Ann Oper Res (2016) 247:523–547

DOI 10.1007/s10479-015-1857-x

Credit default swaps and systemic risk

Rama Cont1 · Andreea Minca2

Published online: 18 April 2015

© Springer Science+Business Media New York 2015

Abstract We present a network model for investigating the impact on systemic risk of
central clearing of over the counter (OTC) credit default swaps (CDS). We model contingent
cash flows resulting from CDS and other OTC derivatives by a multi-layered network with a
core-periphery structure, which is flexible enough to reproduce the gross and net exposures as
well as the heterogeneity of market shares of participating institutions. We analyze illiquidity
cascades resulting from liquidity shocks and show that the contagion of illiquidity takes place
along a sub-network constituted by links identified as ’critical receivables’. A key role is
played by the long intermediation chains inherent to the structure of the OTC network, which
may turn into chains of critical receivables. We calibrate our model to data representing
net and gross OTC exposures of large dealer banks and use this model to investigate the
impact of central clearing on network stability. We find that, when interest rate swaps are
cleared, central clearing of credit default swaps through a well-capitalized CCP can reduce
the probability and the magnitude of a systemic illiquidity spiral by reducing the length of
the chains of critical receivables within the financial network. These benefits are reduced,
however, if some large intermediaries are not included as clearing members.

Keywords Systemic risk · Financial networks · Multi-layered network ·

Credit default swaps · Financial crisis · Intermediation chains · Financial stability

This paper is derived from Chapter 5 of Andreea Minca’s PhD thesis (Minca 2011).

B Andreea Minca
[email protected]
Rama Cont
[email protected]
1 Imperial College London & Laboratoire de Probabilités et Modeles Aléatoires, CNRS-Université
Pierre & Marie Curie, Paris, France
2 School of Operations Research and Information Engineering, Cornell University, Ithaca, NY 14850,
USA

123
524 Ann Oper Res (2016) 247:523–547

1 Introduction

The role played by the multi-trillion dollar market for credit default swaps (CDS) and other
over-the-counter (OTC) credit derivatives in the recent financial crisis, especially in the
demise of AIG, has prompted many discussions on the impact of CDS on financial sta-
bility and systemic risk (Cont 2010; Stulz 2010) and led to various regulatory initiatives,
the most well-known being the requirement of mandatory clearing for standardized CDS
contracts, introduced in the Dodd–Frank Act in the US and in the EMIR framework in
Europe.
Given the cost and effort involved in implementing these central clearing mandates, a
valid question is whether they can be expected to achieve their objective, which is to mitigate
systemic risk in the financial system. While recent studies by regulators tend to answer this
question in the affirmative (Macroeconomic Assessment Group on Derivatives (MAGD)
2013), it is of interest to dispose of independent studies which tackle the question in a more
general setting using a transparent methodology.
The impact of central clearing of OTC derivatives has been studied by Duffie and Zhu
(2011), Cont and Kokholm (2014), Heller and Vause (2012), Sidanius and Zikes (2012),
Amini and Minca (2013), Amini et al. (2014) and Duffie et al. (2014). Most of these studies
consider this problem through the angle of the (aggregate) level of counterparty exposures
and the resulting demand in collateral. The central insight in Duffie and Zhu (2011) is that
the impact of central clearing on the level of exposures crucially depends on the tradeoff
between bilateral netting across derivative classes and multi-netting via the clearing house.
The tradeoff is assessed based on the average exposure under the different netting and clearing
arrangements, and is shown in Cont and Kokholm (2014) to depend on the characteristics of
the network, in particular the correlation of exposures across asset classes and the riskiness
of these exposures.
However, these studies focus on the average (or aggregate) level of counterparty expo-
sures and it is not clear what this metric has to say about systemic risk resulting from the
network’s response to a stress scenario. Indeed, as underlined in recent studies on contagion
in financial networks, what determines the magnitude of contagion stemming from a default
in a counterparty network is not only the aggregate level of exposures but above all the way
these exposures are distributed across links in the network, and their relation to the capital
or liquidity buffer held by financial institutions (Watts 2002; Gai and Kapadia 2010; Amini
et al. 2011; Cont et al. 2013).
A natural framework for tackling these questions is a network model, with networks
becoming increasingly used tools for systemic risk analysis, see Amini and Minca (2013)
and the references therein.
Within the theoretical literature on systemic risk in centrally cleared networks, Amini et al.
(2013) examine the benefits of central clearing using a systemic risk measure that encodes
the network structure, and use this framework to derive solutions to the problem of optimal
design of the CCP waterfall. The impact of partial versus full clearing on the value of the
financial network is studied in Amini et al. (2014).
Within the simulation-based literature, generating networks that capture the real-world
features of OTC networks has remained a challenge. Most existing models aggregate all
bilateral exposures for a given pair of counterparties into a single number, the net exposure
attached to the corresponding link. However, unlike other OTC derivatives, exposures due to
CDS may exhibit large variations (‘jumps’) contingent on a credit event—whether a default
or a downgrade—which may generate large margin calls: these ’contingent liquidity shocks’

123
Ann Oper Res (2016) 247:523–547 525

resulting from CDS transactions are not exogenous to the network but triggered by the default
of nodes within the network.
The structure of the OTC network models—in particular in terms of concentration and
connectivity—needs to reflect the actual structure of OTC exposures between financial insti-
tutions. Although some data sets on bilateral OTC notionals have started to be explored by
regulators (see for example the interesting study by Peltonen et al. 2014), these data sets do
not contain any information on collateral levels and thus do not give any precise information
on exposures which are, of course, the key quantity of interest in any network model. Thus,
there is a need for models which can ‘fill the gaps’ in the data and generate realistic exposure
networks for stress testing and simulation purposes.
The network structure of the OTC market reveals a core of large dealer banks that act
as intermediaries among customers (or end-users) (ECB 2009; Peltonen et al. 2014). Such
dealers may choose in many cases to hedge their exposures to end-users and enter offsetting
contracts with other dealers, leading to intermediation chains. The OTC network can be seen
then as the superposition of these intermediation chains. There is a literature on endogenous
formation of trading paths and intermediation chains in markets with a network structure,
see e.g Babus (2012), Manea (2013), Farboodi (2014), Zawadowski (2013) and Glode and
Opp (2013). The presence of such intermediation chains means that if one large end-user or
intermediary defaults, this default may propagate along the chain of intermediaries through
an illiquidity cascade: when some firms in the intermediation chain do not hold enough
liquidity to cover their payables, counterparties for which those receivables are critical in
order to meet their own payment obligations become illiquid themselves.
The liquidity shocks generated by credit events are further amplified by the large concen-
tration of the CDS market on intermediaries. Indeed, in the CDS market, a few protection
sellers concentrate the large majority of transactions (ECB 2009). These protection sellers
will immediately face a liquidity shortage if the spreads of reference entities across a given
sector move up or down at the same time.
Our first contribution in this paper is to introduce a multi-layered network model, where
each layer contains all CDS exposures references on the default of a given reference entity.
This representation, rather than aggregating the mark-to-market values of all CDS into a
single bilateral exposure, allows to model the change in CDS exposures triggered by a credit
event, which is a necessary step in simulating the response of the network to a credit event
in a stress scenario.
There is also network of non-CDS exposures, by which we understand exposures due to
non-CDS derivatives. The CDS and non-CDS exposures are understood as two classes of
derivatives. There is cross-sectional dependence among them, due to the fact that the set of
banks that act as intermediaries is the same for the two classes of derivatives, and are the
most interconnected banks in the network.
There are of course other exposures among banks that lie outside of our model, in particular
exposures that are not derivatives related and are not subject to margining or collateralization.
A relevant strand of literature on networks, starting from the seminal paper (Allen and Gale
2000), see e.g. Upper and Worms (2002), Furfine (1999), Battiston et al. (2009), Amini et al.
(2011), Babus (2007), Glasserman and Young (2013), Rogers and Veraart (2012), investigates
the transmission of shocks and the contagion in financial networks of general exposures. It
would be interesting to consider the effect of CDS, considered as cross-insurance contracts,
on the stability of such general exposure networks and the overall solvency of the financial
system. In this direction, Giglio (2011) uses a market based approach and derives bounds on
joint default probabilities from bond prices and CDS spreads. In this paper, our focus is on the
liquidity of the financial network of OTC derivatives in which the intermediaries are subject to

123
526 Ann Oper Res (2016) 247:523–547

margining following the change in mark-to-market value of the derivatives-related exposures.

