Professional: On The Same Page

PROFESSIONAL
A GARP Publication DECEMBER 2011

On the Same Page
Principal Financial CRO Greg Elming and
CEO Larry Zimpleman on risk management
as an integrated, collaborative enterprise
ALSO INSIDE
Metrics: a company risk
ranking and a statistical
aid for corporate directors

End-to-end stress testing

1 RISK PROFESSIONAL D E C E M B E R 2 0 1 1 www.garp.org
D E C E M B E R 2 0 1 1
PROFESSIONAL
F R O M T H E E D I T O R
Education in Governance
n the cover of our October issue was the question,
Can risk culture be taught? Lots of people are giv-
ing it the old college try, but the course is too new and
subjective to yield a yes-or-no answer yet. So in this
issue, and in articles online (see the table of contents
for links), we move on to matters of risk governance,
particularly to consider whether corporate boards of directors can make
a constructive difference, and if so, how.
The Credit Crisis Around the Globe: Why Did Some Banks Perform
Better?, a 2010 paper co-authored by Ohio State University professor
and GARP trustee Ren Stulz, did not instill much confdence in board
efforts. Banks that rewarded their shareholders the most in 2006, presum-
ably overseen by shareholder-friendly directors, fared worse during the
subsequent crisis, as earlier risk-taking came back to haunt them.
Out of necessity, directors have embarked on their own risk management
studies and have yet to graduate. Deepa Govindarajan of ICMA Centre
at the University of Reading in the U.K., who advocates more systematic,
proactive oversight of risk appetites, says directors are unable to enunci-
ate formally the risks associated with the strategic choices they pursue.
Thus is the learning curve defned. No one is going to argue that gover-
nance is futile, and boards are more engaged in the risk conversation than
ever before. The Dodd-Frank Wall Street Reform and Consumer Protec-
tion Act requires board-level risk committees at bank holding companies
with at least $10 billion in assets. That may be a welcome indicator of the
enhanced importance of risk management. But, Jim DeLoach of con-
sulting frm Protiviti writes, it will force a realignment of responsibilities
traditionally assigned to audit committees.
Tools and techniques are coming into the breach. As reported in the Up-
front section, GMI is publishing a cautionary list of risky companies, and
Protiviti is touting a risk index that distills the avalanche of data in
board-meeting packages down to a single risk value. But management still
comes down to human collaboration and execution. As James Stewart of
JHS Risk Strategies writes, amid all the metrics and complexity, tradi-
tional leadership roles and responsibilities are arguably more relevant and
important than emerging best practice.
Editor-in-Chief
Jeffrey Kutler
jeff.kutler@garp.com
Executive Editor
Robert Sales
robert.sales@garp.com
Design Director
J-C Suars
jcsuares@nyc.rr.com
Art Director
Jennifer Richman
jenniferrichman@yahoo.com
Risk Professional Editorial Board
Peter Tufano
Dean, Said Business School
University of Oxford
Aaron Brown
Risk Manager
AQR Capital Management
Tim Davies
Director for Risk
Lloyds TSB Bank
Mike Gorham
Director, IT Center for Financial Markets
Stuart Graduate School of Business
David Lee
CRO and Managing Director, Liquid Alternatives
Credit Suisse
Peter Manning
Head of Risk Management
China Construction Bank (London)
Robert Merton
School of Management Distinguished Professor of Finance
MIT Sloan School of Management
Joe Pimbley
Principal
Maxwell Consulting
Riccardo Rebonato
Global Head of Market Risk and Quantitative Analytics
RBS
Peruvemba Satish
Managing Director and Chief Risk Offcer
Allstate Investments
David Shimko
President
Winhall LLC
Victor Yan
Managing Director and Head of Market Risk, China
Standard Chartered Bank
Senior Vice President, Publisher
David S. Greenough
david.greenough@garp.com
Director, Global Advertising and Business Development
Michael Simone
michael.simone@garp.com
International Advertising and Event Sales
Asia: David Chalmers
chalmersdavid@mac.com
Director, Career Center
Mary Jo Roberts
Manufacturing & Distribution
Kristina Jaoude
Production Coordinator
Hanna Park
RISK PROFESSIONAL is published six times a year by the Global Association of Risk
Professionals. It is circulated to all GARP dues-paying members. To subscribe, please go
to https://www.garp.org/news-and-publications/risk-professional.aspx. Basic annual rate:
$150.00. For customer service inquiries, please call: London, UK, +44 207 397 9630; Jersey
City, NJ, +1 201 719 7210. For reprints and e-prints of RISK PROFESSIONAL articles,
please contact PARS International at 212-221-9595.
O
www.garp.org D E C E M B E R 2 0 1 1 RISK PROFESSIONAL 2
D E C E M B E R 2 0 1 1
research, Kelly
joined New York-
based Quantif
in 2009 and has
over 18 years of
fnancial industry
experience. He
has worked for
Citigroup, JPMorgan Chase & Co.,
American International Group and
Credit Suisse First Boston.
The Risk Strategy article on page 38
was written by Gaurav Kapoor,
chief operating offcer of Metric-
stream, a supplier of governance, risk
and compliance systems; and Con-
nie Valencia, principal of Elevate
Consulting, which specializes in pro-
cess improvement, controls and tech-
nology risk management. Kapoor,
a Wharton
School MBA,
spent 10 years
with Citibank
in the U.S. and
Asia, helped
lead Arca-
diaOne from
inception to
acqui si t i on,
founded the
Co mp l i a n -
ceOnline.com
GRC commu-
nity and until
2010 was Palo Alto, California-based
Metricstreams CFO and general
manager of GRC. Valencia brought
Big 4 accounting and consulting ex-
perience to Miami-based Elevate,
where her work includes corporate
governance consultation, policy and
procedure enhancement and Sar-
banes-Oxley compliance.
C O N T R I B U T O R S
President and CEO
Richard Apostolik
rich.apostolik@garp.com
Managing Director, Research Center
Christopher Donohue, PhD
chris.donohue@garp.com
Managing Director, Global Head of Training Services
Alastair Graham
alastair.graham@garp.com
Managing Director, Global Head of Marketing
Kathleen Alcorn
kathleen.alcorn@garp.com
Chief Operating Offcer
Michael Nielsen
michael.nielsen@garp.com
Controller
Lily Pan
lily.pan@garp.com
Board of Trustees
William Martin, Chairman
Chief Risk Offcer, Abu Dhabi Investment Authority
Richard Apostolik
President and CEO, GARP
Kenneth Abbott
Managing Director, Morgan Stanley
Robert Ceske
Chief Risk Manager, Treasury and Global Funding Operations,
GE Capital
Sebastian Fritz
Senior Advisor, Frankfurt School of Finance and Management
Bennett Golub
Senior Managing Director and CRO, BlackRock
Michael Hofmann
Chief Risk Offcer, Koch Industries
Donna Howe
Chief Risk Offcer, Hartford Investment Management
Frederick Lau
Head of Group Risk Management, Dah Sing Bank
Jacques Longerstaey
Executive Vice President, CRO, State Street Global Advisors
Victor Makarov
Head, Market Risk Analytics, Metropolitan Life
Michelle McCarthy
Director of Risk Management, Nuveen Investments
Riccardo Rebonato
Global Head of Market Risk and Quantitative Analytics, RBS
Jacob Rosengarten
Chief Risk Offcer, XL Capital
Robert Scanlon
Group Chief Credit Offcer, Standard Chartered Bank
David Shimko
President, Winhall LLC
Ren Stulz
Reese Chair in Banking and Monetary Economics,
Ohio State University
Peter Tufano
Dean, Said Business School, University of Oxford
Mark Wallace
Managing Director, Chief Risk Offcer,
Credit Suisse Asset Management
RISK PROFESSIONAL, 2011, Global Association of Risk Professionals,
111 Town Square Place, Suite 1215, Jersey City, NJ 07310. All rights reserved.
Reproduction in any form is prohibited without written permission from GARP.
Printed in the United States by RR Donnelly.
A team of six shares credit for the
Credit Risk article on page 19. Five
are analysts and economists from
HSBC, including Alessandra
Mongiardino, a past contributor to
Risk Professional when she was affliated
with Moodys Investors Service. The
sixth is Andrea Serafno (below),
senior econometrician at the U.K.s
Financial Services Authority, who
was at HSBC,
speci al i zi ng
in macroeco-
nomic stress
testing, before
joining the
regulator in
2010.
M o n -
giardino is
now head of
risk strategy
and a mem-
ber of the
risk manage-
ment executive team for HSBC
Bank (HBEU), responsible for the
strategic framework that ensures risk
management is effectively embedded
throughout the bank and across all
risk types. Her HSBC collaborators:
Zoran Stanisavljevic, head of whole-
sale risk analytics; Evguenia Iankova,
senior quantitative economist; Bruno
Eklund, senior quantitative analyst;
and Petronilla Nicoletti, senior quan-
titative economist.
Counterparty risk management is
a current focus of trading and risk
software company Quantif, whose
director of credit products, David
Kelly, contributed the article on that
subject on page 27. Well versed in
derivatives trading and quantitative

