iConViz: Interactive Visual Exploration of the Default Contagion Risk of Networked-Guarantee Loans
iiConViz: Interactive Visual Exploration of the Default Contagion Risk ofNetworked-Guarantee Loans
Zhibin Niu ∗ , Runlin Li ∗ , Junqi Wu * , Dawei Cheng † , Jiawan Zhang ∗ ∗ College of Intelligence and Computing, Tianjin University † Department of Computer Science and Engineering, Shanghai Jiao Tong UniversityFigure 1: System interface of iConViz: (a) Guarantee Network Explorer. It facilitates an overview of and zooming in on levels of detailregarding guarantee networks using a network tessellation layout. It also offers intuitive metaphorical symbols (Contagion EffectBadges) of contagion risk to support the selection of interesting networks. (b) Contagion Effect Matrix. This gives detailed contagionrisk patterns and quantifies severity. (c) Chain Instance Explorer. This supports further narrowing of the search space for instancesof contagion from a financial perspective. (d) Node Instance Explorer. This visualizes the finest-grain information on demand. A BSTRACT
Groups of enterprises can serve as guarantees for one another andform complex networks when obtaining loans from commercialbanks. During economic slowdowns, corporate default may spreadlike a virus and lead to large-scale defaults or even systemic financialcrises. To help financial regulatory authorities and banks managethe risk associated with networked loans, we identified the defaultcontagion risk, a pivotal issue in developing preventive measures,and established iConViz, an interactive visual analysis tool that fa-cilitates the closed-loop analysis process. A novel financial metric,the contagion effect, was formulated to quantify the infectious con-sequences of guarantee chains in this type of network. Based on thismetric, we designed and implemented a series of novel and coordi-nated views that address the analysis of financial problems. Expertsevaluated the system using real-world financial data. The proposedapproach grants practitioners the ability to avoid previous ad hocanalysis methodologies and extend coverage of the conventionalCapital Accord to the banking industry. * e-mail: { zniu, runlinli, wujunqi, jwzhang } @tju.edu.cn, Jiawan Zhang isthe corresponding author. † e-mail: dawei.cheng@sjtu. edu.cn Keywords:
Visualization analytics, Regulatory visualization
NTRODUCTION
Regulatory technology (RegTech) is designed to enhance trans-parency and consistency and address the regulatory challenges facedby financial services providers, including monitoring, reporting, andcompliance obligations [3, 4]. The rapid development of RegTechhas increased awareness of visual analytics and artificial intelligencein this area [32]. The chief economist for the Bank of England,Andy Haldane, imagined the future of RegTech to be a global mapof financial flow that charts spillovers and correlations [21].Networked-guarantee loans, a unique financial and banking phe-nomenon in some countries, are attracting increased attention fromregulators and banks. When enterprises back one another in loanapplications, they form complex directed networks. Highlightedby the multifaceted background of the growth period, the structuraladjustment of the pain period, and early stages of the stimulus period,structural and deep-level contradictions have emerged in the eco-nomic development system. During economic slowdowns, corporatedefault may spread like a virus and lead to large-scale defaults oreven systemic financial crises. Thus, the need for risk management ismore urgent than ever before. Monitoring the worlds financial statusis so complicated that it is usually only after a capital chain rupturethat regulators are able to study cases in depth. Regulators and banks Financial terminology referring to a business failure to fulfill an obliga-tion, especially to repay a loan or appear in a court of law. a r X i v : . [ q -f i n . R M ] A ug re seeking to utilize data-driven and visual analytics approaches tomanaging the risk brought about by networked loans. However, thepublic research output for this problem is still relatively limited.We worked closely with financial experts to conduct research onthe guarantee network loan risk management problem. We identifiedthat the default contagion risk for networked loans is an importantyet unexplored interdisciplinary research problem. Data-driven andvisual analytics-based approaches can provide fresh insights intoassessing contagion risk and possible preventative measures. In thisresearch, we propose iConViz, a novel visual analytics approach, asa means of helping financial experts conduct in-depth analyses ofthe default contagion risk problem. We believe that this is the firststudy to identify and formalize the contagion chain risk managementproblem for networked loans.The main contributions of this work are as follows:• We discover and highlight eight interpretable contagion chainpatterns by analyzing real-world bank loan records. The pat-terns illustrate different contagion characteristics that lay thebasis for quantitative contagion risk assessments (i.e., the con-tagion effect) and our visualization design.• We propose a systematic data-driven and visual analytics ap-proach to the contagion risk problem within the framework ofthe closed-loop analysis process. Our approach offers financialexperts the ability to avoid previously ad-hoc methodologies.• We describe iConViz, an interactive visual analytics tool we de-veloped to analyze the contagion risk problem with networkedloans. Several novel visualization and interaction designs areproposed, such as the guarantee network tessellation layout,flower petal zooming interaction design that overcomes visualclutter, and the contagion effect badge that provides visualsymbols for quantifying contagion risk.
ELATED W ORK
Data and visual analytics technologies are applied extensively infinancial risk management problems in domains such as macro-prudential oversight and fraud detection in online transactions andinvestment [17, 19, 25, 37]. In this section, we first introduce thecapital accord, an important risk management framework for thebanking industry, and then works on interdisciplinary financial riskmanagement and data visualization analytics.
Capital Accord
The Basel Committee on Banking Supervisionissued a series of recommendations on banking laws and regulations(Basel I, II, and III) to enhance the understanding of key supervisoryissues and improving the quality of banking supervision [14]. Theseprinciples have been widely accepted by banks around the world.Under Basel II, a series of parameters, also named as internal ratings-based approach [41], are used to calculate the economic or regula-tory capital of banking institutions. 1) Probability of default (PD).Default probability can be estimated from historical default data,observable prices of credit default swaps, bonds, and options on thecommon stock market identified using machine learning algorithmslike decision trees, logistic regression, support vector machines, neu-ral networks, genetic programming, ensemble methods, and manyother machine learning processes [7, 28]. In credit risk manage-ment, the standard assumption is that a loan is considered in defaultwhen the client is past due on payment by at least three months. 2)Loss-given default (LGD). This refers to the share of an asset thatis lost if a borrower defaults. LGD is facility-specific because suchlosses may be influenced by key transaction characteristics suchas the presence of collateral and the degree of subordination. 3)Exposure at default (EAD). This is defined as the gross exposure ofa facility upon the default of an obligor. 4) Expected loss. This canbe formulated as the product of PD, LGD, and EAD.
