Journal Issue

A network analysis of global banking

Javier Reyes, and Camelia Minoiu
Published Date:
April 2011
  • ShareShare
Show Summary Details

I. Introduction

Network analysis is increasingly recognized as a powerful methodological tool for modeling interactions between economic agents and assessing the resilience of financial systems to systemic risk.2 Financial interconnectedness is also becoming an important concern in macro-financial surveillance and has taken center-stage in discussions on prudential regulation policies.3 Network techniques have been used, among others, to describe the global architecture of cross-border financial flows,4 to analyze financial contagion,5 and to examine the dynamics of payment systems and interbank markets.6 Renewed interest in the application of network analysis tools to analyze economic interconnectedness was spurred by the 2008–09 global financial crisis, which due to its voracity and wide reach, provides a fertile experiment in the evolution of financial networks during times of stress.

In this paper, we explore the properties of the global banking network (henceforth ‘GBN’) using network analysis techniques. Using bilateral data on cross-border bank lending for 184 countries, we describe the topology of the network using different metrics of interconnectedness (such as country centrality and network clustering) and assess its dynamics over 1978–2009. We also document changes in the network around periods of financial stress such as systemic banking crises and sovereign debt crises. Our main findings are as follows. First, network metrics for the GBN tend to be volatile, with a number of structural breaks identifying waves of capital flows. Second, interconnectedness rankings of countries, especially for borrowers, are relatively unstable. Third, connectivity tends to diminish during and after financial crises, with the 2008– 09 crisis standing out as an unusually large perturbation to the network.

Our analysis contributes to the literature in several ways. To our knowledge, this is the first study that analyzes geographical patterns in the GBN using network techniques. Second, unlike earlier studies we focus on flows rather than exposures. While stocks of cross-border claims that capture inter-country exposures can be useful indicators of the potential for contagion, cross-border flows of financial capital reflect liquidity conditions in international markets. As such, they are a potentially informative source of variation during times of financial stress than slowly-evolving stocks of cross-border claims. Third, we focus not only on the presence or absence of bilateral relationships (i.e., the binary network), but extend the analysis to a weighted approach by adjusting the network statistics to reflect the intensity of financial flows across countries. This allows us to better capture the heterogeneity in cross-border financial links.7

The international bank lending market has expanded significantly in recent years.8 Net cross-border bank lending (new loans minus repayments) reached a high of $4.3 trillion in 2007 and plummeted to a negative of almost $1 trillion as the subprime crisis unfolded in 2008. Figures 12 show cross-border bank loans relative to other financial flows (FDI and portfolio investment) over 1980–2008. The figures underscore the importance of cross-border lending as a key source of finance for emerging and developing economies, having reached levels comparable to FDI prior to the 2008–09 crisis. Advanced economies are the main players in the GBN, with flows circulating among them that are roughly ten times higher than those to emerging and developing economies.9 We refer to these economies as the “core” of the network, while emerging and developing countries receiving liquidity from the core form the “periphery.”

Figure 1.Cross-border financial flows to advanced economies

Figure 2.Cross-border financial flows to emerging markets and developing economies

Source: Authors’ calculations based on World Economic Outlook (December 2010) and BIS locational banking statistics.

Note: Total bank lending flows (net) are from BIS reporting countries (with with complete data over 1978Q4–2009Q3) and are calculated by summing up flows across borrowers in each country group. Note that there is somedegree of overlap between bank lending flows and portfolio investment, with cross-border investments by banks in debt securities showing up in both variables.

Why does the network of banking relationships among countries matter above and beyond the actual cross-country bilateral flows? In the wake of the recent crisis it has been argued that network theories can enrich our understanding of the functioning of financial systems by helping model complexity, systemic risk, and the factors that cause seizures in financial markets (Haldane, 2009; Caballero, 2010). For instance, higher interconnectedness in the financial system is believed to improve risk sharing and reduce the risk of contagion (through better ability to absorb shocks) but to also increase it (through a wider outreach of reverberations). Financial systems have been shown to display robust-yet-fragile tendencies (Gai and Kapadia, 2010) and to react differently to shocks depending on the pattern of interconnectedness (Allen and Gale, 2000).10 A thorough knowledge of topological structures of real world financial markets could therefore be useful in developing models that can predict the observed patterns and forecast the reaction of the financial system to shocks.

Network graphs offer a unique perspective on the evolution of interconnectedness in the global banking market between 1980 and 2007 (Figure 3). Both the core-periphery and the core-core network have changed substantially over the last three decades, with the web of cross-border bank relationships expanding markedly as new links were born and cross-border flows increased in magnitude. More banking activity took place with the periphery in 2007 than in 1980, with ECA and MENA being the most integrated regions into the global market in 2007 (Panel A). The network links have been expanding outwards since 1980 particularly towards emerging Europe and international banking centers such as Bahrain, Cyprus, and Mauritius.11 Link proliferation is apparent in the core-core network as cross-border lending activity among advanced economies intensified, too (Panel B). Although Figure 3 clearly depicts the expansion of the GBN, more could be said about the characteristics and dynamics of the network through a thorough analysis of its topology. In what follows, we use network metrics to explore the features of the GBN, while paying special attention to waves of capital flows that took place in recent decades.

Figure 3.Network view of cross-border banking, 1980 and 2007

Source: Authors’ calculations using BIS locational banking statistics (yearly).

Note: The countries represent nodes, while the links between countries represent cross-border bank loans. Thicker and darker colored links indicate larger flows. In Panel B arrows indicate the direction of the flows. When bi-directional flows occur, the connecting links is split into two, each half-link reflecting the magnitude of one flow (hence may have a different color).

The remainder of this paper is structured as follows. In Section II we review previous studies that have borrowed analytical tools from network theory to analyze cross-border financial flows. Section III contains a description of our data and network metrics of interest. The topological properties of the GBN are discussed in Section IV, where we focus on both static and dynamic features of the network and analyze the distributional stability of network statistics, country rankings and ranking dynamics, and within-sample and out-of-sample dynamics. In Section V we document the behavior of the network around financial crises. Conclusions are presented in Section VI.

II. Related studies

Our work relates to a rapidly expanding line of research applying network analysis tools to analyze financial linkages in global markets and their implications for the emergence and management of systemic risk.12 This literature draws on the seminal work of Allen and Gale (2000), who use network theory to model financial interconnectedness and draw implications for system stability. Allen and Gale (2000) relate banking system resilience to shocks to its underlying structure in a stylized four-bank network that is either “complete” (i.e., every bank is connected to every other banks) or “incomplete” (i.e., every bank is connected with fewer than all banks). They find that complete networks are more resilient to shocks due to risk sharing and individual banks bearing a smaller share of the shock, while incomplete networks are more fragile since banks with fewer counter-parties have difficulty diffusing the shock. Using Allen and Gale’s setup, Leitner (2005) shows that interlinkages can be desirable even if they act as conduits for contagion, because they can motivate banks to bail out one another provided that they can coordinate to do so when the threat of contagion arises.

In response to concerns that the network structure considered in Allen and Gale (2000) was too stylized to reflect real-world financial systems, Nier et al. (2007) model a more complex network of banks with interlinked balance sheets to examine financial contagion as a function the bank capitalization, size of cross-exposures, and interconnectedness. They find that bank connectivity has a non-monotonic effect on contagious defaults: at small levels of connectivity, a small increase raises the likelihood of contagion, but in more interconnected networks, higher connectivity improves the ability of the financial system to absorb shocks.

The trade-off between shock absorption and shock diffusion in financial networks is a recurring theme in the economics literature, with complex network structures being seen as both better able to diversify away idiosyncratic risk and more capable of propagating financial distress. Battiston et al. (2010) assess how network density (measured as the number of connecting links) relates to systemic risk in a model of the economy as a credit network, arguing that while higher connectivity allows for improved risk sharing of distress propagation, it also leads to a mechanism of trend reinforcement: when an economic agent suffers a negative shock, her trade partners react by making her conditions even harder. Thus, financial fragility feeds on itself. With higher connectivity and trend reinforcement going hand in hand, a highly interconnected system will be at risk of experiencing avalanches of bankruptcies when negative shocks occur.

