## _wp1174 - References

## Source details

**Canonical URL:** [_wp1174 - References](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2011/_wp1174.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2011/_wp1174.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2011/_wp1174.pdf.json)

---

### Introduction and motivation
- Network analysis used to model interactions between economic agents and assess resilience of financial systems to systemic risk.
- Financial interconnectedness became central to macro-financial surveillance and prudential regulation debates after the 2008–09 global financial crisis.
- The 2008–09 crisis served as an empirical experiment in the evolution of financial networks during stress.

### Scope, data, and approach
- Dataset: BIS locational statistics on exchange-rate adjusted cross-border bank credit, quarterly 1978Q1–2009Q3 for 184 countries.
- Analytical focus:
  - Flows (changes in cross-border exposures) rather than stocks to capture liquidity conditions and variation during stress.
  - Both binary (presence/absence) and weighted (magnitude) network representations.
  - Temporal coverage for main dynamics assessment: 1978–2009.
- Network construction:
  - Nodes: 184 countries; directed links: cross-border lending flows.
  - Sample composition: 15 lenders (BIS reporting countries) and 169 borrowers; core-periphery structure emphasized.
  - Quarterly and annual networks modeled separately; annual data obtained by summing quarters.
- Data preprocessing:
  - Retain positive flows; replace negative flows with zeros.
  - Matrices Mt with entries mtij = flow from lender i to borrower j at time t; binary At with atij = 1 if mtij > 0.
  - Bilateral flows adjusted for inflation using the US CPI.

### Network metrics and definitions
- Degree and strength:
  - Out-degree / In-degree from binary matrix A_t.
  - Out-strength / In-strength from weighted matrix M_t.
  - Relative out-strength and relative in-strength constructed via normalized matrices M̃ and M̂.
  - Maximum in-degree = 15; maximum out-degree = Nt (varies).
- Connectivity and clustering:
  - Connectivity = mt / (15 × Nt).
  - Binary clustering = ratio of complete triangles to possible triangles; ranges 0 to 1.
  - Weighted clustering values complete triangles by the geometric mean of weights; weighted clustering = total value of complete triangles divided by total value of complete and incomplete triangles.
- Average path length: average of the shortest path between all pairs of nodes.

### Main empirical findings
- Volatility and structural breaks:
  - Network metrics for the GBN tend to be volatile with structural breaks identifying waves of capital flows.
  - One-break tests identify breaks in 2001 or 2003; two-break tests find shifts in 1980 or mid-1990s and in 2002–03.
- Connectivity and rankings:
  - Interconnectedness rankings of countries, especially borrowers, are relatively unstable.
  - Connectivity tends to diminish during and after financial crises; 2008–09 crisis is an unusually large perturbation.
- Stylized quantitative magnitudes:
  - Net cross-border bank lending reached $4.3 trillion in 2007 and fell to almost −$1 trillion in 2008.
  - Average cross-border lending per quarter: $293 million overall; $81 million to periphery countries; $212 million to the core.
  - In the full GBN, countries borrow on average from 4.4 lenders up to a maximum of 15.
  - Core countries lend to 48 borrowers on average, of which 8 are in the core and 40 in the periphery.
  - Likelihood that two countries are connected:
    - Full network: ranges between 26 and 44 percent.
    - Within the core: reaches 79 percent before the current crisis (text phrasing).
  - Binary clustering:
    - Full network ≈ 11 percent.
    - Core (bi-directional definition) ≈ 13 percent, reaching 27 percent before the 2008–09 crisis.