Regulatory visualisation
The rapid development of regulatorytechnology has raised awareness of information visualization and vi-sual analytics in this area. We provided regvis.net, a visual bibliogra- phy of regulatory visualization for better indexing the literature [32].In detail, many studies of visualization for macro-prudential over-sight (measures systemic risk both timely and accurately) have beenconducted for the financial sector [37]. The foremost work is bythe volatility laboratory of Nobel Laureate Robert Engle. His groupprovides real-time online measurement, modelling, and forecastingof financial volatility and correlations with systemic risk via classicinteractive charts such as bar charts, Box plots, map charts, andother prime examples [18]. In the cross-sector of machine learningand financial stability, Peter Sarlin published a series of works onvisualized systemic risk analytics in cooperation with the EuropeanCentral Bank, International Monetary Fund, and other internationalbanks [35]. The self-organizing map, a neural network-based unsu-pervised learning method and visualization tool, can be utilized toevaluate and visualize financial stability with the power of simultane-ous clustering and projection [36, 38, 39]. It is also extended with anovel time dimension to decompose and identify temporal structuralchanges in macro-financial data around the global during financialcrises [37]. Fraud detection in transactions is always a primaryconcern of banks [17]. The Advanced Detection System was oneof the earliest risk visualization systems that monitored trades andquotations on the Nasdaq stock index, identifying patterns and prac-tices of behavior of potential regulatory interest [26, 27]. Wirevisemploys specific coordinated keyword visualization for wire trans-actions that can be used to detect suspicious accounts, transactions,behaviors [9]. 3D tree maps have been introduced to monitor real-time stock market and identify unusual trading patterns, suspectedtraders (i.e., attackers), and attack plans [23].Network visualization is employed to represent bank interrela-tions through financial discussion data [12]. The force-directed lay-out is perhaps the most extensively utilized in the financial area. Andthe network centrality measurements such as node degree, between-ness, and closeness, K-core shell are used to measure and visualizethe node or edge importance. For example, Rnnqvist and Sarlincollected text data from online financial forums and generated andvisualized a co-mentioned bank network (i.e., interbank network),with which to quantify the bank interdependence (using centralitymeasurements), such as interbank lending and co-movement in mar-ket data [34]. Xu and Chen discussed criminal network analysis andvisualization, a very close domain. Their insights included that so-cial network analysis can be used to analyze interaction patterns andcriminal networks can be partitioned into subgroups of individualsby the centralities in their network [15,45]. Bottom-up and top-downinteraction are demonstrated can be effectively to reveal financialcrimes such as money laundering and fraud in the financial activitynetwork [16]. Heijmans and others used animation to visualize andanalyze the large transaction networks in the daily Dutch overnightmoney market [22]. There were some other publications mining thesubgraph structures and patterns to interpret the financial meaning.Among them, BitExTract was developed to observe the evolutionof transaction and connection patterns of Bitcoin exchanges fromdifferent perspectives [46]. A ego-centered node-link view depictsthe trading network of exchanges and their temporal transactiondistribution and facilitates the recognition of unique patterns.Guarantee networks consisting of multiple enterprises related bysecured loans were first addressed by the computing community in2015 [30]. Subsequently, a risk management framework for theseguarantee networks was introduced and a visual analytics approachpresented [31]. Since then, intensive research on default risk pre-diction has been performed [10, 11]. However, to the best of ourknowledge, the contagion risk management problem for guaranteenetwork loans has not been adequately addressed. The present re-search is the first attempt to be made by financial computing researchcommunity. One core question is how to assess the risk of contagionintroduced by a network structure. The metric contagion effect,together with a practical visual analysis pipeline, is thus proposedto facilitate a better understanding and more accurate assessment ofontagion risk.
ROBLEM S TATEMENT AND R EQUIREMENT A NALYSIS
In this section, we first present the necessary background of thisunique financial phenomenon, and then report major concerns ob-tained from financial experts; finally, we offer a summary of theanalysis tasks and design rationales based on real-world motivations.
The origin of networked-guarantee loans is illustrated in Fig. 2. It is acommon scenario in which small businesses that wish to obtain loansfrom commercial banks usually lack the security required. In thiscase, they are allowed to seek a guarantee from other businesses. Inpractice, there can be more than one guarantor per loan transaction,and there may be multiple loan transactions for a single guarantor ina given period. Once the loan is approved, the company can usuallyimmediately obtain the full amount of the loan and begin to repay thebank via a regular instalment plan, until the end of the agreement.
Figure 2: A typical loan guarantee process includes four major steps.
We worked closely with two loan assessment experts to betterunderstand the real-world challenges inherent in this system. Expert E a was a senior financial regulatory officer with more than five yearsof experience with the guaranteed loan problem and had publishedseveral important and relevant investigation reports. Expert E b camefrom our partner bank, had ten years of loan approval experience.They divided the default contagions and corresponding responseinterventions into four phases, as shown in Fig. 3. Figure 3: Default contagious phases and response interventions. Thered nodes refer to the default company and dashed nodes/links arethe default contagion chain.
Default contagions are usually triggered by accidental defaultsthat introduce risks. Since these are obligation contracts, the defaultcontagion may spread to adjacent nodes (those providing guarantees).
Anticipating how debt default may spread is critical to introducingthe appropriate response interventions.
An appropriate guaranteeunion reduces the risk of default, but in practice, significant damagefrom contagion can still occur among networked companies. In the case of a down economy, defaults can multiply as large-scalecorporate defaults cause side effects in the network. In such cases,the guarantee network can be alienated from the “joint assistancegroup” as a “breach of contract”. When some companies face opera-tional difficulties, the crisis may set off a domino effect. Default canspread rapidly across the network and put a large number of compa-nies in unfavorable positions. A systemic crisis can result. At thisstage, control and mitigation are imperative. After the eliminationor eradication stage, the guarantee network may need to be split intoseveral smaller networks, with some companies bankrupted and thetransmission risk reduced.Below, we summarize the main concerns of the experts and out-line the requirements for mitigation in the following subsection.
Concern 1:
Basel accord-based risk management systems may notbe well-suited for networked loans, as the network relationship isunique and exceeds the hypothesis [14]. It is urgent and essentialto adapt the old or establish new risk measurements for this type ofproblem.
Concern 2:
Particular attention should be paid to conta-gion risk to prevent the large-scale corporate defaults often broughtabout by networked loans.
Accidental default is usually tolerable,while large-scale defaults or systemic financial crises must abso-lutely be prevented.
However, it is currently unclear how contagionspreads through guarantee networks, how vulnerable the nodes are,or how likely default is to spread. They seek to monitor/assessthe status of such spreads via data-driven and visual analytics ap-proaches to resolving risk and ensuring financial stability.
Below we summarize our analysis tasks, addressing experts’ mostsignificant concerns, supporting the assessment of crisis levels, andgaining insight into precautions preventing potential financial risks.
T.1 Explore networks at different levels of detail, quickly locat-ing networks of interest.
Motivated by expert concerns andtechnical challenges when analyzing massive bodies of guar-antee data, the system should support the quick location ofnetworks of interest and analysis of them according to differentlevels of detail.
T.2 Understand how default contagion may spread across aguarantee network.
The forward approach to understandingthe spread of default is to simulate the situation by establish-ing virus-epidemic models such as in [33]. However, in ourcase this is impossible, due to a shortage of empirical defaultdata. From a data-driven perspective, identifying the potentialcontagion chains and extracting their patterns can also provideinsights useful to the ultimate goal of precautions that avoid orresolve systemic financial risk.
T.3 Analyze instances of contagion chains.
This should sup-port case-by-case risk assessment and evaluation, as there aremultiple instances of contagion chains even with the same pat-tern. Appropriate quantitative indicators are helpful to makingcareful comparisons.
T.4 Provide novel and objective risk measurements/indicatorstailored for networked loans.
The classic Basel accord-basedrisk measurement is not well-suited to the problem, as it isbased on the assumption that these are giant independent play-ers in the market. It is necessary to set up a novel objectiverisk indicator for this type of problem.
T.5 Identify the critical nodes that may be the most destructivein case of risk contagion.
Just like the super-spreaders inepidemiology [40], a few corporations are prone to “super-spreading events”, and given the right conditions can igniteexplosive epidemics. Moreover, such volatility also means thatoutbreaks are more likely to fizzle out if the key nodes areidentified and removed (as preventive measures) quickly.Based on these analysis tasks, a series of design decisions wasmade, as outlined below.
