The Bright and the Dark Side of Cross-Border Banking Linkages
  • 1 0000000404811396https://isni.org/isni/0000000404811396International Monetary Fund

Contributor Notes

When a country's banking system becomes more linked to the global banking network, does that system get more or less prone to a banking crisis? Using model simulations and econometric estimates based on a world-wide dataset, we find an M-shaped relationship between financial stability of a country's banking sector and its interconnectedness. In particular, for banking sectors that are not very connected to the global banking network, increases in interconnectedness are associated with a reduced probability of a banking crisis. Once interconnectedness reaches a certain value, further increases in interconnectedness can increase the probability of a banking crisis. Our findings suggest that it may be beneficial for policies to support greater interlinkages for less connected banking systems, but after a certain point the advantages of increased interconnectedness become less clear.

Abstract

When a country's banking system becomes more linked to the global banking network, does that system get more or less prone to a banking crisis? Using model simulations and econometric estimates based on a world-wide dataset, we find an M-shaped relationship between financial stability of a country's banking sector and its interconnectedness. In particular, for banking sectors that are not very connected to the global banking network, increases in interconnectedness are associated with a reduced probability of a banking crisis. Once interconnectedness reaches a certain value, further increases in interconnectedness can increase the probability of a banking crisis. Our findings suggest that it may be beneficial for policies to support greater interlinkages for less connected banking systems, but after a certain point the advantages of increased interconnectedness become less clear.

I. Introduction

One of the hallmarks of financial globalization has been a growth in cross-border linkages (exposures) among banks. On the positive side, these linkages have been associated with new funding and investment opportunities, contributing to rapid economic growth in many countries (especially in the early part of the 2000s). But the growing financial linkages also have a “dark side”:2 the increased cross-border interconnectedness made it easier for disruptions in one country3 to be transmitted to other countries and mutate into systemic problems with global implications.

The potential harmful consequences of cross-border interconnectedness for domestic banking sector stability have been illustrated rather dramatically during the recent global financial crisis, when shocks to one country’s financial system were rapidly transmitted to many others. One of the upshots of the crisis was that considerable efforts have been devoted to better measuring “systemic importance” of jurisdictions around the world. There is a growing consensus that interconnectedness, together with size, should be a key variable in assessing the systemic importance of a jurisdiction from the viewpoint of financial stability (IMF, BIS, and FSB, 2009; IMF, 2010).

This paper aims to answer the following key question: when a country’s banking system gets more linked to the global banking network, does it become more stable, or less stable? The answer to this question is obviously relevant for policymakers and regulators in individual countries. To the extent that interlinkages help banking stability, should policies and regulations be designed to promote such cross-border interconnectedness? And to the extent that interlinkages are bad for stability, should policies and regulations aim to stop or, at least limit, the growth of such interlinkages?

We examine the above key question in two ways: First, we analyze it conceptually, using simulations. Second, we examine it empirically, based on a range of econometric approaches—parametric as well as non-parametric—that combine data on banking crises around the world with a comprehensive data set on cross-border banking linkages.

To preview the main results, our short answer to the above question is: it depends on the degree of interconnectedness. The relationship between the likelihood of a banking crisis in a country and the degree of integration of that country’s banking sector into the global banking network is far from trivial. We find that in a country whose banking sector has relatively few linkages to other banking sectors, increased cross-border linkages tend to improve that system’s stability, controlling for other factors. In other words, within a certain range, connections serve as a shock-absorber. The system acts as a mutual insurance device with disturbances dispersed and dissipated. Connectivity engenders robustness. Risk-sharing – diversification – prevails. But at some point—which we estimate to be at about the 95th percentile of the distribution of countries in terms of interconnectedness—increases in cross-border links begin to have detrimental effects on domestic banking sector stability. At a yet higher point, when a country’s network of interlinkages becomes almost complete,4 the probability of a crisis goes down again.

One of the novel insights of our paper is that it is important to distinguish whether the cross-border interlinkages are stemming primarily from banks’ asset side or from their liabilities side. We introduce measures that distinguish those two types of interconnectedness (which we call “downstream” and “upstream” interconnectedness), and we find that the impact of changes in interconnectedness on banking system fragility are more significant for liabilities-side (“upstream”) interconnectedness than for asset-side (“downstream”) interconnectedness.

The paper is structured as follows. Section II provides a discussion of the relationship between financial interconnectedness and instability, positioning the paper as a bridge between the network analysis literature and the banking crisis literature. Section III presents the data, in particular the measures of interconnectedness and the measures of financial instability. Section IV provides a simulation exercise. Section V describes the estimation procedures and presents the results of both parametric and non-parametric estimates. Section VI concludes.

II. Financial Stability Consequences of Increasing Interconnectedness

Is greater connectedness to the outside world beneficial or detrimental to a country’s banking system? Despite the obvious relevance of this question to national policymakers and regulators, and despite progress in understanding financial networks in recent years, the available literature does not provide a coherent set of answers.

Two main streams of literature are relevant for answering this question: the network analysis literature and the banking crisis literature. Each of these two streams provides useful insights. The network analysis literature focuses on the transmission of shocks within a network, but does not adequately model the underlying factors that make some banking systems more prone to crises than others in the first place. The banking crisis literature focuses on modeling banking crises in individual countries, but it does not take adequately into account the transmission of shocks via the global banking network.

Our paper aims to bridge these two separate, but closely related, streams of literature. It offers an approach that takes into account both the differences in the underlying soundness of banking system as well as their degree of connectedness with the broader banking network.

A. Network Analysis

Our paper builds on the rapidly growing literature on financial networks. In particular, we build on the insight that financial interconnectedness has two opposing effects. On one hand, linkages may act as channels to propagate shocks to the whole system, that is, they act as “shock transmitters.” On the other hand, through these linkages, shocks can be shared and absorbed by others, that is, financial linkages may act as “shock absorbers.”

In our paper, the basic unit of analysis is a banking sector in a country, and the network being analyzed is the global banking network, i.e. the network of cross-border linkages among the various country-level banking sectors. This is different from much of the banking network literature, in which the basic unit of analysis is a single bank and the network is an interbank network, usually in a single country. Nonetheless, the basic mechanics of shock transmission (via lending exposures) and absorption (via banks’ net worth) are essentially the same.

Early theoretical literature on banking networks emphasized the benefits of interconnectedness for network stability. A key contribution in this regard was Allen and Gale (2000), who related banking system resilience to the completeness of the banking network. Specifically, based on an examination of a stylized four-bank network, they concluded that a “complete” network (one in which every bank is connected to all other banks) is more resilient than an “incomplete” network (one in which some banks are not connected to all other banks), due to both wider possibilities for risk sharing in complete networks and individual banks bearing a smaller share of the shock.

Nier and others (2007) extended Allen and Gale’s work by conducting simulations on a more complex network of banks with interlinked balance sheets. Interestingly, they identified a non-monotonic relationship between bank connectivity and contagious defaults. In particular, they found that the relationship between the number of defaults in a network and the likelihood of interbank exposure is M-shaped: at low levels of connectivity, an increase in interconnectedness raises the likelihood of contagion; at higher levels of connectivity, the resilience of the system improves, declines; and then (in line with Allen and Gale’s calculations) improves again as the network reaches completeness. The paper attributes this non-monotonic shape to the opposing forces of shock absorption and shock diffusion in financial networks.

The notion that complex financial networks are not only better at diversifying away idiosyncratic risk, but also more prone to propagating financial distress is a common theme in a number of recent papers on the subject (e.g., Battiston and others, 2010; Caballero and Simsek, 2009). An important part of the network analysis literature focuses on “cascading” effects in a network (May and Anderson, 1991; Watts, 2002; Kinney and others, 2005). The phenomenon of large but rare cascades triggered by relatively small shocks has been observed in areas as diverse as collective action, the diffusion of norms and innovations, cultural fads, and cascading failures in infrastructure and organizational networks. Watts (2002) reviews the literature and presents a possible explanation of this phenomenon in terms of a network of interacting agents whose decisions depend on actions of their neighbors, based on a threshold rule. His model points out factors that make a network relatively more prone to the occurrence of large “cascades”. In particular, when the network is highly connected, the size distribution of cascades becomes bimodal, implying a kind of instability that is correspondingly harder to anticipate. If the network nodes are characterized by very heterogeneous thresholds, the system is relatively more vulnerable to global cascades; on the other hand, if the nodes are very heterogeneous in terms of their degree of connectivity, the network is relatively less vulnerable.

B. Banking Crisis Literature

A substantial body of literature exists on models of banking crises in individual countries. This includes a range of studies that attempt to identify, at a country level, early warning indicators of banking crises. The findings of this literature are far from conclusive, highlighting a need for further research. The following is a brief summary of the literature (for more, see Breuer, 2004, Davis and Karim, 2008, and Čihák and Schaeck, 2010).