### Network topology, dynamics, and regional heterogeneity
- Waves of globalization:
  - Three waves identified: build-up ending early 1980s debt crisis; surge prior to 1997–98 Asian crisis; large build-up from early 2000s up to the current crisis.
  - Average number of outgoing links per lender rose from about 40 to a peak of 60 countries in 2007, then sharply dropped in 2008.
  - Node strength increased steeply during the 2000s and experienced an unprecedented drop in 2008.
- Distributional shifts and persistence:
  - Lender out-degree preserved shape but shifted right between 1980 and 2007.
  - Borrower in-degree shifted right with a bimodal 2007 distribution; median lender share per borrower (relative out-strength) rose from about 5 percent in 1980 to about 10 percent in 2007.
  - Lender centrality distributions more stable than borrower metrics (KS tests reported in source).
- Country ranking dynamics:
  - Stable top lenders by connectivity: Belgium, France, Germany, Switzerland, United Kingdom; Japan, Switzerland, United States prominent by volume.
  - Borrower centrality evolved: LAC prominent pre–early 1980s; East Asia in 1995; by end of sample BRICs and emerging Europe among most interconnected borrowers.
  - Ranking Stability Index (RSI): lenders stabilized around 0.8 after early 1980s debt crisis; borrower in-degree averaged about 0.2 over full period; strength-based rankings stabilized around 0.7 during the 2000s.
- Regional patterns (2000s):
  - ECA most integrated by in-degree, in-strength, and clustering.
  - EAP shows cyclicality with pre-1990s peaks.
  - MENA ranks fourth by total lending but higher by in-degree and binary clustering.
  - SA increased interconnectedness (driven by India) but volumes remain relatively low.
  - LAC never regained pre-1980s connectivity within sample period.

### In-sample dynamics and long-run distributions
- Structural break windows used for transitional analysis: 1987–98 and 2002–08.
- M-statistics (distributional persistence) — empirical values vs random-network 95% CI:
  - First window (1987‒98):
    - Node degree ω=0: 0.2134 (random CI: (0.0872, 0.1185))
    - Node degree ω=1: 0.5335 (random CI: (0.2534, 0.2991))
    - Node strength ω=0: 0.3583 (random CI: (0.0836, 0.1159))
    - Node strength ω=1: 0.7204 (random CI: (0.2561, 0.3065))
    - Relative node strength ω=0: 0.3539 (random CI: (0.0845, 0.1159))
    - Relative node strength ω=1: 0.7460 (random CI: (0.2569, 0.3026))
  - Second window (2002‒08):
    - Node degree ω=0: 0.2536 (random CI: (0.0774, 0.1177))
    - Node degree ω=1: 0.5996 (random CI: (0.2496, 0.3150))
    - Node strength ω=0: 0.4363 (random CI: (0.0780, 0.1200))
    - Node strength ω=1: 0.8270 (random CI: (0.2486, 0.3117))
    - Relative node strength ω=0: 0.4556 (random CI: (0.0780, 0.1223))
    - Relative node strength ω=1: 0.8449 (random CI: (0.2478, 0.3117))
- Interpretation:
  - M-statistics ω=0 range between 0.2 and 0.5, indicating limited persistence but greater stability than a random network.
  - M-statistics higher in 2002‒08, consistent with consolidation of a group of borrowers.
  - GBN substantially less stable than trade network (trade M-statistics reported as 0.8–0.9 in referenced work).
- Ergodic distributions:
  - First-window ergodic distribution for in-degree unimodal.
  - Second-window ergodic distribution bimodal, implying a long-run landscape with a larger group of highly-connected borrowers under second-wave dynamics.

### The global banking network during financial crises
- Visual and aggregate evidence:
  - Markedly lower intra-core and core-periphery connectivity at end-2008 vs 2007; many links disappear as cross-border flows dry up.
  - Network density and clustering fall during episodes of financial stress: examples include 1987 stock market crash; 1991–92 Scandinavian banking crises and 1992–93 ERM crisis; 1998Q3 LTCM near-collapse; 2000Q2 Internet bubble collapse; 2008Q3 Lehman Brothers bankruptcy.
  - Regional clustering falls during systemic banking crises in LAC and EAP (dates from Laeven and Valencia, 2010 in source).
- Country-level dynamics around crises:
  - Borrower connectedness falls during crises; decline generally begins before the event.
  - Borrowers afflicted by sovereign debt crises do not regain pre-crisis connectedness within five years.
  - Including 2007–08, all lender connectedness measures sharply decline one year after the crisis.
- Panel regressions (country fixed effects; robust standard errors):
  - Banking crises associated with reduced borrower access:
    - Cross-border inflows to afflicted country decrease by $24 million at onset of banking crisis (column 2 wording in source).
    - Average $30 million decrease over subsequent five years (text summary).
  - Lender-side regressions: F tests cannot reject nil effect of banking crises on lender connectedness in sample (limited variation noted).
  - Sovereign debt crises show smaller negative effects on borrowers.
  - Significance notation in source table: *** p<0.01; ** p<0.05; * p<0.1.