DR 1. Scalable and appropriate net-work layout visualization.
The system requires a scalable massive igure 4: Overview of the approach, including: (a) data preprocessing, (b) contagion pattern analysis, and (c) visual analysis loop. The solid arrowis the data processing flow, and the dotted arrow is the visual analysis flow. node network layout to avoid severe visual clutter and sufficientperformance to enable sophisticated interactions. In practice, tar-get users may wish to analyze networks of interest on various lev-els of detail; thus, we need a compact arrangement incorporatingover 3,000 independent networks via interactions such as magnify,filter, and select (T.1-T.5).
DR 2. Focus on contagion chains.
Identify and discover the default contagion patterns in a guaran-tee network (T.2) and propose appropriate contagion risk indica-tors/measurements based on the patterns (T.4), as well as intuitivevisualization to help with efficient evaluation of the assignment ofpriorities.
DR 3. Appropriate symbols and color mapping for in-tuitive metaphors.
Intuitive representation is an essential elementof most visualization systems. Proper visual metaphors help expertsreduce the visual burden and improve their understanding of theactual situation (T.1).
ETHOD
This section describes the method of the visual analytic system.
Motivated by the analytical tasks outlined above, we designed andimplemented iConViz to support financial experts interactively ex-ploring the contagion risk associated with networked loans.Fig. 4 gives an overview of the approach. The process mainlyincluded three steps. 1) Data preprocessing. Raw bank loan recordshave been cleaned and reorganized into the node table (corporationprofiles), edge table (guarantee relations), and contagion chain table.We use solid arrow to illustrate the data processing flow. 2) Conta-gion pattern analysis. We employed an unsupervised learning-basedapproach to extract contagion chain structure patterns, and here pro-pose a novel financial metric for quantifying the contagion risk ofthe chains and networks. The patterns and metrics formed the basisof our visual design and support this type of financial analysis. 3)Visual analysis loop. We designed a visual interface that is closelycoupled with financial risk management tasks and knowledge tosupport closed-loop analysis processing and an iterative level ofdetailed exploration (see the dotted arrows and we give more detailin Section 5.5). We describe the details of the data and contagionpattern analysis in the subsequent subsection and the visualizationdesign and interaction in the section after that.
In this work, we collected ten-year loan data from cooperating com-mercial banks and built the guarantee networks. The names of thecustomers in the records were encrypted. In the record preprocess-ing phase, by joining the tables, we obtained records related to thecorporation ID and guarantee contract. We then constructed theguarantee network. At the data preprocessing step, we cleaned andreorganized the record data into three main tables, as Fig. 4 shows.1) The node table included the corporation profiles (businesstype, size, and registered capital); the primary key was the corpo-ration ID. 2) The edge table was the directed guarantee relationsbetween corporations (i.e., the nodes in the node table). It alsoincluded the guarantee amounts between them as weights of theedges. Fig. 5 gives an overview of the real-world data set, with eachnode representing an enterprise and the link direction representingthe guarantee relationship. More than 20,000 businesses and morethan 3,000 independent networks are visualized. It is clear that thenetworks overlap, and few insights can be drawn. Zooming in, thereis a significant directed subgraph microstructure. Some prime as-pects may be familiar to loan assessment domain experts [10,11,31],but the collective financial properties of those microstructures areunclear and need to be explored.
Figure 5: Overview of the real-world dataset. When zooming in, wecan observe various directed subgraph microstructure. The classicloan guarantee patterns include [31]: (a) direct guarantee; (b) mutualguarantee; (c) revolving guarantee; (d) star shape guarantee; (e) jointliability guarantee. ) Contagion chain table. We defined the contagion chain (i.e.,the chain of contagion) as the subgraph of where the default mightspread. If we can obtain and understand the contagion risk pattern,we may be able to determine how the default risk might spread acrossthe guarantee network and thus implement measures to prevent anypotential occurrence of large-scale defaults. The contagion chaintable was reconstructed from the edge tables (see the solid arrowsin Fig. 4) and used repeatedly throughout this work. The contagionchain, different from the guarantee chain, worked in the directionopposite to the arrows. In practice, the guarantee network was splitinto several subgraphs of contagiousness (noted as contagion chains)when we reversed the directions of the arrows. We applied a breadth-first traversal algorithm and generated a series of subgraphs, storingthem as contagion chain files in the JSON format. Fig. 6 includesan example where (a) is the guarantee network, (b) is the contagionchain when Node A defaults (highlighted with a virus shape icon)and spreads the risk, and (c) gives all possible contagion chains. Itshould be noted that though some of the chain were subgraphs ofother chains, all were analyzed equally because each node could bethe source of outbreak.
Figure 6: Guarantee network and default contagion chains. In thispaper, we utilize the virus-shape node (such as node A in (b)) torepresent the source of outbreak (default crisis).
The contagion chain pattern analysis is the basis of our approach andcritical to understanding the contagion properties of a network. Inthis section, we apply the unsupervised learning approach to extractthe patterns (typical topological structures), interpret their financialmeaning, and quantify the risk brought by each kind of contagionpattern.The contagion chains were subgraphs, so we extracted networkinformation propagation-related attributes to construct the contagionchains features. In this way, each contagion chain was representedby a five-dimensional vector. In detail, the attributes included: (1)the number of nodes and edges of the contagion chain, noted as N ( c i ) and E ( c i ) , respectively. (2) The density of the contagion chain, notedas D ( c i ) . This was defined as the ratio of the number of edges to thenumber of possible edges in a network with nodes. It measures theproportion of possible ties that are actualized among the membersof a network. (3) The average clustering coefficient of the contagionchain, noted C i . This is computed as the average of the clusteringcoefficients C i of all of the nodes in the chain. It is closely related tothe transitivity of a graph and serves as an indicator of a small world.(4) Average shortest path length, noted as l G . This is calculated byfinding the shortest paths between all pairs of nodes and taking theaverage over all paths of the length thereof. It gives the numberof steps it takes to get from one node to another and measures the efficiency of information or mass transport on a network.We chose to use spectral clustering for mining the contagion chainstructure patterns. The approach frequently outperforms traditionalmethods such as k-means or single linkage, especially in graph-based clustering tasks [43]. Moreover, it can be solved efficiently bystandard linear algebra. There are three major steps: Step1: Createthe similarity graph between the contagion chains. We chose to fullyconnect the graph with the Gaussian similarity function to transforma given set x ,..., x n of datapoints with pairwise similarities s i , j orpairwise distances d i , j into a graph G = ( V , E ) . Step 2: Compute thefirst k eigenvectors of the Laplacian matrix to define a feature vectorfor each contagion chain. The graphed Laplacian matrix is definedas L = D − W , where adjacency matrix W is the weight between thevertices of the graph and D is the diagonal degree matrix. It has beenproven that the matrix L has as many eigenvalues of zero as there areconnected components (or clusters), and the corresponding eigen-vectors are the indicator vectors of the connected components. Step3:Run k-means on these features to separate objects into k classes.Projecting the points into a non-linear embedding enhances the clus-ter properties in the data so that they can be easily detected. Inparticular, the simple k-means clustering algorithm has no difficultydetecting the clusters in this new representation with the estimationof k when the eigenvalues are zero. Figure 7: Eight typical contagion chain patterns by unsupervisedclustering. The virus shape nodes represent the source of outbreak(the first node in the contagion chain). It is noted that there could bemultiple source of outbreaks like in P.3 – P.6 due to the existence ofmutual guarantee.