Macroeconomic complexity due to network relationships has been recently captured in models of panic during financial crises. Although not directly modeling the structure of the financial network, Caballero and Simsek (2009) bring to the fore the relevance of the complexity of a bank’s web of relationship by developing a model in which banks assess the health of their trading partners by collecting information about them. When financial stress affects the system, banks find it necessary to collect information not only about their immediate trading partners, but also about the trading partners of those trading partners, and so on. In a complex and highly interconnected network, there comes a point when the information gathering process becomes too costly and is abandoned, with banks withdrawing from loan commitments and illiquid positions, hence spreading the financial crisis. This complexity is endogenized in the model of Caballero and Simsek (2010) where banks facing plummeting asset prices and liquidity positions during severe financial crises, must understand a complex web of inter-connections when making financial decisions. The model thus displays a “complexity externality” in which market seizures occur because banks are increasingly reluctant to buy assets in a confusing and uncertain environment.

These theoretical contributions underscore the importance of documenting the topological properties of real-world economic and financial networks. The most extensively analyzed cross-border network is that of international trade flows (Fagiolo et al., 2009, 2010; Schiavo et al., 2010; Schweitzer et al., 2009). This line of research has documented a trade network with a core-periphery structure and an emerging “rich club” of countries that share high trade intensities and are highly-interconnected (with one another and the periphery) thus bridging different parts of the network. The trade network also appears to be stationary. Financial contagion through the trade network is analyzed in Kali and Reyes (2010), who use network-based measures of connectedness to explain abnormal stock market returns during recent crises, and find that contagion is more likely to occur when the epicenter country is better integrated into the trade network, while more interconnected downstream countries are better able to dissipate the impact.

Our study most closely relates to topological descriptions of global financial networks. Hattori and Suda (2007) analyze the core-periphery network of international bank exposures (stocks of claims) for 215 countries over 1985–2006. They find that the network has become more tightly connected and more clustered over time, displaying higher connectivity and shorter average path length13 in the recent period. Against the backdrop of an increasingly dense network, the authors posit that systemic risk is building up while risk and capital are also being allocated more efficiently.14 A key finding is that network features remain largely unperturbed during major events such as the LTCM near-collapse or the 1997–98 Asian crisis. In contrast, we will document a network of financial flows that is more turbulent than that of financial exposures― one that behaves differently during times of financial stress compared to tranquil periods.

In a similar vein, Kubelec and Sa (2010) and Sa (2010) assemble a large dataset of bilateral cross-border exposures by asset class (FDI, portfolio equity, debt, and foreign exchange reserves) for 18 advanced and emerging market economies, documenting a marked increase in financial interconnectedness over 1980–2005. The financial network has become more clustered with a lower average path length over time, and its central hubs are the United States and the United Kingdom. Comparisons with the international trade network reveal that both networks have experienced increased connectivity over time, despite the fact that trade openness (total trade over GDP) has risen less than financial openness (total assets and liabilities over GDP). Both trade and financial exposures distributions are long tailed,15 with a few countries exchanging large flows while most countries exchange much smaller flows––a pattern that hints at a core-periphery structure.

III. Data and Definitions

A. BIS locational statistics

Our data are the BIS locational statistics on exchange-rate adjusted cross-border bank credit over 1978–2009 (on a quarterly basis: 1978Q1–2009Q3) for 184 countries. Locational statistics are compiled on the basis of residence of BIS reporting banks (in a number of countries)16 and cover “the cross-border positions of all banks domiciled in the reporting area, including positions vis-à-vis their foreign affiliates” (Wooldridge, 2002, p. 80). These positions include loans, deposits, debt securities, and other assets provided by banks. Cross-border flows are then estimated as changes in cross-border exposures (stocks). Since locational cross-border exposures are also reported by currency, changes in stocks are estimated after accounting for exchange rate changes. Thus, exchange-rate adjusted estimates of cross-border flows, which we use in this paper, are considered a better approximation of true flows than unadjusted changes in stocks (Wooldridge, 2002).

Since, the BIS locational banking statistics capture “net flows of financial capital between any two regions channeled through the banking system” (McGuire and Tarashev, 2006, p. 34), they are best suited for analyzing geographical interlinkages, but less so for analyzing the global balance sheets of banks or funding risk (Fender and McGuire, 2010). These location-based data capture activities of all banks (domestic and foreign) operating in a particular location. Thus, to the extent that foreign banks are more active than domestic ones in any given location, the data will reflect that. The BIS locational banking statistics are thus well-suited for studying temporal patterns in financial linkages across countries, the subject of our study, but will have little to say about the drivers of such patterns (Fender and McGuire, 2010).

B. The network

Each of the 184 countries in our dataset is a node within the network. We model the financial flows as a directed network, with nodes being linked through cross-border lending. Banking flows between countries are the links. We work both with quarterly data and annual data (obtained by adding up flows across quarters), and model each quarter or year over the sample period (1978–2009) as a separate network.

The sample comprises 15 lenders (BIS reporting countries) and 169 borrowers. Thus, our full network has a core-periphery structure, with the “core” comprising the economies that act as lenders (and represent a sample of BIS reporting countries), and the “periphery” comprising countries that act as borrowers (non-BIS reporting countries). Since intra-core lending activity dwarfs in magnitude exchanges between the core and the periphery, in the remainder of the paper we focus on the core-periphery network (in which core countries act as lenders only) as opposed to the full network (in which core countries are borrowers as well). We also analyze the core-core network made up solely of “core” countries that are both lenders and borrowers (see the appendix for the list of countries).

From the dataset of bilateral banking flows, we construct our network of interest as follows. We retain all positive flows (corresponding to net increases in cross-border exposure or “net investments”) and replace the negative flows with zeros.17 The resulting matrix––one of net financial investments––thus comprises positive values (i.e., net investments) and zeros (i.e., net repayments or no flows).

C. Network indicators

The network metrics we use to study interconnectedness in the GBN include measures of country centrality (degree and strength) and network density (connectivity and clustering). We describe the network focusing not only on binary country-pair lending relationships, but also through a weighted approach to take into account the magnitude of flows across countries.18

We use bilateral (net) bank lending flows (adjusted for inflation using the US CPI) to build matrices Mt where rows represent lenders and columns represent borrowers. Each entry mijt is the value of the flow from lender i to borrower j at time t. These matrices can be transformed into their binary counterparts (At) where each cell aijt takes value 1 if the flow from country i to country j at time t is positive and 0 otherwise. The dimensions of these matrices (15 ×Nt for 15 lenders and Nt borrowers) change through the sample period as new countries enter the network.

Country centrality

Node degree counts the number of connections (links) for each country (node). Since we work with a directed network, we have in-coming links for borrowers and out-going links for lenders. Therefore, we compute the out-degree (the number of outgoing links) for each country by counting the countries to which it lends (its debtors) and in-degree (the number of incoming links) for each country by counting the countries from which it borrows (its creditors), as follows:

where A(i)t denotes the ith row of matrix At,(A(i)t)′ denotes the ith row of the transpose of matrix At and 1 is a unitary vector with Nt elements. The maximum value for in-degree is 15 as the sample of lenders does not change over time. The maximum number for out-degree (Nt) varies by year as countries enter the network. Note that node degree only uses information from the binary representation of the network19.

Node strength is the total value of flows originating or terminating in a given node. In our case, in-strength for country i(NSitin) is the total amount of cross-border credit it receives, whereas out-strength for country i(NSitout) is the total amount of cross-border credit it lends. Out-strength and in-strength are computed by substituting matrix At for matrix Mt in the node degree formulas presented above. Node strength is the simplest weighted network indicator that captures the intensity of financial relationships among countries.

