See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/311322823

The Impact of Network Topology on Banking System Dynamics: Experimental

Study

Experiment Findings · March 2016

CITATIONS READS

0 882

3 authors, including:

Valentina Y. Guleva Klavdiya Bochenina

ITMO University ITMO University
22 PUBLICATIONS   23 CITATIONS    71 PUBLICATIONS   162 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Supercomputer simulation of critical phenomena in complex social systems View project

Community detection in attributed social networks View project

All content following this page was uploaded by Valentina Y. Guleva on 02 December 2016.

The user has requested enhancement of the downloaded file.

 The Impact of Network Topology
on Banking System Dynamics

Valentina Y. Guleva Abdulmalik Amuda Klavdiya Bochenina

ITMO University ITMO University ITMO University
Kronverkskii, 49 Kronverkskii, 49 Kronverkskii, 49
Saint-Petersburg, Russia, 197101 Saint-Petersburg, Russia, 197101 Saint-Petersburg, Russia, 197101
+7 812 2329704 +7 812 2329704 +7 812 2329704
[email protected] [email protected] [email protected]

ABSTRACT estimation, cascade sizes analysis, measuring the contagion effect

A topology of a banking network influences systemic stability etc. That is they estimate expected losses for arbitrary shock
under fixed banks’ and customers’ policy. Dynamical processes in propagation on a fixed topology. Thus, only a static version of
a network affect topological changes at each iteration (according problem statement is under consideration.
to a given law), so differences in initial topologies may result in We model a banking system as a dynamic network, where nodes
different stability dynamics. Taking into account the influence of represent banks and edges represent interbank lending. Each edge
both nodes’ states dynamics and topology dynamics on the has a weight and a length attributes where a weight is a size of
network fragility, the following ways of initial topology impact interbank lending, and a length is a number of iterations before
can be distinguished: (i) states of nodes after initialization (as a edge deletion. There are two algorithms specified for a bank:
state of a node is influenced by its degree); (ii) nodes’ states (i) an algorithm of interbank links’ addition, (ii) an algorithm of
dynamics; (iii) systemic risk at the fixed iteration. It seems that improvement of a bank’s state if it is close to insolvency or if it
coevolution of network topology and states of nodes leads to the can not satisfy customers’ requests. We use the same algorithms
significant and unpredictable changes of initial conditions. We for all the banks in a network. The flow of customers’ requests is
study the way in which this occurs, and the interrelations between modeled using the additional algorithm (different customers also
initial and resulting states for different types of initial topology. have the same behavior). Despite the fact that we use similar rules
Our results confirm that the dynamic of a borrowing process is for network entities, a system as a whole demonstrates emergent
significantly influenced by topological characteristics (e.g. density behavior which can be quite different for various types of
and heterogeneity of degrees) of underlying interbank network. interbank network topologies. To study how the system evolves
over time, we fix banks’ and customers’ policy and let them
Categories and Subject Descriptors modify a network topology step by step, providing different
Computing methodologies, Network science, Computing simulation results.
methodologies, Agent/discrete models, Applied computing,
Economics. The main part of the study contains formalization of the network
evolution comprising the following: (i) an algorithm for addition
General Terms and deletion of links, (ii) an adaptation algorithm for response to
Algorithms, Economics. changes of customers’ or other banks’ behavior, (iii) the rules of
node’s state modification after topological changes.
Keywords
Banking system, adaptive network, network topology, complex In the experimental part of the study we fix banks’ and customers’
systems. behavioral policy, and vary parameters of network generative
models to observe how the network dynamics and systemic
1. INTRODUCTION stability are affected by those. Thus we get different resulting
Banking systems often demonstrate non-linear behavior [22] so it topologies, and different dynamics after first bankruptcies for a
can be difficult to find a cause of a financial crisis or to determine given number of iterations. Obtained results confirmed the impact
critical values of system’s parameters on the edge between its of a network topology on the patterns of a banking system
stable and unstable states. Network topology invariants (the behavior.
metrics characterizing a structure of interbank interactions) are
one way to indicate the closeness to a crisis. The examples of such 2. LITERATURE REVIEW
metrics are average degree, global clustering coefficient, and There has been a rapid growth of interest in the study of
entropy of Laplacian spectrum. relationship between financial network topology and the systemic
stability which resulted in a lot of publications on national
Influence of network topology on systemic risk in banking payment systems analysis [6, 8, 10, 12, 13]. The major results of
ecosystems was widely studied during previous decade. these studies are addressed in greater detail in review [5]. They
Nevertheless, these studies were mostly focused on loss also provide information about typical values of characteristics of
interbank market topology (e.g. in [2] authors mention that fragility of a banking system. In turn, topological changes affect
interbank networks often demonstrate scale-free topologies). balance sheets’ structure and initiate nodes’ adaptation (i.e.