The so-called first-generation models hypothesize that an adverse macroeconomic environment adversely affects banks’ borrowers, impacting the banks, and setting off bank runs that ultimately lead to bank closures (e.g., Miskhin, 1978); in contrast, the so-called second-generation models focus on depositor behavior, and view banking crises as self-fulfilling prophecies or “sunspot” events reflecting sudden shifts in depositors’ sentiment(e.g., Diamond and Dybvig, 1983). As regards empirical evidence, Gorton (1988) rejects the randomness of bank runs, finding (in long-term U.S. data) a systematic association between bank runs and recessions. Calomiris and Mason (1997), using data from the 1932 Chicago bank panic, cast some doubt on the contagion effects on other institutions that arise from deposit withdrawal: they do not find that such contagion effects lead to insolvency.

Third-generation models focus on the impact of boom and bust on banks’ assets (e.g., Gavin and Hausman, 1996; Hardy and Pazarbaşioğlu, 1998; and Demirgüç-Kunt and Detragiache, 1998). These models point out that in economic booms, banks have incentives to engage in excessive lending against collateral, such as real estate and equities, which appreciates in value. This facilitates a lending boom and increasing leverage in the economy. A subsequent bust results in collapsing asset prices, leading banks to scale back their lending. Ultimately, this translates into an economic slowdown that increases borrower default rates. In contrast to the second-generation models, third-generation models focus on banks’ assets rather than liabilities, and they use (lagged) macroeconomic variables as leading indicators.

Fourth-generation models extend the earlier literature by identifying features of the institutional environment that set the stage for the build-up of macroeconomic imbalances, which subsequently give rise to banking problems. These models accentuate the roles of rule of law and contract enforcement, protection of shareholder and creditor rights, sophistication of supervisory and regulatory frameworks, incentive schemes created by deposit insurance, and the socioeconomic environment (see, e.g., Demirgüç-Kunt and Detragiache, 1998, 2005; Hutchinson and McDill, 1999; Eichengreen and Arteta, 2000; Hutchinson, 2002; Das, Quintyn, and Chenard, 2005; and Buch and DeLong, 2008).

A rapidly growing body of literature has focused on market based indicators, such as the distance to default or the subordinated debt spread, as early warning indicators for banking problems on the micro level (e.g., Gropp, Vesala, and Vulpes, 2004). An advantage of this approach is that it builds upon forward looking information, contained in market prices. Its key disadvantage is its reliance on market prices derived from liquid markets, which limits its applicability when such markets do not exist.

The bottom line is that a clear agreement is yet to be reached in the literature on models and indicators for systemic banking problems. Moreover, importantly from the viewpoint of our paper, this stream of literature tends to examine individual countries in isolation and has not been able to satisfactorily incorporate cross-border linkages and contagion within the global banking network. That is where, primarily, our paper aims to contribute.

III. Overview of Input Data

A. Measuring Interconnectedness

We use network analysis to measure the degree to which a country’s banking sector is connected to the rest of the global banking system. In the recent literature, network analysis has become a key tool for measuring the extent of interactions within a banking network. In our case, the banking network refers to the global architecture of cross-border financial relationships. We consider the global banking network as a set of bilateral claims (links) of different banking systems (nodes) on each other and calculate each system’s interconnectedness (centrality) in the network using BIS locational banking statistics.5 The underlying idea of the analysis is to infer, from the pattern of cross-border linkages among banking sectors, the extent to which a banking sector of a jurisdiction is “central” in the international banking network (von Peter, 2007; Kubelec and Sá, 2010).6

We have identified four possible measures of centrality, based on a review of the network analysis literature. The first one is “degree” centrality, which equals the sum of each banking system’s links to other banking systems and serves a basic measure of centrality. Its advantage is simplicity, but its main drawback is that it implicitly gives all links the same weight, irrespective of size. The second one is “alter-based” centrality (Neal, 2010 and 2011), a recursive measure that takes into account the relative importance of a banking system in the global network, as well as the relative importance of each banking system to which it is connected. In alter-based centrality, individual links are weighted by the partner country’s centrality score and then summed up. The third and fourth measures are, respectively, “alpha” and “beta” centrality (Bonacich and Lloyd, 2001; Bonacich, 1987), which are both derived from Bonacich (1972)’s original eigenvector centrality measure. Similar to alter-based centrality, these two measures take into account the importance of a banking system and the importance of its partners; however, unlike alter-based centrality, they require certain additional assumptions be met (Appendix II).

The subsequent analysis focuses on alter-based centrality as the preferred measure of centrality, given its ease of calculation and its intuitive nature. It is more comprehensive than degree centrality, since it takes into account partner jurisdictions’ centrality, and it is more straightforward than the two Bonacich measures, given that the results for alter-based centrality do not require restrictive assumptions.7 Nevertheless, the four measures demonstrate a high degree of co-movement (Figure 1), and yield broadly similar results when used in regressions.

Figure 1.
Figure 1.

Movements in Centrality Measures for Selected Countries, 1985Q1-2009Q3.

Citation: IMF Working Papers 2011, 186; 10.5089/9781462309269.001.A001

Source: BIS locational banking statistics, authors’ estimates.Note: Downstream measures, rescaled to [0,1] for comparability across measures of different magnitudes.

When examining the developments in alter-centrality over time (Figure 1), it is useful to note that this indicator’s behavior has been far from uniform across countries. While some (e.g., Germany) have seen a relatively steady increase in alter-centrality over the sample period, others (e.g., Japan in the 1990s) have been going through a period of steady decline in alter-centrality, and yet others have seen several ups and downs.

One of the contributions of our paper is that we explicitly distinguish two types of interconnectedness depending on whether the exposures come from the asset side (lending) or the liability side (borrowing). This distinction—a very important one for making appropriate policy decisions—means that we analyze a directed network. That is, links between banking systems provide information about the direction of the relationship rather than merely whether or not a relationship exists.8 We define and calculate:

  • Downstream interconnectedness (or “asset centrality”) as the recursive centrality measure of interconnectedness based on asset exposures for each banking system. The motivation for this comes from calling the asset (credit) exposure of creditor countries vis-à-vis borrowing countries a “downstream” exposure.

  • Upstream interconnectedness (or “liabilities centrality”) as the recursive centrality measure of interconnectedness based on liability exposures for each banking system. The motivation for this comes from calling the funding exposure of borrowing countries visà-vis credit countries an “upstream” exposure.9

B. Measuring Banking (In)stability

We measure the (in)stability of a banking sector in a country as the probability of a banking crisis in that country in a given year: the lower the crisis probability, the more stable the banking system. To derive this variable, we rely on the widely used database of banking crises by Caprio, Klingebiel, Laeven, and Noguera (2005), as updated by Laeven and Valencia (2008). The database covers the universe of 120+ systemic banking crises around the globe since 1970. It is the most complete and most detailed database on banking crises to date. Based on the database, we define a banking crisis dummy variable, equal to 1 if there is a banking crisis in a country i at time t and 0 otherwise.

Scatter plots of the banking crisis variable against the two interconnectedness variables (Figure 2) suggest that there may be a relationship, and it is likely to be far from trivial. In particular, as regards downstream interconnectedness, there is an area in the middle (roughly between 0.3 and 0.6) characterized by low occurrence of banking crises, and there is also an area with high interconnectedness (roughly above 1.0) that has virtually no crisis observations. Similarly, the chart for upstream interconnectedness (liabilities centrality) shows an area with lower occurrence of banking crises in the region between 0.6 and 0.8, and again an area with virtually no crisis observations above 1.0.

Figure 2.
Figure 2.

Banking Crises vs. Interconnectedness: A First Look

Citation: IMF Working Papers 2011, 186; 10.5089/9781462309269.001.A001

Source: BIS locational banking statistics, authors’ estimates.

C. Other Data

The global financial crisis experience has clearly illustrated that the relationship between banking interconnectedness and banking stability is affected by many other variables. While some jurisdictions with high degrees of interconnectedness have been hit severely during the crisis, other highly interconnected jurisdictions managed to withstand this crisis (as well as the previous crises) remarkably well. It is therefore important to control for the other variables that also impact banking stability in a country. Some of these variables have a separate effect on banking stability, while others work in interaction with interconnectedness.

We have therefore compiled a set of other variables, building on datasets identified in the earlier literature. These variables cover the broader macroeconomic and institutional framework as well as features of the banking system such as its financial structure. Appendix I lists the other data compiled for the various control variables used in the analysis.

IV. Simulations

To motivate the empirical analysis, and to illustrate the linkages between the network literature and the banking crisis literature, we have run a simulation exercise. The exercise examined, in a hypothetical network, what kind of relationship one can expect between interconnectedness (centrality) and banking sector stability. Appendix III describes the simulation framework used in this section.

Figure 3 shows simulation results based on a network of 100 banking sectors (100 “countries”), run for 1,000 random realizations of the network and initial conditions. Each node has been assigned a capitalization, and the matrix of interlinkages has been filled with asset and liability values, determining the degrees of centrality for each of the nodes in each of the iteration, as described in the previous sub-section. The results shown in Figure 3 are for net worth set to equal 1 percent of total assets. The relationship between interconnectedness and crisis probability has the same (M-letter) shape for other values of net worth, the difference being that the peaks of the “M” are lower (i.e., crisis probability is lower) for higher values of net worth.10

Figure 3.
Figure 3.