Relative node strength focuses on the relative importance of lenders as providers of financial capital, and respectively, that of borrowers as destinations for financial investment in the network. Borrower j‘s dependence on lender i is the share of inflows it receives from i in her total borrowing. Hence, relative node out-strength increases with the lender’s relative importance as a creditor in the network. Defining a new matrix (M˜t) where each cell (m˜ijt) represents the ratio of inflows received by borrower j from lender i to j‘s total inflows, relative node out-strength is given by:

Similarly, lender i‘s dependence on borrower j as a destination for its financial capital is determined by the share of lending to j in her total lending. Defining matrix (M^t) where each cell m^ijt is the ratio of the inflow received by j from lender i to i‘s total lending, relative node in-strength is given by:

Note that the maximum relative out-strength is equal the total number of borrowers (in a network where one lender provides all the credit) while the maximum relative in-strength is the total number of lenders (in a network in which a borrower receives all the loans).20

Network density

Connectivity. Network connectivity is the number of links that exist between countries (or total node degree) expressed as a share of the total possible number of links. It represents the likelihood of connection between two countries in the GBN. Let mt be the observed number of links (corresponding to positive flows mijt>0 in the matrix Mt or 1’ s in the matrix At). With 15 lenders and Nt borrowers in the core-periphery network, the degree of connectivity is given by (mt15×Nt)..

Unweighted/binary clustering. We consider two clustering coefficients—a binary and a weighted clustering coefficient. The binary measure is the ratio between the total number of complete triangles (in which every country lends to every other country) and the total possible number of such triangles (see Text figure 1). For the core-core network, we take complete triangles to mean bi-directional lending among three countries i, j, and k from the core. For the core-periphery network, complete triangles are those where countries i and j from the core lend to one another and they both lend to periphery country k. (Complete triangles for the two networks are shown in Panel A.) The binary clustering coefficient ranges between 0 and 1, with higher values representing a more clustered network––one that displays a greater share of tripartite relationships. This particular definition also enables us to compute regional clustering coefficients for the core-periphery network by restricting k to belong to a certain region.

Weighted clustering. Our weighted clustering coefficient is inspired by the generalized node clustering coefficient proposed by Opsahl and Panzarasa (2009), which counts complete triangles as defined above and calculates their total value as the geometric mean of their weights (i.e., cross-country flows). The weighted clustering coefficient is obtained by dividing the total value of complete triangles by the total value of complete (three-sided) and incomplete (two-sided) triangles (as shown in Panel B).

Text figure 1.Triangles for clustering coefficients

Notes: Panel A shows complete triangles used to compute binary clustering coefficients. For the core-core network, all countries belong to the core and there are bi-directional links among all three. For the core-periphery network, countries i and j are in the core, while country k is in the periphery. Panel B shows the three types of triangles used in the denominator of the weighted clustering coefficient for both networks.

IV. Results

Here we describe the topology of the network using the measures of connectivity defined above, focusing on the time-evolution of their empirical distributions. The aim is to determine how the web of bilateral bank lending relationships has changed over 1978–2009. Occasionally we also draw comparisons between the GBN and the international trade network (henceforth ‘ITN’), thus contrasting two key channels of financial contagion––financial and trade linkages (Van Rijckeghem and Weder, 2001).

A. Network connectivity

Table 1 gives the summary statistics of our key network indicators for the full, core-periphery, and core-core networks. In the full GNB, countries borrow on average from 4.4 lenders up to a maximum of 15 (that is, all lenders), while the core countries lend to 48 borrowers, of which 8 are in the core and 40 in the periphery. Some $293 million are lent out on average across borders each quarter––$81 million to periphery countries and $212 to the core. The likelihood that two countries are connected through cross-border flows ranges between 26 and 44 percent in the full network, reaching 79 percent within the core before the current crisis. Similarly, the binary clustering coefficient in the full network––the probability that two countries are connected with one another if they both have a relationship with a third country––is about 11 percent. In the core, where clustering is defined more restrictively by requiring that relationships be bidirectional, the probability is 13 percent and has reached 27 percent before the 2008–09 crisis.

Table 1.Summary statistics of network indicators
Obs.MeanSt. Dev.MinMax
Measures of country centrality
Full network
Relative in-strength20,9030.10.305
Relative out-strength1,9049.98.4055
Relative in-strength20,9030.10.304
Relative out-strength1,9048.97.7051
Relative strength1,9051.00.906
Measures of network density
Full network
Binary clustering310.
Weighted clustering310.370.060.250.49
Binary clustering310.
Weighted clustering310.
Binary clustering310.
Weighted clustering310.570.120.230.81
Source: Authors’ calculations using BIS locational banking statistics (quarterly).Notes: Node strength is expressed in 2009 US$ million.
Source: Authors’ calculations using BIS locational banking statistics (quarterly).Notes: Node strength is expressed in 2009 US$ million.

The time-evolution of summary statistics for network indicator confirms that the last three decades have experienced several waves of financial globalization (Figure 4). The average number of outgoing links for each lender steadily increased from about 40 to a peak of 60 countries in 2007 before sharply dropping in 2008. Node strength (the total flows per country) has followed a similarly steep increase during the 2000s, followed by a drop in 2008 that is of historically unprecedented magnitude. Connectivity and clustering indicators for the network as a whole have been volatile, too.21 These features stand in contrast with the ITN, for which network metrics have remarkably stable empirical moments over the past decades (Fagiolo et al., 2010).

Figure 4.Trends in network indicators, 1978–2009

Source: Authors’ calculations using BIS locational banking statistics (quarterly and annual).

Notes: Results are shown for the core-periphery network. In the upper panels the smooth curves are nonparametric local polynomial smoothed estimates. In all panels we superimpose the dates of the two global waves of capital flows discussed in the text: 1987–98 and 2002–08 (as dated in IMF, 2007).

Looking closer at the fluctuations in network metrics, we can identify three global waves of cross-border bank lending. First, the build-up of financial interlinkages between the early 2000s upto to the current crisis stands out for all indicators. Second, a smaller surge in connectivity is evident prior to the 1997–98 Asian crisis. Third, the charts depict part of the build-up of cross-border lending that came to an end with the debt crisis of the early 1980s. The shifts in the topology of the GBN confirm our prior knowledge about the global waves of private capital flows as documented, for instance, in the IMF’s World Economic Outlook (2007).22

Another way to detect shifts in network connectivity over time is by analyzing the shape of the distributions of network metrics. Nonparametric kernel density estimates for degree and relative strength in 1980 and 2007 are shown in Figure 5. On the lender side, the density estimate for out-degree (the number of outgoing links) has preserved its shape over time, but has shifted to the right as lenders have provided financial capital to an increasing number of countries. Similarly, the distribution of in-degree (the number of incoming links) has shifted to the right as borrowers tap into a larger pool of lenders. The bimodality of the 2007 distribution of in-degree suggests that a large group of borrowers are emerging as particularly well-connected, i.e., borrowing from about 10 of the available lenders. The distribution of relative out-strength suggests that most lenders in 1980 provided about 5 percent of flows to any given borrower; by 2007, that number had shifted to about 10 percent, as relative importance has been shared more equitably among lenders. In contrast, the density of relative in-strength reveals a great degree of heterogeneity in terms of the importance of borrowers as destinations for liquidity. Inequality in terms of relative strength among borrowers has risen between 1980 and 2007, hinting at a more polarized periphery. As further documented in the next section, highly-connected borrowers share the periphery with a plethora of less relevant nodes.

Figure 5.Empirical distributions of network indicators

Source: Authors’ calculations using BIS locational banking statistics (annual).

Notes: Results are shown for the core-periphery network. The kernel density estimates use the Epanechnikov kernel and optimal bandwidth (Silverman, 1986).

We wrap up the analysis of the evolution of the empirical distributions through two-sample Kolmogorov-Smirnov tests. The null hypothesis is that the observed empirical distributions (at two points in time) are close enough to conclude that they are drawn from the same data generating process. We undertake the test for each indicator by comparing its empirical distribution function at the beginning of each decade with that in subsequent decades. We summarize the results in Table 2, where we report the proportion of years each decade with empirical distributions of network metrics that are unlike those at the beginning of the decade. Interestingly, the empirical distributions of lender centrality metrics are more stable than those of borrower centrality. Borrower connectivity, strength, and relative strength show much turbulence, with the distributions in 1980 being poor predictors of future ones. For instance, the landscape for borrowers through the 2000s is unlike that in previous decades, but seems unchanged since the year 2000, which suggests that periphery countries have been consolidating their relative positions in the GBN over the last decade.

