide banks and Moreover, he mentioned how network analysis especially non-banks (e.g. non-regulated and simulations, even if “agent-based”, might monolines in CDS markets) with more aligned be less suited to capturing certain market incentives to engage in over-supply of a given imperfections, such as incomplete markets and financial activity or instrument. Overall, based asymmetric or imperfect information. on their evidence, the authors suggest that it might be beneficial if the large negative Lindner agreed that there was an opportunity externalities that arise from the possible demise to exploit today’s computer-power to map of a big player in the CDS network were taken network structures. To this end, he stressed into account when banks are allowed to reduce the importance of getting access to data not capital on assets that have CDS protection. only at the level of individual networks, but especially across networks and across national Finally, welcoming the recent introduction of central counterparty clearing (CCP) in CDS 13 For each trigger bank or non-bank CDS provider, the SRR markets, Markose et al. propose the use of agent- estimates the percentage loss in aggregate core capital resulting based stress-tests to estimate the amount of capital from its collapse. ECB Recent advances in modelling systemic risk using network analysis January 2010 13 borders. Even acknowledging the limits to the combination of different risk controls – mapping complex adaptive systems in a unitary participation requirements, initial and variation framework, the development of a comprehensive margins, and financial resources (i.e. CCP’s network perspective remains a key requirement own capital or clearing fund) – to address the for policy makers and regulators. risks stemming from its participants. For instance, concerning the empirical Concerning the possibility of imposing a price reconstruction of the CDS network provided by on the operations of systemically important Markose et al., Lindner recognised its merit as a players, the subsequent discussion revealed that, good first approximation of bilateral exposures from the point of view of regulators and in the market. However, depending on data overseers, a key operational issue would concern availability, a cross-check with actual bilateral the exact definition of a critical participant exposures among participants, as well as the (should authorities use a binary indicator or inclusion of Europe and/or other markets would should different layers of “criticality” be be important in order to obtain a more reliable considered?) and the way in which network basis for policy implications. connectivity could be taken into account in addition to traditional balance sheet or activity An important point raised by many participants measures (i.e. size and volumes/values). One of was whether network measures could represent the participants put forward the proposal to an appropriate tool to identify systemically integrate existing risk management tools important market players and how these (e.g., CoVaR analysis) with network measures measures could be integrated in the existing for regulatory purposes.14 toolbox of regulators. This is strongly related to the issue of how to address institutions’ systemic A main caveat raised during the discussion importance and, therefore, of how regulators on the identification of key market players could encourage safer topologies. On this aspect, concerned the inadequacy of indicators that Markose et al. argue in their paper that imposing are solely based on participants’ exposures in a a “tax” on the operations of critical players could particular market/instrument. Ignazio Angeloni, be one way of providing financial institutions chairman of the session, pointed out how such with more aligned incentives and hence contain “narrow-view” indicators could actually provide risks. Moreover, concerning the very recent a misleading picture on the criticality of a move towards central counterparty clearing in certain participant. In fact, an institution which CDS markets, Lindner concurred with Soramäki is relatively small in one particular market could and Markose that this is a key example of still be “central” in the network due to its uneven how improvements in the infrastructure and exposure to a large and highly connected player. encouragement from public authorities can Its demise might then still have a large impact affect the robustness of the financial system. on other participants in the system. A CCP reduces counterparty risk, increases The particular usefulness of network tools market liquidity and strengthens transparency. for visualising direct linkages among market However, it also concentrates systemic risk. players and, depending on data availability, This makes the establishment of a strict risk links across different markets is generally management framework and adequate oversight acknowledged. However, some participants by regulators essential. Similar to the need to regulate systemic risk and systemically relevant 14 See IMF (2009), “Global Financial Stability Report”, Chapter II, and M. Brunnermeier et al. (2009), “The Fundamental market players in the financial sector more Principles of Financial Regulation,” Geneva Reports on the broadly, the risk concentration in a CCP requires World Economy, 11. ECB Recent advances in modelling systemic risk using network analysis 14 January 2010 DETAILED SUMMARY OF THE THEMES to the workshop expressed doubts about the scope of network analysis alone to understand the identity of factors driving the expansion of a financial market/instrument over time, or how a certain institution becomes a “key” player for a given market. Such an understanding is critical for regulators. The endogeneity of a market structure – which is the outcome of a dynamic process taking place over time – makes any policy intervention extremely difficult. On this part, while acknowledging this difficulty, Soramäki emphasised his conviction that this should not prevent the regulators from trying to use all the tools available to them to devise mechanisms that have the potential to mitigate risks ex ante and, therefore, to make the financial system safer. ECB Recent advances in modelling systemic risk using network analysis January 2010 15 SESSION II – INTERDEPENDENCIES AMONG settlement and market liquidity, emphasizing INSTITUTIONS, SECTORS AND SYSTEMS possible economic implications that can result from such interlinkages.15 The second session of the workshop, chaired by Paul Mercier, brought together two papers In order to find out how major Fedwire on the theme of interdependencies among participants changed their behaviour in terms institutions, sectors and systems. of delayed settlements, the authors consider two shocks that actually materialised. Firstly, The experience of the recent crisis has shown the