Simulations: Banking Crises vs. Interconnectedness

Citation: IMF Working Papers 2011, 186; 10.5089/9781462309269.001.A001

Source: authors’ simulations.

The figure illustrates the non-linear relationship between stability in a country’s banking system and its interconnectedness with the global banking network. First, for very low levels of connectivity, an increase in connectivity increases the likelihood of a crisis, since connectivity increases the chance of shock transmission. For higher levels of connectivity, increases in connectivity first decrease and then increase the likelihood of a banking crisis. When connectivity is sufficiently high, further increases in connectivity unambiguously decrease contagion as the shock absorption effect starts to dominate and the initial shock is spread over more and more banking systems, each able to withstand the shock received.

The simulation results in Figure 3 are consistent with the earlier literature. In particular, it is in line with the findings of Nier and others (2007), who were the first to identify this kind of a non-monotonic, M-shaped, relationship between bank connectivity and the number of contagious defaults in the network as a whole. The main difference (other than the fact that we analyze a network of banking sectors, rather than a network of banks) is that we focus on the nodes of the network, and examine the likelihood of a default (banking crisis) in each node of the network, instead of counting the (expected) number of defaults in the network as a whole. Another thing to note is that we are modeling a network that is heterogeneous (some nodes are playing much more central role than others) and is far “complete” (numerous links are missing in the network).11 This has important impacts on the observed relationship between interconnectedness and fragility: if the global banking network were close to “complete”, the simulations suggest that we would likely see a drop in its fragility as systems would become more resilient to shocks due to risk sharing.12

The simulation results also seem broadly consistent with the preliminary examination of the data. In particular:

  • The data on banking crises and connectivity (Figure 2) suggest there is a range of middle values of the connectivity (roughly between 0.3 and 0.6 for downstream interconnectedness and between 0.6 and 0.8 for upstream interconnectedness) for which there is a relatively lower occurrence of banking crises. This is consistent with the finding from the simulation that there is a middle area where crisis probability first goes down and then goes up as the probability of links goes up.

  • The raw empirical data (Figure 2) also suggest that when connectedness reaches a certain point (roughly 1.0, both for downstream and upstream interconnectedness), the frequency of crisis observations drops off substantially (very close to zero). Similarly, the simulations (Figure 3) show that as the probability of a link reaches a point of high interconnectedness, crisis probability drops off precipitously (to a number very close to zero).

One part where the simulations may seem, on a quick look, to differ from the empirical data, is the area of very low interconnectedness. There, the simulation model (Figure 4) predicts a low crisis frequency, while empirical data (Figure 2) show a relatively high frequency. However, it needs to be understood that the simulations in Figure 2 focus on the contagious defaults and not on the other factors that may make crisis in a country more or less likely. It is possible that the countries whose banking systems have relatively low degree of connectivity to the global banking network nonetheless have other features that make their systems prone to (domestically-induced) banking crises. To examine this in more detail, we turn to the empirical analysis, which aims to distinguish more precisely the cross-border factors (interconnectedness) from the domestic factors of banking sector fragility.

Figure 4.
Figure 4.
Figure 4.

Nonparametric Analysis Results

Citation: IMF Working Papers 2011, 186; 10.5089/9781462309269.001.A001

Note: The higher- and lower-crisis probability areas have been determined by the nonparametric method described in section V. The polynominal filter of the 3rd order is a simpler version of the multinominal parametric estimates presented in Section IV, and is added here for illustration.Source: BIS locational banking statistics, authors’ estimates.
Figure 5.
Figure 5.

Global Banking Network 1977 vs. 2009

Citation: IMF Working Papers 2011, 186; 10.5089/9781462309269.001.A001

Source: BIS locational banking statistics, authors’ calculations. Note: Triangles represent BIS reporting countries.

V. Estimatinga Financial Stability Model

Empirically, we examine the linkage between banking interconnectedness and domestic banking sector stability using two main complementary approaches: a parametric estimation (probit model) and a non-parametric estimation (threshold approach).

A. Parametric Estimation (Probit Model)

We examine the relationship between cross-border interconnectedness and the probability of a banking crisis in a country using a multivariate probit model for a pooled dataset of 189 banking systems in 1977–2009. The probability of observing a banking crisis in country i in year t is modeled as a function of interconnectedness, a set of macroeconomic and other control variables, as well as interactions between interconnectedness and other variables. The estimated log-likelihood function is

LnL=t=1Ti=1n{P(i,t)ln[F(βX(i,t))]+(1P(i,t))ln[1F(βX(i,t))],(1)

where P(i, t) is the banking crisis dummy variable (equal to 1 if there is a crisis, and 0 otherwise), β is the vector of coefficients, and X is the vector of explanatory variables. As regards interconnectedness, we have tried to include it not only as a linear term, but also as a quadratic and cubic term, to examine the possible non-linearity suggested by the simulations.

The regression results (Table 1)13 suggest that (i) increases in a banking sector’s degree of interconnectedness with the global banking network tend to be associated with reductions the probability of a crisis in that banking sector; (ii) the reduction in crisis probability gets smaller as interconnectedness goes up, keeping other things constant; (iii) when the banking sector’s interconnectedness reaches a certain point, further increases in interconnectedness actually start increasing the probability of crisis in the banking sector; and (iv) at very high degrees of interconnectedness, crisis probability declines again.

Table 1.

Parametric Estimates (Probit Model)

article image

As defined in Laeven and Valencia (2010). The dependent variable takes a value one if there is a banking crisis and the value zero oth

*, **, *** correspond to the 10, 5, and 1 percent significance levels, respectively. We estimate a random-effects probit model Standard errors are given in parenthesis.

These key findings are rather robust across a range of specifications. To demonstrate the robustness of the results, Table 1 includes regressions with different sets of control variables, and it includes regressions both with interactions between interconnectedness and other variables, as well as regressions without the interactive terms. Regression (1) is a basic specification. Regression (2) adds the lagged value of banking crisis dummy variable.14 Regression (3) adds stock price volatility as an additional control variable. Regression (4) adds time dummies while dropping cost-to-income ratio and stock price volatility.15 Regression (5) puts the interactions without the lagged banking crisis or stock volatility. Finally, regression (6) uses downstream interconnectedness, which is not significant when the cubic term is included.

We also examine a specification that takes into account banking sector capitalization (approximated by capital to asset ratio and, alternatively, by regulatory capital to risk weighted assets ratio). Obtaining consistent and long time series on bank capitalization is more challenging than for most of the other variables, resulting in a reduction of the number of observations in the regression. Nonetheless, the results show that bank capitalization is negatively significant and upstream interconnectedness squared is positively significant, supporting our earlier results (these results are not shown in Table 1 to save space, but are available upon request). Additionally, to capture the differences between advanced and other economies, we have considered including a 0/1 dummy variable for advanced economies (World Economic Outlook definition). The variable is not included in Table 1 given colinearity with the level of interconnectedness term and growth of private credit over GDP, but its inclusion without one or both terms yield similar levels of significance of the cubic term of interconnectedness.

Overall, these probit estimates are consistent with the simulation findings of an M-shaped curve illustrating a non-linear relationship between stability in a country’s banking system and its interconnectedness. Specifically, the significant positive sign of the cubic term for upstream interconnectedness implies that for banking systems that are not very connected to the global banking network, increases in interconnectedness are at first associated with a reduced probability of a banking sector crisis. Once the degree of interconnectedness reaches a certain value, further increases in interconnectedness are not associated with improvements in financial stability and can in fact mean increased fragility.16 However, at yet another point, when the banking system becomes highly interconnected, the crisis probability begins to decline again.

From the parameter estimates in Table 1, one can calculate the points at which the effect of increases in interconnectedness on banking crisis probability turns signs. Let us focus here on the point where the relationship between interconnectedness and crisis probability switches signs for the first time, i.e. where it reaches its first local minimum. Solving for this minimum, using the point estimates from specification (5) in Table 1, and keeping the other significant variables at their sample average values, we find that the minimum is reached when upstream interconnectedness equals 0.37.17 This corresponds to the 95th percentile of the distribution of upstream interconnectedness. In other words, 95 percent of the interconnectedness observations in our sample are in the downward sloping portion of the relationship, where interconnectedness reduces crisis probability. For the remaining 5 percent (those with upstream interconnectedness above 0.37), the relationship between interconnectedness and crisis probability is more complex: it is upward sloping at first, only to become downward sloping again. Indeed, these upper 5 percent of observations in terms of interconnectedness include some of the advanced economies whose large and highly interconnected banking sectors have been substantially affected during the recent financial crisis, while also including some highly interconnected banking sectors that have been relatively unharmed.