The CDS are, similarly as in Duffie (2011), Cont and Kokholm (2014), treated as a highly
volatile asset class, for which the hedging chains turn out to be long in the calibrated model.
We show that our network model can accommodate the empirically observed features
of the CDS market in a realistic way: the underlying network of financial institutions has
a core-periphery structure,1 in which we distinguish dealer bank from other market partici-
pants, and the model is flexible enough to reproduce the heterogeneity in market share and
the asymmetries in exposures implied by observed values of gross vs net CDS notionals. We
illustrate the flexibility of the model by calibrating it to the data on the gross and net CDS
protection sold on the top 1000 names first published by The Depository Trust and Clearing
Corporation (DTCC) in 2010 DTCC (2010), as well as on the market share of the dealers.2
Next, we describe how liquidity shocks—in particular those arising from credit events—
propagate through this networks and how, in some situations, may lead to contagion via
an illiquidity cascade. We show that this contagion is concentrated on the subnetwork of
‘contagious links’ which correspond to ’critical receivables’ and exhibit an indicator of the
network’s resilience to this illiquidity contagion, introduced in Amini et al. (2011), which
allows a simulation-free analysis of the degree of resilience of the network to liquidity
shocks. In particular, we show that this indicator can be used to detect the tipping-point, or
phase transition, which is the critical level of liquidity shock which leads to a regime where
illiquidity contagion becomes large scale.
Finally, we apply these concepts to analyze the impact of central clearing of credit default
swaps on financial stability. Using simulations and analytical insights from large-network
asymptotics, we explore the impact of introducing a CDS clearinghouse on the extent of
distress propagation in the network. Our analysis shows that, in a network where other
major OTC derivatives (primarily IR swaps) are cleared, the addition of a CDS clearing
facility enhances network stability and a significantly larger shock is necessary to trigger
a phase transition. On the other hand, in absence of clearing of the other classes of OTC
derivatives, central clearing of CDS may have negative impact on financial stability. The
presence of a well-capitalized CDS clearing facility is shown to increase the resilience
of the network provided that all significant intermediaries are clearing members: in this
case, the CCP significantly shortens the intermediation chains. However, the impact of the
CCP on network stability is weakened if some major intermediaries are not clearing mem-
bers.
This paper is organized as follows. Section 2 is an overview of OTC derivatives transactions
and market structure. Section 3 presents a model for the short term impact of OTC derivatives
cash flows on the liquidity of financial institutions and introduces the notion of ‘illiquidity
cascade’. In Sect. 3.2, we first propose a weighted random graph model for the OTC non-
CDS payables. Section 3.2.2 introduces a multi-layered network model constructed as a
superposition of network of CDS transactions for different reference entities. In Sect. 4.1
we give a criterion linking the network structure to its resilience to illiquidity cascades in a
stress scenario. Section 4 studies the impact of central clearing on the size of the illiquidity
cascade.

1 Although not in the restricted sense of requiring the core and periphery to be complete subnetworks, which
is not realistic.
2 Defined by DTCC as ‘any user that is, or is an affiliate of a user who is, in the business of making markets
or dealing in credit derivative products’ DTCC (2010).

123
Ann Oper Res (2016) 247:523–547 527

2 Over-the-counter markets

A bilateral over the counter (OTC) transaction is one in which two parties transact directly
with each another, rather than passing through an exchange or clearing party. As a result,
any of the parties bears counterparty risk, i.e. the risk that the other party does not fulfill its
obligations. At the inception date of the contract, say time 0, the two parties agree on some
future cash flows between them. Since one party’s inflow is the other party’s outflow, the
contract has opposite value for the two counterparties. Upon the default of one counterparty,
the contract is terminated and a close-out payment equal to the mark-to-market value of the
remaining cash flows is due. If the mark-to-market value is negative for the surviving party,
then the latter will make the full close-out payment. On the other hand, if the mark-to-market
value is positive for the surviving party, only a fraction of the due close-out payment will be
received, so the surviving party suffers a loss.
Counterparty risk is mitigated in several ways. First, when two counterparties hold a
portfolio of derivatives, these derivatives are usually placed under a netting agreement (called
the ISDA Master Agreement). In this case, upon default, a single terminating payment for
all derivatives in the portfolio is due, determined by the mark-to-market net value of all
derivatives in the portfolio. Second, the majority of the contracts are subject to collateral
agreements: with a certain frequency—mostly daily-, the party with negative mark-to-market
value of the portfolio posts collateral to its counterparty (ISDA 2010).
Consider, for example a set of transactions between two parties a and b, consisting of two
derivatives, one with a positive value of 200$mn for b and the other with positive value of
100$mn for a. The whole portfolio has thus a positive value of 100$mn for b. Assume that a
defaults, and that the recovery rate is zero. Without netting and collateral, b would pay to a
100$mn and a would suffer a loss of 200$mn on the derivative with positive value. If netting
is applied, a single terminating payment of 100$mn is due by a, and since a defaults and has
zero recovery rate, this represents the loss of b. If a had previously posted collateral 50$mn
to b, then b seizes this collateral and its loss will be the remaining 50$mn.
As explained in the ISDA Credit Support Documents ISDA (2010) determining the amount
of collateral to be posted:“(i) the [Collateral Taker]’s Exposure plus (ii) the aggregate of all
Independent Amounts applicable to the [Collateral Provider], if any, minus (iii) the aggre-
gate of all Independent Amounts applicable to the Collateral Taker, if any, minus (iv) the
[Collateral Provider]’s Threshold. The term Exposure is defined in a technical manner that in
common market usage essentially means the netted mid-market mark-to-market (MtM) value
of the transactions that are subject to the relevant ISDA Master Agreement. If a Threshold is
applicable to a party, the effect of the Credit Support Amount calculation is that Collateral
is only required to be posted to the extent that the other party’s Exposure (as adjusted by
any Independent Amounts) exceeds that Threshold. An Independent Amount applicable to
a party serves to increase the amount of collateral that is to be posted by that party. This is
to provide a “cushion” of additional collateral to protect against certain risks, including the
possible increase in Exposure that may occur between valuations of collateral (or between
valuation and posting) due to the volatility of mark-to-market values of the transactions under
the ISDA Master Agreement.” Although not a technical term, “variation margin” is used to
refer to the portion of required collateral that relates to the MtM of covered transactions (i.e.
the ”Exposure”).
When an OTC transaction is cleared through a central counterparty, an initial margin needs
to be posted by both parties.

123
528 Ann Oper Res (2016) 247:523–547

OTC derivatives: notional, mark-to-market and daily variations. Table 1 gives an overview
of the notional and gross market values of different types of OTC derivatives. We observe
that interest rate and foreign exchange instruments account for 85 % of the total notional size
of the OTC market, while credit default swaps account for around 5 %. On the other hand,
when looking at the daily variation of the mark-to-market values of these instruments—that
we approximate by the daily variation of spreads and respectively the swap fixed rate—the
picture changes, as shown in Fig. 1 and 2. Turbulent times like the weeks following the failure
of Lehman Brothers on the 15th Sept 2008, showed that the absolute value of the average
5-year CDS spread variation for the high-grade names comprising the CDX index can be
several times larger than the absolute value of the variation of the swap fixed rate. Moreover,
spreads of institutions belonging to the same sector as a failed institution exhibit particularly
large jumps due to cross-sector correlation. Such is the case of General Electric, which is a
component of the CDX index within the sector ‘financials’, whose 5-year spread had a 70 %
jump following the default of Lehman Brothers. Institutions closer in their activity to that of
the failed bank, like other dealer banks, suffered even larger jumps in spreads: the cost of
protection for other dealers doubled over a few trading days in the aftermath of Lehman’s
default (Brunnermeier 2009).
Although non-credit OTC derivatives have a mark-to-market value one order of magnitude
above credit derivatives, CDS may present much larger variations in the mark-to-market
values and as a result for large dealers the variation margins for positions on CDS and
non-CDS derivatives are comparable. This is also documented in the sequel paper (Cont and
Kokholm 2014), who find that the risk per dollar notional of a CDS exposure is typically
three or more times higher than the risk per dollar notional for an interest rate swap contract
with a similar maturity.