impact of the failure of Lehman Brothers that even the failure of relatively small but on liquidity and payment flows is explored. well connected entities can have unforeseeable The injection of liquidity into the financial negative financial consequences through system by the Federal Reserve following the contagious effects. For researchers, this poses bankruptcy of Lehman is considered to be the new challenges as a better understanding of second shock to Fedwire. the structure and the functioning of financial networks is key to preventing risks inherently Focusing on payments settled in the system, present in the system. Bech and Adelstein identify changing liquidity conditions by looking at different patterns Following this line of investigation, the of settlement timing on a daily basis from the presentation given by Morten Bech enhanced end of March 2008 until 1 September 2009. the understanding of settlement behaviour of Settlement liquidity in Fedwire is measured Fedwire participants before, during and after the using data about the degree of daily settlement failure of Lehman Brothers. Network analysis delays, dividing the period of interest provides an adequate toolbox to analyse and into pre-crisis, crisis (Lehman’s default) and visualise the daily changes of settlements in post-crisis periods. the Fedwire network as well as the increased behavioural coordination of its participants. The authors find that prior to the collapse of Lehman Brothers the average settlement time In contrast to the application of network was around 2:30 p.m. whereas it averaged as analysis based on individual payments, late as 3:10 p.m. in the two weeks following Olli Castrén presented a paper looking at this major bankruptcy. During the last sector level interdependencies in the euro area period under consideration, as a result of the financial system. Network analyses of this kind liquidity injection by the Federal Reserve, have so far been conducted on firm level and on settlements were undertaken considerably country level, leaving an unexplored gap at the earlier – on average at 2 p.m. (see Chart 3). intermediate stage. However, settlements in Fedwire typically FEDWIRE SETTLEMENTS vary due to calendar effects. In order to net out The first paper of the session, entitled “Payments, such influences, the authors run a regression crunch and easing” by Morten Bech and using dummies for days that are known to Ian Adelstein, uses network analysis to explore have different settlement timings. The actual the changing pattern of Fedwire settlements delay due to non-calendar effects is then to be following Lehman Brothers’ bankruptcy. found in the regression residuals which provide a net measure of average settlement timing. The authors introduce a threefold concept by distinguishing between market, funding and settlement liquidity. While the focus of the paper 15 Bech and Adelstein point to the failure of Bear Stearns in March 2008 as one prominent example of how lacks in settlement lies in the latter, it also points to existing links liquidity can negatively affect the availability of funding liquidity between settlement and funding as well as for all market players. ECB Recent advances in modelling systemic risk using network analysis 16 January 2010 DETAILED SUMMARY OF THE THEMES Chart 3 Average time of Fedwire Chart 4 Fedwire settlement delays settlements on September 17, 2008 y-axis: time 3:00 3:00 2:30 2:30 2:00 2:00 1:30 1:30 1:00 1:00 Apr. Oct. Apr. Oct. 2008 2009 Source: M. Bech and I. Adelstein (2009). Source: M. Bech and I. Adelstein (2009). The results support previous findings in terms of 75th percentile settlement times across 72 Fedwire time-shifts of settlements prior to, following and participants, they examine the differences in during the period of acute impact of Lehman settlement coordination for three distinct periods, Brothers’ bankruptcy. i.e. pre-crisis, crisis and post-crisis. After operationalising the variables at hand, the Based on correlation matrices, a distribution of authors employ network techniques to visualise correlations for each single period is obtained. the deteriorated degree of liquidity in Fedwire. In line with the findings of the former part of the The analysis is narrowed down to all business paper, it is again the period immediately after days in September 2008 and to a core set of Lehman Brothers’ default that differs from both 16 Fedwire participants. By doing so, the paper the baseline and the post-crisis period, showing a keeps its focus on the actual period of interest, higher degree of correlation in settlement timing. as well as on the turbulences caused by the shock event to other actors in the same financial The robustness of these results is underlined environment. The authors showed that, until by applying the method to a subset of 16 major 12 September 2008, settlement behaviour was Fedwire participants. The results show that these normal. However, this changed dramatically display an even stronger tendency to coordinate throughout the two weeks following behaviour, and they actually seem to be driving 15 September, the day Lehman Brothers filed the heightened coordination in the “crisis” period. for bankruptcy. After that date, significant degrees of delay within Fedwire are observed, To further illustrate their results, Bech and reaching a peak on 17 and 19 September with the Adelstein make use of network techniques to majority of links reflecting overdue payments visually highlight the differences in coordination (see Chart 4). According to the authors, this settlements throughout the three periods under observation can be considered an indicator of consideration. the effect of a systemic shock to the network. EURO AREA FINANCIAL NETWORKS Bech and Adelstein conclude their paper with The second presentation was given by an analysis of the coordination of settlements Olli Castrén on a joint paper with Ilja Kristian among Fedwire participants in the light of Kavonius entitled “Balance sheet contagion and a changing liquidity environment. Using systemic risk in the euro area financial system: daily time series of the correlations of the a network approach”. ECB Recent advances in modelling systemic risk using network analysis January 2010 17 At the beginning of his presentation, Castrén Chart 5 Cross-sector balance sheet gross briefly summarised the work already done on exposures in the euro area financial system the topic at the macro level as well as at the micro level. He pointed to the fact that an 2009 Q2 analysis of accounting-based balance sheet interlinkages at sector level has never been HH ROW conducted before. In addition, he argued that, in order to incorporate elements of risk into the OFI NFC analysis, it is necessary to construct “risk-based balance sheets” which also include volatility of asset values. MFI GOVT The authors use quarterly non-consolidated