Analysis with division of network topologies on pre-crisis and modification of their policies). Varying the topological parameters
post-crisis groups is performed in [20]. during network generation, we develop the capability to observe
the changes in system’s trajectory caused by initial properties of
Banking system stability is rather thought to depend on a network an interbank network, and to make conclusions about an influence
structure. It must be pointed our that interbank links have both of initial topology on a stability of a system.
positive and negative impact, because in the crisis times they can
propagate shock, on the one hand, and absorb it, on the other one 3. BANKING SYSTEM MODEL
[16]. Thus, it is interesting to know if it is possible to optimize a
network structure for system stabilization, and how is it possible 3.1 Network Model
to do. We will consider interbank network as dynamic oriented multi-
graph , where nodes ( ) represent banks (agents)
The research on identifying influential nodes during spreading of and edges ( ) represent interbank lending at the moment t. Each
a dynamical process in a given network is the part of modern edge contains information about source and target vertices, size of
complex network theory. Although it is clear that high-degree debt (edge weight) and deposit’s maturity. Thus,
nodes (hubs) are important as they can spread the process to a
larger number of nodes, recent studies show that the role of hubs
is usually overrated. For example, authors of [15] show that Note that maturity corresponds to edge’s life time, that is why it is
instantaneous states of hubs have a short-lasting effect on the state decremented at each iteration of an evolution of a dynamic
trajectory of the entire system. network:
As it is shown in [6], financial systems inherit typical properties
of complex networks. As a result, there is a significant amount of
works dedicated to applications of general theory of complex Weighted sums of interbank edges form interbank assets and
networks to investigation of banking systems. Cajueiro and Tabak interbank liabilities of banks balance sheets.
in [7] study the topological significance of network nodes in
To represent external assets and liabilities we introduce bank-to-
relation to the Brazilian banking system. In addition to the
traditional metrics (such as degrees and centralities of nodes), they customer network where:
introduce domain-specific measures of node’s importance (e.g.
weighted efficiencies which take into account the sizes of ,
liabilities and exposures). The examples of studying the influence
of network completeness, clustering, average shortest path and is a set of nodes-banks, and is a set of nodes-householders.
other topological features on the banking network stability can be Thus for each bank weighted sums of bank-to-customer edges
found in [17, 18, 19]. Scenarios with some artificial crisis serve correspond to external assets and liabilities.
this purpose [11].
3.2 Bank Model
Exploration of systemic risk influenced by network topologies is A bank state is described by its simplified balance sheet formed
mostly focused on measuring effect of a common shock or of by interbank assets (IA) and liabilities (IL), external assets (EA)
interbank contagion induced by some shock [1, 2, 4, 9]. An and liabilities (EL), available funds (AF), and net worth (NW),
example of shock propagation model is described in [14]. which is the difference between assets ( ) and
Georg [2] compares threat to stability for contagion and common liabilities ( ).
shock, and state that both types of systemic risk require different
system responses. Montagna and Kok [1] show the influence of ,
balance sheet and multilayered network structure on the financial
stability of the whole system. Dynamics in such networks is more
likely to be represented by a contagion and knock-on-effects [14],
than by a network evolution.
The emergence of a complex behavior as a result of interbank
interactions can be considered in the context of multi-agent
modeling. In [1], authors present a dynamic model of a banking
system when each bank is provided with an algorithm for
updating its state according to changing conditions. The impact of
dynamic multilayered network topology on a stability of a system
is explored in [21]. However, there is a lack of studies clarifying
the interplay between the changes of network structure and
iterative adaptation of banks’ and householders’ activity. where C is some initial fund.
In this paper we study the influence of an initial structure of a Available funds get bigger as liabilities get bigger, and become
network which evolves over time on a resulting stability of a smaller with growing assets. In our simplified model a bank is
banking system. This stability is determined both by the initial insolvent if its net worth or available funds are not positive. That
topology of a network and the results of actions of two types of is why each node state is described by these two variables.
agents (banks and customers) with predefined rules of behavior. Operating cost (OC) is some value which is subtracted from
During the process of simulation, an initial topology undergoes available funds at each iteration as an analogue of operational or
significant changes leading to an increase or a decrease of the rental fee.
As assets and liabilities are formed by weighted sums, any then customer loan funds of definite size, otherwise it deposit
topological changes affect changes of nodes’ states. Non- fund of another size. Thus,
satisfactory nodes’ states and elimination of nodes initiate
additional dynamics (as they start a chain of borrows.
Consequently, combination of bank-to-customer network
dynamics and intention of banks to have positive both net worth
and available funds triggers interbank network evolution.

