Are central neighbors better neighbors?
The impact of network centrality on international bank lending after the global financial crisis
J. M. (Jochem) Wissenburg 1983903
j.m.wissenburg@student.rug.nl
University of Groningen, Faculty of Economics and Business Supervisors: Prof. dr. S. Brakman, dr. M.J. Gerritse De Nederlandsche Bank, afdeling Financiële stabiliteit
Supervisors: A.L. Levels, dr. C.A. Ullersma
Abstract:
This thesis takes a network approach to describe the events that unfolded on the interbank market after the collapse of Lehman Brothers. Using data on international claims, this thesis examines whether higher centrality of a country within the global interbank network before the financial crisis has led to greater declines in outward international lending after the crisis. Centrality has had a mostly negative impact on post crisis lending, yet the significance for different centrality indicators is dependent on circumstances such as the time frame, sample size, and inclusion of controls. Therefore, this thesis has provided only moderate evidence for a relationship between ex-ante centrality in the global network and ex-post outward international lending.
2 Index 1. Introduction ... 3 2. Networks in finance ... 6 3. Related literature ... 10 3.1 Contagion in networks ... 10
3.2 Core v. peripheral lending ... 13
4. Methodology and Data ... 17
4.1 Methodology ... 17
4.1.1 Measurement of the dependent variable ... 17
4.1.2 Measuring centrality ... 18
4.1.3 Econometric procedure ... 22
4.2 Data ... 24
4.3 Descriptive statistics ... 25
4.4 Statistical requirements ... 25
5. The interbank network ... 27
6. Results ... 31
6.1 A first statistical illustration ... 31
6.2 Formal econometric specification ... 32
6.3 Sensitivity checks ... 35
6.3.1 Centrality as a proxy for bank characteristics ... 35
6.3.2 Excluding US banks from the analysis ... 37
6.3.3 Alternative time frame ... 38
7. Discussion ... 40
9. Bibliography ... 42
10. Appendices ... 46
Appendix 1: Sample of lending countries ... 46
Appendix 2: Abbreviations of BIS countries from section 5 ... 47
Appendix 3: List of bank controls ... 48
Appendix 4: Statistical requirements ... 49
3
1. Introduction
The financial crisis of 2008 has spurred an interest among policy makers and academics alike to study contagion in financial markets. Contagion can be defined as the risk that a failure or losses in one country cause a spiral of losses and defaults in financial markets or other countries (Claessens and Forbes, 2004). This financial crisis has demonstrated that financial market participants are linked by many direct and indirect channels. According to Haldane (2009), the direct losses from Lehman’s default amounted to USD 5 billion, and as a consequence thereof the IMF estimated that global economic growth would be 5 percent lower than previously expected for 2009. This shows that it is not sufficient to solely look at the performance of single countries to gauge the potential impact on the financial system. To fully comprehend the interrelating aspects of global contagion, one should take into account the position of countries in a network of highly interconnected market participants (Haldane, 2009). For this reason, economists have started to apply network analysis – commonly used in fields such as ecology, epidemiology, and computer science - to financial systems (Börner et al, 2007). Network analysis provides tools for economists to better understand the structure and development of financial markets as well as the importance and connectivity of a country in the network, for example by generating graphs or computing the centrality of nodes (see Caballero, 2015 and Léon and Berndsen, 2014).
To date network analysis has mainly dealt with two separate topics of research. First, scholars have started uncovering topologies and properties of financial structures. These studies have provided insights into the structure and development of financial markets. They emphasize the complex and heterogeneous nature of financial networks and furthermore show that financial networks are composed of many relatively unconnected and few highly connected banks. Second, studies have dealt with contagion in financial networks. These studies indicate that financial networks exhibit a ‘robust yet fragile’ nature (Haldane, 2009). While losses of an important bank may cause a spiral of defaults in the system, failures in the periphery of the financial system hardly lead to destabilizing events.
4
Network analysis deepens the research as it allows the researcher to differentiate between countries that are either highly connected (part of the core) or poorly connected (part of the periphery). Consequently, it is also to be expected that, depending on the position of country in a network, heterogeneous outward lending patterns are observed after the crisis. For example, on the one hand studies have shown that more interconnected countries are more vulnerable to shocks. On the other hand, exactly due to their interconnectedness, these countries also have a more diversified funding inflow (Caballero, 2015).
In this paper, it is hypothesized that countries that were ex-ante more central in the global financial network have exhibited a stronger decline in outward interbank lending to other countries after the financial crisis. The question is relevant from a financial stability perspective because the recent financial crisis has shown how fast a crisis can spread throughout the financial system. To capture these effects on the aggregate level, this paper uses the country as the cross-sectional unit of interest. Thereby it helps to have a greater understanding of the international dynamics of contagion and how different nodes in the financial network can help to spread it.
This study extends on a strand of literature that studies the vulnerability of individual nodes to crises (see e.g. Caballero, 2015; Kali and Reyes, 2010). Generally, these studies show that countries having a more central role in the global financial network are more vulnerable to spread liquidity crises. Second, it adds to studies on post-crisis financial contagion (see e.g. Cetorelli and Goldberg, 2011; Khwaja and Mian, 2008). These studies in general show that larger exposures to a crisis lead to greater declines in outward lending. This study uses Bank of International Settlement (BIS) locational data on international claims between 27 reporting countries and their counterparts. These data reveal interbank flows aggregated on a country-wide basis.
5
international interbank lending behavior through either centrality measures or for example characteristics of a country’s banking system. Most variance within the sample is due to the heterogeneous characteristics of the borrowing countries. This is in line with other studies on post-crisis lending, who as well demonstrated international lending is best explained by the heterogeneous characteristics of borrowing and lending countries captured by fixed effects (De Haas and Van Horen, 2012, Cetorelli and Goldberg, 2011)
6
2. Networks in finance
Networks are theoretical concepts that have been applied to study an interconnected web of nodes. Network analysis does not solely focus on bilateral connections, its aim is to study the overarching structure of all bilateral connections combined. Networks are defined as a set of nodes that are connected by links (see table 1 for a few examples). Network analysis has been applied in research fields as diverse as biology and epidemiology and increasingly as well in economics and finance (Haldane, 2009). In the latter context nodes are represented by economic units such as banks or countries. Links are defined as exposures or flows from one node to another. In the literature networks have been defined in several ways. For example in Caballero (2015), nodes are defined as banks and the linkages between them are syndicated loans. Minoiu and Reyes (2011) define a network of countries connected by international claims. Heijmans et al. (2014) study a network in which banks are nodes that are linked through overnight payment transactions. This shows that network analysis can be a useful tool to study behavior and flows in financial markets. In this study the nodes in the network are defined as a set of countries that are connected by interbank claims.