Concentration in OTC markets Another important empirical observation is the concentration

of the OTC market. Table 2 extracted from OCC (2010) shows the notional positions of the
top five US dealers on different types of OTC derivatives: forwards, swaps, options and credit.
According to this data, the top five US dealers alone hold an OTC derivative global
market share of 46 %. For credit derivatives in particular, their global market share is 71 %.
This demonstrates that the credit derivatives market is significantly more concentrated than
the other OTC markets. For CDS, a comprehensive analysis of concentration is made in ECB
(2009).
According to DTCC data, the total notional amounts of outstanding CDSs sold by dealers
worldwide, represent over 80 % of total protection sold worldwide and a similar percentage
is represented by protection bought by dealers. Also, the interdealer network accounts for
∼75 % of total CDS notional. Remark that no precise number of dealers is given in DTCC
data, they are defined as “any user that is, or is an affiliate of a user who is, in the business
of making markets or dealing in credit derivative products”.
When considering the distribution of the total notional among reference entities, we
observe an important concentration on the top underlying names, as shown in Fig. 3. Peltonen
et al. (2014) arrive at a similar conclusion regarding the concentration of the CDS market
using a more detailed dataset from a different period.
All these observations plead for distinguishing CDS and credit derivatives from non-
credit OTC derivatives (primarily interest rate swaps and other interest rate derivatives) and
accounting for their differences in terms of size, concentration and riskiness in the model.
We now proceed to construct a stylized network model for an OTC market, capable of
accommodating these empirical observations.

123
 Table 1 Amounts outstanding of over-the-counter (OTC) derivatives by risk category and instrument (in billions of US dollars)

Risk category/instrument Notional amounts outstanding Gross market vaues

Jun 2008 Dec 2008 Jun 2009 Dec 2009 Jun 2010 Jun 2008 Dec 2008 Jun 2009 Dec 2009 Jun 2010

Total contracts 672,558 598,147 594,495 603,900 582,655 20,340 35,281 25,314 21,542 24,673
Foreign exchange contracts 62,983 50,042 48,732 49,181 53,125 2262 4084 2470 2070 2524
Forwards and forex swaps 31,966 24,494 23,105 23,129 25,625 802 1830 870 683 925
Currency swaps 16,307 14,941 15,072 16,509 16,347 1071 1633 1211 1043 1187
Options 14,710 10,608 10,555 9543 11,153 388 621 389 344 411
Ann Oper Res (2016) 247:523–547

Interest rate contracts 458,304 432,657 437,228 449,875 451,831 9263 20,087 15,478 14,020 17,533
Forward rate agreements 39,370 41,561 46,812 51,779 56,242 88 165 130 80 81
Interest rate swaps 356,772 341,128 341,903 349,288 347,508 8056 18,158 13,934 12,576 15,951
Options 62,162 49,968 48,513 48,808 48,081 1120 1764 1414 1364 1501
Equity-linked contracts 10,177 6471 6584 5937 6260 1146 1112 879 708 706
Forwards and swaps 2657 1627 1678 1652 1754 283 335 225 176 189
Options 7521 4844 4906 4285 4506 863 777 654 532 518
Commodity contracts 13,229 4427 3619 2944 2852 2213 955 682 545 457
Gold 649 395 425 423 417 72 65 43 48 44
Other commodities 12,580 4032 3194 2521 2434 2141 890 638 497 413
Forwards and swaps 7561 2471 1715 1675 1551
Options 5019 1561 1479 846 883
Credit default swaps 57,403 41,883 36,046 32,693 30,261 3192 5116 2987 1801 1666
Single-name instruments 33,412 25,740 24,112 21,917 18,379 1901 3263 1953 1243 993
Multi-name instruments 23,991 16,143 11,934 10,776 11,882 1291 1854 1034 559 673
of which index products … … … … 7614
Unallocated 70,463 62,667 62,285 63,270 38,327 2264 3927 2817 2398 1788
Memorandum item:
Gross credit exposure 3859 5005 3744 3521 3578

Source: BIS Quarterly Review, December 2010

123
Bold Number Represent: Amounts outstanding by risk category, aggregated over instruments
529
530 Ann Oper Res (2016) 247:523–547

Average 5YR spread relative variation over 125 CDX names Spread return − Gen Elec Cap Corp
0.15 0.7
Average 5YR spread relative variation over 125 CDX nam

0.6

0.1 0.5

0.4

Spread return
0.05 0.3

0.2

0 0.1

−0.05 −0.1

−0.2

−0.1 −0.3
05/08 15/09 01/04 13/05 24/06 05/08 16/09 28/10 09/12 20/01
Date Date

(a) (b)
Fig. 1 Spread variations for CDS. a Average 5-year CDS spread variations for the CDX.NA.IG index. b GE
5-year CDS spread variations

Return of the rate paid by fixed−rate payer on an interest rate swap with five year maturity
0.06

0.04

0.02

−0.02

−0.04

−0.06

−0.08
13/07 05/08 24/08 15/09 05/10
Date

(a) (b)
Fig. 2 Daily variations in OTC derivatives values: IR Swaps vs CDS. a Jump in cost of protection. b Variations
of a 5 year USD swap rate

Table 2 Notional amount of derivative contracts top five holding companies in OTC derivatives December
31, 2010, $ millions
Rank Holding Assets Total OTC Forwards Swaps Options Credit
company derivatives

1 JPM 2,117,605 75,510,099 11,806,979 49,331,627 8,899,046 5,472,447

2 BAC 2,268,347 63,983,932 10,287,375 43,481,989 5,847,866 4,366,702
3 C 1,913,902 45,151,220 6,895,160 28,638,854 7,071,397 2,545,809
4 GS 911,330 43,998,391 3,805,327 27,391,560 8,568,358 4,233,146
5 MS 807,698 41,124,050 5,458,883 27,161,921 3,854,976 4,648,270

Source: OCC’s quarterly report on bank trading and derivatives activities second quarter 2010

3 A network model for contingent OTC exposures

At any time, a snapshot of the OTC market reveals a set of institutions (“banks”) that are
interlinked by their mutual claims. We denote by [n] := {1, . . . , n} the set of banks.
We consider a two period model t = 0, 1, 2. At time 0 banks enter the OTC contracts.
At time 1 the payables are revealed. The payables are random variables that depend on the

123
Ann Oper Res (2016) 247:523–547 531

10 Gross notional of protection

x 10
10

9 Distribution of the Outstanding CDS Notional

3
10
8

Probability Distribution
2
10
Notional

3 1
10

0 0
10
20 40 60 80 100 120 140 160 180 200 0 0.5 1 1.5 2 2.5
Rank of underlying name Outstanding CDS Notional 11
x 10

Fig. 3 Concentration on names: 47 % of the total CDS Notional is written of the top five names and 76 % on
the top ten names

contract specification and the realization of random shocks at time 1. We consider the period
t = 1, 2 as the period during which the cash flows occur. We denote by small letters the
quantities known at time 0 and by capital letters the quantities revealed at time 1. There is
no new information arriving after time 1, so the cash flows occurring during the period 1, 2
are deterministic functions of the quantities known at time 1.

3.1 Payables and receivables

We denote by X : [n]×[n] → R the matrix of receivables at time 1, assumed anti-symmetric.

For i, j ∈ [n], X i+j represents the receivable of i from j and X i−j represents the payable of i to
j and we have X i j = −X ji . Only one of these payables is due, which means that we account
for bilateral netting. Denote L i the liquidity buffer of bank i at time 1. This is constituted
by cash in main currencies or other highly liquid instruments like high-grade government
securities. Most importantly, it includes the total cash that can be raised on the market by the
bank for example using repurchase agreements. This quantity can be though of as the total
capacity to raise liquidity of the institution. Likewise, an initial (exogenous) shock in this
quantity includes large payables outside our network of payables, or a sharp decrease in the
capacity to raise liquidity, such as a run.
As argued in Sect. 2, in case of CDS, variations of their MtM value are more realisti-
cally defined as percentages of the outstanding notional. Therefore, we model the matrix of
receivables X using

1. The network of non-CDS contracts. This network’s features are realized at time 0.
2. r networks of outstanding notional of CDS, where [r ] denotes the set of reference entities.
These networks are realized at time 0.
3. The variations of mark-to-market values of OTC derivatives. These variations are realized
at time 1.