Fig. 7 lists the eight basic contagion chain patterns discoveredby the above approach. The outbreak (or default crisis) starts fromthe node in red and spreads through the guarantee network in eightpatterns. In detail, they are:
P.1 Direct contagion pattern.
Thebasic contagion pattern is where the default can only be spread toits (one) adjacent node and then the contagion is stopped.
P.2 Sin-gle chain contagion pattern.
This usually extends from the directcontagion pattern, with more nodes involved in the contagion chain.The default crisis can only spread across the single chain in thesame direction. The length of the entire single strand is arbitraryand all chains in the structure are categorized according to this pat-tern.
P.3 Mutual contagion pattern.
This describes a situationin which two corporations simultaneously guarantee one another(mutual guarantees) and obtain funds from a bank. Both nodes arefragile because no matter which one encounters the default crisis,the other will be affected.
P.4 Mutual-ext contagion pattern.
Thisis usually an extension of the mutual contagion pattern, where thecontagion chain involves other nodes.
P.5 Loop-mutual contagionpattern.
This is of a loop structure, when three or more nodessimultaneously guarantee one another. Such a structure is quitevulnerable as any default crisis may spread to all of the nodes inthe loop.
P.6 Loop-mutual-ext contagion pattern.
This is usuallyextended from the loop-mutual contagion pattern, when the conta-gion chain involves other nodes.
P.7 Star contagion pattern.
Thedefault crisis of one node will affect many other nodes and thenthe contagion will stop. This can occur when corporations provideguarantees for the same (weak) corporation; the default may spreadto all of the companies that provide support.
P.8 Star-ext contagionpattern.
This is extended from the star contagion pattern, but withore complex structures involved. The default of one node may bespread to several other order nodes.The patterns gradually grow more complicated with the morenodes involved and more complex the guarantee relationships. Inpractice, the contagion chains are combinations of these patterns,and a network may have several instances of the same pattern. Fig. 8gives the example of the pattern (P.8) in two guarantee networks.The default spreads across the chain (in red) and then stops. We alsoobserved that these contagion patterns fell into four types, accordingto the behavior of the contagion : single chain, mutual contagious,loop contagious, and star contagious. We propose here a means ofquantification for contagion risk assessment based on the behaviorof the contagion. Additional details appear in the next section.
Figure 8: Contagion chains in pattern P.8 in guarantee networks. Theclustering algorithm gives each contagion chain a pattern label, andthe first nodes of the chains are the source of outbreak. We utilize thevirus shape node icon to represent the source of outbreak.
ISUALIZATION D ESIGN AND I NTERACTION
In this section, we describe the interface of the iConViz systemthat supports financial experts interactively and iteratively exploringand explaining the contagion risk for networked-guarantee loans.We designed the four coordinated views (see Fig. 1) to facilitatethe closed-loop analysis process and iterative level of detailed ex-ploration, following Shneidermans mantra. Figure 4 illustrates theanalysis procedure. High-level interactions are supported, and allfour main views are coordinated to facilitate various levels of de-tailed exploration of the networks and comparisons of contagionpatterns, instances, and specific businesses. It supports a cycle analy-sis between high-level (massive networks), middle-level (contagionchain-level), and low-level (independent node-level) networks. Suchan analysis procedure allows users to finally understand networksand contagion patterns in iterative ways.
The Guarantee Network Explorer (GNE) view facilities an overviewof and zooming in on a level of detail of a guarantee network, usinga network tessellation layout. It provides intuitive and metaphoricalsymbols of contagion risk (through the Contagious Effect Badge(CEB) described in Section 5.2) to support selection by financialinterest. It usually works as a starting point. A set of interactivetools are provided to enhance the in-depth analysis of each network.
Guarantee network tessellation.
Many of the networks are com-posed of tens or hundreds of nodes, with rare networks composedof thousands of nodes [31]. The naive force-directed graph layoutvisualizes the whole dataset as a hairball and introduces seriousvisual clutter. We designed a grid layout to tessellate the guaran-tee networks, as shown in Fig. 1 (a). In detail, the networks arelaid in order of their complexity (i.e., the number of nodes) for thecommon prejudice that complex networks are more prone to inducelarge-scale corporate defaults (see Concern 2 in Section 3.1).
Interactions.
Rich interactions, including brushing, zoomingin/out, view panning, and dragging are all supported. The zoomingoperation supports navigation of the networks at different levelsof detail. The panning operation enables the viewing of detailednode profiles. The dragging operation facilitates exploration of thenetworks in the canvas. Some more enhanced interaction tools havebeen developed to support in-depth analysis of the networks. Atthe bottom of the GNE view in Fig. 1, from left to right, theseare: 1) expanded view ( ) and 2) risk badge trigger ( ). Therisk badges are visual symbols of the contagion risk of the overallnetwork. When the risk badge trigger button is on, all risk badgesare overlayed on the networks to help experts locate networks ofinterest. More details are provided in the following section. 3) Boxselection ( ). Users can select the network of interest and highlightthe nodes and edges, making the contagion effect matrix (CEM),Chain Instance Explorer (CIE), and Node Instance Explorer (NIE)views display the corresponding content. 4) Edge width trigger ( ).The edge width trigger is proportional to the guarantee amount. Itis useful when analyzing specific networks. We also defined twobuttons. The first is to obtain better performance when loading allnetworks. The second is for use when the width of the edge is notapparent in the network tessellation view. 5) Color palette ( ).Nodes can be colored by a default rate (a graph neural network-based prediction that will be discussed in subsequent research), type,and size of businesses.
Addressing the T.2 and T.4 requirements and based on the behaviorof contagion, we created the
Contagion Effect , a novel financialmetric for quantifying the severity of the risk of contagion. Basedon this core metric, we designed a Contagion Effect Matrix (CEM)view to encode the risks of each network in a matrix manner. It alsoworks as a filter for chain-level analysis. It is further abstracted asthe
Contagion Effect Badge , a visual chart indicating the risk levelsappearing in the GNE view.In finance, the contagion effect explains the impact of a spreadingcrisis in a situation where one shock in a particular economy or re-gion spreads out and affects other sectors [1, 8, 13, 29, 44]. However,as far as we know, there has to date been no measurements quantify-ing the contagion effect for the guarantee network problem. In thisresearch, we identify the contagion risk associated with networkedloans as an important yet unexplored interdisciplinary research prob-lem. As mentioned in the Section 3.1, the classic Basel accord-basedrisk management systems may not be well-suited for networkedloans, as the network relationship is unique and exceeds the hypoth-esis and accidental default is usually tolerable, while large-scaledefaults or systemic financial crises must absolutely be prevented.The extreme case of contagion is the key to preventing any potentialsystemic crisis when the default spread across a network. Thus, twofactors are important: (1). how many corporations a node may affectto the greatest extent possible, and; (2). how frequent different typesof contagion occur. With this information, financial experts canlocate and prioritize networks of interest. We worked with financialexperts to explicitly define the contagion effect of pattern P i as: E i ∝ f i ∗ v i (1)where f i is the frequency and v i is the length of the contagionchain pattern. In practice, networks may have contagion chainpatterns of various compositions, meaning that default can spreadin drastically different ways. The frequency f i describes how muchrisk is induced by pattern P i , and v i describes how many other nodesit may infect at most. The two key factors ( f i and v i ), though notdirectly inspired, essentially follow the principle of the risk matrixin the ISO standard risk assessment techniques [2, 5, 6, 24], wherethe Level o f Risk = Probability × Consequence . Quantifying thecontagion effect arranges the patterns in the CEM. We used the rangef influence (meaning how many other nodes might be impacted)to represent the contagion consequence and level of vulnerability(meaning how frequently this kind of pattern might happen) torepresent the contagion probability.