data from 1999 onwards on euro area financial INS accounts (EAA), based on the methodological framework established in the European System 1999 Q1 of Accounts 1995 (ESA95). By doing so, they HH ROW analyse a closed system of assets and liabilities spread among seven distinct sectors.16 Since OFI NFC these data do not contain any information about the counterparties of the instrument issued by a given sector, the “maximum entropy” technique is used to approximate these allocations. Finally, MFI GOVT matrices of bilateral exposures, reflecting the amounts of assets and liabilities as well as the INS instrument category they belong to, were Sources: O. Castrén and I.K. Kavonius (2009). constructed for inter as well as intra-sectoral balance sheet relationships.17 3. the increasing importance of the other With this information to hand, a complete network financial intermediary sector over the past linking all sectors together by summing up assets ten years. and liabilities for a total of eight instrument categories is obtained. Castrén presented snapshots In fact, networks which are derived from the of these networks of balance sheet gross exposures balance sheet exposures do not only help to in the euro area at instrument level for the first visualise the units of analysis and the links quarter of 1999 and for the second quarter of 2009 between them. Network analysis also offers respectively (see Chart 5). features that allow the modelling and tracing of contagious effects and knock-on events Comparing these snapshots, three main in the system. Making use of this quality, developments become evident: the paper first considers a simplified three sector model and assumes an unanticipated 1. an overall increase in balance sheet net income shock resulting in a deficit for exposures suggesting a higher level of one of the sectors’ profit and loss accounts. interconnectedness in the euro area financial system; 16 The set of sectors consists of the following: households, non-financial corporations, banks, insurance companies and 2. the “hub” position of the banking sector, as pension funds, other financial intermediaries, government, and the rest of the world. revealed by the large weight of its links to 17 The intra-sectoral exposures can thus be found on the matrix counterparties; and diagonal. ECB Recent advances in modelling systemic risk using network analysis 18 January 2010 DETAILED SUMMARY OF THE THEMES Then, mark-to-market accounting is assumed, Chart 6 Sector level distances-to-distress leading to a faster transmission of the shock for the euro area financial system throughout the network, i.e. to the balance sheet of the other sectors. NFC GOV MFI HH OFI ROW To demonstrate how a similar shock could be INS transmitted in the network of sectoral balance 30 30 sheet exposures, the authors introduce a cash-flow shock in the non-financial corporations 25 25 (NFCs) sector that causes a 20% mark-to-market drop in the value of shareholders’ equity. In the 20 20 first round, above all, the NFCs sector itself as well as the other financial intermediaries (OFIs) 15 15 and the government sectors are those most heavily affected. In the subsequent rounds, the most affected sectors are those which hold large 10 10 amounts of equity issued by those sectors which were adversely affected by the initial shock in 5 5 the first round. 0 0 In this context, Castrén emphasised that, 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 in a multi-period framework and when hit by Source: O. Castrén and I.K. Kavonius (2009). a shock, agents are expected to rebalance their accounts by deleveraging, disinvesting or similar actions – a scenario not incorporated in the current The authors find a large discrepancy in impacts analysis. Nevertheless, the presented network can and argue that this could stem from the be used to simulate the effects of such actions non-linear character of the changes in risk once relevant rules and thresholds are specified exposures as a reaction to volatility shocks in and modelling problems have been overcome. sectors that are characterised by high leverage. Furthermore, they state that, owing to these In the second part of the paper, contingent non-linearities in risk exposures, the claims analysis (CCA) is applied to enhance the interconnections may serve as risk-spreading framework by accounting for the accumulation shock amplifiers in a crisis situation, whereas and transmission of risk exposures in the they are assumed to perform the functions financial system, thus overcoming its current of risk sharing and shock absorption in deterministic character. Risk-based balance normal times. sheets at sector level can be constructed by applying CCA, thus making it possible to Concluding his presentation, Castrén emphasised conduct a macro financial risk analysis as the that more research needs to be conducted in propagation of risk exposures across sectors can order to refine propagation mechanisms in such be examined. networks. Castrén and Kavonius use the “distance- MAIN COMMENTS AND DISCUSSION to-distress” measure obtained from CCA The discussant of the second session, (see Chart 6) to capture how risk exposures Goetz von Peter, commented first on the in sectors that were not directly struck by the paper presented by Castrén. He emphasised the initial NFC cash-flow shock considered above innovative approach of conducting a systemic were also affected via contagion across balance risk analysis on the European sector level, a unit sheet items. of analysis not yet examined. ECB Recent advances in modelling systemic risk using network analysis January 2010 19 At the same time, he recommended clarifying the discussant posed an open question on the significance of cross-holdings between some whether delays are indeed the main feature sectors. This brought Von Peter to a general of stress. Other aspects, such as volumes of concern: the high degree of data aggregation. payments and failures, may be equally important In the construction process, existing for assessing tensions. heterogeneity within each sector is averaged out, inevitably resulting in a loss of information. The subsequent discussion by the workshop Consequently, the question of the extent to which participants mainly focused on the changing such balance sheets remain interpretable arises. nature of networks, which needs to be taken In particular, sector-wide balance sheets might into consideration. Furthermore, the use of be misleading, as solvent units are unlikely to aggregated data was discussed, as macro level support failing units in the same sector. networks are well connected by construction, making the application of some network topology Von Peter went on to state that, although the measures somewhat uninformative. It was also authors acknowledge such data limitations, pointed out that network properties become they do not always