3.3 Initialization Maturity is randomly generated by two steps. Firstly, kind of

We generate initial network with the use of NetworkX package 1.
debt is randomly chosen, and secondly, maturity is chosen from
To transform undirected graph (provided by NetworkX) to a
the corresponding range. When parameters for a new edge are
directed one, we replace each undirected edge with two edges
obtained it can be added to the network. External edges are not
having opposite directions. Then we mark each edge with weight
considered in an interbank network, but influence its dynamics
of , and assign interbank assets and liabilities.
with connection over balance sheets.
Edge maturity is generated in two steps. The first one is choosing
the debt type (overnight, short-term, long-term), and the second 3.4.2 Interbank Network Dynamics
step is to choose randomly value for resulting maturity from the First of all, edges with maturity of zero are required to be deleted.
range corresponding to chosen debt type. Then OperationalCosts are subtracted from AvailableFunds, and
the new value is set to it in accordance with net worth value.
Bank balance sheet structure elements are supposed to be related Therefore we have got updated nodes’ states which result in new
with some beneficial proportions. We initialize , interbank edges to be created on a current iteration.
. Net worth is initialized by the definition
New edges’ creation can be triggered by necessity of a bank to
(Section 3.1), , . These coefficients
improve its state, and by bank’s eliminating in the cases of its
were obtained by analysis of annual reports of several national
insolvency. Also we imply that banks have more than one external
banks (see the details in [21]).
and interbank edge.
Customers are assigned to bank initially in amount proportional to
3.4.2.1 Edge Creation Triggered by States
bank’s total assets size. The number of customers for the first
iteration is calculated as: Improvement
This case corresponds to or , where is
a current bank, which is an originator of new edges’ creation. A
bank for interaction is chosen in proportion to its assets’ size.
where is an average number of customers per bank, are the Thereby, the probability of choosing a bank is
total assets of bank , and are the total assets of a network.
Then external assets and liabilities are distributed between
customers to maintain the equal weights of edges.
where is a total amount of assets over all network, and are
3.4 Dynamics the total assets of a bank .
After initialization an algorithm for each iteration can be Edge’s direction depends on the parameter to be improved. If
distinguished into two phases: (i) simulation of customers’ , there is shortcoming of assets, consequently, the
activity, and (ii) simulation of banks’ response. bank is taken as source bank, and bank is target. Otherwise,
is going to be a source bank.
3.4.1 Bank-to-Customer Network — External Edges
Dynamics Maturity is chosen as well as in the Section 3.4.1.
In this phase we iterate over the customers to be active on current If and , then weight of an edge to chosen
iteration, and they do the actions which we describe below. To bank is equals to:
create each edge, we need to determine source and target vertices,
the weight of an edge and its maturity.
A target bank is chosen with probability proportional to its total
asset size like in a mechanism of preferential attachment. for it is going to equal:
Additionally we restrict a number of possible customers’
neighbors with . Therefore, the probability of choosing the bank
is:
In this way banks are enumerating until bank needs to improve
its state. If it still needs this, when all banks have already been
where is the number of banks adjacent to current customer, inquired, it is eliminated from the network.
is probability of choosing a bank among all banks of the
network, is probability of choosing a bank among three 3.4.2.2 Edge Creation Triggered by Node
neighbors of a current customer. Elimination
Edge’s weight corresponds to size of lending and depends on Bank elimination is accompanied with its input and output edges
customer’s account state. If its value is less than some constant, rewriting between its former neighbors. Edges are sorted
according to maturity, and then they are rewritten.
1
https://networkx.github.io/
4. EXPERIMENTS
4.1 Input parameters
Initial network structure was generated by NetworkX package,
written with Python programming language. We have used Erdos-
Renyi (ER) network generative model, Barabasi-Albert (BA)
scale-free model, and Watts-Strogats (WS) generative models
(see, e.g. [23]). ER was run with probability of edge creation in
BA was run with percentage
of attached edges in WS had
percentage of attached neighbors in , and
probability of edge rewriting in .
a)
Bank balance sheet initialization parameters, mentioned in
Section 3.3, are follows:
Parameters for customers
behavior are
Illustrations presented is this section were evaluated for a network
with N = 100 banks and several thousands clients.