The probit estimates highlight that the probability of banking crisis is affected not only by interconnectedness, but also by other factors, some of which also interact with the interconnectedness variable, as shown in estimate (5) in Table 1. The 0.37 “critical point” has been calculated at sample average, so for instance an economy that consistently shows rates of economic growth that are above our sample average, increasing interconnectedness even above 0.37 would still help reduce the probability of banking crisis.18 In other words, for this economy, the “critical point” would occur at a higher level of interconnectedness than 0.37.19

Interestingly, we obtain different results for upstream and downstream interconnectedness—a novel distinction that has not yet been examined in the literature. In particular, we find that the “dark side” of interconnectedness (i.e., the negative effects that dominate when a certain level of interconnectedness is exceeded) is stronger for the upstream measure of interconnectedness. In other words, increasing interconnectedness on the liabilities (borrowing) side is more likely to be become detrimental to banking stability than increasing interconnectedness on the asset (creditor) side. Therefore, financial turmoil originated in creditor countries and “flowing upstream” via borrowing countries’ funding channels could be more devastating than financial turmoil originated in borrowing countries and “flowing downstream” to the creditors.

Finally, the probit estimates also provide useful information on the impact of other variables that also need to be taken into account when considering the relationship between interconnectedness and banking sector stability. The control variables have the expected signs. For instance, the lagged ratio of growth of private sector credit to GDP is significant and exhibits a positive sign. This is consistent with the early warning literature suggesting that lending booms tend to lead to banking crises. The results also show that real GDP growth is negatively related to the occurrence of a banking crisis, which highlights the importance of the state of the real economy in determining the heath of the banks. As expected stock market volatility is also positively related to banking crises, confirming that stock market prices are a good warning indicator. A higher cost-to-income ratio, which could be interpreted as reflecting the banking system’s inefficiency, is associated with a significantly higher probability of a banking crisis. The lag of banking crisis is significant, highlighting the persistence of crisis. The ratio of the monetary aggregate M2 to foreign exchange reserves is used as a proxy for sudden capital outflows, to control for the relationship between these outflows and banking sector problems in countries with fixed exchange rates.

B. Nonparametric Estimation (Threshold Approach)

To cross-check and complement the parametric (probit) estimates, we have also estimated nonparametric models aimed at signal extraction. As with the parametric exercise, the purpose of the nonparametric estimation is to examine the impact of cross-border interconnectedness on the likelihood of a banking crisis occurring in a country.

Nonparametric estimation (threshold approach) does not impose distributional assumptions upon the data, and the inferences drawn from such estimation may therefore be considered more robust than parametric models, such as probit models. An additional benefit is the ability of the nonparametric tests to illustrate the classification accuracy of the relevant variables over different threshold levels. Their main drawbacks include computational difficulties, especially when analyzing interaction among numerous indicators with potentially multitudes of possible thresholds.

The nonparametric approach aims to identify a combination of threshold values for interconnectedness and other variables (capitalization, economic growth, etc.—the same ones as in the parametric exercise) such that they enable us to identify situations of banking crises (i.e., observations for which banking crisis dummy equals to 1) as accurately as possible. Of course, the definition of “as accurately as possible” needs to take into account both Type I errors (false alarms) and Type II errors (missed crises), with the latter being much more costly. Our non-parametric algorithm therefore seeks to minimize the number of Type II errors, but it is a constrained minimization. Otherwise, without constraints, the result of the minimization would be a set of thresholds that would identify all or most or our sample as potential crisis observations (i.e., we would have very large Type I errors).

The nonparametric algorithm therefore solves for a combination of thresholds such that it minimizes Type II errors (missed banking crises) subject to the constraint that the ratio of predicted banking crises to total observations is not larger than the actual frequency of crises observations. That percentage of actual crises observations in our sample is about 7.4 percent. Therefore, our algorithm seeks to identify a set of thresholds (on interconnectedness and other variables) such that as many of the actual crises (i.e. observations for which the banking crisis dummy equals 1) are “predicted” (i.e., all the thresholds are breached for those observations) as possible, but the number of such predicted crises is not larger than 7.4 percent of all observations. We are using a numerical algorithm that examines various combinations of thresholds in a step-wise fashion and for each calculates share of “predicted” crises and the number of Type II errors.

The data used in the nonparametric analysis is the same dataset as used in the parametric estimates. It is a pooled dataset on 189 banking systems in 1977–2009 (see Appendix I for description). We have focused on pairs and triplets of variables, with one or two thresholds on each variable. This reflects computational complexities (as well as the challenges of presenting visually the results in more than two—and definitely in more than three— dimensions).

Overall, the results of the nonparametric estimates are consistent with the simulations and with the earlier findings from the parametric (probit) estimates. In particular, Figure 4(a) shows, for upstream interconnectedness, that the optimum results in terms of the noise/signal ratio (as described above) are achieved for a combination of thresholds that identifies two “higher crisis probability” areas (one for interconnectedness roughly below 0.5 and one between 0.8 and 1.3) and two “low crisis probability” areas (one roughly between 0.5 and 0.8 and one above 1.3).20 Figure 4(b) shows the results for downstream interconnectedness, which are qualitatively broadly similar, but statistically much weaker.

VI. Concluding Remarks

When a country’s banking system becomes more linked to the global banking network, does it get more or less stable? Our answer is: it depends on the degree of interconnectedness. Based on model simulations as well as an econometric estimation based on a comprehensive global dataset, we find that in banking systems that are not very connected to the global banking network, increases in interconnectedness are associated with a reduced probability of a banking sector crisis. Once the degree of interconnectedness reaches a certain value, further increases in interconnectedness do not improve financial stability and can in fact increase fragility. When the interconnectedness reaches close to complete network, interconnectedness starts again reducing likelihood of crisis.

The “dark side” of interconnectedness (i.e., the negative effects that dominate when a certain level of interconnectedness is exceeded) is stronger for the upstream measure of interconnectedness. In other words, increasing interconnectedness on the liabilities (borrowing) side is more likely to be become detrimental to banking stability than increasing interconnectedness on the asset (creditor) side. Therefore, financial turmoil originated in creditor countries and flowing upstream via borrowing countries’ funding channels could be more devastating for financial stability.

Our findings are potentially quite relevant for policymakers and financial sector regulators. Our calculations suggest that up to a point, it may be beneficial for policies and regulations (including supervision) to support greater interlinkages between the domestic banking sector and foreign banks. Above that threshold, the benefits of greater interlinkages are less clear; in fact, our calculations suggest that further growth in interlinkages can at that point become detrimental for banking stability. The calculations also indicate that the potential negative effects of interconnectedness can be compensated for by other factors (which can also be influenced by policies), such as greater capitalization of the banking sector. A fuller examination of the impacts of policies influencing financial interlinkages and financial stability is an important topic for further research.

Appendix I. Description of Data Sources and Transformations

  • GDP growth: the geometric rate of growth of real GDP, between the earliest data in the sample and 2000, in constant 2000 USD. Source: World Bank World Development Indicators.

  • Trade/GDP: total exports and imports of goods and services as a percentage of GDP. Source: World Bank World Development Indicators.

  • FDI/total investment: the ratio of foreign direct investment net inflows to gross fixed capital formation. Source: World Bank World Development Indicators.

  • Lending interest rate: the rate charged by banks on loans to prime customers. Source: World Bank World Development Indicators.

  • Money and quasi money growth: the average annual growth rate in M2. Source: World Bank World Development Indicators.

  • Real effective exchange rate index (2000=100). Source: World Bank World Development Indicators.

  • Stocks traded/GDP: the total value of shares traded during a year as percentage of GDP. Source: World Bank World Development Indicators.

  • Central government debt/GDP. Source: World Bank World Development Indicators.

  • Fiscal balance/GDP: cash surplus or deficit as percentage of GDP. Source: World Bank World Development Indicators.

  • Inflation: average annual CPI inflation. Source: World Bank World Development Indicators.

  • Current account balance/GDP. Source: World Bank World Development Indicators.

  • Gross national income: GNI calculated by the Atlas method (using current US dollars). Source: World Bank World Development Indicators.

  • Gross domestic product. Source: World Bank World Development Indicators.

  • Domestic credit to private sector/GDP. Source: World Bank World Development Indicators.

  • Exchange rate regime: annual coarse classification of regime. Source: Ilzetzki, Reinhart, and Rogoff (2008).

  • Polity score: Revised combined polity score (Polity2) from the Polity IV project; higher values indicate more democratic systems. source: http://www.systemicpeace.org/polity/polity4.htm.

  • Political risk rating: a means of assessing the political stability of a country on a comparable basis with other countries by assessing risk points for each of the component factors of government stability, socioeconomic conditions, investment profile, internal conflict, external conflict, corruption, military in politics, religious tensions, law and order, ethnic tensions, democratic accountability, and bureaucracy quality. Source: The PRS Group International Country Risk Guide (ICRG).