Table 2.Kolmogorov-Smirnov tests
Relative out-

198019902000Relative out-

Source: Authors’ calculations using BIS locational banking statistics (annual).Notes: The table reports the proportion of years that display “statistically different” empirical distribution of the network indicators in each decade compared to the year indicated as column head. For instance, the figure 25 in the out-degree panel says that in 25 percent of years (i.e., 2 years) through the 2000s, the empirical distribution of out-degree was different than that in 1980. The statistical closeness of empirical distributions is assessed using Kolmogorov-Smirnov tests at the 5 percent level of significance.
Source: Authors’ calculations using BIS locational banking statistics (annual).Notes: The table reports the proportion of years that display “statistically different” empirical distribution of the network indicators in each decade compared to the year indicated as column head. For instance, the figure 25 in the out-degree panel says that in 25 percent of years (i.e., 2 years) through the 2000s, the empirical distribution of out-degree was different than that in 1980. The statistical closeness of empirical distributions is assessed using Kolmogorov-Smirnov tests at the 5 percent level of significance.

B. Country rankings, dynamics, and regional heterogeneity

Two questions arise from the patterns discussed in the previous section. The first is which countries and regions have been or currently are the most interconnected in the GBN. The second is whether the relatively stable shapes of the empirical distributions conceal turbulence of country rankings, which may be possible if countries swap places in terms of connectivity from one year to another.

We begin by reporting the top ten lenders in terms of node degree and strength (Table 3) for 1980, 1995, and 2007. The most globally connected lenders are Belgium, France, Germany, Switzerland, and the United Kingdom, with Japan, Switzerland, and the United States joining the top ranks in terms of absolute and relative volumes (Panel A). The rankings are relatively stable for lenders, but less so for borrowers, with liquidity appearing to follow different geographical patterns (Panel B).23 Before the debt crisis of the early 1980s, LAC countries (Argentina, Brazil, Chile, Mexico, and Venezuela) were the most central borrowers in the network. In 1995 they gave way to the fast-growing East Asian countries, while the BRICs began their own ascending path. By the end of the period, the BRICs had become the most interconnected borrowers alongside emerging Europe (e.g., the Baltics, Poland, Romania, and Ukraine).

Table 3.Country rankings by degree of interconnectedness
Panel ALendersPanel BBorrowers
BelgiumNetherlandsFranceEgyptThailandRussian Fed.
Out-strength UKJapanUKIn-strengthMexicoThailandRussian Fed.
LuxembourgLuxembourgAustriaRomaniaSouth AfricaTurkey
RelativeUKUKUKRelativeMexicoThailandRussian Fed.
GermanySwitzerlandUSVenezuelaRussian Fed. China
AustriaAustriaBelgiumAlgeriaSouth AfricaIndia
Source: Authors’ calculations using BIS locational banking statistics (annual).Notes: Panel A refers to our sample of BIS reporting countries only. Panel B refers to non BIS-reporting, middle-income countries.
Source: Authors’ calculations using BIS locational banking statistics (annual).Notes: Panel A refers to our sample of BIS reporting countries only. Panel B refers to non BIS-reporting, middle-income countries.

To obtain a broader view of ranking dynamics for all counties, we also calculate pair-wise Spearman rank-correlation coefficients for consecutive years ρt,t1(X) and define a Ranking Stability Index (RSI) for each network indicator s the time-average of the Spearman coefficients, as follows:

The RSI, which has the usual properties of a correlation coefficient, is useful in detecting shape-preserving ranking turbulence, which occurs when the empirical distributions of network indicators do not change but countries swap places in terms of centrality in the network.

The RSIs for degree and strength are depicted in Figure 6. Lender rankings seem more stable than borrower rankings, becoming less volatile after the debt crisis of the early 1980s and stabilizing at about 0.8. In contrast, borrower rankings based on the number incoming links are more volatile (0.2 on average over the full period), and even negative in the late 1970s. This suggests that the relatively stable empirical distributions of node in-degree (seen in Figure 5) conceal degree great deal of rankings turbulence. In contrast, rankings based on strength have become more stable since the early 1990s, leveling off at about 0.7 during the 2000s. This pattern suggests that a group of strong borrowers has emerged and is consolidating its place as a central destination for cross-border lending in the recent years. Closer inspection of borrower rankings reveals that this strong group comprises primarily the BRICs and emerging European countries (Figure 7). Reyes et al. (2008) show that through the 1990s the BRICs have moved from the periphery into the core of the ITN as well, reaching the top in terms of network centrality (Reyes et al., 2008).). Nevertheless, the ITN has relatively less volatile country rankings than the GBN.

Figure 6.Ranking stability indices

Source: Authors’ calculations using BIS locational banking statistics (annual).

Notes. The results are based on the sub-sample of borrowers present throughout the sample period.

Figure 7.Average rankings for the BRICs and emerging Europe

Source: Authors’ calculations using BIS locational banking statistics (annual).

Notes. The countries grouped under ‘emerging Europe’ comprise Latvia, Lithuania, Poland, Romania, and Ukraine. In both panels we superimpose the dates of the second global wave of capital flows discussed in the text: 2002–08.

Global averages of network metrics conceal interesting regional heterogeneity (Figure 8). During the 2000s, the ECA region (including emerging Europe) has been the most integrated region into the GBN, with the highest level of connectivity (in-degree), average inflows per country (in-strength), and clustering (complete triangles with countries in the core). Average and total flows for EAP display a strong cyclicality, with a peak in cross-border lending to this region occurring prior to the Japanese crisis in the early 1990s. For other regions, network metrics tell a different story compared to total flows. The MENA region comes fourth based on total lending, but ranks higher up in terms of in-degree and binary clustering. Although during the 2000s, SA (driven by India) has become more interconnected by tapping into a larger pool of lenders, cross-border lending volumes remain relatively low. Further, total lending to LAC falls markedly in the wake of the early-1980s debt crisis but surges prior to the 2001 Argentine crisis and the recent downturn. However, prior to the end of the sample period, the LAC region never attains the degree of connectivity it enjoyed prior to the debt crisis of the early 1980s. Similarly, EAP experienced three surges in cross-border lending, but these are relatively muted in network connectivity terms. In all charts, the downturn in network connectivity and clustering during the 2008–09 crisis stands out as an unusually large perturbation to the network.

Figure 8.Regional heterogeneity: degree, strength, and clustering

Source: Authors’ calculations using BIS locational banking statistics (annual).

Notes: Results for North America/Oceania are omitted from all charts for visual tractability. In the bottom right panel, the figures for 2009 are upto Q3 inclusive.

C. In-sample dynamics and out-of-sample evolution

We now turn to characterizing the distributional dynamics of the GBN over 1978–2009. We have found the fluctuations in network features roughly coincide with previously documented global waves of private capital flows. To formalize the dates of the waves, we carry out unit root tests with one and two structural breaks in the mean (Clemente, Montanes, and Reyes, 1998) for all four indicators (average degree, average strength, network connectivity and network clustering). The results (Table 4) show that one-break tests identify a break in 2001 or 2003, while two-break tests find the first mean shift either in 1980 (before the debt crisis) or mid-1990s (before the Asian crisis) and the second one in 2002–03. In what follows we ignore the first break in the early 1980s since our data is insufficient to gauge the build-up of that wave, and restrict our attention to the behavior of the GBN during the more recent windows 1987–98 and 2002–08 (dated as in IMF, 2007).

Table 4.Unit root tests for empirical moments of network indicators
1 breakp-value2 breaksp-value




Binary clustering20030.016199620030.6790.010
Weighted clustering20040.142199620020.1210.031
Source: Authors’ calculations using BIS locational banking statistics (annual).Notes: The results pertain to the core-periphery network. Years in boldface identify the structural breaks that are statistically significant at the 5 percent level.
Source: Authors’ calculations using BIS locational banking statistics (annual).Notes: The results pertain to the core-periphery network. Years in boldface identify the structural breaks that are statistically significant at the 5 percent level.