C O N T E N T S
F E A T U R E S C O V E R S T O R I E S
13 CRO-CEO DIALOGUE:
On the Same Page at the Principal
Few companies can trace an enterprise risk management heri-
tage back as far as the 1970s, as does the Principal Financial
Group. The Des Moines, Iowa-based insurer and asset manager
also boasts a chief executive offcer and a chief risk offcer with
nearly 70 years of combined service to the organization. In an
exclusive joint interview, CEO Larry Zimpleman and CRO
Greg Elming describe their implementation of ERM and its
proft-enhancing and loss-mitigating paybacks.
19 CREDIT RISK/STRESS TESTING:
A Robust End-to-End Approach
Alessandra Mongiardino and colleagues at HSBC and the
Financial Services Authority put forward a factor-model
approach to stress testing that can identify corporate portfolios
that are particularly vulnerable to stress, improve the accuracy
of default-probability and loss-given default forecasts and
produce transparent results.
27 How the Financial Crisis
Changed Counterparty Risk
In the face of dramatically increased capi-
tal requirements for counterparty credit
risk, banks have created credit value adjust-
ment desks to centralize the monitoring
and hedging of counterparty exposures in
a more active and timely way. Despite the
improvements in this area, banks still face
data and analytical challenges, according to
David Kelly of Quantif.
38 A Strategy-Based Approach
A better alignment of risk management
with company goals and objectives can
take into account all risks, including black
swans, while enabling balanced, focused
and more cost-effective analyses and re-
sponses to potential threats, write Gaurav
Kapoor and Connie Valencia.
13
19
C
o
v
e
r

P
h
o
t
o

b
y

2
0
1
1

S
c
o
t
t

S
i
n
k
l
i
e
r
D E P A R T M E N T S E X C L U S I V E L Y O N L I N E
1 From the Editor
Education in Governance
2 Contributors
7 UPFRONT
Ten North American companies with
GRC issues appear on research frm
GMIs inaugural Risk List; Protivitis
single-number risk index snapshot;
risk managers slow road to the top job;
wind power and weather derivatives;
and can a World Currency Unit stabilize
forex volatility?
22 QUANT CLASSROOM
Attilio Meucci shows how to estimate
VaR and other risk numbers for arbi-
trary portfolios by emphasizing histori-
cal scenarios similar to todays environ-
ment.
22
8
10
WEBCASTS
Risk Appetite, Governance
and Corporate Strategy
The Ramifcations of the
Dodd-Frank Act: Derivatives
VIEWPOINTS
Should a Board of Directors Have
A Separate Risk Committee?
Demonstrating the Effectiveness
Of Compliance and Ethics Programs
Risk Culture and Tone at the Top
Increasingly accountable to regulators and shareholders on risk manage-
ment matters, executives and directors should not lose sight of tradition-
al goal-setting, leadership and communications as they become conver-
sant in the technicalities and analytics of fnancial market complexity,
according to consultant James Stewart.
Governance and the Risk Appetite Statement
Any investor weighs risk-return trade-offs, expressed in terms of risk ap-
petite. But the differences between individual investment decisions and
those of a collective board acting on behalf of stakeholders need to be
better understood, says ICMA Centres Deepa Govindarajan.