In the sequel we approximate the mark-to-market value of a CDS contract by the contract
notional times the spread of the reference entity, i.e., for reference entity k ∈ [r ], m k Sk gives
the mark-to-market value (from the point of view of the buyer of protection), before netting,
of the CDS contracts on k.

123
532 Ann Oper Res (2016) 247:523–547

Let e denote the gross exposures on non-CDS contracts and (m k )k∈[r ] the networks of
CDS outstanding notionals for each reference entity. Denote by (ΔSk )k∈[r ] and ΔMt M the
spread variations and respectively the variation of the MtM values of non-CDS derivatives.
The matrix of receivables at t = 1 is given by
   t +
 
(k) (k)
X := e ΔMt M + m ΔSk − e ΔMt M + m ΔSk , (1)
k k
t
where the operator denotes matrix transposition.
A bank is said to be fundamentally liquid if, assuming all other banks pay their payables
in full, it can meet its payables in full
  
Li + X i+j − X i−j = L i + X i j ≥ 0. (2)
j j j

It is said to be fundamentally illiquid otherwise.3

If the bank is fundamentally illiquid, it will default on its payment obligations and an
illiquidity cascade may emerge.

Assumption 1 Upon default, we consider that in the short run recovery rates are 0.

The second assumption is standard in the literature on cascading defaults in financial networks
(Amini et al. 2011).
A bank becomes illiquid due to contagion if its liquidity buffer is such that it depends
on its receivables from the other banks in order to meet its own payment obligations. Such a
situation can arise for highly ‘leveraged’ banks, i.e. banks that are well hedged and holding
little liquidity.
Consider the example of an institution A that buys protection from an institution B on a
reference entity k for a total notional m (k) . Institution B will hedge its exposure to the default
of the reference entity by buying protection from an institution C on the same notional amount
as it sold protection on to A, and so on, until reaching an institution D which is a net seller of
protection. All the intermediary institutions are well hedged and have little incentive to keep a
high liquidity buffer, especially if counterparties have high ratings (i.e. are deemed as having
small probability of default). On the other hand, payables may be particularly large following
jumps in the spread of the reference entity. If the end net seller of protection defaults, then
there is potential of domino effects along the chain of intermediaries.

Definition 1 (Illiquidity cascade) Starting from the liquidity buffer s at time 1, {L i }1≤i≤n ,
and the matrix of receivables {X i j }1≤i< j≤n , the illiquidity cascade can be determined as
follows
– The set D0 of initially illiquid banks is defined as
⎧ ⎫
⎨  ⎬
D0 = i ∈ [n] | L i + Xi j < 0
⎩ ⎭
j

3 Such a situation may arise from large jumps in mark-to-market values of net OTC derivatives payables,
stemming for example from large correlated jumps in the spreads of reference entities of CDS. Institutions
with large unilateral positions are particularly prone to this kind of illiquidity. Nonetheless, our model allows
for a bank to become fundamentally illiquid via an exogenous shock like a run by short term creditors.

123
Ann Oper Res (2016) 247:523–547 533

 
– For r ≥ 1, set Dr = {i ∈ [n] | L i + j X i j − j∈Dr−1 X i+j < 0}, the set of institutions
which become illiquid after r iterations.
We obtain an increasing sequence of defaults D0 ⊂ D1 , . . . ⊂ Dn−1 . The set Dn−1 represents
the final set of illiquid banks.
First, note that the liquidity buffer L and the matrix X of receivables tend to be negatively
correlated. Again, let us take the example of a CDS protection seller. In a first approximation,
the jump in the negative position of the seller is given by the jump in the spread of the
reference entity. As pointed out by Cont and Kan (2011), the spread variation exhibits positive
autocorrelation, volatility clustering and heteroscedasticity. Their empirical distribution of
the spread variation is heavy tailed. Moreover, empirical data shows that spread variations
are correlated across certain classes of reference entities. It follows that a large value for
derivatives payables is very likely to occur after a period of increases in spreads, which had
the effect of fragilizing the liquidity buffer of the seller. This is a typical example of wrong-
way risk, exacerbated if this seller concentrates positions on several correlated reference
entities.
Second, one should not ignore that large downward jumps in market values may also cause
contagion. In case banks use rehypothecation,4 there is no guarantee that a party with negative
exposure will receive back its excess collateral in case the (absolute value) of the exposure
diminishes. This may cause the party to become illiquid if it is part of an intermediation
chain. Whereas the danger of rehypothecation has been pointed out in relation to this kind
of over-collateralization (Singh and Aitken 2009), one should keep in mind that the risk of
over-collateralization incurred by one party is the dual of the risk of under-collateralization
incurred by the other party.

3.2 A network model for OTC exposures

We now detail the model for exposure matrices e and (m (k) )k∈[r ] . The construction of the OTC
network with vertex set [n] = {1, . . . , n} is centered around the fact that a small subset of
these institutions, among the largest and most interconnected, act primarily as intermediaries
between other institutions, so that generally they are counterparties to off-setting contracts.
We refer to these institutions as dealers or intermediaries.
The model is based on the following parameters:
– The aggregate gross value of the OTC market (source: BIS);
– The non-credit derivatives market share for the top 10 dealers (source: OCC);
– The credit derivatives market share for the top ten dealers (source: DTCC);
– The gross CDS protection bought on the top 1000 reference entities (source: DTCC);
– The net CDS protection bought on the top 1000 reference entities (source: DTCC).
This data is more informative for CDS derivatives than for non-CDS derivatives, for which
only the aggregate gross value is available. Arguably, this can lead to multiple network
structures for non-CDS exposures that are consistent to the data. A large variety of network
structures have been used in the literature on financial contagion and the results turn out to
depend on the network structure.
Precisely for this reason, in Sect. 3.2.1 we introduce a network model whose properties are
chosen to match empirically observed properties which are shared by real financial networks.
Key among these properties are heterogeneity of the connectivity, and the exposures across

4 Rehypothecation refers to an institution posting as collateral to its creditors the collateral that it received
from its debtors.

123
534 Ann Oper Res (2016) 247:523–547

nodes. Models based on heterogenous random graphs such as Cont et al. (2013) were cited
in Yellen (2013) as an example of realistic settings.
The model we propose in this paper for non CDS exposures is a parsimoniously parame-
terized random network model, which allows for heterogeneity in connectivity and exposures
across nodes. There are only five parameters, governing the distribution of degrees and the
relative sizes of exposures. Its empirical features, in terms of heterogeneity, degree distrib-
ution and disassortativity, quantitatively match empirical studies on interbank networks in
Cont et al. (2013). Other empirical studies, e.g., Boss et al. (2004), report parameters in the
same range.
The data on CDS derivatives is more informative, in particular for each reference entity
there is a high ratio of gross to net protection bought on that reference entity. Moreover,
the market share of each dealer is known. Recently, Glode and Opp (2013) rationalize the
existence of intermediation chains that stand between buyers and sellers of assets in OTC
markets. Hedging chains can be though of as the signature of buying and selling the asset,
in our case CDS protection for a standardized notional on a reference entity, through a chain
of intermediaries. Since the gross amounts are one order of magnitude above net amounts,
the resulting hedging chains are necessarily large. This considerably decreases the number
of network structures that are compatible with the data. In Sect. 3.2.2, we construct such a
network structure compatible with a data.

3.2.1 A matching model for non-CDS exposures

We detail the construction of the network e of non-CDS exposures using a weighted version
of Blanchard’s random graph model (Blanchard et al. 2003).
We assume that we are given a sequence of out degrees (di+ )i∈[n] . Our goal is to construct a
network e such that for any i ∈ [n], di+ represents the sum of the elements on the column i
(this has the financial interpretation of the number of banks to which bank i is exposed)

di+ = ei j .
j

We assume that the empirical distribution of the out-degree approximates a power law
with tail coefficient γ + (real financial networks have been shown to exhibit these degree
distributions, Cont et al. 2013):
Condition 2
+ +1
lim #{i | di+ = j} ∼ j γ . (3)
n→∞
In Blanchard et al. (2003), a random graph is chosen among all graphs that have these out-
degrees using a sequential matching: an arbitrary out-going edge will be assigned to a node
with probability proportional to the power α of the node’s out-degree. For α > 0, one obtains
positive correlation between in and out-degrees. The main theorem in Blanchard et al. (2003)
states that the marginal distribution of the resulting out-degree approximates a distribution
+
with a Pareto tail with exponent γ − = γα , provided 1 ≤ α < γ + :
− +1
lim #{i | di− = j} ∼ j γ .
n→∞

We now extend this model to account for the heterogeneity of weights. The intuition behind
our construction can be given by rephrasing the Pareto principle: 20 % of the links carry 80 %
of the mark-to-market value of non CDS derivatives. Therefore, we will distinguish between
two types of links.