7 Table 1: Network models
Network Connections Related literature Example
Random Randomly determined following a probability distribution
See e.g Erdös en Renyi (1960), Watts and Strogatz (1998), Allen and Gale (2000)
Scale-free On basis of preferential attachment following a certain probability distribution leading to a heterogeneous network
See e.g. Barabási and Albert (1999), Boss et al. (2004), Wetherhilt et al. (2010), Léon and Berndsen (2014)
Core-periphery
Node heterogeneity, with core banks
performing an
intermediary role for peripheral banks leading to a hierarchy among nodes
See e.g. Craig and von Peter (2010) Van Lelyveld and in ’t Veld (2012), Lux (2015), and In ’t Veld et al. (2014)
1
One of the earliest models in network theory is the random model, discussed in Erdös and Renyi (1960). In these models it is assumed that connections between nodes emerge randomly. The number of nodes in such networks is fixed. This creates a structure wherein each node is as connected as every other node in the network. Random networks are thus fairly homogeneous networks.
Yet, there is overwhelming evidence in the literature that financial networks tend towards a heterogeneous shape. Therefore, random models have been deemed more or less useless to explain dynamics in financial networks (Van Lelyveld and In ‘t Veld, 2012). Instead, one part of the literature acknowledges the scale-free (SF) nature of interbank
1
8
markets, see for example Boss et al. (2004) for Austria, León and Berndsen (2014) for Colombia, and Wetherhilt et al. for the UK (2010). Another strand of literature argues that financial markets display a core-periphery (CP) division, see for example Van Lelyveld and In ‘t Veld (2012), in ‘t Veld et al. (2014), and Craig and von Peter (2010). A third body of literature explains heterogeneous network structures from a historical and geographical perspective. For example Kindleberger (1974) discusses the role of money market centers. These are countries that have an important intermediary role in the conduct of global finance. Finally, Von Peter (2007) shows that money center banks have a vital function in market making. They deal in a wide variety of markets and financial products.
Although there are (theoretical) differences between the highly stylized SF, CP models (see table 1) and studies on money center banks and markets, all of the above mentioned studies are similar in the sense that they acknowledge a heterogeneous network structure with a well-connected core and a large group of unconnected nodes in the periphery.
9
10
3. Related literature
In section 2 it was shown that international banking networks follow a hierarchical pattern with many unconnected and few highly connected nodes. This study deals with the impact of this structure for international financial contagion. Financial contagion occurs when a shock or crisis dissipating from one country negatively affects other countries. These shocks are, amongst others, transmitted via the international bank lending network. A financial shock caused by for example a failure of a bank in one country may influence the financial conditions in another. International bank lending thus provides a base for analysis of the effects of financial shocks. Taking the international network structure into account when studying contagion may provide interesting new insights. Other empirical studies have discussed interbank contagion mostly in the context of liquidity shortages, risk, or for instance the distance between two countries. Until now, the literature has to lesser extend discussed the role of a country’s particular position in a network for contagion. A network perspective to international contagion might help to better understand the non-linear dynamics of contagion and how different nodes in the financial network can help to spread it.
3.1 Contagion in networks
The contagion literature has identified multiple channels through which crises occurring in one node can impact conditions in another. The focus in this paper is on contagion caused by financial linkages and liquidity shortages that are transmitted from one country to another (Claessens and Forbes, 2004), which can be analysed by using international bank lending data in order to compute the network. This cascade of losses and defaults as caused by financial linkages was for example shown in a seminal paper by Allen and Gale (2000). They demonstrate that a default in one part of the financial system may lead to a series of failures of connected nodes.
11
local crisis saw a greater decline in lending as well. In a similar vein, Cetorelli and Goldberg (2011) demonstrate that countries that were more intensely connected to the US saw a greater decline in international bank lending after the subprime crisis of 2007-2008. A third contribution comes from De Haas and Van Horen (2011). They show that countries that are more distant from a specific bank have seen greater decline in financial flows from that bank after the crisis. Finally Tintchev (2014) demonstrates the role that risk plays in lending. However, none of these studies has addressed network dynamics in the study of interbank contagion. They fail to explain what Haldane (2009) calls non-linear dynamics in financial networks. For example, why is it that the Pakistani shock as described in the Khwaja and Mian study is only locally felt, while the shock in the USA as described by Cetorelli and Goldberg had a global impact? A network approach helps to understand how the position of country in the global interbank network explains post-crisis lending dynamics, for example how aggregate lending changed after the financial crisis of 2007.
12
financial crisis, such as falling asset prices that increasingly deteriorate banks’ balance sheets (Summer, 2013).
While the abovementioned network studies have dealt with the question in what case contagion is likely in networks (and have provided a rather nuanced image on default contagion) a third strand has researched whether highly connected countries or banks in a financial network are more resilient to crises and contagion than relatively unconnected nodes. In contrast to simulations discussed above, this literature approaches networks from the level of the nodes. In that context, Kali and Reyes (2010) demonstrate for a trade network, that countries that have been better connected in the global trade network are also better able to dissipate exogenous financial shocks than peripheral countries. Furthermore they show that shocks dissipating from a central node have a larger impact than those coming from a peripheral epicenter. Caballero (2015) on the contrary argues that increased financial connectivity of nodes in a network is a determinant for the occurrence of crises. According to Caballero, countries that rely more on inward flows (measured through network statistics) in the global financial system, face more financial crises. Yet, he also shows that countries wherein banks play an intermediary role (the core), experience fewer banking crises. Caballero argues that banks in these countries have more liquidity sources that they can tap into. In addition to that, Chinazzi et al. (2012) find that high interconnectedness of a country diminishes the effect of a crisis as it “allows adverse shocks to dissipate quicker.” Interesting is however that Hale et al. (2014) argues that intermediary banks in the international network are more exposed to shocks in network as they have more connections. He shows that network centrality is negatively related to bank performance, especially in case that a bank has many crisis exposures. This interesting dichotomy in the literature with respect to network centrality and crises provides a basis for further research into the role of centrality on post-crisis contagion.
13
financial crisis. Up to date, the heterogeneity in international bank lending has not been researched from the perspective of networks. A network approach therefore helps to explain if countries that have a different position in a network also have seen different outward lending patterns after a crisis.
3.2 Core v. peripheral lending
This discussion next turns to literature that provides more arguments as to why central countries in the global network are more likely to reduce lending after a major financial crisis. Although the focus in this thesis is primarily on international banking flows, some literature is cited that takes banking systems or individual banks as a point of reference. The aim of this part is to provide insights into the advantages and disadvantages of being in the center of a financial network with respect to contagion.