Contagion Effect Matrix.
This matrix was designed to visualizethe contagion effect of the patterns in a network. As Fig. 9 shows,the column is the length v of the contagion chain pattern. The rowsare the frequency f i of the contagion chain pattern, where we directlyused the count of instances of the pattern. In this view, all of thepatterns are spatially arranged into four quadrants according to thebehavior of the contagion. Each quadrant is encoded with a colordesignating the risk level (consistent with the color specifications inthe financial sector). These colors form the basis of the risk badge.Each cell also displays the count of the instances of this pattern fora selected network in the top-left corner. Figure 9: Contagion effect matrix, where the patterns are arrangedby the contagion behavior. Each quadrant is given a color accordingto its contagion risk level. The red, orange, yellow, and green colorcorrespond to high, middle, low, and safe risk levels. This is consistentwith the convention settings in the credit rating business. Along theRange of Influence axis is the node number of the patterns. Accordingto the node number, P.1 = P.3, and P.2 = P.4 = P.5, thus, we use theextra letters a, b, and c to distinguish between them.
We categorized these into four kinds of contagion behavior bythe range of influence and vulnerability level. Q.I: For chain-likepatterns (P.1 and P.2), the default can only spread across a singlechain. Usually, such nodes and defaults will not lead to massivedefaults. It is relatively easy to break the contagion path by removingthe key node on the chain. Q.II: For mutual patterns (P.3 and P.4),the defaults can infect one another (P.3 and P.4). Such patterns arevulnerable because of mutual guarantees. Q.III: For loop-mutualpatterns (P.5 and P.6), the default may spread more easily than inchain-like and mutual patterns, due to the existence of loop-mutualguarantees. Q.IV: For star-like patterns (P.7 and P.8), the defaultin the center chain position may affect all supporting corporations.Such kinds of contagion may distress large numbers of companies.The layout of the CEM is meaningful. First, the patterns on theleft half are more vulnerable to crisis due to the existence of theloop-mutual pattern and broader range of influence, as more nodesare involved in a network than in the right-half patterns. Expertsmay, in practice, wish to have more guarantee patterns in the righthalf of the CEM to provide stabler situations. Second, the patternsin the different quadrants can be converted into one another in real situations, providing clues to useful decomposition strategies whensplitting a complex network into pieces to avoid potential systemicrisks. For example, mutual patterns are more basic than loop-mutualpatterns, and when splitting a network during a risk outbreak, wecan remove the nodes with a Q.III pattern and generate patterns inQ.II, or even in Q.I.
Contagious Effect Badge
Contagion chain patterns pervade eachof the guarantee networks. Quantifying the proportional compositioncan help to identify the type of contagion risk. The patterns can be in-tegrated effectively with financial indicators such as those in the Cap-ital Accord for better risk-level assessments of guarantee networks.We explicitly define the contagion score as a four-dimensional vector [ EDA , pq , pq , pq , pq ] , where EDA is the total amount of expo-sure of the nodes in the network and pq j is the percentage share ofinstances of this kind of contagion behavior.The CEB is designed as a four-slice pie chart to symbolize therisk levels, based on the contagion score. The size of the CEB isproportional to the relative exposure risk ratio (compared to themaximum exposure of all guarantee networks). The portion of eachshare of the CEB is encoded as pq j to chart the contagion behaviors.The badge can be overlayed through the functional button onto thenetworks in the GNE view, allowing users to quickly locate networksof interest (see Fig. 1 (a)). The size and color of the CEB encodesthe relative exposure ratios and contagion types. Note that the risklevels are not consistent with the complexity of the networks, andusers need only choose the network of interest via the CEB. Thiswas emphasized in the training session during the case studies. The Chain Instance Explorer (CIE) is a tailored financial coordinatesystem for middle-level (contagion chain) risk analysis. Fig. 10 (a)shows the financial coordinate system designed for this research,where the y-axis is the exposure and x-axis is the total guaranteeamount of the chain. Each node in the financial coordinate systemrepresents a chain instance. In practice, multiple nodes may havethe same exposure and guarantee amount (refer to Section 3.1) andocclude one another. Thus, we propose a flower-petal-zooming visualization design for the coincident object selection problem.
Figure 10: Chain Instance Explorer: (a) financial coordinate sys-tem, (b) visualization of contagion chains when multiple nodes arecoincident, and (c) dense example case for flower-petal-zoominginteractions.
In detail, each chain instance is designed as a petal-shaped node inhe financial coordinate system. When multiple petals come together,they form a flower. The number in the center shows the count ofcoincident chains. Each petal is clickable for the user to select thechain instances in other views. The design intension is to avoidpossible visual clutter by iterative brush and zooming interactionssupported for selection in extremely dense cases (see Fig. 10 (c)).The CIE is coordinated with the CEM and GNE, enabling financialexperts to explore contagion chains case-by-case.
In order to facilitate low-level (i.e., node-level) detail on demandfor the finest grain analysis of guaranteed loans, we provide thenode instance explorer. It is composed of the following five tabs(see Fig. 11). In detail, they are: 1) Node projection tab. Thisprovides an overview of the companies by similarities in their guar-antee network structures. The nodes are first represented as vectorsby node2vec [20] and then projected by t-SNE visualization [42].Box brush interaction is supported for a coordinated analysis. 2)Financial distributions tab. Four important financial statistics (i.e.,exposure, registered capital, business size, and business type) areprovided. The histograms are visualized by cross-filters to enable afurther fine-tuning that includes or excludes records. More sophisti-cated indicators may be included in the full system when deployed.3) Overall picture tab. Repulsed bubbles (corporations) are laidalong their exposure axis (i.e., their amount of debt) to prevent vi-sual clutter. The bubble sizes are proportional to the corporationsregistered capital. When a user clicks a bubble, the contagion chainis displayed in the GNE. 4) Detailed view by business size and type.Both are charts refined by business size and type to the overview tab.
Figure 11: Node Instance Explorer is composed of the (a) nodeprojection view; (b) financial distribution view; (c) overall picture view,consisting of bubbles alongside exposures to give an overview; (d)detailed view by business size and type, which is similar to the overallpicture view but omitted due to space limitations.
Coordinated interactions can draw out hidden knowledge from theiConViz system. High-level interactions are supported, and all fourmain views are coordinated to facilitate various levels of detailed net-work exploration and comparisons of contagion patterns, instances,and specific businesses. Fig. 4 illustrates the analysis procedure.The system supports a cycle analysis between high-level (massivenetworks), middle-level (contagion chain-level), and low-level (in-dependent node-level) networks. Such an analysis procedure allowsusers to eventually understand networks and contagion patterns initerative ways. As Fig. 4 shows, the main analysis loop includes four steps. Theusers start with the GNE view to choose the network of interestwith the guidance of the CEB or from the CEM view. They arethen able to observe the overall contagion patterns and obtain a “bigpicture” of the data. For example, in the first step from the GNEview, the user can click on the risk badge trigger ( ) to overlay theCEB onto the network. A user can then compare the sizes (or radii)of the CEBs to choose the network of interest based on the levelof exposure, or compare the proportions of each colored sector tochoose the most interesting network based on the contagion behavior.Next, the user can brush-select a network and use the CEM viewto see in detail the number of each chain pattern. In the L2 step, auser can click on a cell in the CEM view to see the detailed chaindistribution in terms of exposure and total guarantee amount. Whenthe user finds an interesting chain (see the red node in the CIE viewin Fig. 4), they can click on it to see the detailed chain in the GNEview (the L3 step); the source of the break node is marked by a virus-shaped icon ( ). They can also use the NIE view (the L4 step)to see corporation-level details such as the distribution of financialinformation.