123
Ann Oper Res (2016) 247:523–547 535

– Links of type A represent a percentage a of the total number of links and carry a percentage
a of the total mark-to-market value.
– All other links are said to be of type B.
We can now define the random graph model that we use to model the non-CDS mark-to-
market values which is a weighted version of Blanchard’s random graph model (Blanchard
et al. 2003).
Definition 2 (Heterogeneous exposure network) Let (di+ )i∈[n] be a sequence of out-degrees,
assumed to verify Condition (2). For every node i, its di+ out-going links are partitioned into
di+,A links of type A and di+,B links of type B:

di+ = di+,A + di+,B . (4)

 
We denote m A := i=1 n
di+,A and by m B := ii=1 di+,B their respective numbers. Denote
A mB
F A : Rm+ → [0, 1] and F : R+ → [0, 1] the joint probability distributions functions
B

for the weights carried by links of type A and B respectively. The probability distribution
functions F A and F B are assumed to be invariant under permutation of their arguments
(exchangeability).
The graph is generated then as follows:
– Generate the weighted subgraph of links of type A by Blanchard’s algorithm with degree
sequence (di+,A )i∈[n] and parameter α > 0.
– Draw m A random variables from the joint distribution F A . Assign these exchangeable
variables in arbitrary order to the links of type A.
– Proceed similarly for the links of type B.
The tail coefficient γ + is calibrated to the dealers’ market share in OTC derivatives in
Table 2. We let α = 1. The topology of the non-CDS exposure network is governed by the
following parameters: γ + = 2, α = 1, a = 5 %.
Denoting by T the total gross mark-to-market value of non-CDS derivatives, the exposures
are governed by the cumulative distribution functions F A and F B . We generate the weights
of type A as the differences of the order statistics of m A i.i.d. random variables, uniformly
distributed in the interval [0, a · T ]. We take F B as the distribution of m B i.i.d. random
variables drawn from the Pareto distribution with tail coefficient γ L . The exposure sequence
is governed by the following parameters: T = 3.5$tn, a = 80 %, γ L = 1.1.

3.2.2 A multi-layered network model for CDS exposures

We denote by [r ] the set of reference entities.5 We denote by [I ] ⊆ [n] the set of interme-
diaries, which are assumed to be the nodes with the highest connectivity in the network of
exposures e.
For each k ∈ [r ], we denote by gr oss(k) and respectively by net (k) the aggregate gross
and net notional written on the name k. The gross notional represents the sum over all
contracts referencing node k. The net notional represents the sum over all nodes of the net
protection sold on name k. Note that the latter sum is equal to the sum over all nodes of
the total protection bought on name k. The ratio of the net and gross notional on a name k
provides a measure of the average length of intermediation chains of contracts with k as a
reference entity. Each intermediation chain links two customers with opposite positions and

5 These may be sovereigns, corporates, banks, etc.

123
536 Ann Oper Res (2016) 247:523–547

w w ... w w
s il−1 i1 b

Fig. 4 Hedging chain for one CDS contract on reference entity k. The length of the chain is l =  grnet
oss(k)
(k) 

passes through a number of dealer banks, with zero net position (see Fig. 4). We will refer
to contracts between dealers as intermediary contracts and to contracts between a customer
and a dealer as end contracts.
The following algorithm generates an OTC network calibrated to a a given set of
– net and gross CDS notionals (gr oss(k), net (k))k∈[r ]
– market shares
for each dealer.

Algorithm 3 (Network of notional) Let w the standard notional of a single CDS contract.
For every reference entity k ∈ [r ]:
1. Let  grnet
oss(k)
(k)  − 1 the number of dealers between two end contracts on reference entity
n;
2. Let  netw(k)  the number of end contracts in which a customer is a buyer of protection;
(k)
3. For each pair of nodes i and j, set m i j ← 0 the number of contracts on reference entity
k in which i sells protection to j;
4. For each contract 1, . . . ,  netw(k) , iterate:
– Choose a sequence (b, i 1 , . . . , i  gross(k) −1 , s) consisting of the end-user acting as
net (k)
buyer of protection on name k, the intermediary dealer banks and the end-user
acting as seller of protection on name k.
The intermediary i 1 is chosen on [I ] from the probability distribution p conditioned
on being different from the reference entity k and any subsequent intermediary is
chosen on [I ] from probability distribution p conditioned on being different from the
reference entity k and the previous intermediary in the sequence.
The customer b is chosen over the set [n] \ [I ] and the end-user s is chosen uniformly
over [n] \ [I ], conditioned on being different from b. The resulting hedging chain is
shown in Fig. 4.
(k) (k)
– Set m i j ← m i j + 1, for any two consecutive i and j in the sequence S.

For each dealer bank i ∈ [I ] the number of contracts in which i acts as protection seller is
 (k)
given by j∈[n] m i j .

Remark 1 An alternative construction could be to consider that each market participant

hedges a proportion ρ < 1 of its exposures. In this case, an initial transaction of notional w
leads to a chain of secondary transactions with notionals ρw, ρ 2 w, . . . , ρ l−1 w. In this case
the total gross notional along the chain is 2w(1 − ρ l )/(1 − ρ) and the net notional is 2w, so
the length of the chain l can be calibrated to the ratio of net to gross notional, which is equal
to (1 − ρ l )/(1 − ρ).6

6 We thank one of the referees for pointing out this alternative construction.

123
Ann Oper Res (2016) 247:523–547 537

In−degree Distribution Out−degree Distribution

3 3
10 10

P(Out−degree ∈ [k−1, k))

P(In−degree ∈ [k−1, k))

2 2
10 10

1 1
10 10

0 0
10 10
0 1 2 3 0 1 2 3
10 10 10 10 10 10 10 10
k k

(a) (b)
Non CDS Exposure Size Distribution 12 Concentration of the non−CDS OTC Market
x 10
15
Top 10 dealers Remaining institutions
P(Non CDS Exposure Size > k)

−2
10
10
US Dollar

−3
10
5

−4
10 0
9 10 11 1 2
10 10 10
k

(c) (d)
Fig. 5 Features of the OTC (non-CDS) exposure network. a The distribution of out-degree has a Pareto
tail with exponent 2. b The distribution of the in-degree has a Pareto tail with exponent 2. c Cross-sectional
distribution of exposures. d Concentration on the top dealers

The alternative construction and our construction would be equivalent if we allowed the
end users in the hedging chains to belong to the set I . In this case a dealer bank would
not always act as an intermediary, and the proportion ρ of its unhedged positions would
correspond to the frequency with which it appears as an end user in the hedging chains.
Note that in our construction the CDS and non-CDS networks are linked since the set of
intermediaries I that determines the CDS exposures is chosen to be the set of nodes with the
highest connectivity of non-CDS exposures.

3.3 Calibration results for a sample OTC network

The features of a sample of the random network e of non-CDS exposures are shown in Fig. 5.
Based on this sample, a CDS network generated from the model of Sect. 3.2.2 has, by
construction, the same features as given by the empirical data: the top five CDS dealers sell
protection totaling 65 % of the CDS outstanding notional, the top ten sell protection totaling
87 % of the outstanding notional. Also, as shown by Fig. 6, the subnet of CDS contracts sold
by the top ten dealers to other top ten dealers is a complete network. This network represents
in our calibrated sample 76 % of the total outstanding notional.