explain what this means for increasingly non-linear the more aggregated the the results. In addition, the application of the data are. This leads to interpretational biases, as maximum entropy technique by construction the underlying structure between the network leads to a complete network, obviating the components is not correctly captured. Another inclusion of statements on degree distribution suggestion concerned the inclusion of robustness or the like. tests in both papers. This can be done, for example, by altering the definition of nodes and In review of the paper presented by Bech, edges in order to check whether similar results Von Peter underlined the insights gained in are obtained. studying a unique event (the Lehman Brothers bankruptcy) in the data-rich environment of Fedwire. He commended the focus on settlement liquidity as an interesting choice and praised the inclusion of behavioural aspects. However, further exploration is needed on the selection of parameter thresholds in the paper, as they appear to be of a rather arbitrary nature. Improvements can also be made by stating precisely whether it was the type of payment or the type of participant that led the authors to conclude whether or not there was discretion in when the payment needed to be sent. More importantly, Von Peter questioned the use of correlations to quantify coordinated delays, as they only deliver information about the tendency of participants to move together, whether late or early. Furthermore, he felt that further exploitation of the network structure that the authors had constructed would be desirable. For instance, considering transitive relationships (e.g. clustering) would help explain why delayed incoming payments would lead a bank to also delay sending payments. Additionally, ECB Recent advances in modelling systemic risk using network analysis 20 January 2010 DETAILED SUMMARY OF THE THEMES SESSION III – INTERBANK CREDIT, MARKETS Chart 7 Average number & value of AND LIQUIDITY MANAGEMENT IN LARGE simulated contagious defaults per day VALUE PAYMENT SYSTEMS bank accounts The speakers of the third session, chaired by transfer accounts Daniela Russo, presented two applications 20 20 of network analysis to large value payment 16 16 systems (LVPS). Both papers were motivated by the need for the national central banks 12 12 (in this case, the Oesterreichische Nationalbank 8 8 and De Nederlandsche Bank) to gain a better understanding of the robustness of the domestic 4 4 payment system. To this end, both papers analyse 0 0 the topology of the payment network and perform a set of simulations in order to assess its stability 2.4 2.4 when a highly connected participant is removed 2.0 2.0 from the system (in the Dutch case, which was 1.6 1.6 presented by Iman van Lelyveld) or with an 1.2 1.2 operational incident at one participant’s account 0.8 0.8 (in the Austrian case, which was presented by Claus Puhr). These exercises allow the authors 0.4 0.4 to assess the relevance of contagion for the 0.0 0.0 domestic LVPS (respectively ARTIS for Austria Sources: C. Puhr and S.W. Schmitz (2009). and TOP for The Netherlands) and the systemic importance of some players. Also, transfer accounts cause significantly more THE AUSTRIAN LVP SYSTEM contagion than bank accounts (due to their The failure of a large domestic bank prior to the centrality in the network) while, unexpectedly, crisis motivated Claus Puhr and his co-author, operational shocks on days with higher Stefan W. Schmitz, to start an exploratory transaction activity cause lower contagion.19 analysis of ARTIS liquidity data using different This last and somewhat counterintuitive result is econometric techniques. Their aim was to possibly related to the uncovered decreasing eventually extract some early warning signals time trend in the number of simulated contagious from actual payments data. In the paper presented defaults per day over the period from at the workshop, “Structure and stability in November 2005 to November 2007. This seems payment networks: a panel data analysis of to suggest that the Austrian system has become ARTIS simulations”, network topology and more stable over these two years. counterfactual simulations are used to quantify the contagious impact of unsettled payments In the last part of the paper, the authors use a panel resulting from an incident at an individual bank data analysis to assess the relative significance of level (namely from the inability of a participant network topology indicators in explaining the to submit payments for the whole day). high variation of contagion – as measured by (i) number of banks with unsettled payments; The results of 63 different operational stress 18 A “contagious default” occurs when a bank that does not receive scenarios, for the period from November 2005 a payment is in turn unable to send payments for that day. until November 2007, reveal that only a few 19 “Transfer accounts” are ARTIS accounts held by other accounts are systemically important in terms of Eurosystem central banks at the Oesterreichische Nationalbank. All national TARGET components are directly linked by transfer number and value of contagious defaults that accounts. All transactions to and from the respective country and they might cause per day (see Chart 7).18 Austria are routed via these accounts. ECB Recent advances in modelling systemic risk using network analysis January 2010 21 (ii) number of unsettled payments, at the end of connected nodes correspond to the largest Dutch the day, due to an operational incident at another banks. For instance, one of the most connected participant and (iii) value of unsettled payments, hubs is a clearing institution. According to the at the end of the day, due to an operational incident presenter, this clearly indicates that network at another participant – both in the measures do provide an additional tool to assess cross-section (i.e., among ARTIS participants) the criticality of a participant from a systemic and across days.20 The results show that, out of point of view, and to better evaluate the more than a hundred indicators at network and performance of the system. node level, the best measures for the identification of systemically important accounts in ARTIS are The paper provides new evidence of the the number and volumes of (contagious) defaulted influence of the chosen time frame in the analysis payments that a bank can cause. That is, following of network properties. In fact, due to finality of an incident at one participant’s account, payments in RTGS systems, links are extremely the ensuing liquidity loss and the level of aggregate short-lived. This implies that the chosen time liquidity available in the system offer the most horizon is crucial when assessing the results. convincing explanations.21 In the case of TOP, a ten-minute slice of recorded flows is already sufficient to