These two windows provide an interesting setting for an exercise about out-of-sample dynamics. Over the past three decades the global financial architecture has been constantly reshaped by financial and debt crises, and different regulatory and macroeconomic policy regimes. Here we seek to map the in-sample dynamics of the GBN observed within the two waves into long-term dynamics. We wish to determine how the global financial architecture would have looked in the long run if in-sample dynamics were the same as in 1987–98 and 2002–08, respectively. We proceed in two steps. First, we analyze the within-window transitional dynamics using stochastic kernel estimates and a network-wide statistic. We then estimate long-run, limiting distributions for the network indicators.

Transitional dynamics can be visualized using stochastic kernel estimates, which we present in contour plots. These depict the conditional density estimate for a random variable that governs the transition of countries from one quantile of a network statistic distribution to another quantile. For illustration, we compute the 1-year stochastic kernel estimates for in-degree and relative in-strength (Figure 9). For the number of incoming links, most of the probability mass is concentrated around the 45-degree line in the middle range, meaning that countries in the middle quantiles of the distribution tend to stay in the same quantile from one year to another. In contrast, countries in the upper quantiles tend to ascend towards higher quantiles, which means that once a country establishes borrowing relationships with many lenders, in the next period chances are that it will borrow from even more lenders. This pattern is clearer still during the second wave of capital flows. For relative in-strength a “middle-class” of globally-interconnected countries seems to be emerging in the second wave.

Figure 9.Stochastic kernel density estimates

Source: Authors’ calculations using BIS locational banking statistics (annual).

Notes: The results are based on the sub-sample of countries present through the sample period.

Another measure of the persistence of distributional dynamics is the empirical M-statistic, which aggregates country-level probabilities of moving across quantiles (of network statistic X) from one year to the next. Assuming that the distributional dynamics of our network statistics can be approximated well by 10 quantiles (or deciles), we construct a Markov transition matrix by estimating 1-step transition probabilities using the maximum likelihood estimator for the probability of migrating from one decile to another: 24

where nijt1,t is the number of countries transitioning from the ith to the jth decile (of each network metric X) between t-1 and t and nit is the number of countries in decile i at time t. Since the network indicators considered present structural breaks which identify distinct waves of capital flows (1987-98 and 2002-08), we compute our measure of distributional dynamics––the M-statistic––within each window as the time-average of:

where parameter ω = 0 corresponds to countries staying in the same quantile and ω = 1 corresponds to countries moving one quantile up or down.25 The M-statistic is bounded between 0 and 1, with higher values indicating a higher probability that countries stay in the same or adjacent quantile in consecutive years.

Distributional dynamics display little persistence as our empirical M-statistics are relatively low (Table 5), ranging between 0.2 and 0.5 for parameter value ɷ=0 (staying in the same quantile). For the second window (2002–08) the M statistics are higher, which is consistent with the emergence of a group of borrowers that have been consolidating their position in the network. We also report 95 percent confidence intervals for the M-statistics corresponding to a “random” network.26 The M-statistics for the GBN always fall to the right of the confidence interval upper bounds for the random network, suggesting that the GBN is more stable than a random network. Nevertheless, they are lower than the values of 0.8 and 0.9 documented for the trade network (Fagiolo et al., 2010), hinting at in-sample dynamics of financial flows that are more turbulent than trade links.

Table 5.Empirical M-statistics
First window: 1987–98ω = 0ω =1
Node degree0.2134

(0.0872, 0.1185)

(0.2534, 0.2991)
Node strength0.3583

(0.0836, 0.1159)

(0.2561, 0.3065)
Relative node strength0.3539


Second window: 2002–08ω=0ω =1
Node degree0.2536

(0.0774, 0.1177)

(0.2496, 0.3150)
Node strength0.4363

(0.0780, 0.1200)

(0.2486, 0.3117)
Relative node strength0.4556

(0.0780, 0.1223)

(0.0780, 0.1223)
Source: Authors’ calculations using BIS locational banking statistics (annual).Notes: 95 percent confidence intervals for the M statistics for a “random” network are shown in parentheses (see text for explanations).
Source: Authors’ calculations using BIS locational banking statistics (annual).Notes: 95 percent confidence intervals for the M statistics for a “random” network are shown in parentheses (see text for explanations).

Finally, we investigate the implications of these transitional dynamics for the out-of-sample evolution of the GBN by estimating the ergodic distributions for in-degree, taking the first year of each window as the base year.27Figure 10 depicts the long-run stabilizing tendency of the in-degree distribution arising from the transitional dynamics observed within each wave of capital flows (and taking the mid-year as the base). The limiting distribution of country connectivity is unimodal in the first wave and bimodal in the second wave. A more “diversified” long-run financial landscape emerges based on second wave transitional dynamics––in which there are more highly-connected borrowers than previously. This is consistent with the 1990s being dominated by a single region––East Asia––as the leading destination of cross-border bank lending, while the 2000s witnessed the emergence of a number of economies (such as the BRICs and emerging Europe) as key borrowers of financial capital.28

Figure 10.Ergodic distributions for node degree

Source: Authors’ calculations using BIS locational banking statistics (annual).

Note: The ergodic distributions are computed based on transitional dynamics in each window, after restricting the sample to the countries present in the dataset throughout the full period. The base year is 1992 for the first window and 2005 for the second window.

V. The global banking network during financial crises

We close our exploration of the topological features of the GBN by describing its behavior before, during, and after financial crises. Cetorelli and Goldberg (2010) have identified cross-border bank linkages as a key transmission channel of the 2007–08 US subprime crisis to emerging market countries, arguing that domestic loan supply in afflicted economies contracted due to the collapse of direct cross-border lending by foreign banks and a general weakening of balance sheets of both foreign affiliates and domestic banks caused by shortages of interbank and cross-border liquidity. To provide evidence of cross-border lending acting as a potential conduit of the negative shock both inside the core and towards the periphery, we present network visualizations for 2007Q4 and 2008Q4—just before and after the crisis (Figure 11). There is markedly lower intra-core and core-periphery connectivity at end-2008, with links disappearing as cross-border banking flows dried up (Hoggarth et al., 2009; Milesi-Ferretti and Tille, 2011).

Figure 11.Global banking network before and during the 2008-09 crisis

Source: Authors’ calculations using BIS locational banking statistics (quarterly).

Notes: The countries represent nodes, while the links between countries represent cross-border bank loans. Thicker and darker colored links indicate larger flows. In Panel B arrows indicate the direction of the flows. When bidirectional flows occur, the connecting links is split into two, each half-link reflecting the magnitude of one flow (hence may have a different color).

To determine how periods of financial distress correlate with features of the GBN, we plot network connectivity and clustering against synchronized recession dates (Figure 12). These include the 1987 stock market crash, the 1991–92 Scandinavian banking crises and 1992–93 ERM crisis, the 1998Q3 LTCM near-collapse, the 2000Q2 Internet bubble collapse, and the 2008Q3 Lehman Brothers bankruptcy. Many links die and network clustering diminishes during episodes of financial stress. This pattern is even clearer when we move to the regional level to consider regional network density during the crises of Latin America and East Asia (Figure 13).

Figure 12.Network density during financial crises

Source: Authors’ calculations using BIS locational banking statistics (annual).

Notes: Highlighted areas represent synchronized recession dates (IMF, 2007), namely the 1987Q3 stock market crash, 1991–92 Scandinavian banking crises, 1992 ERM crisis, 1998Q3 LTCM near-collapse, 2000Q2 Internet bubble collapse, and 2008Q3 Lehman Brothers bankruptcy.

Figure 13.Regional clustering for LAC and EAP regions during financial crises

Source: Authors’ calculations using BIS locational banking statistics (annual).

Notes: For LAC we superimpose dates for the onset of systemic banking crises for Argentina, Brazil, and Mexico. For EAP we superimpose dates for the onset of systemic banking crises for Japan, Indonesia, Korea (Republic of), Malaysia, Philippines, Thailand, Vietnam, and Mongolia. Crisis dates have been taken from Laeven and Valencia (2010).