  • Government stability: a measure of the government’s ability to stay in office and carry out its declared program(s), depending upon such factors as the type of governance, cohesion of the government and governing parties, approach of an election, and command of the legislature. Source: The PRS Group International Country Risk Guide (ICRG).

  • Corruption: a measure of corruption within the political system. Source: The PRS Group International Country Risk Guide (ICRG).

  • Financial risk rating: a means of assessing a country’s ability to pay its way by financing its official, commercial and trade debt obligations. Risk points are assessed for each of the component factors of foreign debt as a percentage of GDP, foreign debt service as a percentage of exports of goods and services (XGS), current account as a percentage of XGS, net liquidity as months of import cover, and exchange rate stability. Source: The PRS Group International Country Risk Guide (ICRG).

  • Financial Reform Index: an index of financial sector reform; components include interest rate controls, entry barriers/pro-competition measures, banking supervision, privatization, international capital flows, and security markets. Source: Abiad, Detragiache, and Tressel (2008).

  • Capital account openness: an index of legal restrictions on international financial transactions. Source: Chinn and Ito (2008).

  • Banking crises: indicator of systemic banking crises during the period. Source: Laeven and Valencia (2010).

  • Currency crises: indicator of currency crises during the period. Source: Laeven and Valencia (2008).

  • Public debt/GDP. Source: Abbas, Belhocine, ElGanainy, and Horton (2010).

  • Foreign currency rating: rating of external long-term sovereign debt. Source: Fitch Ratings.

  • Creditors’ rights: an index aggregating creditor rights, following La Porta and others (1998). A score of one is assigned when each of the following rights of secured lenders are defined in laws and regulations: First, there are restrictions, such as creditor consent or minimum dividends, for a debtor to file for reorganization. Second, secured creditors are able to seize their collateral after the reorganization petition is approved, i.e. there is no “automatic stay” or “asset freeze.” Third, secured creditors are paid first out of the proceeds of liquidating a bankrupt firm, as opposed to other creditors such as government or workers. Finally, if management does not retain administration of its property pending the resolution of the reorganization. Source: Djankov, McLiesh and Shleifer (2007).

  • Deposit insurance: indicates existence of deposit insurance. Source: Demirgüç-Kunt, Karacaovali, and Laeven (2005).

  • Stock indices: MSCI USD equity indices; for those countries with no MSCI index, equity indices available in Bloomberg were used. Source: Bloomberg.

  • Stock index volatility: a measure of the risk of price moves for a security calculated from the standard deviation of day to day logarithmic historical price changes. The 360-day price volatility equals the annualized standard deviation of the relative price change for the 360 most recent trading days closing price. Source: Bloomberg.

  • Foreign exchange reserves. Source: IMF World Economic Outlook.

  • Centrality measure: see Appendix II. Source: calculated from BIS Locational Banking Statistics.

  • Deposit money bank assets/(deposit money + central) bank assets: ratio of deposit money bank claims on domestic nonfinancial real sector to the sum of deposit money bank and central bank claims on domestic nonfinancial real sector. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Liquid liabilities/GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Central bank assets/GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Deposit money bank assets/GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Other financial institutions assets/GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Private credit by deposit money banks/GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Private credit by deposit money banks and other financial institutions/GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Bank deposits/GDP: demand, time and saving deposits in deposit money banks as a share of GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Financial system deposits/GDP: demand, time and saving deposits in deposit money banks and other financial institutions as a share of GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Bank credit/bank deposits: private credit by deposit money banks as a share of demand, time and saving deposits in deposit money banks. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Liquid liabilities (mil. 2000 USD): absolute value of liquid liabilities in 2000 US dollars. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Bank overhead costs/total assets: accounting value of a bank’s overhead costs as a share of its total assets. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Net interest margin: accounting value of bank’s net interest revenue as a share of its interest-bearing (total earning) assets. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Bank concentration: assets of three largest banks as a share of assets of all commercial banks. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Bank ROA: average Return on Assets (net income/total assets). Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Bank ROE: average Return on Assets (net income/total equity). Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Bank cost-income ratio: total costs as a share of total income of all commercial banks. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Bank Z-score: Z-score is estimated as (ROA+equity/assets)/sd(ROA); the standard deviation of ROA, sd(ROA), is estimated as a 5-year moving average. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Life insurance premium volume/GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Non-life insurance premium volume/GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Stock market capitalization/GDP: value of listed shares to GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Stock market total value traded/GDP: total shares traded on the stock market exchange to GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Stock market turnover ratio: ratio of the value of total shares traded to average real market capitalization. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Listed companies per 10K population: number of publicly listed companies per capita. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Private bond market capitalization/GDP: private domestic debt securities issued by financial institutions and corporations as a share of GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Public bond market capitalization/GDP: public domestic debt securities issued by government as a share of GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • International debt issues/GDP: international debt securities (amount outstanding) as a share of GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Loans from non-resident banks (net)/GDP: international debt securities (net issues) as a share of GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Loans from non-resident banks (amount outstanding)/GDP: offshore bank loans relative to GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Offshore bank deposits/domestic bank deposits: offshore bank deposits relative to domestic deposits. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • Remittance inflows/GDP: net remittance inflows as a share of GDP. Source: Beck, Demirgüç-Kunt, and Levine (2010).

  • CAR: Regulatory capital to risk-weighted assets. Source: FinStats 2011, World Bank.

  • Capitalization: Two measures of bank capital to assets. Source: FinStats 2011, World Bank and Monetary Statistics based on the Standardized Report Forms, IFS, IMF.

  • NPL ratio: NPLs to total gross loans. Source: FinStats 2011, World Bank.

  • Provisions to NPLs: Provisions to non-performing loans. Source: FinStats 2011, World Bank.

  • Stock volatility: 360-day price volatility for equity indices. Source: Bloomberg.

Appendix II. Measuring Interconnectedness

A Primer

With the increased interconnectedness of the global financial network and the related spillover effects, recognizing and defining the importance of each jurisdiction within the network based on their relationships with other jurisdictions is becoming increasingly important. Borrowing methodology familiar to sociology, biology, and other disciplines where network analysis plays an important role, we explore several measures of interconnectedness (centrality) for the global banking network.

The most basic of these measures, and easiest to understand, is degree centrality. It is defined as the sum of each network member’s (node’s) connections with all other members (nodes) of the network. That is,

ai=jaij(1.1)

Where aij are elements of matrix A, representing relationships between each node. For binary matrices (elements equal to 1 if a connection exists, 0 otherwise), degree will merely be the number of existing connections between node i and all other nodes. While for valued matrices, degree will be the summed values of all existing connections.

A more complex measure of centrality is Bonacich’s eigenvector centrality. Since its introduction in 1972, it has been a widely used measure to gauge the importance of individual nodes within a network. The basic premise underlying this measure is that a node’s importance within the network is recursively related to the importance of the nodes to which it is connected. Where A is a matrix of relationships between nodes, this can be expressed as

xi=a1ix1+a2ix2++anixn(1.2)

With aij defining the relationship that exists between nodes i and j, and x representing the centrality score of each node. Since it is unlikely that this system of equations has a nonzero solution, it can be generalized to assume that each node’s centrality score is proportional to, rather than equal to, the weighted sum of the other nodes’ scores:

λxi=a1ix1+a2ix2++anixn(1.3)

In matrix notation this is expressed as

λx=Ax(1.4)

λ is a scalar and an eigenvalue (characteristic root) of matrix A, while x is an eigenvector (characteristic vector) of matrix A. Each element of the eigenvector, related to each node in the network, provides a relative measure of each node’s importance within the network.

In 1987, Bonacich introduced a related measure21, x^ or beta centrality, which allows for variation in the degree and direction of dependence of each node’s score on other nodes:

x^i=j(α+βx^)Aij(1.5)

In matrix notation, this can be written as:

x^=α(IβA)1A1(1.6)

Again, A is a matrix representing the relationships existing between each node of the network; β reflects the degree to which a node’s score is a function of other nodes’ scores. Small values of β put more emphasis on direct connections between nodes, while larger values take into account more distant connections. That is, small values emphasize the local structure, while larger values put more emphasis on the network as a whole. It can be seen from equation that as β approaches 0, x^, approaches the degree measure (i.e., merely the sum of each node’s immediate connections to other nodes). While this measure allows for more flexibility in defining the relationship between nodes, there are certain assumptions that must be met for meaningful results. More specifically, while not essential for equation, Bonacich’s measure is interpreted under the assumption that β does not exceed the value of the absolute value of the inverse of the largest eigenvalue of matrix A. That is β<|1λ|.

when this condition is met, x^ is an infinite sum:

x^=αk=0βkAk+11=α(A1+βA21+β2A31+)(1.7)

However, if this condition is not met, equation does not converge and x^ loses some of its interpretation.

A more specific form of beta centrality, known as alpha centrality, has also been introduced. In matrix notation this measure can be expressed as

x¯=(IαA)1e(1.8)

When x¯ is the centrality score, α reflects the relative importance of endogenous versus exogenous factors in the determination of centrality, and e represents exogenous factors. Since this measure is almost identical to beta centrality, and also relies upon the abovementioned assumption being met, it yields similar results.