characterise the THE DUTCH LVP SYSTEM structure of the whole system. The second speaker, Iman van Lelyveld, presented “Interbank payments in crisis”, a joint In the second part of the paper, the authors work with Marc Pröpper and Ronald Heijmans study the vulnerability of the system to the on the topology of the domestic LVP system removal, one by one, of the ten most highly (TOP) and on the broader implications network connected participants. Looking at the impact topology might have for financial stability. of these removals on network properties and on settled volumes and values makes it possible to First, the paper presents an intraday analysis of measure indirectly the role of highly connected transactions processed and values transferred players (either banks or clearing institutions) in through TOP from June 2005 to May 2006. the stability of the network. The authors then study the changes in the structure of the network over time in terms of Finally, the authors try to investigate the effects commonly used network measures. These are of the recent financial crisis on the payment size (number of active nodes); connectivity system by monitoring traditional activity between banks (the ratio of actual to possible measures as well as network properties in the links); clustering (the probability of two period from June 2006 to December 2008. neighbours of a node also sharing a link); Consistent with the results of the previous paper, and network correlations (whether nodes that indicators in the Dutch payment system network make payments to many counterparties also also seem to add relatively little to the analysis receive payments from many). once volumes and values are taken into account. They find that the Dutch network is small 20 Number and value refer, respectively, to the total number and in terms of both nodes and links, compact total value of payments that could not be settled by banks that (with banks that are on average only two steps did not experience an operational incident. 21 The authors selected 44 indicators at network level and 71 at apart) and sparse in terms of connectivity over node (stricken bank) level. Both univariate and multivariate the period under analysis, for all the different analysis showed the existence of a significant correlation between node-indicators and (contagious) unsettled payments. time-snapshots used (1 hour, 1 day, or 1 year). In particular, higher node-degree and connectivity and lower Moreover, it is characterised by a few highly average path length are significant in explaining higher contagion. connected nodes linked to several nodes with Among network-level indicators, betweenness centrality – the average of all individual nodes’ betweenness centralities (see the relatively few connections. Interestingly, on definition in footnote 9) – turns out to be particularly helpful in a short time scale, not all of these most highly predicting contagious defaults in the Austrian interbank market. ECB Recent advances in modelling systemic risk using network analysis 22 January 2010 DETAILED SUMMARY OF THE THEMES Chart 8 The impact of the failure of Lehman nodes do not react to the simulated triggering Brothers 1) event. Holthausen argued that until adaptation in behaviour in response to shocks is not x-axis: period contemplated by network models the latter will not y-axis: temp represent an appropriate tool for the assessment of 15 15 systemic risk and of systems’ resiliency to shocks. This is because the current models “miss” the kind of strategic, non-cooperative, and self-reinforcing 10 10 feedback loops that are crucial in the development of a financial crisis. 5 5 Another key issue raised in the discussion is the need for improvements in data and information- sharing across national borders and across 0 0 today’s numerous interdependent systems and 0 100 200 300 400 500 markets. These are critical in order to gain a Source: I. Van Lelyveld et al. (2009). thorough understanding of interactions existing 1) Developments of gross turnover and of a set of network measures from 1 January 2007 to 31 December 2008. in the global financial system and, therefore, to extend the network framework currently used The monitoring exercise reveals the absence of for the analysis of payment systems to the study any noticeable disruptions in the Dutch payment of broader questions about financial stability. system until the migration to TARGET2 was achieved in September 2007. A drastic change Beyond the lack of any adaptation in behaviour in the reconstructed network is clear after the following a shock, Holthausen questioned the collapse of Lehman Brothers (see Chart 8). appropriateness of some of the assumptions However, the migration to TARGET2 does not on which the presented papers rely on allow for an appropriate disentanglement of (e.g. the inability of a troubled institution crisis effects. to send any payment on a given day or the absence of strategic delays in settlement). MAIN COMMENTS AND DISCUSSION In order to build a meaningful link between the Cornelia Holthausen, discussant for this session, analysis of network properties and systemic highlighted the scope for further comparison of stability, elements such as the identity of systems that, however heterogeneous in terms of market players, changes in the set of the most volumes processed and number of participants, active banks over time or the potential scale might nonetheless share a common structure. of financial obligations which are not reflected The striking similarity between the results on in actual payments should not be overlooked. the Austrian LVPS and those obtained from This is of the utmost importance in making analyses looking at the US Fedwire system network research results a reliable basis from suggests that comparisons among payment which to draw regulatory implications. systems might be especially useful as a source of policy recommendations for the enhancement Holthausen made an additional general remark of network stability. to papers in this emerging field of financial network research, namely the need to identify A key comment made by the discussant about more clearly the scope of the analysis at hand both presentations concerned the absence of for policy recommendations. On this point, behavioural assumptions. Convincing behavioural Puhr mentioned the importance of good aspects are excluded from a standard simulation business continuity arrangements, especially analysis, where only the static consequences at the most important/connected nodes, as one of each simulated scenario are considered and of the main implication of the presented paper. ECB Recent advances in modelling systemic risk using network analysis January 2010 23 In the case of an operational failure at one not only by a shock per se, but even more so by account, allowing for alternative ways of settling existing interdependencies