### Conceptual implications and policy-relevant insights
- Network theories highlight trade-offs:
  - Higher interconnectedness can improve risk sharing but increase contagion and produce robust-yet-fragile outcomes depending on topology.
  - Interlinkages may be desirable for diffusion of risk but can act as conduits for contagion when coordination or information frictions exist.
- Complexity and information frictions:
  - In highly interconnected networks, information gathering about counterparties can become prohibitively costly, generating a “complexity externality” and potential market seizures.
- Relevance for stress testing and macroprudential surveillance:
  - Topological features of real-world networks can inform models used to predict shock propagation and to incorporate first- and second-round effects in stress tests.

### Conclusions and open questions
- Summary conclusions:
  - The GBN is relatively unstable with structural breaks, volatile country rankings (especially borrowers), and cycles of expansion and contraction in density.
  - The 2002–08 wave saw BRICs and emerging Europe consolidate as highly-interconnected borrowers.
  - The 2008–09 global financial crisis produced an unusually large perturbation to the GBN.
- Open research questions:
  - How does network topology relate to emergence of systemic risk?
  - How do different networks (banking, FDI, trade, remittances) compare and interact empirically?
  - What is the empirical relationship between connectedness and amplification or diffusion of shocks?
  - Analysis of net-repayments networks and the leading-indicator capacity of network measures remains for future work.

*Source: _wp1174 - References*

### References .............................................................................................................

### _wp1174 - References

### Introduction and motivation
- Network analysis is increasingly recognized as a powerful methodological tool for modeling interactions between economic agents and assessing the resilience of financial systems to systemic risk.
- Financial interconnectedness has become an important concern in macro-financial surveillance and prudential regulation policy discussions.
- Renewed interest in network analysis for economic interconnectedness was spurred by the 2008–09 global financial crisis, which provides a fertile experiment in the evolution of financial networks during times of stress.

### Scope, data, and approach
- The paper explores properties of the global banking network (GBN) using network analysis techniques and bilateral data on cross-border bank lending for 184 countries.
- The analysis focuses on flows rather than stocks, arguing that cross-border flows reflect liquidity conditions and provide informative variation during times of financial stress.
- Both binary (presence/absence of links) and weighted (intensity of financial flows) network approaches are used to capture heterogeneity in cross-border financial links.
- The temporal coverage for the main dynamics assessment is 1978–2009.

### Main empirical findings
- Network metrics for the GBN tend to be volatile, with a number of structural breaks identifying waves of capital flows.
- Interconnectedness rankings of countries, especially for borrowers, are relatively unstable.
- Connectivity tends to diminish during and after financial crises, with the 2008–09 crisis standing out as an unusually large perturbation to the network.

### Stylized facts and quantitative magnitudes reported
- Net cross-border bank lending (new loans minus repayments) reached a high of $4.3 trillion in 2007 and plummeted to a negative of almost $1 trillion in 2008 as the subprime crisis unfolded.
- Cross-border lending flows among advanced economies are roughly ten times higher than those to emerging and developing economies (advanced economies constitute the “core”; emerging and developing recipients constitute the “periphery”).

### Geographic and topological observations
- The core-periphery and core-core networks changed substantially between 1980 and 2007: the web of cross-border bank relationships expanded markedly as new links were born and cross-border flows increased in magnitude.
- In 2007, more banking activity occurred with the periphery than in 1980; ECA and MENA were the most integrated regions into the global market in 2007 (based on textual description of network graphs).
- Network links expanded outwards since 1980, particularly towards emerging Europe and international banking centers such as Bahrain, Cyprus, and Mauritius.
- Link proliferation also occurred within the core as cross-border lending among advanced economies intensified.