YSTEM E VALUATION
We performed two case studies and one set of expert interviews toevaluate the iConViz system. First, we invited experts E a and E b to conduct case studies to validate the effectiveness of the systemin terms of contagion risk assessment. We then conducted a setof in-depth interviews with a broader set of expert users (i.e., riskcontrol managers and specialists) from a cooperative bank. Experts E a and E b were organized as a team to perform the casestudies. Both had a deep understanding to the real-world financialchallenges introduced by guarantee networks. They were also highlyinvolved in our research, as specified in Section 3. They approved thedesign for the functional view and visual analytics pipeline. Belowis a description of their analysis procedure and conclusions. Initially, the experts were attracted to the contagion risk matrixbecause it was centrally positioned and featured bright colors (seeFig. 12, view C.1.a). They familiarized themselves with the red-green credit risk color setting. By default, the numbers alongside therows indicating the vulnerability levels were arranged by range ofinfluence. They carefully reviewed the numbers in the contagion riskmatrix (see Fig. 9) and gave overview-level comments on contagionrisk (T.2, T.5).The main insights provided by the experts were as follows.(a)
The main contagion risk comes from the star-like and loop-mutual patterns in this dataset (T.2).
There was a high number ofinstances in P.5 to P.8 (as can be observed as 385, 538, 628, and1,590 on the vulnerability axis), meaning that the revolving and jointliability guarantees (see Fig. 5 (c), (e), and Fig.7 for more detail [31])are common practice, while other patterns such as the star-shapedguarantee (Fig. 5(d)) are relatively stable from the contagion riskperspective. Expert E a explained that in practice, the star-like guar-antee is usually shaped by a professional guarantee company, thejoint liability guarantee may be formed by subsidiaries and parentcompanies, and the revolving guarantee often emerges from businesspeers of a similar size. Expert E b remarked that different resolutionstrategies should be adopted for different kinds of guarantee andcontagion patterns. For example, in the P.7 and P.8 contagion in-stances, in order to prevent risk contagion, the bank should removethe outbreak nodes (see Fig. 7) so that the network will be brokenup into several smaller independent groupings. The P.5 and P.6contagion instances are a bit more sophisticated, for there might be igure 12: The analysis begins with an overview of contagion risk tofacilitate selection and comparison of networks of interest. Zoom infor the numbers. several potential outbreak nodes. That analysis would need to becoordinated with other iConViz views.(b) There are more instances of the P.7 (observed 628 times)and P.8 (observed 1,590 times) patterns than of the P.5 (observed385 times) and P.6 (observed 538 times) patterns (T.5).
Expert E a explained that this was caused by the fact that in practice, there aremore joint liability guarantees than revolving guarantees. He addedthat in preventing large-scale corporate default, it is more importantfor risk control managers to cut contagion chains of the P.7 and P.8patterns to prevent more businesses being affected.(c) The risk is not proportional to the complexity of the guaranteenetwork.
Although the 11 guarantee networks (see Fig. 12, viewC.1.a) all had complicated network structures, from the size andcomposition of the badge colors, the experts were able to obtain anoverview of how serious the risk was and what types of risk existedin the networks (T.1). For example, the 11 networks had similarnode numbers, but the radii of the CEBs obviously varied. Thismeant that different networks had quite different levels of exposure(EDA for the Contagious Effect badge). Banks should focus firston networks with larger CEB radii, as they have more bank debt(T.4). For instance, the CEB of G1 in C.1.b was obviously largerthan that of G2 in C.1.c, meaning that there was more exposure inthe former guarantee network and thus it should have attracted moreattention from banks. More cases can be observed in Fig. 1 (T.1).Moreover, the G1 and G2 networks were composed of a similarnumber of nodes but radically different network structures (see theCEBs). The number in the top-left corner of the CEM (see Fig. 12)confirmed that the G1 network in C.1.b had many more star-likepatterns ( was observed 19 times) and loop-mutual patterns ( was observed 8 times). These types can induce a significant risk ofcontagion during economic downturns (T.2, T.5). Expert E a affirmedthe result after serious consideration, and was pleasantly surprised bythe discovery. It makes sense in practice, he explained, because thesource of the risk can vary greatly and the classic Basel accord-basedrisk factors (see Sections 2 and 3.1) cannot accurately describe therisk, especially contagion risk brought about by network relation-ships. The fundamental issue for risk management is to minimizeany potential bank losses. He further affirmed the effectiveness ofthe contagion score (T4). Expert E b also confirmed that differentfrom the common perception, networks with more nodes are notnecessarily inherently more risky, as network structure and finan-cial information are also critical to such assessments. Both expertsemphasized that this was an important discovery, correcting a mis-understanding in the financial sector that more complex networksare inherently more dangerous.
The GNE view demonstrates that the CEB can effectively rep-resent risk composition and provide evidence for the selection ofanalytical priorities (T.1). The conclusions drawn by the financial ex-perts confirmed that the contagion score can provide measurementsthat perfectly meet the needs of financial professionals (T.4), validat-ing the usefulness of the CEM and extending the visual symbolismof the CEB for quantifying how default contagion may spread acrossguarantee networks (T.2, T.5).
The experts then shifted their attention to chain-level contagion risk(T.1, T.3, T.5). Fig. 13 shows the interactions and results. The expertsbrushed and selected the G1 network because it had a relatively largeproportion of contagion risk, highlighted in red ( ). The networkwas composed of various contagion chains: 19 instances of the P.8pattern, eight instances of the P.6 pattern, and four instances of theP.7 pattern (T.2, T.5). Expert E a explained that this meant that thenetwork would be vulnerable during a crisis and might influencemany nodes. The numbers can thus provide the risk behavior andguide the choice of preventative measures with regards to possiblelarge-scale corporate default. The experts needed to analyze thecontagion chains case by case because their financial informationwas all different and thus they demonstrated different levels of risk.Therefore, the experts chose to analyze the instances in pattern P.8. Figure 13: Contagion chain-level risk analysis. The dotted arrowgives the interaction operation. The virus-shaped nodes in viewsC.2.c and C.2.d were the source nodes for the outbreak. They wereautomatically labeled by the back-end algorithm.
As Fig. 13 shows, in the CIE view, the experts observed thatthough there were multiple similar chains, some resided on the x-axis, meaning there was no risk exposure (i.e., all the funds wererepaid to the bank). There were some nodes in the upper part of theoordinate system that required analysis. Thus, the experts zoomedin on the upper part (the dotted rectangle in the CIE view) andclicked on the nodes to view the chains (view C.2.b). The contagionchains highlighted in red in C.2.c and C.2.d attracted the attention ofthe experts (T.1), who explained that both were of a complex struc-ture composed of star-ext (see Fig. 7). Such shapes are multifacetedand difficult to extract via traditional ad hoc examination. Since bothtypes of chain are evidence of vulnerability and have a high rangeof influence if the default spreads, they need to be addressed, espe-cially during an economic down period. For example, the expertssuggested that from the contagion risk perspective, we could retainfinancial stability by providing individual financial aid or cuttingthe network at the outbreak node (the virus-shaped node ), as itwas in a pivotal position. In practice, preventive measures should beconsidered in a more sophisticated fashion. The NIE view providesmore practical financial information for decision making.