123
538 Ann Oper Res (2016) 247:523–547

11
x 10 CDS Outstanding Net Notional
3.5
Net protection bought
Net protection sold
3

2.5
Net notional

1.5

0.5

0
5 10 15 20 25 30
Dealer
(a) (b)
Fig. 6 Concentration in the CDS market. a The dealer structure of the CDS market: first ten largest dealers
sell/buy 87/88 % of the total CDS Notional. b Dealer to dealer network: complete network representing 76 %
in terms of outstanding CDS notional

Gross CDS Protection Net CDS Protection

14 12
10 10

13
10
11
10

12
10
Data

Data

10
10
11
10

9
10
10
10

9 8
10 10
9 10 11 12 13 14 8 9 10 11 12
10 10 10 10 10 10 10 10 10 10 10
Calibrated Calibrated

Fig. 7 Calibration to DTCC Data DTCC (2010)

The results of the calibration to DTCC data on the net and gross notional sold on the
top reference entities are show in Fig. 7. Since the top reference entities are not necessarily
financial institutions, the notionals of protection sold on nonfinancial names—the data also
includes sovereigns—have been aggregated (in Fig. 7 the point with the highest notional
corresponds aggregates the non-financial reference entities).

4 Does central clearing of CDS reduce systemic risk?

The purpose of this section is to analyze the impact of central clearing on an OTC network.
This network is constructed as a sample of the random network introduced in the previous
section. When the complete network is observed at time 0, we may use the resilience indicator
to assess the transmission of distress under a stress scenario.

123
Ann Oper Res (2016) 247:523–547 539

4.1 Resilience to illiquidity cascades in stress scenarios

We now explain the methodology for stress-testing the OTC network: we consider a stress
scenario ω defined by
– a vector of CDS spread variations across reference entities ΔS = (ΔSk (ω))k∈[r ] ,
– the variation of the mark-to-market value of the non-CDS derivatives ΔMt M =
ΔMt M(ω),
– liquidity reserves of financial institutions L = (L i (ω))i∈[n] .
The matrix of receivables in the stress scenario ω is then given by (1). The magnitude of the
illiquidity cascade can be investigated on the network X either directly using simulation or
using the asymptotic analysis given in Amini et al. (2011), as we now explain.
As in Amini et al. (2011), we note that contagion takes place primarily along a sub-network
defined by ’contagious links’ which, in this case, can be understood as ’critical receivables’:

Definition 3 (Critical receivables) Let i be a node  that does not become fundamentally
illiquid under the stress scenario ω, i.e. L i (ω) + j X i j (ω) > 0. We say that, under the
stress scenario ω, there is a critical cash flow between k and i if

+
X ik > Li + Xi j , (5)
j

i.e. node i cannot meet its payables if node k is illiquid. We write in this case
c
k → i.
c
We now define w(i) = #{k ∈ [n] | k → i} the number of ‘contagious links’ of i, i.e. the
number of liquidity inflows the absence of which would render i illiquid. If w(i) = 0 this
means the node i has sufficient liquidity to cover the scenario where any single inflow is
cancelled: this should be the case in a ’normal’ regime, and this is indeed the focus of current
regulatory requirements on liquidity ratios. Nevertheless, in a stress scenario we can have
w(i, ω) > 0 i.e. contagious links may appear due to liquidity shocks.
Moreover, we denote by
Ci+ = #{ j ∈ [n] s.t. X i j > 0},
the in-degree of a node i, given by the number of its in-flows, while its out-degree of a node
i is the number of its out-flows
Ci− = #{ j ∈ [n] s.t. X i j < 0}.
Clearly we have that the total number of linkages in the system is given by
 
Ci+ = Ci− .
i∈[n] i∈[n]

Following Amini et al. (2012), we consider the following indicator of the network’s
resilience to contagion:

Definition 4 (Resilience to contagion in a stress scenario) The resilience indicator of the

CDS network in stress scenario ω is defined as
1 
ν̂(ω) := 1 −  − Ci− w(i, ω). (6)
j∈[n] C j i∈[n]

123
540 Ann Oper Res (2016) 247:523–547

A key result shown in Amini et al. (2011) is that a negative value of ν̂(ω) indicates that
there exists a large subnetwork of banks strongly interlinked (there is a path from any node
to another node) by critical receivables. As such, any default of a node in this subnetwork
triggers the illiquidity of the whole subnetwork.

4.2 The stress scenario

Starting from the OTC networks, the matrix of receivables X is determined in our example
according to the following stress scenario:
– The gross market values of credit default swaps having as a reference entity one of the
dealers have an absolute jump equal to 15 % of the notional;
– The gross market values of credit default swaps having as a reference entities other financial
institution aside dealers, has an absolute jump equal to 10 % of the notional;
– The gross market value of credit default swaps on other reference entities has an absolute
jump equal to 5 % of the notional;
– The gross market value of the other OTC derivatives confounded decreases by 5 %.
The receivables we consider result from margin calls which are linked to changes in market
values of outstanding derivatives positions. As opposed to changes in asset values, which
are simply accounting losses but may not result in any cash flows, such margin calls need
to be settled in cash, so they correspond to draws on liquidity buffers. The magnitude of the
shocks in stress scenarios are calibrated as is the practice in derivatives clearinghouses when
computing margin levels, by computing the standard deviation of the underlying contract
over a horizon which is the margin horizon (typically 2–5 days). Here we have chosen
a representative contract in each asset class (5 year USD swap, 5 year CDS index) and
computed the shock size from the observed standard deviation of the contract value during
the period following Lehman’s default, where the daily variations were plotted in Figs. 1 and
2 (see Cont and Kan 2011; Cont and Kokholm 2014 for similar calculations and examples).
For any bank i, the liquidity buffer at time 0 is assumed to be the minimal liquidity buffer
such that no bank has any critical receivables in the event of a jump equal to a percentage
γ = 5 % of the MtM value of non-CDS derivatives and respectively of the CDS notionals:
⎛ ⎞+
 
i = γ · ⎝ e ji − ei j + max(ei j − e ji )+ ⎠
j
j j
⎛ ⎞  +
 (k)
 (k)
 (k)
 (k)
+γ ·⎝ m ji − mi j ⎠ + γ · max mi j − m ji , (7)
j
j k j k k k

and the liquidity buffer at time 1 is

L i = i (1 − ),
where is an exogenous liquidity shock assumed constant over all banks between time 0
and time 1.
Clearly, since our stress scenario is more severe, critical receivables will appear in the
system. We investigate the role of central clearing in mitigating the transmission of distress
via these critical receivables, under several clearing configurations. Concerning the CDS, we
consider three case studies:
1. The case where CDS are not centrally cleared;

123
Ann Oper Res (2016) 247:523–547 541

2. The case where CDS are centrally cleared with a set of 20 dealers;
3. The case where CDS are centrally cleared but only a reduced set of ten dealers have
access to the clearing house.
Concerning the other derivatives, in their majority IR derivatives, we compare the following
cases:
1. The case without central clearing;
2. The case of a dedicated clearing house;
3. The case of joint clearing with CDS.
Note that the definition of the liquidity buffer given by (7) is independent of any netting
across derivative classes. The reason for this is that different cases of clearing strongly affect
the netting opportunities, whereas we need precisely a definition of the liquidity that would
serve for a common base for comparing these cases.
On the other hand, the liquidity buffer of the clearing houses is defined by taking into
account the possibilities of netting across derivative classes. Also, more precaution is taken:
not only the clearing house is not allowed to have critical receivables (which implies that
the CCP withstands the default of any of its members), but it must withstand the default of
the two members to which it has the largest exposure. Note also that the cash inflows of any
clearing house equal its outflows. So, for a clearing house c, the liquidity buffer at t = 0 is
given by
 
c = 2 · γ · max(e(c, j) − e( j, c) + m (k) (i, c) − m (k) (c, i))+ , (8)
j
k k

and we assume no exogenous liquidity shock for the clearing houses between time 0 and
time 1
L c = c .