among the systems at least the largest payments would greatly in which it operates. As a consequence, the same reduce the systemic impact of the incident. player will behave differently in each system, even if it faces no liquidity hoarding or other In agreement with the discussant, and strategic motivation. notwithstanding his confidence as regards the contribution network theory can make to the analysis of the functioning of “the plumbing” of the financial system, Van Lelyveld expressed some scepticism when it comes to the use of networks for the purpose of studying the vulnerability of the system. In fact, for broader financial stability questions, more information is needed about what motivates participants’ payment decisions, especially in reaction to a shock, and about the way changes in agents’ choices might eventually reinforce one another in a non-cooperative way. Van Lelyveld pointed out how strategic behaviour is probably less relevant in small networks, where participants know each other. In the ensuing discussion the importance of tailoring existing measures to the specific application and objective at stake was highlighted. The discussion following the first session had highlighted the need to set up a careful categorisation of which particular measures are the most appropriate for the analysis of each specific type of shock. The discussions in the current session converged on the idea that the choice of the most appropriate time window for each specific issue at hand is an additional important issue to be considered in simulation exercises using financial networks.22 Concerning the expansion of the scope of network analysis to study financial stability, the chairman, Daniela Russo, pointed out how interdependencies across different systems and markets are potentially more important for financial stability than interdependencies within the system. This is the case because these types of links have the potential to dramatically change the behaviour of market participants. 22 In particular, some participants argued that the resilience of Russo pointed out how, especially in a crisis the system to shocks is probably best analysed using a one-day snapshot, while shorter time windows would be more appropriate situation, the behaviour of a player who is active for capturing behavioural aspects and the sudden evaporation of in many different systems might be affected trust that characterises financial crises. ECB Recent advances in modelling systemic risk using network analysis 24 January 2010 DETAILED SUMMARY OF THE THEMES SESSION IV – SYSTEM-LEVEL LIQUIDITY EFFECTS important features of the current paper. The AND NETWORKS IN EARLY WARNING MODELS model used in the paper mainly consists of three distinct layers that are interconnected through The presentations of the fourth session, chaired cross-holding exposures of loans and equities: by Mauro Grande, dealt with balance sheet interconnections between economic entities 1. a core of interacting domestic banks and the potential risk stemming from these constituting a complete network; links when shocks occur. Both papers have a similar analytical scope as they disentangle the 2. a set of international banks, typically well web of claims and obligations present in the connected to their immediate neighbours; and financial system in order to gain insights into the contagious effects that can be rooted in tight 3. a group of firms operating independently of financial relationships. They differ, however, each other but borrowing both from domestic substantially regarding the units of analysis and and international banks. the aggregation level at which the analysis is carried out. The linkages among these entities can be summarised in a single restricted matrix, The first paper, presented by Sujit Kapadia, representing a large part of banks’ balance sheet takes into account the intricacy of financial items.23 Initially, a macro shock hits the system systems as it examines the relationship between leading to corporate defaults that trigger credit firms, domestic banks and international banks. losses for both types of banks under The chosen network approach captures a large consideration, potentially causing their default. portion of the links between financial agents, To let banks compensate for the capital loss a feature not often found in existing network suffered, the possibility of fire sales is models. incorporated into the model. This distress sale of assets might lead to mark-to-market losses In contrast, the second presentation by Juan Solé which can trigger further fire sales in the system, and Marco Espinosa examined consolidated provoking an even larger negative impact on the claims and liability relationships across national participants of the system. On the other hand, banking systems. In their paper, simulations financial entities primarily suffer credit losses as based on idiosyncratic shocks are analysed, a result of a bank default, an event that can have leading to the identification of systemically knock-on effects, again leading to further important as well as particularly vulnerable defaults of other banks (see Chart 9). banking systems. Furthermore, the contagion paths, and thus the spreading of risk throughout Concluding the explanation of the basic model the system, are explored using network and its specific features, the authors describe techniques. the calibration of the model, using data from 17 UK banks, 120 foreign banks and 50.000 FINANCIAL RELATIONSHIPS BETWEEN FIRMS, firms, stemming from various sources. DOMESTIC BANKS AND INTERNATIONAL BANKS The first presentation was given by Sujit Kapadia In a baseline scenario, an idiosyncratic rather on “Complexity and crisis in financial systems”, than aggregate shock was considered, driving, on a joint paper with Kartik Anand, Simon Brennan, average, 220 firms into bankruptcy. This causes Prasanna Gai and Matthew Willison. an average asset loss of 0.15% to domestic banks and 0.12% to international banks, which Kapadia started with a brief overview of network does not threaten the stability of the system. theory concepts and their applications in 23 The matrix takes on a restricted form because it is assumed that economics, stating that tipping point properties firms do not lend and that financial institutions only hold equity and fat-tailed loss distributions are particularly in firms and not in each other. ECB Recent advances in modelling systemic risk using network analysis January 2010 25 Chart 9 Mapping shocks to systemic risk Macro shocks Corporate defaults Credit losses for banks (domestic and international) Bank capital falls Credit Mark-to-market Bank Asset fire losses for losses for defaults sales banks banks Inter-bank Idiosyncratic shocks network to banks Aggregate loss distribution Source: S. Kapadia et al. (2009). As a second step, the smallest shock that can 4% and 