14 Figure 1: Aggregated losses of banks by size class
Figure 1 illustrates the impact of international liquidity shocks on market-funded banks and deposit-funded banks in the Eurozone. The size of the banks is depicted by the colors in the graph, whilst the vertical axis shows the aggregate losses per size class in a certain year (horizontal axis). While losses were greatest for the large global banks at the onset of the crisis when liquidity markets dried up, losses were more equally distributed once the economic crisis really kicked in.
15
Still, some studies have pointed into other directions. In these studies it is argued that being in the center of a network provides a country or bank with additional sources of liquidity. As mentioned in section 3.1, Caballero (2015) shows that intermediary countries in the global financial network are overall more resilient to crises. He argues that these banks have a more diversified funding base that allows them to deal with liquidity crises more easily. In that context, Craig et al. (2013) argue that “having a large portfolio of established lending relationships seems to help banks meeting their liquidity requirements more efficiently.” In case of a curb on liquidity in markets, these banks are thus more likely to still fulfill their requirements, making them potentially more stable lending partners. Through network analysis we may observe which effect was stronger in the case of the global financial crisis.
A second set of arguments relates to regulation and subsequent behavior of banks. Since the last crisis there has been an increasing focus on too-big-too-fail (TBTF) banks. On the one hand it could be argued that these kinds of banks are better monitored and thus more trusted by markets (Qi, 2008). This would provide additional sources of stable funding for core countries, which could in turn lead to more stable outgoing flows. Other research finds that the implied stability of core countries often leads to a safe haven affect after crises. However, on the contrary, Van Rijckeghem and Weder (2003) show that developed countries often simultaneously reduce their funding to different peripheral countries when a crisis occurs, in search for more secure countries.
16
Finally, insights into contagion through core countries are also provided for in the context of monetary policy. Leon et al. (2014) show an important role for so-called super-spreaders in the conduit of monetary policy. Super-super-spreaders are those banks that make most use of a central bank’s monetary policy and further divide these funds in the market. Leon et al. argue that this set of core banks plays an important role in the dispersion of liquidity operations of the central bank. However, they also suggest that this may point to a role of these spreaders in case of contagion. Second in the context of monetary policy transmission, Cetorelli and Goldberg (2012) introduce the argument that internationally active banks are insulated from domestic shocks that are due to monetary policy as they have access to liquidity through their various branches around the globe. They furthermore show that this insulation is a source of contagion: globally active banks significantly reduce lending to other countries when domestic liquidity is curbed. Since it also to be expected that global banks are positioned in those countries that are more central in the international financial network, an analysis through the use of network statistics is able to capture these different dynamics.
17
4. Methodology and Data
4.1 MethodologyIn this thesis it is tested whether the heterogeneous lending patterns between countries after the financial crisis can be explained by the ex-ante centrality of the lending country in the global interbank market. It is hypothesized that countries that were more central before the crisis, have seen greater declines in outward international bank lending after the crisis. Therefore the methodology in this thesis is based on studies that have earlier discussed contagion through the international bank lending channel (see e.g. Cetorelli and Goldberg, 2011, Khwaja and Mian, 2008, and De Haas and Van Horen, 2011). These studies are similar to this thesis in the sense that they discuss the impact on post-crisis lending of a particular ex-ante conditions. For example, Cetorelli and Goldberg discuss the impact of ex-ex-ante vulnerability to the dollar market on post-crisis lending. Therefore these studies have been deemed useful as this thesis is concerned with the role ex-ante centrality. Other specifications that were considered are those of Hale (2014), Kali and Reyes (2010) or Caballero (2015). Yet, these papers have dealt with contagion in the context of stock market returns, bank performance, or the incidence of crises and were therefore deemed less useful. The focus of this paper is on contagion through the international bank lending channel.
4.1.1 Measurement of the dependent variable
The dependent variable in this research is the change in the log of average exposures between country i and j in the pre-crisis period (2006q2 – 2007q2) and the post-crisis period (2008q3 – 2009q2). Taking log differences is a common method in studies on post-crisis lending.23 First it allows researchers to deal with the skewness that is so often observed in exposure data. Second it helps to correct for relative differences in size of loans.
2
Khwaja and Mian look at the change in log loan amount from banks to firms before and after a crisis in Pakistan. De Haas and Van Horen take the change in log volume of loans between a pre-crisis and post-crisis period from banks to countries. Cetorelli and Goldberg compare average lending growth in the pre-crisis and post-crisis period and take log difference as well.
3 De Haas and Van Horen (2011) research contagion as well by taking ‘sudden stop’ as a dependent variable,
18
The crisis periods were defined in the following papers by De Haas and Van Horen (2011), and Cetorelli and Goldberg (2011). Both have roughly defined the pre-crisis period from mid-2006 to mid-2007 and the post-crisis period from fall 2008 to the summer of 2009.4 These studies purposely leave out the period in between. Generally it is argued that market functioning was normal up to 2007q2 (when the commercial paper market dried up), and turmoil started after the crash of Lehman in 2008q2. As was shown in the literature section (figure 1) the crisis on the global banking markets was most imminent in 2007-2008, therefore this thesis first follows the definition by Cetorelli and Goldberg (2011) who have defined the pre-crisis period from 2006q2 to 2007q2 and the post-crisis period from 2008q3 to 2009q2. The two periods allow for comparison of exposures in normal and crisis times. Later, in a sensitivity test, an alternative time setting is applied.
4.1.2 Measuring centrality
The independent variables in this research are different centrality indicators, combined providing overall image of the centrality of a country. Network statistics that measure the centrality of a country in a financial network are generally referred to as centrality indicators. Centrality indicators measure a country’s ‘de facto integration’ in global financial markets (Caballero, 2015). In the network in this thesis, nodes are represented by countries, the links among them by international claims. Each link between two nodes thus represents a direction and a certain weight. In network terminology the shortest link between two connected nodes in a network is called a geodesic path. The network statistics are computed from the ex-ante global interbank market in 2006q2. 5
Centrality is an ambiguous concept and can be captured by a wide variety of indicators. In this thesis a selection of these statistics have been used (Caballero (2015) uses over 20 different statistics throughout his paper). Caballero identifies four types of centrality indicators that measure connectivity of a node in the network. The first type of centrality indicators are degree measures. Degree measures compute the number of links going in and
4 Tintchev (2014) initially defines a crisis period from September 2007 to June 2010 and a post-crisis period
from September 2010 to September 2011. However as was shown in the literature section (figure 1) the crisis on the global banking markets was most imminent in 2007-2008.
5
19
out of a node. In this thesis, the measures degree and in-degree have been incorporated. 6 Degree reflects the total number of incoming and outgoing flows from and to a node. So each degree-point represents one financial flow in or out of a country. In-degree only measures the amount of incoming flows. While degree is a broad measure of connectivity, in-degree indicates the dependence on incoming flows. In-degree and degree reflect the immediate vulnerability of a node to a shock that dissipates through the network. Figure 2 serves as example by showing node degree scores within the nodes. The advantage of degree measures is that they are easily computed and that the extensive margins are relatively stable.