4.3 Simulation results

Using the stress scenarios considered in the previous section, we now assess the impact of
central clearing on the sample network presented in the previous section. We will compute the
resilience indicator given by Definition 4. Figure 8 shows the size of the illiquidity cascade
in the stress scenario as a function of a varying exogenous liquidity shock . Recall that for
every bank i, the liquidity shock is defined as a fixed percentage of the liquidity buffer i
given in (7) and this percentage is constant over all banks. In all cases, we relate the size of
the illiquidity cascade to the resilience indicator.
These results show that, as in Amini et al. (2012), as the resilience indicator becomes
negative a ’phase transition’ occurs: the sub-network of critical receivables acquires a giant
component along which contagion can take place, exposing the system to a possible large
scale illiquidity cascade. Since the resilience indicator indicates the tipping point where
the onset of contagion occurs, we can use it to assess the effect of central clearing on the
mitigation of systemic risk.
The results show, that in absence of central clearing of the other classes of derivatives,
CDS clearing does not impede the phase transition. It is the large size of the jump in IR swaps
(recall the stress scenarios consider a jump equal to 5 % of the MtM value of IR swaps) that
plays a dominating role here and the system cannot withstand even a small liquidity shock.
However, when IR swaps are centrally cleared, CDS clearing has an important impact on
impeding the phase transition. Both in the case where the IR swaps are cleared in a dedicated

123
542 Ann Oper Res (2016) 247:523–547

Number of illiquid banks − Without IR CH Resilience to illiquidity measure − Without IR CH

2000 0.6
Without CDS CH
1800 CDS CH (top 20 dealers)
0.4 CDS CH (top 10 dealers)
1600

Resilience to illiquidty measure

1400 0.2
Number of illiquid banks

1200
0
Without CDS CH
1000 CDS CH (top 20 dealers)
CDS CH (top 10 dealers)
−0.2
800

600 −0.4

400
−0.6
200

0 −0.8
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Liquidity shock Liquidity shock

(a) (b)
Number of illiquid banks − With Dedicated IR CH Resilience to illiquidity measure − With Dedicated IR CH
2000 0.6
Without CDS CH
1800 CDS CH (top 20 dealers)
0.4 CDS CH (top 10 dealers)
1600
Resilience to illiquidty measure

1400 0.2
Number of illiquid banks

1200
0
Without CDS CH
1000 CDS CH (top 20 dealers)
CDS CH (top 10 dealers)
−0.2
800

600 −0.4

400
−0.6
200

0 −0.8
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Liquidity shock Liquidity shock

(c) (d)
Number of illiquid banks − With IR/CDS CH Resilience to illiquidity measure − With IR/CDS CH
2000 0.6
Without CDS CH
1800 CDS CH (top 20 dealers)
0.4 CDS CH (top 10 dealers)
1600
Resilience to illiquidty measure

1400 0.2
Number of illiquid banks

1200
0
Without CDS CH
1000 CDS CH (top 20 dealers)
CDS CH (top 10 dealers)
−0.2
800

600 −0.4

400
−0.6
200

0 −0.8
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Liquidity shock Liquidity shock

(e) (f)
Fig. 8 The number of illiquid banks and the resilience indicator for different clearing cases

CH and the case of joint clearing, when a CH for CDS with the top 20 members is introduced,
the phase transition occurs for a significantly larger liquidity shock. Without CDS clearing,
a liquidity shock of 3 % induces a phase transition. With CDS clearing with 20 members, the
onset of contagion occurs when the liquidity shock reaches 12 %.
We observe that the benefits of central clearing decrease when less members are allowed in
the CH. We can explain this in the following way. Recall that the CDS network is constructed

123
Ann Oper Res (2016) 247:523–547 543

by introducing the intermediation chains. Introducing a CCP compresses the intermediation

chains consisting exclusively of clearing members. For example if a intermediation chain
has length 10 and all intermediaries are members of the CH, then this chain will have length
two after compression. On the other hand it suffices for only one of the intermediaries to be a
non-member of the CH for the benefits to decrease. It follows that not just the size of notional
of protection sold/bought should be a criterion to allow a member in the clearing house, but it
should also be accounted for the number of intermediation chains passing through this node.
A significant dealer is a node that has both an important market share but more importantly,
is present as an intermediary in a large number of the intermediation chains.

4.4 Robustness

We have so far investigated the impact of central clearing on one sample of the random network
of OTC derivatives. A critical point concerns the robustness of these results with respect to
the realization of the random network. To this end, we draw 5000 samples and compute the
number of illiquid banks for each realized network, in the same cases investigated previously.
The comparative results are shown in Fig. 9. Each point in these graphs corresponds to one
realization of the network.
The exogenous liquidity shock is fixed at = 9 %, as a percentage of the liquidity buffer i
given in (7). This is within the range that we considered in Fig. 8. It a severe shock scenario, but
repo runs may lead to even higher levels for these shocks, see e.g., Duffie (2011). We expect
that our comparison results would hold for higher levels of the exogenous liquidity shock.

– In Fig. 9a we consider the case when IR derivatives are uncleared. We compare, for each
network realization, the number of illiquid banks in the case of no CDS clearing and the
case of CDS clearing when the top 20 dealers are included. We note that for a fair amount
of network realizations, introducing central clearing for CDS has a negative impact on
financial stability. We conclude that in absence of clearing of IR derivatives, there is a
significant probability that CDS clearing has a negative impact, measured by the number
of illiquid banks.
– In Fig. 9b we compare, under the scenario without IR clearing, the performance of CDS
clearing when the clearinghouse has the top ten dealers vs the top 20 dealers. While for
the majority of the realized networks, having more clearing members increases financial
stability, there is a significant number of realized networks for which the opposite is
true. This is not surprising since clearing one class of derivatives when the other class is
uncleared has an ambiguous effect: the CCP collapses the chains of critical receivables
but reduces the bilateral netting opportunities with the other asset class.
– In Fig. 9c, d we consider the case when IR derivatives are cleared through a clearing
house that clears only IR derivatives (dedicated IR CCP). We compare, for each realized
network, the number of illiquid banks in the case without CDS clearing and the case with
CDS clearing when the top 20 dealers are included. We find that clearing of CDS enhances
the network stability, in particular in the case where more significant dealers (20 vs. 10
dealers) are members of the clearinghouse. This means that for our choice of parameters,
the positive effect of the CCP on the collapse of the chains of critical receivables dominates
the reduction of bilateral netting opportunities with the other asset class.
– In Fig. 9c, d we consider the case when IR and CDS are cleared in the same clearing house.
Similarly to the case with a dedicated IR clearinghouse, a CDS clearinghouse increases
network stability. Unlike in case 4.4 above, there is no ambiguous effect of CDS clearing
when all significant members are allowed in the CCP, there are benefits both from the

123
544 Ann Oper Res (2016) 247:523–547

Number of illiquid banks − Without IR CH Number of illiquid banks − Without IR CH

4 4
10 10

3 3
10 10

CDS CH (top 10 dealers)

Without CDS CH

2 2
10 10

1 1
10 10

0 0
10 10
0 1 2 3 4 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10
CDS CH (top 20 dealers) CDS CH (top 20 dealers)

(a) (b)
Number of illiquid banks − With Dedicated IR CH Number of illiquid banks − With Dedicated IR CH
4 4
10 10

3 3
10 10
CDS CH (top 10 dealers)
Without CDS CH

2 2
10 10

1 1
10 10

0 0
10 10
0 1 2 3 4 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10
CDS CH (top 20 dealers) CDS CH (top 20 dealers)

(c) (d)
Number of illiquid banks − With IR/CDS CH Number of illiquid banks − With IR/CDS CH
4 4
10 10

3 3
10 10
CDS CH (top 10 dealers)
Without CDS CH

2 2
10 10

1 1
10 10

0 0
10 10
0 1 2 3 4 0 1 2 3 4
10 10 10 10 10 10 10 10 10 10
CDS CH (top 20 dealers) CDS CH (top 20 dealers)

(e) (f)
Fig. 9 Number of illiquid banks for different clearing cases. Liquidity shock: 9 %

collapse of of the chains of critical receivables, while no bilateral netting possibility is

removed. We expect that this result is robust to the choice of the parameters governing the
network structure. As evidenced by Fig. 9c, there is one case among 5000 where central
clearing decreases network stability and this case occurs when not all significant dealers
are members of the clearinghouse. Whereas the probability of a negative impact of a CDS

123
Ann Oper Res (2016) 247:523–547 545

clearing house cannot be excluded, we find that in the majority of network realizations
(99,98 % of cases) the CDS clearing house decreases the probability of a system wide
illiquidity contagion.