How country connectivity in the GBN behaves during financial crises is depicted in Figures 1415, where we plot average degree and strength (across countries) around systemic banking and sovereign debt crisis episodes.29 The averages are shown for 5 years before and after the onset of the crisis. Without exception, network indicators of borrower connectedness in the GBN fall during crises, although the decline generally begins before the event.30 The pattern holds up for lenders despite the paucity of systemic banking crises in the core prior to 2007. For sovereign debt crises, we only show the evolution of connectedness for borrowers, and document the same deleterious impact. Borrowers afflicted by debt crises do not attain pre-crisis connectedness levels within the following five years. When incorporating 2007–08 crises into the sample and restricting the window to −5/+1 years around the episode (Figure 15), all connectedness measures for lenders sharply decline one year after the crisis, reflecting the unusually large impact of the 2008–09 episode on the GBN.

Figure 14.Interconnectedness before and after financial crises

Source: Authors’ calculations using BIS locational banking statistics (annual) and Laeven and Valencia (2010) database for systemic banking and sovereign debt crises dates.

Notes: Results are based on the full network. The left panels include systemic banking and sovereign debt crises that occurred between 1985 and 2003 and are at least 10 years apart, which allows for a 5-year non-overlapping window around them. (Countries with two crises within 10 years are dropped, but the results are robust to retaining either crisis.)

Figure 15.Interconnectedness before and after systemic banking crises

(including 2007–08 episodes)

Source: Authors’ calculations using BIS locational banking statistics (annual) and Laeven and Valencia (2008) database for systemic banking crises.

Notes: Results are based on the full network. Data for 2009 is upto Q3 inclusive.

We formalize the analysis by estimating a simple panel specification that traces the evolution of network measures around crisis dates. Specifically we regress country-level indicators of network centrality (degree, strength, and relative strength) on a set of dummies for crisis and post-crisis years while controlling for country fixed effects. The results (Table 6) indicate that banking and debt crises are associated with reduced borrower access to capital markets as measured by network degree (number of links) and strength (total flows). Cross-border inflows to the afflicted country decrease by $24 million at the onset of the crisis, and $30 million on average over the subsequent five years (columns 1−3). On the lending side, the F tests of joint significance of coefficient estimates on the lags cannot reject a nil effect of banking crises on lender connectedness (columns 4−6), which may be explained by the lack of variation in the sample. The negative effects of sovereign debt crises are also evident for borrowers (columns 7−9), but are smaller in magnitude. These results complement the literature on access to capital markets in the aftermath of debt crises (e.g., Ozler, 1993; Arteta and Hale, 2008; Fuentes and Saravia, 2009). Furthermore, by looking beyond lending volumes, they represent a deeper characterization of access to global financial markets in the aftermath of crises.

Table 6.Interconnectedness during financial crises: Regression estimates
Systemic banking crisesSovereign debt crises
In-degreeIn-strengthRelative in-

Out-degreeOut-strengthRelative out-

In-degreeIn-strengthRelative in-

Year of crisis-0.75***









1 year later-1.66***









2 years later-1.19***









3 years later-0.86***









4 years later-0.67***









5 years later-0.45*









F test joint significance:
test statistic and p-value60.
Number of countries868686868686414141
Source: Authors’ estimations using BIS locational banking statistics (annual)Notes: Robust standard errors in parentheses (*** p<0.01; ** p<0.05; * p<0.1). The F-test of joint significance is for all coefficient estimates on lagged variables.
Source: Authors’ estimations using BIS locational banking statistics (annual)Notes: Robust standard errors in parentheses (*** p<0.01; ** p<0.05; * p<0.1). The F-test of joint significance is for all coefficient estimates on lagged variables.

VI. Conclusions

The potential usefulness of network techniques in analyzing systemic risk has taken center-stage in academic and policy debates in the aftermath of the 2008–09 global crisis. Nevertheless, little is known about the structural properties and time-evolution of the network of cross-country financial linkages, which are key to understanding how the global financial system reacts to shocks, and whether and where systemic risk may emerge. In this paper we have analyzed the properties of the global cross-border bank lending market (for 184 countries over 1978–2009) using network metrics of interconnectedness such as centrality, connectivity, and clustering. We have also sought to determine how geographical linkages changes around financial crises.

We found that the global banking network is relatively unstable; we have identified structural breaks in network connectivity and centrality, and documented volatile interconnectedness rankings for countries, especially borrowers. In the 2002–08 wave of global capital flows, the BRIC countries and high-growth Europe emerge and consolidate their position as highly-interconnected borrowers. Network density expands and contracts, following the cycle of capital flows. Furthermore, country centrality falls at the onset and in the aftermath of banking and sovereign debt crises. While the global banking network appeared more stable in the second half of the 2000s than in earlier periods, the 2008–09 global financial crisis stands out as an unusually large perturbation.

A number of questions emerge from our analysis. While it has been established that cross-border bank lending is a key channel of transmission of financial crises, how the topology of financial networks relate to the emergence of systemic risk remains underexplored. How do the properties of the different networks–– banking, FDI, trade, and remittances––compare and how do countries’ degrees of connectedness interact in different webs of relationships? What is the empirical relationship between connectedness and the way in which shocks get amplified or diffused? These and related questions remain interesting avenues to explore in future work.


    Allen, F. and A.Babus, 2009, “Networks in finance,” in P.Kleindorfer and J.Wind(Eds.), Network-based Strategies and Competencies, pp. 367-382, Wharton School Publishing.

    • Search Google Scholar
    • Export Citation

    Allen, F. and D.Gale, 2000, “Financial contagion,” The Journal of Political Economy, Vol. 108(1), pp. 133.

    Alessandri, P., Gai, P., Kapadia, S., Mora, N. and C.Puhr, 2009, “Towards a framework for quantifying systemic stability,” International Journal of Central Banking, Vol. 5(3), pp. 47-81.

    • Search Google Scholar
    • Export Citation

    Arteta, C. and G.Hale, 2008, “Sovereign debt crises and credit to the private sector,” Journal of International Economics, Vol. 74(1), pp. 53-69.

    • Search Google Scholar
    • Export Citation

    Barrat, A., Barthelemy, B., Pastor-Sattoras, R. and A.Vespignani, 2004, “The architecture of complex weighted networks,” Proceedings of the National Academy of Sciences, Vol. 101(11), pp. 3747-3752.

    • Search Google Scholar
    • Export Citation

    Battiston, S., Gatti, D. D., Gallegati, M., Greenwald, B. and J.Stiglitz, 2010, “Liaisons dangereuses: Increasing connectivity, risk sharing, and systemic risk,” NBER Working Paper No. 15611 (Cambridge, MA: National Bureau of Economic Research).

    • Search Google Scholar
    • Export Citation

    Caballero, R., 2010, “Macroeconomics after the crisis: Time to deal with the pretense-of-knowledge syndrome,” Journal of Economic Perspectives, Vol. 24(4), pp. 85-102.

    • Search Google Scholar
    • Export Citation

    Caballero, R. and A.Simsek, 2010, “Fire sales in a model of complexity,” MIT Department of Economics, unpublished manuscript (Cambridge, MA: Massachusetts Institute of Technology).

    • Search Google Scholar
    • Export Citation

    Caballero, R. and A.Simsek, 2009, “Complexity and financial panics,” NBER Working Paper No. 14997 (Cambridge, MA: National Bureau of Economic Research).

    • Search Google Scholar
    • Export Citation

    Cechetti, S.G., Fender, I., and P.McGuire, 2010, “Toward a global risk map,” BIS Working Paper No. 309 (Basel: Bank of International Settlements).

    • Search Google Scholar
    • Export Citation

    Cetorelli, N. and L. S.Goldberg, 2010, “Global banks and international shock transmission: Evidence from the crisis,” IMF Economic Review, forthcoming.

    • Search Google Scholar
    • Export Citation

    Chan-Lau, J. A., 2010, “Balance sheet network analysis of too-connected-to-fail risk in global and domestic banking systems,” IMF Working Paper No. 10/107 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    Clemente, J., Montanes, A., and M.Reyes, 1998, “Testing for a unit root in variables with a double change in the mean,” Economics Letters, Vol. 59, pp. 175-182.