Figure 2: Network with node degree scores 7
The disadvantage of degree measures is that they do not take the whole network structure into account and the size of the flows does not matter. For example a node that has only two links which are very considerable in size, would attain a low level of degree. For this issue, weighted degree and weighted in-degree were included. These measures correct for the size of a flow. Weighted in-degree is a measurement of a country’s dependence on borrowing in the network, while weighted degree measures dependence on both borrowing and lending (Caballero, 2015). The problem with degree statistics is that they are in essence ‘local measures’. Degree measures only take direct linkages of a country into regard.
A second type of measures was therefore included that captures the intermediary role of a node, such as betweenness and clustering coefficient. In this thesis betweenness was
6 All centrality measures were calculated with network plotting tool gephi, which provides algortithsm for each
measure. See Bastian and Heymann (2013). This tool has as well been employed by Caballero (2015).
7
Images drawn from lecture on a network centrality at Bryn Mawr College,
20
included.8 Betweenness measures how often a node is on the shortest path between two other nodes. Figure 4 illustrates this statistic. The node in the middle of the network is on the shortest path in all possible connections and therefore attains score 10. Other nodes are not on any shortest path and therefore attain score 0. In a financial context, betweenness allows the researcher to identify which countries have an important intermediary role. In the research by Caballero (2015), from a wide range of network statistics, betweenness attains the highest level of significance in describing crises. Using betweenness as well, Hale (2014) shows that centrality is significant and negatively related to bank performance. Therefore betweenness was deemed to be a useful statistic.
Figure 4 : Network with node betweenness scores 9
Like Caballero (2015), two other types of measures have been excluded from this research. The first type bases centrality on the length of geodesic paths between nodes, an example is closeness. Yet according to Caballero the computation of these measures is most likely to be biased and thereby doesn’t reflect the true extend of a node’s centrality. Therefore they have been excluded in this research as well. Secondly some measures take the whole network structure into account when computing node centrality (e.g. PageRank, eigenvector
8
Centrality measures clustering coefficient, hub, and authority were as well considered. Clustering coefficient measures the importance of a node in particular cluster of the network. Authority and hub relative dependence on inward and outward flows. However as the BIS network only includes a limited amount of lending
relationships from the reporting countries, the network computed on the basis thereof does not generate local clusters. This results in biased clustering coefficients. Furthermore hub and authority were nearly perfectly correlated with in-degree due to the data structure as well.
9
Images drawn from a lecture on network centrality at Bryn Mawr College,
21 centrality). However Caballero argues that these measures have limited explaining value in directed network. For example, PageRank assigns a probability to each node of being reached after starting in a random node and crossing a certain number links. 10 Thus nodes having a large number of inward flows from nodes that are themselves highly connected, receive a higher PageRank. However, for the network in this study it is found as well that PageRank has limited distinctive value. The covariance with weighted degree measures is over 0.99. Therefore like in Caballero, these kinds of measures were excluded from this research.
By including two relatively simple network statistics (degree and in-degree), two weighted measures that reveal dependence on lending and borrowing (weighted degree and in-degree) and a measure that computes the intermediary role of a node (betweenness), this research aims at capturing network centrality in a broader way. Table 2 lists the network characteristics that have been included in the analysis and briefly summarizes their characteristics. Section 5 gives a more elaborate insight into the structure of the global financial network. Furthermore it provides a table with the rankings and scores that countries hold within the global interbank network on the basis of these network characteristics.
Table 2: Network statistics
10
PageRank was first used to measure the ’systemic’ relevance of webpages, but has well been used in finance (see e.g. Battiston et al., 2012).
Statistic Description
Degree Number of interbank flows into and out of a country
In-degree Number of interbank flows into a country, it measures the
relative dependence of a country on incoming flows
Weighted-degree Number of interbank flows into and out of a country,
corrected for size of flows
Weighted-in-degree Number of interbank flows into country, corrected for size of flows
Betweenness Node betweenness centrality measures how often a node
22 4.1.3 Econometric procedure
As a first statistical illustration a two-sample t-test is performed to see whether there are differences in lending from central and peripheral countries.11 Therefore two groups have been defined on the basis of their degree levels with a cut-off between high and low level of degree at the median.
As mentioned earlier, the main econometric specification (1) is a common empirical specification used in several studies on post-crisis lending (Khwaja and Mian, 2008, De Haas and Van Horen, 2011, and Cetorelli and Goldberg, 2011). The dependent variable in this case is the log-difference in exposures from country i to j between the defined crisis periods. The independent variables are the centrality measures. A fixed effect estimator on the borrower side is included to control for individual-specific, time-invariant characteristics (Hill et al. 2012). The dataset in this study includes multiple lending relationships to one and the same country. Therefore fixed effect estimators in this case allow the researcher to control for host-country heterogeneities that are invariant for different lending relationships (De Haas and Van Horen, 2011). Some countries have been affected by the crisis more heavily than others, and therefore it is to be expected that their borrowing patterns are different as well. By including a fixed effect estimator, it is possible to separate those factors that are truly due to centrality and the variation that is caused by host-country characteristics. This is a common practice in the literature. Studies on post-crisis lending demonstrate that the largest part of variance is actually captured by the inclusion of a fixed-effect estimator in regressions.12
This leads to the following specification:
(1) ∆𝐿𝑖𝑗 = 𝛽1𝐶𝑖 + 𝛽𝑗+ 𝜀𝑖𝑗
Subscript i denotes the lending country whereas subscript j the borrowing country. 𝐶𝑖 are the centrality measures defined above. 𝛽𝑗 is the fixed effect estimator for the borrowing country.
11
Cetorelli and Goldberg (2011) have first performed a parametric test to identify whether a group of countries that had an ex-ante low vulnerability to the dollar market saw greater decline in lending than a group with an ex-ante high vulnerability. In this research instead a two-sample t-test is performed to see if differences are present and statistically significant.
12
23
Furthermore several sensitivity tests are included in the analysis. First, a sensitivity test is performed whereby bank characteristics are used as controls. The following controls were implemented. The ratio of private credit by deposit money banks to GDP (pcrdbgdp) gives an indication of the overall size of the banking section. The ratio of a bank’s overhead costs to total assets (overhead) and the return on assets (roa) illustrate the efficiency of a country’s banking sector. The z-score (z-score) is a measure of risk, and therefore shows the ex-ante stability of a banking sector. The log of total international claims (linclaims) is a measure of size and integration in the global financial network. Cetorelli and Goldberg (2011) use similar controls.13 In the literature on post-crisis lending, several other types of controls have been applied. De Haas and Van Horen (2011) study the impact of distance on lending after a crisis and therefore control for specific aspects that define the relationship between two countries. Van Rijckeghem and Weder (2003) use macroeconomic controls for the borrowing country. Since this thesis has applied fixed effect estimation, it is not possible to put time-invariant variables on the borrowing side of the equation. However since centrality is most likely to be determined by the characteristics of banks in a country, in this thesis controls that characterize the banking systems have been used.