### Conceptual implications
- Network theories can enrich understanding of financial-system functioning by modeling complexity, systemic risk, and factors that cause market seizures.
- Higher interconnectedness may both improve risk sharing and increase contagion risk—financial systems may display robust-yet-fragile tendencies and react differently to shocks depending on network topology.
- Detailed knowledge of topological structures of real-world financial markets is useful for developing models that predict observed patterns and forecast system reactions to shocks.

### Paper structure (high-level)
- Section II: Review of previous studies using network theory to analyze cross-border financial flows.
- Section III: Description of data and network metrics.
- Section IV: Topological properties of the GBN—static and dynamic features, distributional stability of network statistics, country rankings and ranking dynamics, within-sample and out-of-sample dynamics.
- Section V: Behavior of the network around financial crises.
- Conclusions follow.

*Source: _wp1174 - References*

### Section VI.

### Section VI.

### II. RELATED STUDIES
- Context: Rapidly expanding literature applying network analysis tools to analyze financial linkages and implications for systemic risk. 12
- Seminal model:
  - Allen and Gale (2000): stylized four-bank network; contrast between “complete” and “incomplete” networks.
  - Core finding: complete networks are more resilient to shocks due to risk sharing; incomplete networks are more fragile because banks with fewer counterparties have difficulty diffusing shocks.
  - Leitner (2005): interlinkages can be desirable as conduits for contagion if they motivate coordinated bailouts.
- Extensions and more complex network models:
  - Nier et al. (2007): model of interlinked bank balance sheets examining contagion as a function of bank capitalization, size of cross-exposures, and interconnectedness.
    - Finding: bank connectivity has a non-monotonic effect on contagious defaults—at small levels of connectivity, a small increase raises likelihood of contagion; in more interconnected networks, higher connectivity improves shock absorption.
  - Battiston et al. (2010): assess network density (measured as the number of connecting links) in a credit network.
    - Argument: higher connectivity improves risk sharing but also produces trend reinforcement where trading partners react to a negative shock by making conditions harder, creating a mechanism for avalanches of bankruptcies in highly interconnected systems.
- Complexity and information frictions:
  - Caballero and Simsek (2009, 2010): models where banks assess trading-partner health by collecting information; in complex, highly interconnected networks information gathering becomes too costly and may be abandoned, leading to withdrawals from loan commitments and market seizures.
    - Introduces notion of a “complexity externality” where market seizures occur because banks are increasingly reluctant to buy assets in a confusing and uncertain environment.
- Empirical network topology studies:
  - Trade networks: Fagiolo et al. (2009, 2010); Schiavo et al. (2010); Schweitzer et al. (2009).
    - Documented features: core-periphery structure; emerging “rich club” of countries with high trade intensities and high interconnection; apparent stationarity of the trade network.
    - Kali and Reyes (2010): use network measures of connectedness to explain abnormal stock returns during crises; find contagion more likely when epicenter country is better integrated into trade network; more interconnected downstream countries better dissipate impacts.
  - Financial exposure networks:
    - Hattori and Suda (2007): analyze core-periphery network of international bank exposures (stocks of claims) for 215 countries over 1985–2006.
      - Findings: network became more tightly connected and more clustered over time; higher connectivity and shorter average path length13 in the recent period.
      - Posited that systemic risk is building up while risk and capital are being allocated more efficiently. 14
      - Note: network features largely unperturbed during events such as the LTCM near-collapse or the 1997–98 Asian crisis.
    - Contrast claimed by present study: the network of financial flows documented in this paper is more turbulent than that of financial exposures and behaves differently during times of financial stress compared to tranquil periods.
    - Kubelec and Sa (2010) and Sa (2010): dataset of bilateral cross-border exposures by asset class (FDI, portfolio equity, debt, and foreign exchange reserves) for 18 advanced and emerging market economies over 1980–2005.
      - Findings: marked increase in financial interconnectedness over 1980–2005; network became more clustered with lower average path length over time; central hubs are the United States and the United Kingdom.
      - Both trade and financial exposures distributions are long tailed15, hinting at core-periphery structure.
    - Song et al. (2009): using CPIS data for 2001-06, show distribution of global financial exposures well approximated by a Weibull distribution.
- Additional notes:
  - Martinez-Jaramillo et al. (2010): simulate contagion in a banking system and derive measures of system fragility based on topological features of interbank exposures network.
  - Network structures incorporated into stress tests to monitor first- and second-round effects of macroeconomic shocks (e.g., Espinosa-Vega and Sole, 2010).