24% across institutions. This softens the bring down the entire system is considered. The tipping point property of the loss distribution simulation shows that the system collapses in but, when the average size of collapsing firms is 0.4% of the cases when (on average) 2700 firms given a sufficiently large value, the entire system default whereas in 99.6% of cases it does not. may still default as in the previous simulations. These findings underline the tipping point property which is characteristic for such The presentation went on to draw a link to the networks when put under stress, i.e. a sudden current crisis and to highlight the increased increase in distress in the loss distribution.24 The vulnerability of complex financial systems to authors find that adding fire sales to the scenario Lehman Brothers-type system-wide breakdowns. increases the vulnerability of the system to much Additionally, the model also emphasizes the smaller macro shocks. potentially amplifying effects of mark-to- market losses, an observation which has been In the next step, the authors relax the assumption of homogeneity across banks in terms of the 24 Kapadia pointed out that, although in this scenario a 100% loss given default for inter-bank loans was considered, this sort of sizes of their capital buffers (4% for all banks) bi-modal loss distribution prevails even after this assumption and instead allow the buffers to vary between is relaxed. ECB Recent advances in modelling systemic risk using network analysis 26 January 2010 DETAILED SUMMARY OF THE THEMES made also in the context of the most recent following the transmission of the initial shock turmoil. In this regard, Kapadia draw attention throughout the network. In addition, it is possible to declining capital buffers and increasing to identify systemic players within the network, leverage in recent years. The authors argue i.e. those banking systems whose failures cause that these developments, among others, can be immense stress to their counterparties in the partially held responsible for making the system network. The financial distress triggered by more vulnerable to instability. The introduction those players can trigegr not only to the default of systemic capital requirements might therefore of other banking systems, but also the collapse of deserve more consideration, also in the light of all systems under consideration. Another feature recent experiences. of the analysis is that capital impairment on a country-to-country basis can be easily displayed At the end of his presentation, as a potential in a matrix form, providing useful information avenue for future research, Kapadia pointed to about actual counterparty risks. the need to incorporate liquidity risk into the modelling of systemic risk in financial systems. As a result of the first simulation, which applies the first dataset, the UK and the US banking ASSESSING CROSS-BORDER LINKAGES systems are identified as important entities The second paper presented dealt with the causing three and four rounds of contagion topic “Network analysis as a tool to assess respectively, as well as 44.6% and 80% loss of cross-border financial linkages” and was presented all capital in the system (see Chart 10). Due to by Juan Solé from the IMF, on behalf of his co-authors Marco Espinosa and Kay Giesecke. 25 The data is taken from the BIS consolidated banking statistics (www.bis.org) which provides quarterly data on immediate borrower basis (on-balance-sheet items) and on ultimate risk At the beginning of his presentation, Solé basis (including risk transfers). stressed the potential of network analysis to become an important tool for cross-border Chart 10 Induced banking system failures surveillance. For regulators, it provides a metric to identify institutions that are potential sources of contagion. Furthermore, it can help to track contagion paths and offers a metric that can be credit channel credit and funding channel used to find out when and whether a financial 18 18 entity is “too connected to fail” in times of 16 16 financial stress. 14 14 In order to present the methodological 12 12 framework of the paper, Solé introduced a 10 10 stylised bank balance sheet identity that makes 8 8 it possible to follow “movements” of balance sheet items when a shock event occurs. In the 6 6 paper, the authors first consider a pure 4 4 idiosyncratic credit shock and then extend the 2 2 analysis to a credit-plus-funding shock. These simulations are carried out for two different 0 0 1 2 3 4 5 6 datasets. The first one contains solely on- 1 Finland balance-sheet items whereas the second one 2 France 3 Germany adds elements of risk transfer.25 4 Netherlands 5 United Kingdom 6 United States In both simulations, the main aim is to trace the Source: J. Solé (Workshop presentation, 2009). path of contagion among banking systems by ECB Recent advances in modelling systemic risk using network analysis January 2010 27 their close linkages to the countries from which financial institutions, a better understanding the shocks were assumed to originated, Belgium, and monitoring of direct and indirect linkages is the Netherlands, Sweden and Switzerland needed. Network analysis is one tool to assess this are found to be the most vulnerable to these problem. However, as some participants voiced particular shocks. Their banking systems fail doubts concerning its empirical practicability, in at least three out of the fifteen hypothetical Solé referred to Chapter II of the April 2009 simulations in which they were themselves not IMF Global Financial Stability Report, where considered as the trigger country. basic models of this kind are outlined.27 The analysis of the second simulation – the Furthermore, he emphasised that information credit-plus-funding event – largely confirms the about off-balance-sheet items and non-bank results from the first exercise.26 Again, the financial institutions, as well as other financial banking systems of the United Kingdom and the entities, need to be better incorporated into United States take on systemic roles, triggering network analysis. However, since this is often even more hypothetical defaults than before. not possible due to data limitations, he appealed Surprisingly, the collapse of the French banking for more joint surveillance as well as data system now induces three hypothetical defaults sharing between countries in the future. compared with none in the former scenario (see Chart 10). Solé et al. argue that this might MAIN COMMENTS AND DISCUSSION reflect the important role of this country as In his discussion of the two papers, liquidity provider in the system and that it shows Diego Rodríguez Palenzuela underlined the the usefulness of including scenarios that renewed importance of network analysis as account for different types of stress. The authors a scientific field and highlighted the value