4.2 Metrics
To indicate changes in network topology we estimate average
degree, average clustering coefficient, average shortest path and
number of edges. Several topological characteristics (like number
of edges and average degree) were calculated with account of the
fact that our model uses multi-digraph to represent the system.
Algebraic graph theory proposes graph Laplacian spectrum as
feature for network structure characterization [25], and argue that
isospectral graphs have similar topology. In this work we also use
invariant described for Laplacian spectrum in [26].
To distinguish reasons of bankruptcies we used state-of-nodes- b)
oriented features [24]. Those consider proposal velocity of
approaching bankruptcy, using information about node’s states
dynamics on previous iterations. Finally, we used a number of
bankruptcies as the main systemic robustness sign.

4.3 Watts-Strogatz Generative Model

A generative process in Watts-Strogatz model in divided on two
phases: (i) ordered — initially all vertices are placed in ring, and a
predefined number of links is added from each vertex to its
nearest neighbors; (ii) rewiring — each of the edges is rewired
with a given probability. A combination of these two phases
allows regulating the values of average clustering coefficient and
average shortest path. When the probability of rewiring is small,
the neighbors of each vertex tend to be connected more densely
with each other resulting in high clustering coefficient, but
absence of remote links leads to high average distance. The
increase of rewiring probability leads to decrease of both
clustering and average distance between the vertices.
We denote as K a number of edges for each vertex in ordered c)
phase of network’s generation, and as P a percentage of rewired
edges in rewiring phase. In our experiments we study the
dynamics of banking system stability for different combinations
of K and P. The selected results are shown in Figures 1–5.
First of all one can clearly see (Figure 1a-d) that a system is
destroyed more quickly for sparse networks ( and ).
The probability of edge rewiring influence the results only for
. For this value of K, the first bankruptcy occurs about 150
iteration for and , and about 100 for and
(i.e. when a sparse network becomes closer to random
network, it achieves critical state earlier).
 transformation of adjacency matrix (and spectrum) of a graph. For
the configurations with high initial clustering coefficient, an
evolution of a network leads to its decrease; to the contrary, an
initially low clustering coefficient grows with the evolution of a
network. Figure 3 shows a gradual increase in average shortest
paths.

d)

Figure 1. Dynamics of bankruptcies and entropy

for Watts-Strogatz initial network.
Entropy of Laplacian spectrum reflects changes in network
structure, which are the most significant for for
and the least significant for . Figure 1c shows that the
value which was reached with violet graph can not be reached by Figure 3. Dynamics of average shortest path
other networks configurations. Networks generated with other for Watts-Strogatz initial network (P = 60)
parameters are not modified so seriously, despite the parameters
The dynamics of topological characteristics is more likely to be
of bank policy were fixed. Comparison of amplitudes of entropy
attributable to the processes of addition and deletion of edges
values for different rewiring probabilities shows that range of
(which change a number of links) rather than simple edge
their variability tend to increase with a growth of P for both sparse
rewiring. It is illustrated in Figure 4 which show a number of
and dense networks. Intuitively, it could be explained by the fact
edges for different values of K and P = 60.
that the disturbances in more ordered networks usually fade more
quickly that in chaotic ones.
The plots of topological characteristics (Figires 2, 3) show sharp
decrease until about the 30th iteration because we placed a
restriction in the model that customers cannot have more than
three banks to interact with. In the beginning they tend to create
new links with new banks, but approximately to the 30th iteration
all customers already have three neighboring banks.

Figure 4. Dynamics of edges’ count

for Watts-Strogatz initial network (P = 60)
Figures 2–4 also demonstrate that the values of parameters for a
networks with seem to converge in
some point near the 300th iteration, and further show similar
dynamics, while graphs with and show
fluctuations.