    • Search Google Scholar
    • Export Citation

    Degryse, H., Elahi, M. A., and M. F.Penas, 2010, “Cross-border exposures and financial contagion,” International Review of Finance, Vol. 10, pp. 209–240.

    • Search Google Scholar
    • Export Citation

    Espinosa-Vega, M. and J.Sole, 2010, “Cross-border financial surveillance: A network perspective,” IMF Working Paper No. 10/105, (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    European Central Bank, 2010a, “Financial Stability Report,” June2010, Section IV Special Features Financial networks and financial stability, pp. 155-160 (Frankfurt: European Central Bank).

    • Search Google Scholar
    • Export Citation

    European Central Bank, 2010b, “Recent advances in modeling systemic risk using network analysis,” January (Frankfurt: European Central Bank). Available on the Internet: (accessed on January3, 2011).

    • Search Google Scholar
    • Export Citation

    Fagiolo, G., Reyes, J., and S.Schiavo, 2010, “The evolution of the world trade web: A weighted network analysis,” Journal of Evolutionary Economics, Vol. 20(4), pp. 479-514.

    • Search Google Scholar
    • Export Citation

    Fagiolo, G., Reyes, J., and S.Schiavo, 2009, “World-trade web: Topological properties, dynamics, and evolution,” Physical Review E, Vol. 79(3), pp. 79–33.

    • Search Google Scholar
    • Export Citation

    Fender, I. and P.McGuire, 2010, “Bank structure, funding risk and the transmission of shocks across countries: Concepts and measurement,” BIS Quarterly Review, September, pp. 63–79.

    • Search Google Scholar
    • Export Citation

    Fuentes, M. and D.Saravia, 2009, “Sovereign defaulters: Do international capital markets punish them?” Journal of Development Economics, Vol. 91(2), pp. 336–347.

    • Search Google Scholar
    • Export Citation

    Fujii, M. and M.Takaoka, 2010, “Structure of the financial system and the contagion effect––A network approach,” Financial Review, Vol. 3(101), pp. 98–118.

    • Search Google Scholar
    • Export Citation

    Gai, P. and S.Kapadia, 2010, “Contagion in financial networks,” Proceedings of the Royal Society A, Vol. 466, No. 2120, pp. 2401–2423.

    • Search Google Scholar
    • Export Citation

    Georg, C.-P., 2010, “The effect of the interbank network structure on contagion and financial stability,” Global Financial Markets Working Paper No. 12 (Halle, Germany: University of Jena).

    • Search Google Scholar
    • Export Citation

    Goyal, S., 2007, Connections: An Introduction to the Economics of Networks. Princeton University Press.

    Haldane, A., 2009, “Rethinking the financial network,” Speech delivered at the Financial Student Association Conference in Amsterdam onApril28.

    • Search Google Scholar
    • Export Citation

    Hattori, M. and Y.Suda, 2007, “Developments in a cross-border bank exposure network,” Bank of Japan Working Paper No. 07-E-21.

    Hoggarth, 3G., Mahadeva, L. and J.Martin, 2009, “Understanding international bank capital flows during the recent financial crisis,” Bank of England Financial Stability Paper No. 8 (London: Bank of England).

    • Search Google Scholar
    • Export Citation

    International Monetary Fund, 2010a, “Integrating Stability Assessments under the Financial Sector Assessment Program into Article IV Surveillance: Background Material,” Background paper prepared by the Monetary and Capital Markets Department, August27 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    International Monetary Fund, 2010b, “Understanding financial interconnectedness,” Staff paper prepared by the Strategy, Policy, and Review Department and the Monetary and Capital Markets Department, October4 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    International Monetary Fund, 2009, “Assessing the systemic implications of financial linkages,” Global Financial Stability Report, April2009, Chapter 2 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    International Monetary Fund, 2007, “Managing large capital inflows,” World Economic Outlook, October2007, Chapter 3 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    Iori, G., de Masi, G., Precup, O.V., Gabbi, G., and G.Caldarelli, 2008, “A network analysis of the Italian overnight money market,” Journal of Economic Dynamics and Control, Vol. 32(1), pp. 259-278.

    • Search Google Scholar
    • Export Citation

    Kali, R. and J.Reyes, 2010, “Financial contagion on the international trade network,” Economic Inquiry, Vol. 48(8), pp. 1072-1101.

    Kubelec, C. and F.Sa, 2010, “The geographical composition of national external balance sheets: 1980–2005,” Bank of England Working Paper No. 384.

    • Search Google Scholar
    • Export Citation

    Laeven, L. and F.Valencia, 2010, “Resolution of banking crises: The good, the bad, and the ugly,” IMF Working Paper No. 10/146 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    Laeven, L. and F.Valencia, 2008, “Systemic banking crises: A new database,” IMF Working Paper No. 08/224 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    Leitner, Y., 2005, “Financial networks: Contagion, commitment, and private sector bailouts,” Journal of Finance, Vol. 60(6), pp. 2925–2953.

    • Search Google Scholar
    • Export Citation

    Martinez-Jaramillo, S., Perez, O.P., Embriz, F. A., and F. L. G.Dey, 2010, “Systemic risk, financial contagion, and financial fragility,” Journal of Economic Dynamics and Control, Vol. 34, pp. 2358–2374.

    • Search Google Scholar
    • Export Citation

    McGuire, P. and N.Tarashev, 2006, “Tracking international bank flows,” BIS Quarterly Review, December, pp. 27–40.

    Milesi-Ferretti, G.-M. and C.Tille, 2011, “The great retrenchment: International capital flows during the global financial crisis,” Economic Policy, forthcoming.

    • Search Google Scholar
    • Export Citation

    Nagurney, A., 2003, “Innovations in financial and economic networks,” in Financial and Economic Networks: An Overview, Chapter 1, pp. 1–25. Edward Elgar Publishing.

    • Search Google Scholar
    • Export Citation

    Nier, E., Yang, J., Yorulmazer, T. and A.Alentorn, 2007, “Network models and financial stability,” Journal of Economic Dynamics and Controls, Vol. 31(6), pp. 2033-2060.

    • Search Google Scholar
    • Export Citation

    Opsahl, T. and P.Panzarasa, 2009, “Clustering in weighted networks,” Social Networks, Vol. 31(2), pp. 155-163.

    Ozler, S., 1993, “Have commercial banks ignored history?” American Economic Review, Vol. 83(3), pp. 608–620.

    Reyes, J., Schiavo, S., and G.Fagiolo, 2008, “Assessing the evolution of international economic integration using random walk betweeness centrality: The cases of East Asia and Latin America,” Advances in Complex Systems, Vol. 1(5), pp. 685–702.

    • Search Google Scholar
    • Export Citation

    Sa, F., 2010, “Evidence on financial globalization and crises: Geographic/bilateral external balance sheets,” Trinity Collge, University of Cambridge Working Paper No. 1038 (Cambridge, UK: University of Cambridge).

    • Search Google Scholar
    • Export Citation

    Schiavo, S., Reyes, J., and G.Fagiolo, 2010, “International trade and financial integration: A weighted network analysis,” Quantitative Finance, Vol. 10(4), pp. 389-399.

    • Search Google Scholar
    • Export Citation

    Schweitzer, F., Fagiolo, G., Sornette, D., and F.Vega-Redondo, 2009, “Economic networks: What do we know and what do we need to know?” Advances in Complex Systems, Vol. 12(4 & 5), pp. 407–422.

    • Search Google Scholar
    • Export Citation

    Scott, J., 2009, Social Network Analysis: A Handbook, 2nd edition, SAGE Publications, London, Thousand Oaks, New Delhi.

    Silverman, B. W., 1986, Density Estimation for Statistics and Data Analysis, Chapman & Hall.

    Song, D.-M., Jiang, Z.-Q., and W.-X.Zhou, 2009, “Statistical properties of world investment networks,” Physica A, Vol. 388(12), pp. 2450–2460.

    • Search Google Scholar
    • Export Citation

    Soramaki,K., Bech, M. L., Arnold, J., Glass, R. J., and W. E.Beyeler, 2007, “The topology of interbank payment flows,” Physica A, Vol. 379, pp. 317–333.