### III. DATA AND DEFINITIONS
A. BIS locational statistics
- Data source: BIS locational statistics on exchange-rate adjusted cross-border bank credit over 1978–2009 (on a quarterly basis: 1978Q1–2009Q3) for 184 countries.
- Definition and scope:
  - Locational statistics compiled on basis of residence of BIS reporting banks and cover “the cross-border positions of all banks domiciled in the reporting area, including positions vis-à-vis their foreign affiliates” (Wooldridge, 2002, p. 80).
  - Positions include loans, deposits, debt securities, and other assets provided by banks.
  - Cross-border flows estimated as changes in cross-border exposures (stocks); because exposures are reported by currency, changes in stocks are estimated after accounting for exchange rate changes—resulting in exchange-rate adjusted estimates of cross-border flows used in this paper.
  - Exchange-rate adjusted estimates considered a better approximation of true flows than unadjusted changes in stocks (Wooldridge, 2002).
- Suitability and limitations:
  - BIS locational banking statistics capture “net flows of financial capital between any two regions channeled through the banking system” (McGuire and Tarashev, 2006, p. 34), making them well-suited for analyzing geographical interlinkages and temporal patterns in financial linkages across countries.
  - Less suitable for analyzing global balance sheets of banks or funding risk (Fender and McGuire, 2010).
  - Location-based data capture activities of all banks (domestic and foreign) operating in a location; if foreign banks are more active in a location, the data will reflect that activity.
  - For comparison: BIS consolidated statistics are based on the nationality of reporting banks and net out intragroup positions (see Waysand, Ross, and Guzman, 2010).

B. The network
- Network construction:
  - Nodes: 184 countries in the dataset; model the financial flows as a directed network where nodes are linked through cross-border lending and banking flows between countries are the links.
  - Temporal granularity: work with quarterly data and annual data (annual data obtained by adding up flows across quarters); each quarter or year over 1978–2009 modeled as a separate network.
- Core-periphery structure:
  - Sample composition: 15 lenders (BIS reporting countries) and 169 borrowers.
  - Resulting structure: full network has a core-periphery structure with the “core” comprising economies that act as lenders (sample of BIS reporting countries) and the “periphery” comprising countries that act as borrowers (non-BIS reporting countries).
  - Empirical focus: because intra-core lending activity dwarfs in magnitude exchanges between the core and the periphery, remainder of the paper focuses on the core-periphery network (core countries act as lenders only) rather than the full network (where core countries are borrowers as well).
  - Also analyze the core-core network composed solely of “core” countries that are both lenders and borrowers.
- Definitions:
  - Average path length: defined as average of the shortest path between all pairs of nodes in the network; it measures the “degrees of separation” among nodes.13

*Source: _wp1174 - Section VI.*

### appendix for the list of countries).

### _wp1174 - appendix for the list of countries)

### Data construction and preprocessing
- Dataset: bilateral banking flows constructed from BIS locational banking statistics (sample period 1978‒2009 as used throughout the document).
- Treatment of flows:
  - Retain all positive flows (net increases in cross-border exposure or “net investments”).
  - Replace negative flows with zeros (net repayments or no flows).  
  - Resulting matrix is one of net financial investments comprising positive values and zeros.
- Matrix construction:
  - Matrices Mt where rows represent lenders and columns represent borrowers; each entry mtij is the flow from lender i to borrower j at time t.
  - Binary counterpart At: atij = 1 if mtij > 0, otherwise 0.
  - Dimensions: 15tN for 15 lenders and Nt borrowers; Nt varies through time as new countries enter the network.
- Inflation adjustment: bilateral (net) bank lending flows adjusted for inflation using the US CPI.