Figure 2. Dynamics of average clustering coefficient

for Watts-Strogatz initial network (P = 60)
Average clustering and average shortest path explain entropy
dynamics. The changes in topological characteristics show the
dynamics of addition and deletion of edges resulting in
 Figure 5. Average rate of change of net worth Figure 6. Dynamics of bankruptcies
for Watts-Strogatz initial network (P = 40) for Barabasi-Albert initial network
DynamicsOfNodeStates feature reflects the average speed with
which a bank approaches unsatisfactory state in terms of net
worth, or remotes from it. Negative velocity indicates
approaching. In Figure 5 one can see that in this series of
experiments banks have positive dynamics for their net worth.
This usually corresponds to negative available funds’ dynamics.
Figure 5 shows that peaks have varying amplitude for different
densities of networks, and time of appearance grows with the
increase of the value of
The results for other values of P (which are not illustrated in
Figures 2–5) demonstrate dynamics similar to the presented here.

4.4 Barabasi-Albert Generative Model

Barabasi-Albert model is a most popular mechanism to generate
networks with scale-free degree distribution via the preferential
attachment. The process of network’s creation is iterative, and
each vertex is linked to already existing vertices with a probability
proportional to their degree.
Figure 7. Potential of interaction for
In a NetworkX software package a BA model has a single Barabasi-Albert initial network.
parameter – a number of links which are added for each new
vertex. In our experiments, we calculate this value as a fixed 4.5 Erdős–Rényi Generative Model
percentage A of a number of vertices (e.g. if A = 30, for N = 100 a Erdős–Rényi model is a simplest model of random graph. It
number of links will be equal to 30, and for N = 200 it will be assumes that there is a given probability of an edge between any
equal to 60). The networks with smaller values of A are more pair of vertices. Degree distribution for this types of networks
sparse that the networks with larger values of this parameter. follows a binomial law. We denote as P a probability of edge
which is a parameter of this model.
For BA initial configuration we oppositely got more stable
systems for sparse networks, and more fragile systems for dense All graphs in the Figure 7 show bankruptcies near the same
networks. Figure 6 shows the most early bankruptcies (before the values, but a system with is destroyed earlier. It can be
100th iteration) for . At the same time, for first explained by the fact that for an Erdős–Rényi network with 100
bankruptcies occured after the 250th iteration. It should be noticed nodes a threshold probability for the appearance of giant
that for the cases of and component is , so is border-line value in terms of
dynamics of bankruptcies looks rather similar. network connectivity. Figure 8 also shows that the system is the
most fragile for this value. Systems with show
The second feature presented in the Figure 6 is Potential of
similar dynamics in bankruptcies, and converge to some value
Interaction. It reflects summarized proposal impact of edges in the
near the 350th iteration.
network (for the detailed description of a metric see [24]). Each
edge is implied to have possibility of both positive and negative As for velocities, networks with have higher amplitude,
effect. Figure 6 displays that for there is high possibility and they are well-distinguished with parameters variations.
of negative impact near the 50th iteration. For this occurs
only about the 350th iteration. Thus, we can summarize that the
systemic risk increases with increase of a number of attached
edges in a Barabasi-Albert generative model.
 5. CONCLUSIONS
In this paper, we study the impact of initial network topology on
dynamics of time-evolving banking system. The experiments were
carried out for Erdős–Rényi, Barabasi-Albert and Watts-Strogatz
network generative models with varying parameters. We show
that for fixed banks' behavioral policies different topologies
influence the resulting stability of a system with different
intensities.
For the Erdős–Rényi model our experiments show that sparse
networks are more fragile for fixed bank policies. At the same
time, it seems that there is some range of probabilities of edge’s
creation when all configurations show similar dynamics. For two
different types of networks the results illustrate that sparse
networks are more robust for Barabasi-Albert, and more fragile
for Watts-Strogatz models. For particular ranges of parameters, it
was shown that different initial topologies have invariants
Figure 8. Dynamics of bankruptcies converging to some similar values during modeling.
for Erdős–Rényi initial network Further extensions of this study can include theoretical
investigation of bounds of initial parameters providing the
convergence of topological characteristics to a fixed value. This
investigation should include exploration of interplay between
banks’ and customers’ policies, parameters of simulation and
parameters of network topologies.