5 Conclusions

We have introduced a multi-layered network model, with a core-periphery structure, for

studying illiquidity contagion in OTC derivatives markets. Our model accounts for the het-
erogeneity in gross and net notional exposures and market share of market participants in each
asset class, as well as the presence of intermediation chains which characterize OTC markets.
In such a setting, liquidity shocks may generate contagion along an intermediation chain.
Our model provides a framework for studying the magnitude and dynamics of illiquidity
cascades in OTC markets, in stress scenarios formulated in terms of liquidity shocks and
variations of the contracts’ mark-to-market value. This analysis may be done using simulation
of cash flows in a stress scenario, but also using a simulation-free approach based on insights
from asymptotic analysis of large networks (Amini et al. 2011): introduce an indicator of the
resilience of the network to liquidity shocks that highlights the role of ‘critical receivables’
i.e. receivables on which an intermediary depends to meet its own short-term obligations,
and show that this indicators allows to detect phase transitions, i.e., the tipping point when
the liquidity shock generates large-scale contagion in the financial system.
Finally, we used this framework allows to analyze the impact of CDS clearing on systemic
risk.
Our analysis shows that, while in absence of a clearing facility for interest rate swaps,
an additional clearing facility for CDS does not appear to enhance financial stability, when
interest rate derivatives (mainly swaps) are centrally cleared—as is currently the case—and
under adequate capitalization, a CDS clearinghouse can contribute significantly to financial
stability by enhancing the resilience of the OTC network to large liquidity shocks, provided
all significant dealers are members of the clearing house. However, the impact of the CCP on
network stability is weakened if some major intermediaries are not clearing members. These
results are shown to be robust with respect to the network realization.
These results show the importance of taking into account the structure of the network
and the network induced ‘tail events’ when discussing the benefits of central clearing for
systemic risk. Our study provides a comprehensive framework for studying the impact of
central clearing of OTC derivatives on systemic risk and is potentially applicable to other
data sets and products.

Acknowledgements We thank three anonymous referees for helpful comments that improved the presentation
of the paper. We thank seminar participants at the Mathematical Modeling of Systemic Risk Workshop (Paris,
June 2011), the IMF Workshop on Systemic Risk Monitoring (2010), the Capital Markets Function Seminar
(NY Fed, December 2011), the Research in Options Conference (Buzios 2011), Information and Econometrics
of Networks Workshop (Washington DC, 2012) and the Symposium on Critical Challenges at the Interface of
Mathematics and Engineering (2012) for helpful discussions.

References
Allen, F., & Gale, D. (2000). Financial contagion. Journal of Political Economy, 108(1), 1–33.
Amini, H., Cont, R., & Minca, A. (2011). Resilience to contagion in financial networks. Mathematical Finance
(to appear) http://dx.doi.org/10.1111/mafi.12051.

123
546 Ann Oper Res (2016) 247:523–547

Amini, H., Cont, R., & Minca, A. (2012). Stress testing the resilience of financial networks. International
Journal of Theoretical and Applied Finance, 15(1), 1250006–1250026.
Amini, H., Filipovic, D., & Minca, A. (2013). Systemic risk with central counterparty clearing. Swiss Finance
Institute Research Paper No. 13–34. http://ssrn.com/abstract=2275376.
Amini, H., & Minca, A. (2013). Mathematical modeling of systemic risk. In E. Kranakis (Ed.), Advances in
network analysis and its applications (Vol. 18, pp. 3–26). Mathematics in industry Springer: Berlin.
Amini, H., Filipovic, D., & Minca, A. (2014). To fully net or not to net: Adverse effects of partial multilateral
netting. Swiss Finance Institute Research Paper No. 14-63. http://ssrn.com/abstract=2516974.
Babus, A. (2007). The formation of financial networks. RAND (forthcoming).
Babus, A. (2012). Endogenous intermediation in over-the-counter markets. Available at SSRN.
Battiston, S., Gatti, D. D., Gallegati, M., Greenwald, B., & Stiglitz, J. E. (2009). Liaisons dangereuses:
Increasing connectivity, risk sharing, and systemic risk. Preprint http://www.nber.org/papers/w15611.
Blanchard, P., Chang, C.-H., & Krüger, T. (2003). Epidemic thresholds on scale-free graphs: The interplay
between exponent and preferential choice. Annales Henri Poincaré, 4(suppl. 2), S957–S970.
Boss, M., Elsinger, H., Summer, M., & Thurner, S. (2004). The network topology of the interbank market.
Quantitative Finance, 4, 677–684.
Brunnermeier, M. K. (2009). Deciphering the liquidity and credit crunch 2007–2008. Journal of Economic
Perspectives, 23(1), 77–100.
Cont, R. (2010). Credit default swaps and financial stability. Financial Stability Review (Banque de France),
14, 35–43.
Cont, R., & Kan, Y. H. (2011). Statistical modeling of credit default swap portfolios. http://ssrn.com/
abstract=1771862.
Cont, R., & Kokholm, T. (2014). Central clearing of otc derivatives: Bilateral vs multilateral netting. Statistics
and Risk Modeling, 31(1), 3–22.
Cont, R., Moussa, A., & Santos, E. B. (2013). Network structure and systemic risk in banking systems. In J.-P.
Fouque & J. Langsam (Eds.), Handbook of systemic risk. Cambridge: Cambridge University Press.
DTCC. (2010). Trade information warehouse data. http://www.dtcc.com/products/derivserv/data/index.php.
Duffie, D. (2011). How big banks fail and what to do about it. Princeton: Princeton University Press.
Duffie, D., Scheicher, M., & Vuillemey, G. (2014). Central clearing and collateral demand. Working Paper.
Duffie, D., & Zhu, H. (2011). Does a central clearing counterparty reduce counterparty risk? Review of Asset
Pricing Studies, 1, 74–95.
ECB. (2009). Credit default swaps and counterparty risk. Technical report, European Central Bank.
Farboodi, M. (2014). Intermediation and voluntary exposure to counterparty risk. Working paper, University
of Chicago.
Furfine, C. H. (1999). The microstructure of the federal funds market. Financial Markets, Institutions &
Instruments, 8(5), 24–44.
Gai, P., & Kapadia, S. (2010). Contagion in financial networks. Proceedings of the Royal Society A, 466(2120),
2401–2423.
Giglio, S. (2011). Credit default swap spreads and systemic financial risk. Proceedings of the Federal Reserve
Bank of Chicago, 5, 104–141.
Glasserman, P., & Young, H. P. (2013). How likely is contagion in financial networks? Journal of Banking
and Finance (forthcoming).
Glode, V., & Opp, C. (2013). Adverse selection and intermediation chains. Working paper.
Heller, D., & Vause, N. (2012). Collateral requirements for mandatory central clearing of over-the-counter
derivatives. BIS Working Paper 373, Bank for International Settlements.
ISDA. (2010). Market review of OTC derivative bilateral collateralization practices. http://www.isda.org/c_
and_a/pdf/ISDA-Margin-Survey-2010.
Macroeconomic Assessment Group on Derivatives (MAGD). (2013). Macroeconomic impact assessment of
OTC derivatives regulatory reforms. Technical report, Bank for International Settlements.
Manea, M. (2013). Intermediation in networks. Working Paper.
Minca, A. (2011). Mathematical modeling of financial contagion. PhD thesis, Universite Paris VI (Pierre et
Marie Curie).
OCC. (2010). OCC’s quarterly report on bank trading and derivatives activities. Second Quarter.
Peltonen, T. A., Scheicher, M., & Vuillemey, G. (2014). The network structure of the CDS market and its
determinants. Journal of Financial Stability, 13, 118–133.
Rogers, L., & Veraart, L. (2012). Failure and rescue in an interbank network. Management Science (forthcom-
ing).
Sidanius, C., & Zikes, F. (2012). OTC derivatives reform and collateral demand impact. Financial Stability
Paper 18, Bank of England.

123
Ann Oper Res (2016) 247:523–547 547

Singh, M., & Aitken, J. (2009). Deleveraging after Lehman—Evidence from Reduced Rehypothecation. SSRN
eLibrary.
Stulz, R. M. (2010). Credit default swaps and the credit crisis. Journal of Economic Perspectives, 24(1), 73–92.
Upper, C., & Worms, A. (2002). Estimating bilateral exposures in the german interbank market: Is there a
danger of contagion? Deutsche Bundesbank Discussion Paper No 09/02.
Watts, D. J. (2002). A simple model of global cascades on random networks. Proceedings of the National
Academy of Sciences of the USA, 99(9), 5766–5771.
Yellen, J. (2013). Interconnectedness and systemic risk: Lessons from the financial crisis and policy
implications. Board of Governors of the Federal Reserve System at the American Economic Associ-
ation/American Finance Association Joint Luncheon San Diego, California.
Zawadowski, A. (2013). Entangled financial systems. Review of Financial Studies, 26, 1291–1323.

123