Further sensitivity tests included the alteration of the initial time frame for specification (1). It might be that the results are driven by the specific time frame that was initially chosen. Third a sensitivity test with specification (1) is performed without banks in the US. The US was the origin country of the financial crisis and thereby is likely to have distorted the results. 14
13
Cetorelli and Goldberg (2011) include the ratio of offshore deposits to normal deposits and return to equity as well. Data is missing on offshore for Austria, Belgium and the UK, this variable is excluded from this study since results are likely to be biased. Moreover, linclaims also provides a measurement of overall external exposure of a banking sector.
Furthermore, Cetorelli and Goldberg include return to equity as well. This variable was excluded in this thesis since there is a great deal of collinearity between roe and roa in the models. Inclusion of the variables individually in the model led to expected positive significant results for both variables, while including them together resulted in mixed and insignificant signs. Possibly both indicators capture similar effects. They have been both included in the collinearity table in the appendix and have a covariance of above 0.7.
14
24
4.2 Data
Data for the dependent variable was drawn from the BIS international banking statistics. This database compiles data on international banking flows from a group of reporting countries. In this study the locational statistics on international claims are used. These statistics measure international positions on the basis of location of the office, regardless of the nationality of the bank. The network statistics mentioned in section 4.1 are as well computed from this network. Nodes in the network are thus represented by countries, and the linkages among them are international claims.
Lending patterns were assessed for 27 BIS reporting-countries. It is a an extended sample on other studies that have used aggregated country data to study post-crisis contagion. For example Van Rijckeghem and Weder (2003) use a set of 11 developed countries to study post-crisis lending. Cetorelli and Goldberg (2011) study lending patterns of 17 developed countries.15 In this study the sample includes all developed countries that report to the BIS. Furthermore developing countries that report were added to the sample as well. Developing countries are most often in the periphery of the network and thus have lower centrality scores (see section 5). Thereby the diversity with regard to centrality indicators in the sample is increased. The sample is presented in appendix 1.
Several studies on post-crisis lending research financial flows from a group of developed countries to a set of developing countries (Van Rijckeghem and Weder, 2003, and Cetorelli and Goldberg, 2011). In this study, this division has not been made, since it is likely that lending relationships between the reporting countries do as well include valuable information, therefore lending relationships with 177 counterparty countries are studied. The total number of lending relationships that is studied in this sample is 2258.16
15
Most studies do not include offshore centers in their sample, therefore they have been excluded in this research as well. Offshore centers that were excluded are the Bahamas, Bahrein, Bermuda, Cayman Islands, Isle of Man, Jersey, Macao, and Panama (IMF). Developed countries such as Switzerland and Luxemburg were included in other studies as well. For the variable cap, France and Korea lacked data until 2008, therefore for these observations data for 2008 was used.
16 Fixed effects require that borrowing countries have more than two relationships. Furthermore only those
25
4.3 Descriptive statistics
Data from the BIS International banking statistics on international lending flows is used in this thesis. The dependent variable was calculated by taking the change log between the average exposures in the pre-crisis and post-crisis period. The ex-ante network statistics were taken for 2006q2, the first quarter of the pre-crisis period. Most data on bank characteristics of banks was taken from Financial Development and Structure Dataset (Beck, Demirgüç-Kunt, Levine, 2000). Bank capitalization ratios were taken from the World bank database. Generally controls are used for the year prior to the crisis-window. The crisis period in this thesis is 2007q3-2008q2 (Kali and Reyes, 2010). Therefore controls, like the centrality indicators were taken for 2006. Due to more limited availability of these data the sample size was somewhat smaller for these variables. It excludes Taiwan and Chile.
Table 3: Summary statistics
Variable Obs Mean Std. Dev. Min Max
Lending 2258 .1475 .143 -8.71 5.99 Degree 2258 126.38 34.43 50 184 Indegree 2258 32.63 4.21 20 36 Weighted degree 2258 14.21 18.71 .244 74.37 Weighted indegree 2258 6.91 9.75 .058 39.41 Betweenness 2258 22.40 19.59 .26 70.05 Cap 2100 5.48 1.99 3 11.7 Zscore 2100 15.12 8.42 2.55 35.41 ROE 2100 13.60 7.43 1.93 28.12 ROA 2100 .915 .68 .087 4.57 Overhead 2100 1.56 1.31 .25 9.22 Pcrdbgdp 2100 114.83 39.99 14.67 174.76 Linclaims 2100 12.77 1.43 8.97 15.07 17 4.4 Statistical requirements
The data were checked for multi-collinearity, heteroscedasticity, and normality, in order to be able to make proper inferences. The results are presented in appendix 4. Collinearity between controls is mostly well below 0.8, and should therefore not pose much of a problem. Yet some of the centrality measures and linclaims seem to suffer from collinearity as their covariances are rather high. This would suggest a strong relationship between the centrality measures and
17
26
the total size of assets of a banking sector abroad. Furthermore a VIF test was provided for the controls. VIF scores were below 10, therefore multicollinearity is not present among controls. Second, the data were tested for normality. The log transformation of the dependent has already solved a large part of the skewness in the data. Still some tests have been performed to test for normality. These tests indicate that the dependent variable is still non-normal. Therefore, residual plots are posted in the appendix. These plots indicated that the residuals for the regression are approximately normal with mean zero. Hence, the dependent variable was deemed suitable for further analysis.
Furthermore, the plots do show that there is substantial heteroscedasticity in the data. Therefore heteroscedastic robust standard errors should be in place. The problem with normal robust standard errors in panel data is that they are likely to be biased. Since the data are grouped there might be correlation between different observations for one country. It is more appropriate to use cluster robust standard errors in that case, since normal standard errors are likely to deliver inconsistent estimates. Tests were conducted to assess whether group-wise heteroscedasticity is indeed present among the observations of borrowing countries, these were conclusive.18 All were significant with a p-value of 0.19 Therefore cluster robust standard errors are in place. The use of cluster robust standard errors is quite common in the model setting that was chosen (see De Haas and Van Horen, 2011 and Khwaja and Mian, 2008).
Since this panel data was structured in a way that there are multiple relationships for one country, but not over time, it is not necessary to check for autocorrelation. Autocorrelation requires different observations over time for one country (Hill et al., 2012).
Finally Hausman tests were performed for each separate regression, to assess whether fixed or random models were more appropriate. Results are presented in appendix 5. The results of the random effects models are presented in appendix 5 as well and where needed discussed in the results section.