We conducted a set of expert interviews with target expert users (tworisk control managers and five specialists from one bank) to furtheridentify potential problems, validate the design decisions, and assessthe systems effectiveness. All interviewees were familiar with thedataset and had rich experience with loan assessment. We walkedthem through the system and then let them explore independentlythroughout the study process. We encouraged them to think about thesystem and ask questions during the meeting. We interviewed themand collected their comments and feedback regarding the systemseffectiveness, visualization design, and usability in order to identifypotential issues and refine the system accordingly.
Effectiveness:
The system received high endorsement from allinterviewees. Since the risk factors was extended from the widelyaccepted Capital Accord risk management system, the Risk Badgesand Contagion Risk Matrix were easily understood and accepted bythe users. The target expert users remarked that it was very usefulto be able to quantify the contagion risk and explore the hazardsassociated with different guarantee networks in such a user-friendlymanner. Previously, they had no choice but to adopt a bottom-upapproach using Excel or SQL queries (i.e., begin with one node andfollow the vine to analyze the neighbors). They could usually onlyanalyze relatively simple guarantee networks and were not providedwith the bigger picture. Thus, it was difficult for them to immersethemselves in the massive body of data and analyze their topics ofinterest. The iConViz system granted them the ability to avoid thisanalysis strategy. They did not get lost in an ocean of network data.On the contrary, they quickly located their guarantee networks ofinterest and performed in-depth analyses of the risk patterns. Theyalso expressed interest in deploying the system in the actual riskmanagement procedure for their bank after the study.We identified several usability issues through the expert userinterviews. For example, in our initial version, the system couldonly visualize the chains in the GNE view when the users clickedon a flower/petal. The chain gave the potential path across whichthe default might spread. The expert users needed to draw the graphstructure on paper and then analyze it. This was inconvenient, so inour updated iConViz system, we automatically labeled the source ofthe outbreak node using a virus-shaped icon ( ) so that the expertusers had no problem identifying the key node. The second issue wasproposed by a risk control manager who said that these days, deeplearning-based risk prediction is popular in bank risk managementsystems. He suggested that the current iConViz system mainly focuson understanding the existing data and patterns. We will incorporatesome deep learning prediction results into our future visual analyticssystem. We were glad to receive this constructive suggestion andplan to bring prediction results into the iConViz system.
Visualization Design:
The experts expressed approval of theoverall design. The light-yellow background inspired a sense of spirituality and encouraged creativity. What they most appreciatedwas the Risk Badge idea. They felt it to be a powerful abstraction ofwhat they wanted to understand. With such intuitive symbols, theydid not need to use the mouse wheel to flip through Excel, page bypage, and get lost in the data (their words). Moreover, they likedthe compact and informative design of the views. They approved ofthe design of the grid layout for tessellating the guarantee networks,describing it as clear, concise, and providing all of the necessaryinformation. The petal and flower design of the CIE (see Fig. 10)attracted their interest, though during their initial attempts, someignored this feature. They praised its combined beauty and func-tionality. However, some users also suggested that it would be moreintuitive to give an overview of all the networks on one screen. Atthat time, they needed to use the mouse to drag the canvas to seeall of the networks. There was no slider or other hint in the GNEview. Similarly, we also identified some visualization issues fromthe feedback. For example, initially, all of the networks in the GNEview had fixed lengths and widths, and it could be difficult to choosethose of most interest. We added a button (expanded view ) sothat the user can now choose a full-screen view. Another issue raisedby a risk control manager was that he believed the CEM designcould be more compact, as it seemed sparse in the middle of thescreen. However, he emphasized that the numbers in the CEM wereimportant, as quantitative descriptions are essential in finance.
System Interactions:
The expert users believed that the interac-tions in the views presented by our system were useful and could helpthem explore and analyze issues important to their work. The inter-actions among multiple coordinated views facilitated the closed-loopanalysis process and iterative level of detailed exploration. Moreover,there were several functional buttons and selection/zooming interac-tions supported in each view. One expert user said that these werepowerful but made the system a bit complex. However, he agreedthat it was difficult to make a proper trade-off between complexfunctions and powerful analytical abilities. For example, withoutthe walkthrough training he would have had no idea how to use thecoordinated views to perform the analysis loop, so he suggested thattraining would be necessary for any future pilot study or deployment.Several risk specialists discovered that due to the sophisticated inter-actions and massive amount of data being visualized, the systemsoperation was not always smooth, and this was its main imperfec-tion. This was because we had to make a trade-off between highlyinteractive software and the amount of information that could berendered on a single page. Future work will keep improving theinteraction performance.
ONCLUSION
In this research, we report our progress on risk management in thenetworked-guarantee loan problem. We believe that this is the firststudy to identify and formalize the contagion chain risk managementproblem for networked loans. This new research avenue providesrefreshing opportunities for both the computing and financial com-munities. In our work, a novel financial metric – the contagioneffect, is formulated to quantify the infectious consequences of guar-antee chains in a network. Based on this metric, we designed andimplemented a series of novel and coordinated views to facilitateanalysis of this financial problem. Experts evaluated the system us-ing real-world financial data. We also conducted experts’ interviewto further collect feedback and improve our system. The resultsdeepen our understanding and ability to assess the potential riskof contagion in complex network structures, hopefully preventingpotential large-scale corporate defaults. A CKNOWLEDGMENTS
This work was supported by National Natural Science Foundationof China (NO.61802278) and foundation of Key Laboratory of Arti-ficial Intelligence, Ministry of Education, China (AI2019004).
EFERENCES [1] R. Aloui, M. S. B. A¨ıssa, and D. K. Nguyen. Global financial crisis,extreme interdependences, and contagion effects: The role of economicstructure?
Journal of Banking & Finance , 35(1):130–141, 2011.[2] L. Anthony (Tony) Cox Jr. What’s wrong with risk matrices?
RiskAnalysis: An International Journal , 28(2):497–512, 2008.[3] D. W. Arner, J. Barberis, and R. P. Buckey. Fintech, regtech, andthe reconceptualization of financial regulation.
Nw. J. Int’l L. & Bus. ,37:371, 2016.[4] D. W. Arner, J. Barberis, and R. P. Buckley. Fintech and regtech ina nutshell, and the future in a sandbox.
Research Foundation Briefs ,3(4):1–20, 2017.[5] Artem Chernyshev. Risk matrix, 2018. [Online; accessed 22-July-2020].[6] Artem Chernyshev. The iso 31000 standard risk management: princi-ples and guidelines, 2020. [Online; accessed 22-July-2020].[7] B. Baesens, T. Van Gestel, S. Viaene, M. Stepanova, J. Suykens,and J. Vanthienen. Benchmarking state-of-the-art classification al-gorithms for credit scoring.
Journal of the operational research society ,54(6):627–635, 2003.[8] D. G. Baur. Financial contagion and the real economy.