### Network indicators — definitions and computation
- Binary and weighted approaches used to capture both existence and magnitude of links (see Barrat et al. (2004) citation in source).
- Country centrality
  - Node degree (binary representation A t):
    - Out-degree: number of outgoing links (countries to which a country lends).
    - In-degree: number of incoming links (countries from which a country borrows).
    - Maximum in-degree = 15 (sample of lenders fixed); maximum out-degree = Nt (varies by year).
    - Formal expressions (as in source):
      - out_t_i = A_t(i·) · 1
      - in_t_i = (A_t(i·)') · 1
  - Node strength (weighted representation M t):
    - Out-strength = total cross-border credit a country lends.
    - In-strength = total cross-border credit a country receives.
    - Computed by substituting M t for A t in the degree formulas.
  - Relative node strength:
    - Relative out-strength (lender importance to borrower j): construct matrix M̃ where m̃tij = inflows to borrower j from lender i divided by j’s total inflows; relative out-strength = M̃_t · 1.
    - Relative in-strength (lender i’s dependence on borrower j): construct matrix M̂ where m̂tij = inflow to j from i divided by i’s total lending; relative in-strength = (M̂_t' ) · 1.
    - Notes on maxima: maximum relative out-strength equals total number of borrowers; maximum relative in-strength equals total number of lenders.
- Network density
  - Connectivity: observed number of links mt expressed as share of total possible links = mt / (15 × Nt).
  - Binary (unweighted) clustering:
    - Ratio of total number of complete triangles to total possible number of such triangles.
    - Core-core complete triangle: bi-directional lending among three core countries i, j, k.
    - Core-periphery complete triangle: core countries i and j lend to one another and both lend to periphery country k.
    - Binary clustering ranges between 0 and 1.
  - Weighted clustering:
    - Inspired by Opsahl and Panzarasa (2009): counts complete triangles and values them by the geometric mean of their weights (cross-country flows).
    - Weighted clustering = total value of complete triangles divided by total value of complete and incomplete triangles.