    • Search Google Scholar
    • Export Citation

    Van Rijckeghem, C. and B.Weder, 2001, “Sources of contagion: Is it finance or trade?” Journal of International Economics, Vol. 54(2), pp. 293-308.

    • Search Google Scholar
    • Export Citation

    Waysand, C., Ross, K., and Guzman, 2010, “European financial linkages: A new look at imbalances,” IMF Working Paper No. 10/295 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    Wooldridge, P. D., 2002, “Uses of the BIS statistics: An introduction,” BIS Quarterly Review, March2002.

List of countries used in the analysis
Core (Sample of BIS

reporting countries)1/
Periphery (Other countries)2/
AustriaAfghanistanCzech RepublicKuwaitParaguayTurkey
BelgiumAlbaniaCôte d’IvoireKyrgyz RepublicPeruTurkmenistan
FranceArgentinaDominican Republic LebanonPortugalUnited Arab Emirates
ItalyAustraliaEl SalvadorLibyaRussiaVanuatu
JapanAzerbaijanEquatorial GuineaLithuaniaRwandaVenezuela
LuxembourgBahamasEritreaMacau SARSamoaVietnam
NetherlandsBahrainEstoniaMacedonia, FYRSao Tomé and PrincipYem en
SwedenBangladeshEthiopiaMadagascarSaudi ArabiaZambia
United KingdomBelarusFinlandMalaysiaSerbia
United StatesBelizeGabonMaldivesSeychelles
BeninGambiaMaliSierra Leone
Bosnia and Herzegovina GreeceMauritiusSlovenia
BotswanaGrenadaMexicoSolomon Islands
BruneiGuineaMoldovaSouth Africa
BulgariaGuinea-BissauMongoliaSouth Korea
Burkina FasoGuyanaMontenegroSpain
BurundiHaitiMoroccoSri Lanka
CambodiaHondurasMozambiqueSt. Lucia
CameroonHong Kong, SARMyanmarSt. Vincent
Cape VerdeHungaryNamibiaSudan
Central African Republic IcelandNepalSurinam
ChadIndiaNetherlands Antilles Swaziland
ChileIndonesiaNew ZealandSyria
Comoros IslandsIsraelNigeriaThailand
Congo, Rep.JamaicaNorwayTimor Leste
Congo, DRCJordanOmanTogo
Costa RicaKazakhstanPakistanTonga
CroatiaKenyaPanamaTrinidad and Tobago
CyprusKiribatiPapua New GuineaTunisia

Includes BIS reporting countries with complete data from 1978Q4 onwards.

Includes all countries for which the sample BIS reporting countries provide data.

Includes BIS reporting countries with complete data from 1978Q4 onwards.

Includes all countries for which the sample BIS reporting countries provide data.


International Monetary Fund (Minoiu) and University of Arkansas Sam M. Walton College of Business (Reyes). Part of this work was undertaken while Javier A. Reyes was a Visiting Scholar at the IMF Institute (October 2010). We are grateful to Eugen Tereanu for his invaluable contribution at the early stages of this project. We thank the Bank of International Settlements (BIS) for providing us with data on BIS locational banking statistics on a bilateral basis and to Patrick Imam, Tümer Kapan, Patrick McGuire, Swapan-Kumar Pradhan, Marc Quintyn, Mahvash S. Qureshi, Stefano Schiavo, and INSAC seminar participants for useful comments and discussions. All errors are our own.


For recent contributions, see, Nier et al. (2007), Alessandri et al. (2009), and Cechetti et al. (2010). Goyal (2007) provides an excellent introduction to network theory. For reviews of its applications in economics and finance, see Nagurney (2003) and Allen and Babus (2009).


Haldane (2009), ECB (2010a, 2010b), and IMF (2009, 2010a, 2010b).


See, e.g., Soramaki et al. (2007) and Iori et al. (2008) for toplogical descriptions of interbank markets.


The importance of analyzing topological features of real-world networks in both binary and weighted terms has been highlighted, e.g., in Fagiolo et al. (2010) who show that in the binary international trade network countries with many trade partners typically trade with countries with few partners. However, in the weighted trade network where trade intensity is taken into account, countries that exchange large trade volumes have many trade partners, but only trade intensively with few well-connected partners.


We use interchangeably the terms global, international, or cross-border lending to mean cross-border (external) loans extended by banks in a given country to institutions in other countries.


Through the paper we assume that the cross-border banking statistics we use (further discussed in Section III.A) are a good first-order approximation for actual cross-border banking flows among countries. Nevertheless, the coverage is incomplete because only banks from some 15 advanced economies have submitted data to the BIS since the early 1980s. Currently, 43 countries report their locational statistics to the BIS.


See also Georg (2010) and Fujii and Takaoka (2010) for analyses of how network topology influences contagion in interbank markets and a hypothetical financial market, respectively.


This mostly reflects activities by internationally active foreign banks (see, e.g., McGuire and Tarashev, 2006).


A recent foray into network models of systemic risk include Martinez-Jaramillo et al. (2010) who simulate contagion in a banking system and derive measures of system fragility based on the topological features of the interbank exposures network. Furthermore, network structures have been incorporated into stress tests of financial systems to monitor the first- and second-round effects of macroeconomic shocks (see, e.g., Espinosa-Vega and Sole, 2010).


Average path length is defined as average of the shortest path between all pairs of nodes in the network. It measures the “degrees of separation” among nodes.


Using similar data on cross-border financial exposures, Degryse et al. (2010) use simulations to examine the risk and speed of potential contagion in the global financial network over 1999–2006. They find that a shock that affects the liabilities of one country may have rippling effects through the system. Furthermore, the network topology in 2006 suggests that a shock affecting emerging Europe, Turkey or Russia would affect most countries in the world.


Using data from the Coordinated Portfolio Investment Survey (CPIS) for 2001-06, Song et al. (2009) show that the distribution of global financial exposures are well approximated by a Weibull distribution.


In contrast, another often-used source of cross-border financial data are the BIS consolidated statistics, which are based on the nationality of the reporting bank, and net out intragroup positions. For a comparison between consolidated and locational banking statistics, see, e.g., Waysand, Ross, and Guzman (2010).


The analysis of the “net repayments” network is left for future research.


Node degree captures the extent to which a country is well connected or “in the thick of things.” A high node degree simply means that the node has a large neighborhood of local contacts, be it lenders or borrowers, and it relatively prominent in its neighborhood. Thus, node degree is an indicator of “local centrality.” More sophisticated indicators (such as shortest path length) measure the node’s strategic significance in the overall network, or its “global centrality” (see Scott, 2009, pp. 82-93.) We are limited in undertaking such an analysis, however, by the core-periphery structure of the dataset, with countries in the periphery not being interlinked except via the core.


By construction, average RNSout (across borrowers) is the average number of borrowers per lender, while average RNSin (across lenders) is the average number of lenders per borrower.


Interestingly, network connectivity for the GBN based on cross-border flows appears less persistent than that for the GBN based on cross-border exposures. The reason is that the likelihood of links dying in the network of cross- border exposures is substantially lower than in our network (see Hattori and Suda, 2007).


The concept of “private” capital flows refers to changes in foreign assets and liabilities of the recipient domestic private sector (IMF, 2007, pp. 3).


In constructing borrower rankings we exclude high-income countries. Including them brings to the top of interconnectedness countries such as Australia, Finland, Greece, Portugal, and Spain.


Note that we are not assuming time-homogeneity of the transition probabilities.


For ɷ>=0, the M statistic is the trace of the 1-step transition matrix(divided by 10 while for ɷ>=1, the M statistic is the average of all the entries in the main diagonal and those lying one entry to the right and one entry to the left of the main diagonal.


The “random” network is obtained by keeping the distribution of flows unchanged and reshuffling the links across countries.


The conclusions are qualitatively the same if we consider alternative years as the base years for the ergodic distributions.


How the 2008–09 crisis has affected this trend is an interesting question to explore as more recent data becomes available.


See Laeven and Valencia (2008, 2010) for definitions.


The analysis of network measures’ capacity to act as leading indicators of crises is left for future research

Other Resources Citing This Publication