18
27
5. The interbank network
The network shown below is based on similar ex-ante data (2006q2) on international interbank flows that is used for the regressions in section 6. As is illustrated by the figures and tables presented in this section, the interbank market exhibits similar characteristics that have been observed in other financial networks as well.
Figure 2 presents a sample of the global interbank network. The nodes represent countries and the linkages among them are the aggregated claims that one country holds on another. The presented network exhibits a similar hierarchy that has been discussed in earlier studies. It is divided between a highly connected core in the center of the network and a less densely connected periphery.
Figure 5: The global interbank lending network in 2006q2
20
20
28
The most important financial centers such as Great Britain (GB), Spain (ES), the Netherlands (NL), France (FR), Japan (JP) and Germany (DE) are located in the middle of the network. On the outskirts of the network, we see smaller and mostly less developed countries such as Poland (PL), Hungary (HU), and Czech Republic (CZ).
Figure 3 provides further evidence for the hierarchical nature of the global interbank market. It displays the distribution of all connections in the interbank network. The distribution of degree is clearly skewed to the right. The largest part of connections between countries are to and from only a handful of nodes in the network.
Figure 6: Degree distribution in the interbank network in 2006q2
29 Figure 7: In-degree distribution in the interbank network in 2006q2
Although the effect is less profound, there still is an observed skewness in the frequency of in-degree levels. Overall, it can be concluded that the interbank network as described in this research displays similar characteristics as have been found in empirical studies on financial markets.
30 Table 4: Rankings and scores of countries on centrality indicators in 2006q2
21
There is substantial diversity in the rankings and especially in relative scores of countries for the different indicators. Table 4 shows that with regard to degree levels, differences in connectivity between countries are much less profound than for network statistics that are weighted or incorporate the whole network structure. These latter measures do especially reveal the skewness in connectivity in interbank markets. For example, betweenness computes how often a node is on the shortest path in-between two other nodes. Table 4 shows that the UK is nearly 1.5 times more often on the shortest path between two nodes than Germany. However with regard to degree levels both countries only differ by 11 connections. The difference is even more profound between the UK and the Netherlands. Betweenness is thus much better able to capture the skewness in interbank data. Furthermore the differences in rankings, but mostly in relative scores rationalize a multidimensional approach towards centrality and the inclusion of these various measures in the regressions in section 6.
21
*The Cayman Islands were excluded from this table, since they are not included in the regression analysis. ** Scores for weighted (in-) degree and betweenness are not transformed in this table.
Degree
UK 186 Germany 37 UK 7507229 UK 4012048 UK 704,87
Germany 175 Switzerland 37 US 3644805 US 1947180 France 566,58
France 174 Canada 37 Germany 2694927 Germany 1034715 Germany 471,41
Switzerland 148 UK 36 France 2434654 France 1090969 Belgium 250,94
Belgium 143 France 36 Switzerland 1535839 Switzerland 548499 Canada 237,51
Italy 139 Belgium 36 Netherlands 1203959 Netherlands 636048 Switzerland 236,36
Netherlands 136 Italy 36 Japan 1196682 Japan 474641 Italy 231,55
Japan 131 Netherlands 35 Luxemburg 1000856 Luxemburg 426483 Taiwan 200,61
Austria 127 Japan 35 Italy 834485 Italy 543569 Denmark 149,99
Luxemburg 125 Austria 35 Belgium 796546 Spain 250131 Netherlands 145,55
31
6. Results
First, a two-tailed t-test is presented as an illustration of the impact of centrality on post-crisis lending. Secondly, the formal specification (1) is performed. Then, three sensitivity tests are performed. First, specification (1) is complemented with controls for bank characteristics for the lending country. Second, a specification is portrayed which excludes the US from the sample. The US was the originating country of the subprime mortgage crisis and might therefore have distorted results in specification (1). Third, specification (1) is performed with an alternative time frame.
6.1 A first statistical illustration
The t-test was done to give a first illustration as to whether there are difference between ex-ante central and peripheral countries. The t-test is portrayed in table 5.
Table 5: Two sample t-test for high and low degree group
Group 11 Obs Mean Std. Err. Std. Dev. [95% Conf. Interval]
1 1152 0.188 0.041 1.395 0.107 0.269
2 1106 0.105 0.044 1.471 0.019 0.192
diff 0.083 0.06 -0.036 0.201
t = 1.3672 p = 0.1717 22
As can be observed from table 5, group 2, the core, saw an on average 8% smaller lending growth compared to group 1, the periphery, in the post-crisis period (2008q3 – 2009q2) compared to the pre-crisis period (2006q2 – 2007q2). However, the corresponding p-value is 0.17 and therefore the difference between the two groups is insignificant. The test thus provides only weak evidence that there are differences in lending growth between the ex-ante central and peripheral countries in the global interbank market.
22
32
6.2 Formal econometric specification
For more nuanced and robust findings it is however necessary to perform a regression analysis. Therefore table 6 portrays results of OLS estimates and the fixed effect estimation from specification (1).
Like in Khawja and Mian (2008) and Cetorelli and Goldberg (2011), the OLS results are presented next to the fixed effects estimations to demonstrate the impact of the inclusion of fixed effects. Thus OLS results are displayed for illustrative reasons and best interpreted alongside the fixed effect estimation. OLS provide biased estimates in this case, since the complete effect of the post-crisis lending shock is due to centrality in this case. It does not account for the demand-side shock to lending. Most centrality measures are significant at the 1% cutoff. The disparity between OLS and the fixed effect estimations shows that a large part of the post-crisis lending shock was caused by a demand shock from the borrowing country. Thus the variance in the data is best described by individual country characteristics that are not captured by centrality indicators. This is in line with what for example Cetorelli and Goldberg (2011) find as well. In the fixed estimation, most centrality indicators are negative yet insignificant. Only betweenness attains the 10% level of significance. As an illustration the fixed effect estimation tells that a 10 point higher betweenness leads to an approximately 2% lower growth in post-crisis lending. On average it could be argued that centrality has had a negative effect on ex-post lending, yet the evidence for this effect is not very strong. Hausman tests were performed and were mostly significant and therefore fixed effects were deemed more appropriate. The results are presented in appendix 5. Although the hausman tests are significant, the signs weighted degree and indegree turn significant in the random effect models.
A few further issues stand out from these tables below. First with regard to the results, in-degree is positive, but insignificant. This could point to a positive relationship between having more inward flows and stable post-crisis outward lending, as for example has been argued by Craig (2013). More likely however is that results for in-degree are biased due to limited variance in in-degree levels. For example, weighted in-degree instead has a negative sign.