### Results — network topology, dynamics, and regional heterogeneity
- Overview and key summary statistics (from Table 1 and textual summary):
  - In the full GBN, countries borrow on average from 4.4 lenders up to a maximum of 15.
  - Core countries lend to 48 borrowers on average, of which 8 are in the core and 40 in the periphery.
  - Average cross-border lending per quarter: $293 million overall; $81 million to periphery countries; $212 million to the core.
  - Likelihood that two countries are connected:
    - Full network: ranges between 26 and 44 percent.
    - Within the core: reaches 79 percent before the current crisis (text phrasing).
  - Binary clustering:
    - Full network ≈ 11 percent.
    - Core (bi-directional definition) ≈ 13 percent, reaching 27 percent before the 2008–09 crisis.
- Time-evolution and waves of globalization:
  - Three global waves of cross-border bank lending identified:
    - Build-up from early 2000s up to the current crisis (largest wave).
    - Smaller surge prior to the 1997–98 Asian crisis.
    - Build-up ending with the debt crisis of the early 1980s.
  - Trends (Figure 4, textual description):
    - Average number of outgoing links per lender rose from about 40 to a peak of 60 countries in 2007, then sharply dropped in 2008.
    - Node strength increased steeply during the 2000s and experienced an unprecedented drop in 2008.
    - Connectivity and clustering indicators were volatile overall; in contrast, ITN metrics remained remarkably stable (Fagiolo et al., 2010 reference in source).
- Distributional shifts (Figure 5 and KS tests in Table 2):
  - Lender-side out-degree preserved shape but shifted right (lenders lending to more countries in 2007 vs 1980).
  - Borrower in-degree distribution shifted right with a bimodal 2007 distribution indicating a large group borrowing from about 10 lenders.
  - Relative out-strength: median lender share per borrower rose from about 5 percent in 1980 to about 10 percent in 2007.
  - Relative in-strength: greater heterogeneity among borrowers in 2007 vs 1980 — increased inequality and a more polarized periphery.
  - Kolmogorov-Smirnov (two-sample) tests:
    - Empirical distributions of lender centrality metrics more stable than borrower metrics.
    - Borrower connectivity, strength, and relative strength show much turbulence; distributions in 1980 were poor predictors of future decades (details in Table 2).
- Country rankings and stability (Table 3, Figures 6–7):
  - Top globally connected lenders: Belgium, France, Germany, Switzerland, United Kingdom (stable lenders); Japan, Switzerland, United States join top ranks by volume.
  - Borrower centrality shifted over time:
    - Pre–early 1980s debt crisis: LAC countries (Argentina, Brazil, Chile, Mexico, Venezuela) were central borrowers.
    - 1995: East Asian countries prominent.
    - By end of sample: BRICs and emerging Europe (Baltics, Poland, Romania, Ukraine) among most interconnected borrowers.
  - Ranking Stability Index (RSI) based on average Spearman correlations:
    - Lender rankings stabilized around 0.8 after early 1980s debt crisis.
    - Borrower rankings based on in-degree averaged about 0.2 over full period, sometimes negative in late 1970s — indicating substantial year-to-year ranking turbulence.
    - Rankings based on strength stabilized around 0.7 during the 2000s.
- Regional heterogeneity (Figure 8, textual description):
  - During the 2000s, ECA (emerging Europe) most integrated region by in-degree, in-strength, and clustering.
  - EAP displays cyclicality with peaks prior to early 1990s Japanese crisis.
  - MENA ranks fourth by total lending but higher by in-degree and binary clustering.
  - SA (driven by India) increased interconnectedness in the 2000s but volumes remain relatively low.
  - LAC: total lending falls after early-1980s debt crisis, surges prior to 2001 Argentine crisis and recent downturn, but never reaches pre-1980s connectivity levels within sample period.
  - Downturn in connectivity and clustering during the 2008–09 crisis is an unusually large perturbation.

### In-sample dynamics and out-of-sample evolution
- Structural break tests (Clemente, Montanes, and Reyes, 1998) for four network indicators (average degree, average strength, connectivity, clustering):
  - One-break tests identify a break in 2001 or 2003.
  - Two-break tests find first mean shift in 1980 or mid-1990s and second in 2002–03.
  - Analysis focuses on windows 1987–98 and 2002–08 (dated as in IMF, 2007).
- Transitional dynamics:
  - Stochastic kernel (1-year) estimates for in-degree and relative in-strength:
    - In-degree: most probability mass around 45-degree line in middle range — middle-quantile countries tend to remain; upper-quantile countries tend to ascend further (stronger in second wave).
    - Relative in-strength: emergence of a “middle-class” of globally-interconnected countries in the second wave.
- M-statistic (distributional persistence measure using decile transition matrices):
  - Construction: 1-step transition probabilities p̂1,t ij = n1,t ij / n_t i, aggregated into M statistic for ω = 0 (stay same decile) and ω = 1 (stay same or move one decile).
  - Empirical M-statistics (Table 5):
    - First window (1987‒98):
      - Node degree ω=0: 0.2134 (random-network 95% CI: (0.0872, 0.1185))
      - Node degree ω=1: 0.5335 (random CI: (0.2534, 0.2991))
      - Node strength ω=0: 0.3583 (random CI: (0.0836, 0.1159))
      - Node strength ω=1: 0.7204 (random CI: (0.2561, 0.3065))
      - Relative node strength ω=0: 0.3539 (random CI: (0.0845, 0.1159))
      - Relative node strength ω=1: 0.7460 (random CI: (0.2569, 0.3026))
    - Second window (2002‒08):
      - Node degree ω=0: 0.2536 (random CI: (0.0774, 0.1177))
      - Node degree ω=1: 0.5996 (random CI: (0.2496, 0.3150))
      - Node strength ω=0: 0.4363 (random CI: (0.0780, 0.1200))
      - Node strength ω=1: 0.8270 (random CI: (0.2486, 0.3117))
      - Relative node strength ω=0: 0.4556 (random CI: (0.0780, 0.1223))
      - Relative node strength ω=1: 0.8449 (random CI: (0.2478, 0.3117))
  - Interpretation:
    - M-statistics range between 0.2 and 0.5 for ω=0, indicating limited persistence.
    - M-statistics higher in 2002‒08 consistent with consolidation of a group of borrowers.
    - GBN more stable than a random network but substantially less stable than the trade network (trade M-statistics reported as 0.8–0.9 in Fagiolo et al., 2010).
- Ergodic (long-run) distributions for in-degree (Figure 10):
  - First window ergodic distribution is unimodal.
  - Second window ergodic distribution is bimodal — implying a long-run financial landscape with more highly-connected borrowers under second-wave dynamics.
  - Consistent interpretation: 1990s dominated by East Asia; 2000s saw emergence of BRICs and emerging Europe as key borrowers.