Finally, we consider a newer measure of centrality introduced by Neal (2010), known as alter-based centrality. This measure is also a recursive measure of centrality, but rather than relying upon calculation of eigenvalues and eigenvectors or certain assumptions being met, alter-based centrality is calculated for node i as the sum of each node j’s degree centrality weighted by the connection between i and j:

x˜i=jaijaj(1.9)

Where ij, aij is the relationship that exists between nodes i and j and aj is node j’s degree centrality. This measure, therefore, takes into account the relative importance of nodes to which i is connected and the strength of the connection between nodes; however, it does not rely upon eigenvectors, does not require selection of a beta or alpha value, and does not require any assumptions be met, making it more intuitive and mathematically transparent than its predecessors. While the measures are not exactly the same, they generally tend to show a high degree of similarity (see for example Figure 1).

Data and Network Matrices

Interbank network matrices for each year are constructed using year-end stock data from the Bank for International Settlements’ quarterly locational banking statistics, covering the period 1977Q4 through 2009Q3. Data are reported by 33 jurisdictions22:

article image

Jurisdictions report asset and liability bilateral positions, expanding the dataset to around 200 countries when counterparties are considered, and enabling us to create directed networks capturing the path of the bilateral relationships. It should be noted that while this provides an approximation of the global banking network, it is an incomplete picture since positions among nonreporting countries are not captured; nonreporting countries have a maximum of 66 links (assets + liabilities) in any given year with the reporting countries.

Since we are interested in both the direction and the size of exposures, we characterize the networks using valued matrices rather than binary ones (which would indicate only the existence of a link but not its strength). Given that matrix (network) size varies throughout our sample period and BIS data is reported in current USD, links are expressed as a fraction of matrix totals (that is, as a fraction of total interbank positions for each year). This allows for comparability across years and adjusts for increased interconnectedness throughout the sample (Figure 5).23

A panel time series is constructed after alter-based centrality scores are calculated for each jurisdiction for each year of the sample.

Appendix III. Simulation Framework

The simulation framework developed in this appendix builds on earlier work in this area, in particular Nier and others (2007), which in turn builds on Eboli (2004). In Eboli’s (2004) model of a banking system, nodes (individual banks) are connected to a ‘source’, where the initial shock is generated, and every node is assigned a ‘sink’ where the losses are directed to – the bank’s net worth or capital. When losses reach a node, they are absorbed by the sink or, if they are large enough, they trigger a default and flow further through the network via links. Nier and others (2007) identify external assets as the source of shocks and add depositors to the model as the ultimate sink (recipient of loss). They also define a probability distribution that governs whether or not a bank is exposed to another bank through an interbank exposure. They can therefore study varying degrees of connectivity, ranging from non-connected to complete structures in a systematic way.

Our simulations differ from those in Nier and others (2007) in several respects. In particular, instead of focusing on a banking system in a single country, in which individual banks are the nodes, we are studying the global banking system, where each node represents a country-level banking sector and each link represents a directional lending relationship between two nodes. Another important difference is that the primary focus of Nier and others (2007) is examining the relationship between the degree of interconnectedness in the network as a whole and the overall number of defaults in the network; in contrast, our primary focus is the relationship between the centrality of a node within the network and the likelihood of default in that particular node.

The network is based on the following two exogenous parameters that describe the random graph: the number of nodes (i.e., individual countries), N, and the probability pij that a banking sector in country i has lent to a banking sector in country j. For the beginning of the analysis, the probability pij is assumed to be equal across all (ordered) pairs (i, j). The simulation engine then delivers realizations of this graph that conform to the specified parameters and that exhibit a set number (Z) of realized links. For any realization of the random graph, we populate the individual country banking sector balance sheets in a manner consistent with country-level and global balance sheet identities, broadly in line with the approach used by Nier and others (2007) for individual bank data. In the following text, which describes this in more detail, lower-case letters are used for variables at the country level, capital letters for variables at the global level, and Greek letters are for ratios.

Individual banking sector’s assets, denoted by a, include external assets (investors’ borrowing), denoted by e, and interbank assets (other banking sectors’ borrowing), denoted by i. Thus, for banking system i, we have ai = ei + ii, where i = 1,…,N. Banking sector’s liabilities, denoted by l, are composed of net worth of the banking system, denoted by c, its customer deposits, denoted by d, and its interbank borrowing, denoted by b. Hence for a banking system i, we have li = ci + di + bi, where i = 1,…,N. And as a balance sheet identity, we have ai = li for i = 1,…,N.

To construct balance sheets for individual banking sectors, we first decide on the total external assets of the global banking system, denoted by E. These external assets represent total loans made to ultimate investors and thus relate to the total size of the flow of funds from savers to investors through the banking systems. Next, we decide on the percentage of external assets in total assets (A) at the system level, denoted by β = E/A. Note that the aggregate assets of the global banking network can be written as A = E + I, where I represents the aggregate size of interbank exposures. Hence, for a given aggregate amount of external assets, E, the aggregate ratio of external assets in total assets, denoted by β, delivers both the size of total assets A and the aggregate size of interbank exposures I. That is, we have A = E/β and I = θA, where θ = (1−β) is the percentage of interbank assets in total assets. Dividing the total interbank assets by the total number of links Z, we arrive at the country-level size (the weight) of any directional link, denoted by w (w = I/Z), which determines how much one country’s banking lends to another. Hence, using w and the structure of the network, we can calculate ii and bi.

To determine the size of each country banking sector’s external assets, we assume that its balance sheet satisfies some basic restrictions. In particular, for any banking sector to be able to operate, we require that its external assets are no less than its net interbank borrowing, that is, we have eibiii. To ensure that this constraint is fulfilled, we apply the following two-step algorithm. First, for each country’s banking system, we fill up its external assets so that its external assets plus interbank lending equal its interbank borrowing, i.e. ei*+ ii = bi, where ei* is the level of bank i’s external assets we got after this first step. Second, whatever is left in aggregated external assets is evenly distributed among all banks. Note that total external assets equal E. Hence, a total of (E-Σei*) units of external assets have not been allocated to the individual banking systems’ balance sheets yet. In the second step, we distribute this amount equally among all N banking systems. Now, let us denote ei** the amount to be allocated to each individual bank, (E-Σei*)/N. Hence, we have ei= ei*+ ei**. This completes the asset side of the bank balance sheet as well as interbank borrowing b on the liability side. The remaining components are net worth, c, and deposits, d, on the liability side. Net worth is determined as a fixed proportion γ of total assets at bank level, that is, ci = γai. And, consumer deposits take up the remainder to meet the bank’s balance sheet identity, that is di = aicibi.

This completes the construction of the global banking system and of each constituent banking sector’s balance sheet. All possible global banking systems constructed in this way can be described by four structural parameters (γ, θ, p, N, E), where γ denotes net worth as a percentage of total assets, θ is the percentage of interbank assets in total assets, p is the probability of any two nodes being connected, N is the number of banks, and E is total external assets of the banking system.

We focus on the consequences of an idiosyncratic shock affecting the value of a banking sector’s external assets in one country and spreading through the global banking network. While it may be possible for a shock to affect several (or all) countries at the same time may also be relevant in practice, idiosyncratic shocks are a cleaner starting point for studying knock-on defaults due to interbank exposures and liquidity effects. Aggregate shocks amount to reducing the net worth of all banks at the same time. If this shock is big enough to bring down any bank, in our set-up this will typically lead to all banks defaulting with little scope for further analysis. One can thus think of aggregate shocks as potentially affecting all banks’ net worth to a point where none of the banks are yet in default. We then study the consequences of idiosyncratic shocks for any given positive aggregate net worth.

For any given realization of the global banking system, we shock one country banking sector at a time, by wiping out a certain percentage of its external assets (the ‘source’ of the shock). Let si be the size of the initial shock. This loss is first absorbed by the banking system’s net worth ci, then its interbank liabilities bi and last its deposit di, as the ultimate ‘sink’. That is, we assume priority of (insured) customer deposits over bank deposits which, in turn, take priority over equity (net worth). If the banking sector’s net worth is not big enough to absorb the initial shock, the banking sector defaults and the residual is transmitted to creditor banking sectors through interbank liabilities. And in case these liabilities are not large enough to absorb the shock, some of the losses are born by depositors. Formally, if si > ci, then the banking sector defaults. If the residual loss (si −ci) is less the amount bank bi that banking sector i has borrowed from other banking sectors (i.e., through the cross-border interbank market), then all the residual loss (si −ci) is transmitted to creditor banking systems. However, if (si −ci) > bi, then all of the residual cannot be transmitted to creditor banking systems and depositors receive a loss of (si −cibi).