also demonstrate that the incorporation of the in exploring different approaches within the funding channel into the model increases the field. Influences from innovative concepts in overall vulnerability of the network in terms of economics are needed, given the shortcomings of defaults and capital impairments. the established paradigm in economic theory in terms of foreseeing the depth of the recent crisis. However, looking at the transmission of a shock using the second dataset, i.e. including risk In this regard, he commended the efforts made by transfers, the results change. The resilience to the authors to take into account the complexity shocks of Belgium, Sweden and Switzerland of the financial system, as both papers provide improves relative to the previous case. a contribution for a better understanding and Furthermore, the French banking system becomes monitoring of systemic risk. Regarding the first more important, inducing three hypothetical paper, Rodríguez Palenzuela welcomed the failures in both scenarios. Additionally, the effort to incorporate heterogeneous bank balance relevance of the German banking system in the sheets, as this is a first step away from the credit-plus-funding-shock simulation increases prevailing undifferentiated maximum entropy dramatically. Its collapse hypothetically causes technique. He also complimented the authors five other banking systems to fail. These new for showing how macroeconomic shocks, asset findings lead the authors to conclude that, market liquidity and network structure can although the data on risk transfers used in the cause system-wide credit losses and contagion paper is of bilateral nature only, the additional via interaction. insights gained from its use are noteworthy. 26 The authors assumed a 50% haircut in the fire sale of assets and a 65% roll-over ratio of interbank debt (M. Espinosa, J. Solé, At the end of his presentation Solé concluded and K. Giesecke (forthcoming) “Network analysis as a tool to with reflections on the policy implications assess cross-border financial linkages”, page 20). 27 This chapter in the IMF Global Financial Stability Report of of the presented work. First, he mentioned April 2009 was also written by J. Chan-Lau, M. Espinosa-Vega, that, with an increasing interconnectedness of K. Giesecke and J. Solé. ECB Recent advances in modelling systemic risk using network analysis 28 January 2010 DETAILED SUMMARY OF THE THEMES As regards possible improvements, the system risk and the unexplored nature of market discussant suggested the exploitation of real data failures. Until these challenges are dealt with, to conduct stochastic rather than deterministic it is hard to see network analysis to be applied parameter calibrations, given that assuming, more broadly for policy calibration. for example, zero recovery rates is rather restrictive. This is also true for the assumption In the subsequent discussion, comments that bank debt is completely illiquid. focused on the exploratory character of network By construction, this rules out the possibility analysis. that banks might be able to soften the impact of a shock through the replacement of debt In particular, the question whether and on what with equity. basis capital surcharges can be imposed upon systemically important firms or banks remains Another suggestion was made on the price a challenge for policy makers. Adding to this, impact of fire sales as well as the selection of the Solé stressed the point that financial institutions trigger point that leads banks to start the selling are usually not aware of how interconnected they of assets in the model. These distress parameters are. Consequently, they do not fully internalise – are modelled to be constant in the paper, whereas by setting aside additional capital buffers – the a varying adjustment depending on the state of network externalities they might cause. the economy would be more suitable. The same is true for the parameter reflecting the amount In addition, attention was drawn to the fact that of firms defaulting, since the underlying process network analysis, based on bank balance sheet driving the bankruptcies in the models is not models, often does not cause many players to fully elaborated. Hence, an early warning signal default in the simulations unless large shocks derived from the framework of the analysis are considered. However, as could be observed would be rather rigid due to the underlying during the recent crisis, contagion leading other constant values of its parameters. Furthermore, financial institutions to come close to bankruptcy Rodríguez Palenzuela questioned whether a does not necessarily need to be based on a large log-normal distribution can correctly capture initial shock. Therefore, future models need the fat tails correctly. He suggested using other to account for this, e.g. by incorporating risks distributions to install a more flexible structure stemming from high leverage. in the model. Based on these comments, Kapadia pointed to As regards the second paper, the discussant the need for more research on liquidity risk to praised its usefulness in assessing cross-border be conducted, as this has been a main feature financial stability risk using aggregated data. in the current crisis. Confirming the importance However, Rodríguez Palenzuela proposed of liquidity, Espinosa emphasised that they had incorporating country-specific default already included such risk in their simulations. In probabilities for first and second rounds instead the analysis, this indeed led to more contagious of simple country defaults. In his view, this defaults and made banking systems generally would permit a more efficient analysis of more vulnerable to shocks. vulnerabilities. Moreover, a cross-check using other indicators or approaches to examine the Furthermore, a consensus emerged in the robustness of the results was recommended. discussion that difficulties remain concerning the communication to decision makers as they The discussant concluded his presentation may not be fully familiar with network analysis pointing to the need for further elaboration in the and hence are frequently not convinced of its field of network analysis, especially regarding the usefulness. Nevertheless, network modelling unspecified role of time, the lack of fundamental can help to identify entities that are “critical” theorems, the definition of a central measure for for the stability of financial systems, providing ECB Recent advances in modelling systemic risk using network analysis January 2010 29 decision makers with arguments for policy discussions. In this context, the important role of the research community in making network analysis a useful tool for policy advice and to adapt it properly to supervisors’ toolboxes was underlined. ECB Recent advances in modelling systemic risk using network analysis 30 January 2010