Journal ofBanking & Finance , 36(10):2680–2692, 2012.[9] R. Chang, M. Ghoniem, R. Kosara, W. Ribarsky, J. Yang, E. Suma,C. Ziemkiewicz, D. Kern, and A. Sudjianto. Wirevis: Visualization ofcategorical, time-varying data from financial transactions. In
VisualAnalytics Science and Technology, 2007. VAST 2007. IEEE Symposiumon , pp. 155–162. IEEE, 2007.[10] D. Cheng, Y. Tu, Z. Ma, Z. Niu, and L. Zhang. Risk assessment fornetworked-guarantee loans using high-order graph attention represen-tation. In
Proceedings of the 28th International Joint Conference onArtificial Intelligence , pp. 5822–5828. AAAI Press, 2019.[11] D. Cheng, Y. Zhang, F. Yang, Y. Tu, Z. Niu, and L. Zhang. A dynamicdefault prediction framework for networked-guarantee loans. In
Pro-ceedings of the 28th ACM International Conference on Informationand Knowledge Management , pp. 2547–2555, 2019.[12] A. Constantin, T. A. Peltonen, and P. Sarlin. Network linkages topredict bank distress.
Journal of Financial Stability , 35:226–241,2018.[13] M. F. da Silva, ´E. J. d. A. L. Pereira, A. M. da Silva Filho, A. P. N.de Castro, J. G. V. Miranda, and G. F. Zebende. Quantifying the con-tagion effect of the 2008 financial crisis between the g7 countries (bygdp nominal).
Physica A: Statistical Mechanics and its Applications ,453:1–8, 2016.[14] C. di Basilea per la vigilanza bancaria.
International convergenceof capital measurement and capital standards: a revised framework .Bank for International Settlements, 2004.[15] W. Didimo, G. Liotta, F. Montecchiani, and P. Palladino. An advancednetwork visualization system for financial crime detection. In , pp. 203–210, 2011.[16] W. Didimo, G. Liotta, F. Montecchiani, and P. Palladino. An advancednetwork visualization system for financial crime detection. In
Visual-ization Symposium (PacificVis), 2011 IEEE Pacific , pp. 203–210. IEEE,2011.[17] M. Dumas, M. J. McGuffin, and V. L. Lemieux. Financevis. net-avisual survey of financial data visualizations. In
Poster Abstracts ofIEEE Conference on Visualization , vol. 2, 2014.[18] R. Engle. Monitoring Risk with V-Lab. http://vlab.stern.nyu.edu/ . [Online; accessed 17-March-2020].[19] M. D. Flood, V. L. Lemieux, M. Varga, and B. W. Wong. The appli-cation of visual analytics to financial stability monitoring.
Journal ofFinancial Stability , 27:180 – 197, 2016. doi: [20] A. Grover and J. Leskovec. node2vec: Scalable feature learning fornetworks. In
Proceedings of the 22nd ACM SIGKDD internationalconference on Knowledge discovery and data mining , pp. 855–864.ACM, 2016.[21] A. G. Haldane. Managing global finance as a system.
Bank of England,Speech at the Maxwell Fry Annual Global Finance Lecture, Birming-ham University , 29, 2014.[22] R. Heijmans, R. Heuver, C. Levallois, and I. Van Lelyveld. Dynamic visualization of large transaction networks: The daily dutch overnightmoney market.
Social Science Electronic Publishing , 2014.[23] M. L. Huang, J. Liang, and Q. V. Nguyen. A visualization approachfor frauds detection in financial market. In , pp. 197–202. IEEE, 2009.[24] I. ISO. 31010: Risk management–risk assessment techniques.
Event(London). Geneva , 552, 2009.[25] D. H. Jeong, W. Dou, H. Lipford, F. Stukes, R. Chang, and W. Ribarsky.Evaluating the relationship between user interaction and financial visualanalysis. In , pp. 83–90, 2008.[26] J. D. Kirkland, T. E. Senator, J. J. Hayden, T. Dybala, H. G. Goldberg,and P. Shyr. The nasd regulation advanced-detection system (ads).
AIMagazine , 20(1):55, 1999.[27] R. A. Leite, T. Gschwandtner, S. Miksch, S. Kriglstein, M. Pohl,E. Gstrein, and J. Kuntner. Eva: Visual analytics to identify fraudulentevents.
IEEE Transactions on Visualization and Computer Graphics ,24(1):330–339, 2018.[28] F. Louzada, A. Ara, and G. B. Fernandes. Classification methodsapplied to credit scoring: Systematic review and overall comparison.
Surveys in Operations Research and Management Science , 21(2):117–134, 2016.[29] E. G. Mendoza and V. Quadrini. Financial globalization, financialcrises and contagion.
Journal of monetary economics , 57(1):24–39,2010.[30] X. Meng, Y. Tong, X. Liu, Y. Chen, and S. Tan. Netrating: Credit riskevaluation for loan guarantee chain in china. In
Pacific-Asia Workshopon Intelligence and Security Informatics , pp. 99–108. Springer, 2017.[31] Z. Niu, D. Cheng, L. Zhang, and J. Zhang. Visual analytics fornetworked-guarantee loans risk management. In
Pacific VisualizationSymposium (PacificVis), 2018 IEEE , pp. 160–169. IEEE, 2018.[32] Z. Niu, R. Li, J. Wu, Y. Xue, and J. Zhang. regvis.net–a visual bibli-ography of regulatory visualization. arXiv preprint arXiv:2007.03573 ,2020.[33] R. Pastor-Satorras and A. Vespignani. Epidemic spreading in scale-freenetworks.
Physical review letters , 86(14):3200, 2001.[34] S. R¨onnqvist and P. Sarlin. Bank networks from text: interrelations,centrality and determinants.
Quantitative Finance , 15(10):1619–1635,2015.[35] P. Sarlin. Visual Systemic Risk Analytics. http://vis.risklab.fi/ . [Online; accessed 17-March-2020].[36] P. Sarlin. Self-organizing time map: An abstraction of temporal multi-variate patterns.
Neurocomputing , 99:496–508, 2013.[37] P. Sarlin. Macroprudential oversight, risk communication and visual-ization.
Journal of Financial Stability , 27:160–179, 2016.[38] P. Sarlin and D. Marghescu. Visual predictions of currency crises usingself-organizing maps.
Intelligent Systems in Accounting, Finance andManagement , 18(1):15–38, 2011.[39] P. Sarlin and T. A. Peltonen. Mapping the state of financial stability.
Journal of International Financial Markets, Institutions and Money ,26:46–76, 2013.[40] M. Small, C. Tse, and D. M. Walker. Super-spreaders and the rateof transmission of the sars virus.
Physica D: Nonlinear Phenomena ,215(2):146–158, 2006.[41] H. Thomas and Z. Wang. Interpreting the internal ratings-based capitalrequirements in basel ii.
Journal of Banking Regulation , 6(3):274–289,2005.[42] L. van der Maaten and G. Hinton. Visualizing data using t-sne.
Journalof Machine Learning Research , 9:2579–2605, 2008.[43] U. Von Luxburg. A tutorial on spectral clustering.
Statistics andcomputing , 17(4):395–416, 2007.[44] X. Wen, Y. Wei, and D. Huang. Measuring contagion between energymarket and stock market during financial crisis: A copula approach.
Energy economics , 34(5):1435–1446, 2012.[45] J. Xu and H. Chen. Criminal network analysis and visualization.
Communications of The ACM , 48(6):100–107, 2005.[46] X. Yue, X. Shu, X. Zhu, X. Du, Z. Yu, D. Papadopoulos, and S. Liu.Bitextract: Interactive visualization for extracting bitcoin exchange in-telligence.