6. REFERENCES
[1] Montagna, M., Kok, C. 2013. Multi-layered interbank model
for assessing systemic risk. Kiel Working Paper, No. 1873.
[2] Georg, C. P. 2013. The effect of the interbank network
structure on contagion and common shocks Journal of
Banking & Finance. – 2013. – Т. 37. – №. 7. – С. 2216-2228
[3]
[4] Lenzu S., Tedeschi G. Systemic risk on different interbank
network topologies //Physica A: Statistical Mechanics and its
Applications. – 2012. – Т. 391. – №. 18. – С. 4331-4341.
[5] Bougheas S., Kirman A. Complex financial networks and
systemic risk: A review. – Springer International Publishing,
Figure 9. Dynamics of an average shortest path 2015. – С. 115-139.
for Erdős–Rényi initial network
[6] M. Boss, H. Elsinger, M. Summer, and S. Thurner. Network
topology of the interbank market. Quantitative Finance,
4:677–684, December 2004
[7] Cajueiro D. O., Tabak B. M. The role of banks in the
Brazilian Interbank Market: Does bank type matter?
//Physica A: Statistical Mechanics and its Applications. –
2008. – Т. 387. – №. 27. – С. 6825-6836.
[8] H. Degryse and G. Nguyen. Interbank exposures: an
empirical examination of contagion risk in the belgian
banking sector. International Journal of Central Banking,
3(2), June 2007
[9] E. Nier, J. Yang, T. Yorulmazer, and A. Alentorn. Network
models and financial stability. Journal of Economic
Dynamics & Control, pages 2033–2060, March 2007
[10] K. Soramaki, M. Bech, J. Arnold, R. Glass, and W. Beyeler.
The topology of the interbank payment flows. FRBNY Staff
Rep, (243), March 2006
[11] D. W. Diamond and P. H. Dybvig. Bank runs, deposit
insurance and liquidity. Federal reserve bank of Minneapolis
Quartely Review, 24(1):14–23, Winter 2000
Figure 10. Average rate of change of net worth and available
funds for Erdős–Rényi initial network
 [12] H. Degryse and G. Nguyen. Interbank exposures: an [19] Eboli M. Systemic risk in financial networks: a graph
empirical examination of contagion risk in the belgian theoretic approach //Universita di Chieti Pescara. – 2004.
banking sector. [20] Ismail O. R. H. AN AGENT–BASED COMPUTATIONAL
[13] R. Cont and E. B. Santos. Banking system topology and MODEL FOR BANK FORMATION AND INTERBANK
systemic risk: the brazilian network empirical analysis. NETWORKS. – 2012.
Working Paper, September 2009. [21] [This title was eliminated to maintain anonymous review
[14] May R. M., Arinaminpathy N. Systemic risk: the dynamics process]
of model banking systems //Journal of the Royal Society [22] Purica I. Nonlinear Dynamics of Financial Crises: How to
Interface. – 2010. – Т. 7. – №. 46. – С. 823-838. Predict Discontinuous Decisions. – Academic Press, 2015.
[15] Quax R., Apolloni A., Sloot P. M. A. The diminishing role [23] Chung F. R. K., Lu L. Complex graphs and networks. –
of hubs in dynamical processes on complex networks Providence : American mathematical society, 2006. – Т. 107.
//Journal of The Royal Society Interface. – 2013. – Т. 10. –
№. 88. – С. 20130568. [24] [This title was eliminated to maintain anonymous review
process]
[16] Gai P., Kapadia S. Contagion in financial networks
//Proceedings of the Royal Society of London A: [25] Mohar B. et al. The Laplacian spectrum of graphs //Graph
Mathematical, Physical and Engineering Sciences. – The theory, combinatorics, and applications. – 1991. – Т. 2. – №.
Royal Society, 2010. – С. rspa20090410. 871-898. – С. 12.
[17] Amini H., Minca A. Inhomogeneous Financial Networks and [26] Ye C. et al. Thermodynamics of Time Evolving Networks
Contagious Links //Available at SSRN. – 2014. //Graph-Based Representations in Pattern Recognition. –
Springer International Publishing, 2015. – С. 315-324.
[18] Silva T., Souza S., Tabak B. Network structure analysis of
the Brazilian interbank market. – Central Bank of Brazil,
Research Department, 2015. – №. 391.

Columns on Last Page

Should Be Made As Close
As Possible to Equal
Length

View publication stats