All creditor banking systems receive an equal share of the residual shock, which in turn, is first absorbed by their net worth. If the net worth is larger than the shock transmitted, the creditor banking system withstands the shock. Otherwise, the creditor banking system “defaults” (has a banking crisis); the residual loss is transmitted through interbank liabilities first, and in the case that these liabilities are not large enough to absorb the shock, the remaining loss is borne by depositors. The part that is transmitted through the interbank channel may cause further rounds of contagious default. The transmission continues down the chain until the shock is completely absorbed.

Formally, let k be the number of creditor banking sectors and let banking sector j be one of those that have lent to banking sector i, the banking sector that has been hit by the initial shock. Hence, if (sici) < bi, then banking sector j has a loss of sj = (sici)/k. If sjcj, then banking sector j withstands the shock. Otherwise, the creditor banking sector defaults and again the residual loss is transmitted through interbank liabilities, and so forth.

References

  • Abbas, S. A., N. Belhocine, A. ElGanainy, and M. Horton, 2010, A Historical Public Debt Database, IMF Working Paper 10/245 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • Abiad, A., E. Detragiache, and T. Tressel, 2008, A New Database of Financial Reforms, IMF Working Paper 08/266 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • 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. 4781.

    • Search Google Scholar
    • Export Citation
  • Allen, F. and D. Gale, 2000, Financial Contagion, Journal of Political Economy, Vol. 108, No. 1, pp. 133.

  • Allen, F. and D. Gale, 2001, Comparing Financial Systems, MIT Press.

  • Allen, F., B. Thorsten, E. Carletti, P.R. Lane, D. Schoenmaker, and W. Wagner, 2011, Cross-Border Banking in Europe: Implications for Financial Stability and Macroeconomic Policies, Centre for Economic Policy Research.

    • Search Google Scholar
    • Export Citation
  • Barth, J. R., G. Caprio Jr., and R. Levine, 2004, Bank Regulation and Supervision: What Works Best?” Journal of Financial Intermediation, 13, pp. 205248.

    • 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
  • Beck, T., A. Demirgüç-Kunt, and R. Levine, (2010), A New Database on Financial Development and Structure, Financial Sector Discussion Paper No. 2, The World Bank, September (updated version of a 1999 database).

    • Search Google Scholar
    • Export Citation
  • BIS, 2010, Research on Global Financial Stability: The Use of BIS International Financial Statistics, CGFS Papers no. 40, June.

  • Bonacich, P., 1972, Technique for Analyzing Overlapping Memberships, Sociological Methodology, Vol. 4, pp. 176-185.

  • Bonacich, P., 1987, Power and Centrality: A Family of Measures, American Journal of Sociology, Vol. 92, No. 5 (March), pp. 1170-82.

  • Bonacich, P., and P. Lloyd, 2001, Eigenvector-like Measures of Centrality for Asymmetric Relations,” Social Networks, Vol. 23, pp. 191201.

    • Search Google Scholar
    • Export Citation
  • Breuer, J. B., 2004, An Exegesis on Currency and Banking Crises, Journal of Economic Surveys, 18 (3), pp. 293320.

  • Buch, C. M., and G. DeLong, 2008, Do Weak Supervisory Systems Encourage Bank Risk-Taking?, Journal of Financial Stability, 4, pp. 2339.

    • 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
  • Calomiris, C. W., and J. R. Mason, 1997, Contagion and Bank Failures during the Great Depression: The June 1932 Chicago Banking Panic,” American Economic Review, 87 (5), pp. 86383.

    • Search Google Scholar
    • Export Citation
  • Caprio, Gerard, Daniela Klingebiel, Luc Laeven, and Guillermo Noguera, 2005, Appendix: Banking Crisis Database, in Patrick Honohan and Luc Laeven (eds.), Systemic Financial Crises: Containment and Resolution (Cambridge, U.K.: Cambridge University Press).

    • Search Google Scholar
    • Export Citation
  • Chinn, M., and H. Ito, 2008, A New Measure of Financial Openness, Journal of Policy Analysis, vol. 10, no. 3, pp. 309322.

  • Cihak, M. and K. Schaeck, 2010, How Well Do Aggregate Prudential Ratios Identify Banking System Problems?, Journal of Financial Stability, vol. 6, issue 3, September.

    • Search Google Scholar
    • Export Citation
  • Cole, A. and J. Gunther, 1998, Predicting Bank Failures: A Comparison of On-and Off-site Monitoring Systems, Journal of Financial Services Research, Vol. 13, no. 2.

    • Search Google Scholar
    • Export Citation
  • Das, U. S., M. Quintyn, and K. Chenard, 2004, Does Regulatory Governance Matter for Financial System Stability? An Empirical Analysis, IMF Working Paper 04/89 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • Davis, E.P. and Karim, D., 2008, Comparing Early Warning Systems for Banking Crises, Journal of Financial Stability, 4, pp. 89120.

  • Demirgüç-Kunt, A. and E. Detragiache, 1998, The Determinants of Banking Crises in Developing and Developed Countries, IMF Staff Papers, Vol. 45, No. 1.

    • Search Google Scholar
    • Export Citation
  • Demirgüç-Kunt, A. and E. Detragiache, 2005, Cross-Country Empirical Studies of Systemic Bank Distress: A Survey. IMF Working Paper 05/96 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • Demirgüç-Kunt, A., B. Karacaovali, and L. Laeven, 2005, Deposit Insurance around the World: A Comprehensive Database, World Bank Policy Research Working Paper 3628, June.

    • Search Google Scholar
    • Export Citation
  • Diamond, D. W., and P. H. Dybvig, 1983, Bank Runs, Deposit Insurance, and Liquidity Journal of Political Economy, 91 (3), pp. 401419.

    • Search Google Scholar
    • Export Citation
  • Djankov, S., C. McLiesh, and A. Shleifer, 2007, Private Credit in 129 Countries, Journal of Financial Economics, vol. 84, pp. 299-329.

  • Eboli, M., 2004, Systemic Risk in Financial Networks: A Graph Theoretic Approach, mimeo, Universita di Chieti Pescara.

  • Edison, H., 2002, Do Indicators of Financial Crises Work? An Evaluation of an Early Warning System, International Journal of Finance and Economics, vol. 8, issue 1.

    • Search Google Scholar
    • Export Citation
  • Eichengreen, B. and C. Arteta, 2000, Banking Crises in Emerging Markets: Presumptions and Evidence, Paper C00’115, Center for International and Development Economics Research (Berkeley: University of California).

    • Search Google Scholar
    • Export Citation
  • Estrella, A., S. Park and S. Peristiani, 2000, Capital Ratios as Predictors of Bank Failure, Federal Reserve Bank of New York Economic Policy Review, vol. 6, no. 2.

    • Search Google Scholar
    • Export Citation
  • Gai, P, and S. Kapadia, 2010, Contagion in Financial Networks, Bank of England Working Paper No. 383 (London: Bank of England).

  • Gavin, M. and R. Hausmann, 1996, The Roots of Banking Crises: The Macroeconomic Context, In: Hausmann, R. and L. Rojas-Suárez, (eds.), Banking Crises in Latin America, (New York: Inter-American Development Bank), pp. 27 -75.

    • Search Google Scholar
    • Export Citation
  • Gorton, G., 1988, Banking Panics and Business Cycles, Oxford Economic Papers, New Series, 40 (4), pp. 75181.

  • Gropp, R., J. Vesala, and G. Vulpes, 2004, Market Indicators, Bank Fragility, and Indirect Market Discipline, Federal Reserve Bank of New York Economic Policy Review, 10 (2), pp. 5363.

    • Search Google Scholar
    • Export Citation
  • Hale, G., 2011, Bank Relationships, Business Cycles, and Financial Crisis, FRBSF Working Papers No. 14, May.

  • Hardy, D. and C. Pazarbaşioğlu, 1998, Leading Indicators of Banking Crises: Was Asia Different? IMF Working Paper 98/91 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • Herrmann, S. and D. Mihaljek, 2010, The Determinants of Cross-Border Bank Flows to Emerging Markets: New Empirical Evidence on the Spread of Financial Crises, BIS Working Papers No. 315, July (Basel: Bank for International Settlements).

    • Search Google Scholar
    • Export Citation
  • Huang R. and L. Ratnovski, 2010, The Dark Side of Bank Wholesale Funding, IMF Working Paper 10/170 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • Hutchinson, M. M., 2002, European Banking Distress and EMU: Institutional and Macroeconomic Risks, Scandinavian Journal of Economics, 104 (3), pp. 36589.

    • Search Google Scholar
    • Export Citation
  • Hutchinson, M. M., and K. McDill, 1999, Are All Banking Crises Alike? The Japanese Experience in International Comparison, Journal of the Japanese and International Economies, 13, pp. 15580.

    • Search Google Scholar
    • Export Citation
  • IMF/BIS/FSB, 2009, Guidance to Assess the Systemic Importance of Financial Institutions, Markets and Instruments: Initial Considerations, October.

    • Search Google Scholar
    • Export Citation
  • International Monetary Fund, 2010, “Integrating Stability Assessments Under the Financial Sector Assessment Program into Article IV Surveillance: Background Material”, available at http://www.imf.org/external/np/pp/eng/2010/082710a.pdf.