### The global banking network during financial crises
- Visual evidence (network plots for 2007Q4 vs 2008Q4, Figure 11):
  - Markedly lower intra-core and core-periphery connectivity at end-2008; links disappear as cross-border banking flows dry up.
- Network density and crises (Figure 12 and regional clustering Figure 13):
  - Many links die and clustering diminishes during episodes of financial stress (examples: 1987 stock market crash; 1991–92 Scandinavian banking crises and 1992–93 ERM crisis; 1998Q3 LTCM near-collapse; 2000Q2 Internet bubble collapse; 2008Q3 Lehman Brothers bankruptcy).
  - Regional clustering falls during systemic banking crises in LAC and EAP (crisis dates taken from Laeven and Valencia, 2010).
- Country-level connectedness around crises (Figures 14–15):
  - Average degree and strength (5 years before and after crisis onset) show:
    - Borrower connectedness falls during crises; decline generally begins before the event.
    - Lender connectedness also falls for crises that include 2007–08; paucity of systemic banking crises in the core prior to 2007 limits detection earlier.
    - Borrowers afflicted by sovereign debt crises do not regain pre-crisis connectedness within five years.
    - Including 2007–08, all lender connectedness measures sharply decline one year after the crisis (unusually large impact).
- Panel regressions (Table 6) — country-level indicators regressed on crisis dummies with country fixed effects:
  - Banking and debt crises associated with reduced borrower access to capital markets:
    - Cross-border inflows to afflicted country decrease by $24 million at onset of banking crisis (column 2 wording in text).
    - On average $30 million decrease over subsequent five years (textual summary).
  - Lender side: F tests cannot reject nil effect of banking crises on lender connectedness in sample (columns 4–6); limited variation may explain lack of significance.
  - Sovereign debt crises show smaller negative effects on borrowers (columns 7–9).
  - Regression notes: Robust standard errors; significance denoted in source table (*** p<0.01; ** p<0.05; * p<0.1).

### Conclusions and open questions
- Main findings:
  - The global banking network (GBN) is relatively unstable: structural breaks in connectivity and centrality, volatile country rankings (especially borrowers).
  - In the 2002–08 wave, BRIC countries and high-growth Europe emerged and consolidated as highly-interconnected borrowers.
  - Network density expands and contracts with cycles of capital flows; country centrality falls at and after banking and sovereign debt crises.
  - The 2008–09 global financial crisis is an unusually large perturbation to the GBN.
- Open questions for future research (as stated in source):
  - How does the topology of financial networks relate to emergence of systemic risk?
  - How do properties of different networks (banking, FDI, trade, remittances) compare and interact?
  - What is the empirical relationship between connectedness and amplification or diffusion of shocks?
  - Analysis of net-repayments network and leading-indicator capacity of network measures left for future work.

*Source: Appendix for the list of countries of IMF Working Paper _wp1174 (dataset and analysis content as provided in the supplied PDF unit).*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2011/_wp1174.pdf_