33
higher R² with significant centrality indicators as well. The difficulty of describing all variance in the data is due to large differences in post-crisis lending growth rates to individual countries. Centrality measures alone are unable to capture all this variance, since these variables simply display an average growth rate, which is approximately in the middle of both the extreme positive and negative values. This results in a low R². The fixed effect estimation is able to deal with these heterogeneities among destination countries by internalizing the country specific effects that explain the variation for each destination country. This is exemplified by the high R² in the fixed effect estimations. Other studies have as well shown that large parts of the variance in international lending data are the result of between country heterogeneity (for example in Cetorelli and Goldberg, 2011; and De Haas and Van Horen, 2011). Many of the explaining variables in these studies are significant but fail to describe large parts of the variance in the data. Apparently, most of the variance in lending growth is due to unobserved characteristics that are captured through fixed effect estimation.
Table 6: OLS and fixed effect estimations for specification (1)23 Model 1 2 3 4 5 6 7 8 9 10 Degree -0.00196** -0.000180 (0.000875) (0.000799) Indegree 0.00115 0.0104 (0.00716) (0.00655) Weighted Degree -0.00482*** -0.00210 (0.00161) (0.00134) Weighted In degree -0.00882*** -0.00403 (0.00309) (0.00252) Betweenness -0.00475*** -0.00214* (0.00154) (0.00123) Constant 0.395*** 0.110 0.216*** 0.208*** 0.254*** (0.115) (0.236) (0.0378) (0.0369) (0.0457)
Borrower fixed effects No No No No No Yes Yes Yes Yes Yes
Observations 2258 2258 2258 2258 2258 2238 2238 2238 2238 2238
R-squared 0.002 - 0.004 0.004 0.004 0.326 0.327 0.327 0.327 0.327
p-value 0.0254 0.873 0.00277 0.00432 0.00201 0.823 0.116 0.120 0.112 0.0843
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
23
6.3 Sensitivity checks
The results above indicate a generally negative, yet mostly insignificant relationship between pre-crisis centrality and post-crisis lending. For proper inference some robustness checks are however required.
6.3.1 Centrality as a proxy for bank characteristics
In a first alternative setting several bank characteristics were complemented to specification (1). It might be that the centrality of a country hinges to a large extend on the characteristics of particular banks in that country. However as the results in table 8 demonstrate, most centrality indicators are once again negative and most of them are now significant due to the inclusion of controls. This indicates that centrality does not function as a proxy for a country’s banking system characteristics. Rather, the controls have complemented the model and enlarged the significance of centrality indicators.
The results for the controls are mixed, only some are significant. ROA yields positive and some significant results. This would suggest that better performing countries in terms of returns on assets, also saw greater outward lending after the crisis. Moreover, overhead has mostly negative and had some significant signs, indicating as well that countries with more efficient banks have seen lower declines in lending. Most interesting and counterintuitive are the significant and positive results for linclaims. While centrality seems to lead to greater declines in outward lending, linclaims yields mostly positive results. This might be caused by the large covariance between linclaims and centrality. Nonetheless, the inclusion of linclaims in the model has not reduced the sign or significance of any of the centrality indicators. The variables for capitalization (cap), risk (z-score) and overall size of the banking sector (pcrdbgdp), did not yield any interesting results.24 In conclusion, it can be argued that the variables that measure centrality do not suffer from the inclusion of bank controls, therefore centrality does not function as a proxy for bank controls. Finally it should be noted that the models below seem to again suffer from the aggregation of data, as control variables yield mixed and mostly insignificant results. It seems very hard to capture the heterogeneous dynamics that took place after the crisis through simultaneous aggregation of both independent and dependent variables. Since hausman tests pointed to the prevalence of the
24
36
random effects model over the fixed effect model, the former were presented in the appendix. Results were however similar to those presented below.
Table 8: Fixed effect estimation for specification (1) with bank controls
Model 1 2 3 4 5 Degree -0.00369* (0.00220) Indegree 0.0212 (0.0225) Weighted degree -0.00533** (0.00261) Weighted indegree -0.00880** (0.00443) Betweenness -0.00395** (0.00195) Linclaims 0.0824* -0.0143 0.0649* 0.0545* 0.0486* (0.0450) (0.0358) (0.0350) (0.0320) (0.0291) Cap -0.0290 0.00890 0.0144 0.0130 -0.0135 (0.0226) (0.0265) (0.0224) (0.0225) (0.0191) Zscore 0.00480 0.00327 0.00255 0.00288 0.00353 (0.00404) (0.00449) (0.00435) (0.00427) (0.00418) Pcrdbgdp -0.000710 -0.000954 -0.000399 -0.000341 -0.000763 (0.000839) (0.000927) (0.000863) (0.000874) (0.000842) ROA 0.186*** 0.197*** 0.153** 0.162*** 0.169*** (0.0576) (0.0577) (0.0614) (0.0599) (0.0590) Overhead -0.0728* -0.0927** -0.0677 -0.0708* -0.0651 (0.0412) (0.0418) (0.0413) (0.0410) (0.0413)
Borrower fixed effects Yes Yes Yes Yes Yes
Observations 2100 2100 2100 2100 2100
R-squared 0.346 0.346 0.346 0.346 0.346
p-value 0.0947 0.348 0.0424 0.0484 0.0438
*** p<0.01, ** p<0.05, * p<0.1 Robust standard errors in parentheses
25
25 Cluster robust standard errors were applied, since tests for clustered heteroscedasticity were significant with
37 6.3.2 Excluding US banks from the analysis
Second, a sensitivity test is performed that excludes US banks from the main regression (1). According to Cetorelli and Goldberg (2011), data on international lending after the financial crisis may be biased if US banks are included. After all, the crisis originated from the US housing market. Therefore the results below exclude the US banking sector. As can be observed results are not much different from the formal regression. Most centrality indicators are negative, yet insignificant again. However, the hausman tests for models with degree and weighted in-degree were insignificant. Therefore the random effects model was more appropriate for those models. As can be observed in appendix 5, both show up to be significantly different from zero in those models.