    • Search Google Scholar
    • Export Citation
  • Ilzetzki, E., C. Reinhart, and K. Rogoff, 2008, “Exchange Rate Arrangements Entering the 21st Century: Which Anchor Will Hold?This paper includes updates to the exchange rate regime classifications in Reinhart, C., and K. Rogoff, 2004, “The Modern History of Exchange Rate Arrangements: A Reinterpretation,” Quarterly Journal of Economics 119(1):1-48.

    • Search Google Scholar
    • Export Citation
  • Kaminsky, G., S. Lizondo and C. Reinhart, 1998, Leading Indicators of Currency Crises, IMF Staff Papers, 45.

  • Kaminsky, G. and C. Reinhart, 1999, The Twin Crisis: The Causes of Banking and Balance of Payments Problems, American Economic Review, vol. 89 (3).

    • Search Google Scholar
    • Export Citation
  • Kinney, R, Crucitti, P, Albert, R and Latora, V, 2005, Modeling Cascading Failures in the North American Power Grid, The European Physics Journal, B 46 (2005) 101.

    • Search Google Scholar
    • Export Citation
  • Kubelec and , 2010, The Geographical Composition of National External Balance Sheets: 1980-2005, Bank of England Working Paper No. 384, March (London: Bank of England).

    • Search Google Scholar
    • Export Citation
  • Laeven, L. and F. Valencia, 2008, Systemic Banking Crises: A New Database, IMF Working Paper 08/224 (Washington: International Monetary Fund).

    • 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 10/146 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • May, R.B and R.M. Anderson, 1991, Infectious Diseases of Humans, Oxford University Press.

  • Miskhin, F., 1978, The Household Balance Sheet and the Great Depression, Journal of Economic History, 38, pp. 91837.

  • Minoiu, C. and J. A. Reyes, 2011, A Network Analysis of Global Banking: 1978–2009, IMF Working Paper No. 74 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • Neal, Z., 2010, Differentiating Centrality and Power in the World City Network, Urban Studies, Vol. 1–16.

  • Neal, Z., 2011, Comparing Beta Centrality and a Modified Degree Centralitypresented at Sunbelt XXXI Conference, St. Pete Beach, Florida, February.

    • 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. 20332060.

    • Search Google Scholar
    • Export Citation
  • Rose, A. and M. Spiegel, 2009, Cross-Country Causes and Consequences of the 2008 Crisis: Early Warning, NBER Working Paper no. 15357.

  • Sachs, A., 2010, Completeness, Interconnectedness and Distribution of Interbank Exposures – A Parameterized Analysis of the Stability of Financial Networks, Deutsche Bundesbank Discussion Paper Series 2: Banking and Financial Studies No 08/2010.

    • Search Google Scholar
    • Export Citation
  • Von Peter, 2007, International Banking Centres: A Network Analysis, BIS Quarterly Review, December.

  • Watts, D., 2002, A Simple Model of Global Cascades on Random Networks’, Proceedings of the National Academy of Sciences 9: 5766-5771.

  • Wooldrige, J.M., 2005, Simple Solutions to the Initial Conditions Problem in Dynamic, Nonlinear Panel Data Models with Unobserved Heterogeneity,” Journal of Applied Econometrics, No. 20, pp. 3954.

    • Search Google Scholar
    • Export Citation
1

The paper benefitted from discussions with Dimitri Demekas and comments by Eugenio Cerutti, Stijn Claessens, and Camelia Minoiu. We thank Zachary Neal and Goetz von Peter for sharing their codes and unpublished papers. Any remaining errors are ours.

2

Literature suggests that various aspects of finance have both a ‘bright side’ and ‘dark side’ (see, e.g., the discussion of the ‘dark side’ of bank wholesale funding in Huang and Ratnovski, 2010).

3

Throughout the paper, the term “country” or “jurisdiction” refers to a geographic entity for which banking and other statistics are reported separately.

4

A complete network is one in which all nodes are connected with each other.

5

BIS locational statistics are based on the residency principle: data is collected on the positions of all banking offices located within each reporting jurisdiction, on a gross basis, regardless of nationality. This is consistent with balance of payment and external debt methodology and therefore the rest of the variables used in our analysis. The BIS data also include consolidated banking statistics, showing worldwide consolidated international financial claims of domestically owned banks (that is, the statistics are compiled using the nationality principle and interbank positions are netted out). Given that foreign subsidiaries are not wholly dependent on their parent bank and there exists the possibility of contagion from child to parent (and vice versa), and the fact that our banking crisis variable does not distinguish between countries’ crises involving only domestically owned banks versus those involving domestic and foreign banks, we have used the locational statistics, which essentially consider all banking offices as separate entities.

6

For other recent contributions to the network analysis in finance, see e.g. Nier and others (2007), Alessandri and others (2009), Hale (2011), and Minoiu and Reyes (2011).

7

The data on the interconnectedness measure used in this paper are available from the authors upon request.

8

In a similar vein, Minoiu and Reyes (2011) study directed networks using BIS locational data. Our upstream/downstream degree variables are similar to their in/out strength variables in that they both consider valued, rather than binary, network matrices. The main difference is that their analysis focuses on the dynamics and evolution of network metrics (flows rather than exposures), while our focus is on the impact of interconnectedness on financial stability. Appendix II has details on our calculations of centrality variables.

9

The BIS banking statistics include banks’ assets and liabilities vis-à-vis residents in foreign currencies, so exposures to other jurisdictions could be overestimated in some cases.

10

In addition to the random graph simulations shown in Figure 3, we have also run simulations in which there are two groups of nodes: those with a high degree of connectivity (“money centers”) and those with lower connectivity. This is similar to the “money center” networks examined by Sachs (2010), except that we are examining this in the context of the global country network rather than a single-country interbank market. Compared to the random graphs, crises tend to be more frequent in the “money center” networks, consistent with the findings of Sachs (2010). The general shape of the relationship between interconnectedness and stability remains as shown in Figure 3. The additional results are available upon request.

11

For the 33 main jurisdictions for which BIS comprehensive cross-border exposure data are available, the network is 92 percent “complete” in the sense of Allen and Gale (2000), i.e. there are 92 percent of links out of all the theoretically possible links among those 33 jurisdictions. For the full sample of 200 jurisdictions, a precise determination is difficult due to missing values, but based on our estimates, the network is less than half “complete” (reflecting the much lower degree of network completeness outside the 33 reporting jurisdictions).

12

Allen and Gale (2000) and the related literature emphasize that a “complete” financial network (one with full links among all banking sectors around the world) is more resilient to shocks than “incomplete” networks. However, Allen and Gale (2000) focus on a highly stylized small system with 4 individual banks. In our simulation, we had a global banking network of 100 banking sectors that were interconnected, but did not form a “complete” network.

13

Table 1 shows the results with the cubic term of interconnectedness. Similar results are obtained with a quadratic term.

14

This follows the study by Wooldridge (2005), showing that standard random-effects probit estimates can be used on dynamic panel datasets.

15

When cost-to-income ratio is excluded, the number of observations increases by about 1,000. With the addition of time dummies, the estimates become highly significant.

16

As with any similar probit estimations, we are careful not to interpret our results in a casual way.

17

This calculation is meaningful only for upstream interconnectedness, given that the point estimate for downstream interconnectedness is insignificant.

18

This is because the interaction of economic growth with interconnectedness reduces the probability of banking crisis, as indicated in estimate (5) in Table 1.

19

Note that we are using here the alter-based centrality measures, the theoretical maximum for which is 25. In our sample, the maximum observed value is 1.75 for the liability measure and 1.37 for the asset measure. Note also that the maximum observed values declined somewhat in the 1990s and 2000s, as the global banking network became bigger and more multi-nodal (rendering individual nodes less ‘central’ to the network).

20

As a side note, the comparison with a simple polynomial filter (order 3) illustrates that the nonparametric approach emphasizes reducing Type II errors (missed crises)—it includes among the “higher probability” even areas that, based on the polynomial filter, could be identified as medium-probability.

21

Bonacich’s 1987 measure is related to Katz’s 1953 measure. See Bonacich (1987) for details.

22

Beginning of reporting to BIS varies by jurisdiction. Additional confidential data has also been included for some off-shore financial centers.

23

For example, a $100 link is given a value of .5 in a 2x2 matrix totaling $200. If the network in the next year is an 8x8 matrix totaling $5000, a $100 link is relatively less important, receiving a value of only .02. If however, the 8x8 matrix totals only $200, a $100 is still relatively important within the network despite the existence of additional links, and would be given a value of .5. Note that we do not make adjustments for, or distinguish between, increases in matrix size due to additional BIS reporting countries and increases due to other increases in interconnectedness.

The Bright and the Dark Side of Cross-Border Banking Linkages
Author: Ms. Sònia Muñoz, Mr. Ryan Scuzzarella, and Mr. Martin Cihak