Table 9: Fixed effect estimation for specification (1) excluding US banks26
Model 1 2 3 4 5 Degree -0.000125 (0.000801) Indegree 0.0102 (0.00663) Weighted degree -0.00214 (0.00134) Weighted indegree -0.00417 (0.00253) Betweenness -0.00207* (0.00125)
Borrower fixed effects Yes Yes Yes Yes Yes
Observations 2162 2162 2162 2162 2162
p-value 0.876 0.116 0.112 0.101 0.0987
R-squared 0.326 0.327 0.327 0.327 0.327
*** p<0.01. ** p<0.05. * p<0.1 Robust standard errors in parentheses
26 Cluster robust standard errors were applied again, since tests for clustered heteroscedasticity were
38 6.3.3 Alternative time frame
A third check is performed with an alternative time frame, using original specification (1). The original time frame was based on studies by Cetorelli and Goldberg (2011) and De Haas and Van Horen (2011). The time frame used in their studies might have fit the specific data used in their studies. While they focus on overall lending to non-financial firms as well, this study’s focus is on the interbank market. Therefore figure 5 portrays a plot of aggregated growth of international claims on the interbank market in percentages from 2006q4 to 2009q4. This figure shows that the global interbank market saw a rapid increase in the run up to the crisis and a sharp decline in overall exposures between 2008q2 and 2009q2. Therefore the next regression applies specification (1) but instead compares the average exposures in the run-up to the crisis from 2007q1 to 2008q1 with the average exposures from 2008q2 to 2009q2. This fixed effect estimation is reported in table 10. The regression yields similar results as the former above, with mostly negative and insignificant results. The Hausman tests indicated that the random effects model was more appropriate in most cases. However the results in appendix 5 are not much different from those in table 10, only weighted degree now attains significance as well.
Figure 5: Aggregated growth of international interbank claims in %
-8 -6 -4 -2 0 2 4 6 8 10 12 14
39 Table 10: Fixed effect estimation for specification (1) with an alternative time frame.
Model 1 2 3 4 5 Degree -0.000102 (0.000619) Indegree 0.00238 (0.00547) Weighted Degree -0.00152 (0.000945) Weighted In degree -0.00315* (0.00179) Betweenness -0.00104 (0.000925)
Borrower fixed effects Yes Yes Yes Yes Yes
Observations 2314 2314 2314 2314 2314
R-squared 0.141 0.141 0.142 0.142 0.142
p-value 0.869 0.664 0.111 0.0796 0.264
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
27
27 Cluster robust errors for the borrowing countries were applied once again. Plots for normality were similar as
40
7. Discussion
This study is first in exploring whether centrality in the global interbank network is related to post-crisis international interbank lending. It was hypothesized that countries that are ex-ante central in the global financial network have seen greater declines in outward international interbank lending after the financial crisis. In the literature section differences between core and peripheral countries were discussed. Core countries often rely on market funding. On the contrary, in peripheral countries, banks are more involved in deposit-based funding. This heterogeneity provides a basis for intermediation in international financial networks by the core. Yet, it is also a source of vulnerability for core countries to shocks that curb financing on global capital markets.
In the empirical section, it was tested whether the centrality of countries in the global interbank network was determinant for post-crisis lending patterns. In general a negative relationship between ex-ante centrality and post-crisis lending was found. The t-test was insignificant but showed a negative relationship between ex-ante centrality and post-crisis international lending. The signs of centrality measures were as well mostly negative and yielded more significant results, even when controls were added for several bank characteristics. These results are an indication that on average, centrality in the global interbank network led to greater declines in outward lending after the financial crisis. This is in line with literature that has shown that countries that are relatively more exposed to a crisis see greater declines in outward lending (Khwaja and Mian, 2008, Cetorelli and Goldberg, 2011). Most interesting were the results with regard to betweenness. Caballero shows that countries with higher betweenness are better able to dissipate shocks. Hale (2014) demonstrates that banks that play an intermediary role see worse performance. This thesis adds that countries that have an intermediary role have generally seen greater declines in outward lending after the financial crisis. Finally, the bank controls showed that countries with better performing banks (in terms of higher ROAs and lower overhead) have seen higher lending as well after the crisis.
41
This leads to the important conclusion that large part of the deviations in interbank lending data is due to unobserved heterogeneities in the borrowing country. This is in line with what other studies have shown as well (Cetorelli and Goldberg, 2011, De Haas and Van Horen, 2011).
This may be due to a few reasons. First, BIS data is aggregated for the lending as well as the borrowing country. Thereby it was not possible to make inferences on the basis of bank-to-bank relationships. Country aggregated data may miss out on the effects that take place at the bank level. Furthermore the BIS data is not an ideal database to study networks as it only includes the lending relationships of a limited amount of reporting countries. Therefore most of the lending flows in the BIS data have only one direction: from reporting to counterparty country. This limits the power of network analysis as the computed network is incomplete. Thirdly, this thesis has relied on the BIS locational database, it could be that the consolidated statistics that aggregate lending on the basis of nationality rather than residence would have yielded different or better results.
Finally, the lack of clear outcomes may be owed to the difficulty of describing events on the interbank market during and right after the financial crisis. While other studies have included bank lending to non-financial corporations as well, this thesis focus was solely on interbank market. However, the initial default of Lehman may have caused a distrust on the interbank market that is not easily described by centrality indicators or bank controls. This was for example discussed in Summer (2013). He argues that network analyses fail to incorporate the dynamics that are not caused by direct financial linkages, for example due to amplification mechanisms which are often at play during crises.
42
9. Bibliography
Acharya, Viral V., Anginer, Deniz, and Warburton, Joseph, 2014. The End of Market Discipline? Investor Expectations of Implicit Government Guarantees. Unpublished Working paper.
Albert, Réka, Hawoong Jeong and Albert-László Barabási, 2000. Error and attack tolerance of complex networks. Nature 406 (6794), 378-382.
Allen, Franklin, and Gale, Douglas, 2000. Financial Contagion. Journal of Political Economy 108(1), 1–33.
Arregui, Nicolas, et al, 2013. Addressing Interconnectedness: Concepts and Prudential Tools. In: IMF working paper WP/13/199. IMF.
Bastian, M., Heymann, S., Jacomy, M. 2013. Gephi: an open source software for exploring and manipulating networks. In: International AAAI Conference on Weblogs and Social Media. Version 0.8.2, https://gephi.org/
Battiston, Stefano, Michelangelo Puliga, Rahul Kaushik, Paolo Tasca, Guido Caldarelli. 2012. DebtRank: Too Central to Fail? Financial Networks, the FED and Systemic Risk. Scientific Reports 2 (541).
Barabási, Albert László and Réka Albert, 1999. Emergence of Scaling in Random Networks. Science, 286(5439), pp. 509-512.
Baum, Christopher F, 2006. Stata tip 38: Testing for groupwise heteroscedasticity. The Stata Journal 6 (4), 590–592.
Beck, Thorsten, Demirgüç-Kunt, Aslı and Levine, Ross, 2000. A New Database on Financial Development and Structure. In: World Bank Economic Review 14, 597-605.
Börner, Katy, Sanyal, Soma and Vespignani, Alessandro, 2007. Network Science. In: Annual Review of Information Science & Technology 41 (1), Chapter 12, 537-607. Boss, Michael, Elsinger, Helmut, Summer, Martin, and Thurner, Stefan, 2004. The Network
Topology of the Interbank Market. Quantitative Finance 4(6), 677-684.
Caballero, Julian, 2015. Banking crises and financial integration: Insights from network science. Journal of International Financial Markets, Institutions & Money 34(2015), 127-146.