C O N T E N T S
With so much uncertainty in the global financial markets, its more important than ever to have the right skill set.
Because Certified FRMs have proven they have the knowledge to manage a broad cross-section of financial risk, they are
qualified to help solve todays complex challenges. Thats why Certified FRMs are in demand by the leading banks, consulting
firms, corporations, asset management firms, insurance firms, and government and regulatory bodies worldwide. And
thats why theres never been a better time to be an FRM.
Registration for the May 2012 FRM Exam opens on December 1.
To learn more, visit www.garp.org/frm
Creating a culture of risk awareness.
TM
2011 Global Association of Risk Professionals. All rights reserved.
Theres never been a
better time to be an FRM.
Naming Names
isk managers, investors
and regulators seeking
guidance about pub-
licly-held companies
with troublesome risk profles
are getting some help from the
New York-based corporate gov-
ernance research organization
GMI.
The independent provider of
environmental, social and gov-
ernance (ESG) ratings on 4,200
actively traded companies world-
wide released its inaugural GMI
Risk List in October. It consists
of 10 North American corpora-
tions in alphabetical order that
exhibit specifc patterns of risk .
. . representing an array of ESG
and accounting transparency is-
sues.
The methodology places em-
phasis on non-fnancial consid-
erations that are increasingly
seen as factors contributing to
corporate failures. Companies
that lag tend to have compen-
sation or pay-per-performance
issues, questionable board inde-
pendence, ownership structures
that disadvantage large numbers
of public shareholders, or poor
environmental, health and safety
risk disclosures. GMI bases its
assessment on the likelihood of
such problems resulting in such
negative outcomes as regulatory
actions, shareholder litigation
and material fnancial restate-
ments.
Anybody focused on risk
management efforts, including
contingency planning, supply
R
A top 10 of risky situations
By Katheri ne hei res
chain risk or environmental,
social and corporate governance
risks, will fnd this research of
great interest, says Jack Zwingli,
CEO of GMI, which was formed
in December 2010 with the
merger of GovernanceMetrics
International, the Corporate Li-
brary and Audit Integrity.
Zwingli notes that in the last
decade, meltdowns such as those
of Adelphia Communications
only heightened investor, regula-
tor and risk manager interest in
monitoring and understanding
cultural and qualitative as well
as fnancial issues that can drive
companies into bankruptcy, reg-
ulatory discipline or costly litiga-
tion.
Multiple Concerns
One GMI Risk List company, for
example, was cited for the fact
that investors outside the found-
ing family had little power to infu-
ence the companys course, while
GMI research showed it may also
be inappropriately capitalizing
assets and smoothing earnings.
Another had a board deemed to
was highlighted for lack of atten-
tion to environmental risk while
also capitalizing some expenses
to infate reported earnings.
Aside from the top 10 re-
lease, GMI offers more in-depth
analytics about each companys
weaknesses through its GMI An-
alyst research platform, including
a complete list of the red fags or
factors seen as having a material
impact on investment returns.
Not more than 5% of the list of
4,200 companies GMI tracks fall
into the riskiest category, receiv-
ing the lowest grade of F.
It has been one factor after
another that has only served to
heighten awareness of the impor-
tance of these non-fnancial is-
sues at corporations and the need
to measure and tighten these
types of risks, says Zwingli.
Kimberly Gladman, GMIs
director of research and risk ana-
lytics, states, Fiduciaries increas-
ingly consider an evaluation of
ESG issues a core part of their in-
vestment due diligence. Our Risk
List research is intended to assist
investors in identifying particular
ESG factors that may correlate
with negative events.
GMI literature points out that
ESG scrutiny is on the rise, for
example, among institutional in-
vestors and other signatories to
the United Nations Principles for
Responsible Investment, which
currently represents over $25 tril-
lion in managed assets.
Combined Competencies
Zwingli says GMI is particularly
well positioned to supply risk-
related research because its own
scope and synergies have broad-
ened: The Corporate Library was
known mainly for its expertise in
governance issues, Governance-
The Risk List
UPFRONT
Company Exchange Ticker Industry
ApolloGroupInc. NASD APOL Personal Services
ComstockResourcesInc. NYSE CRK Oil and Gas
Exploration/Production
ComtechTelecomm.Corp. NASD CMTL Communications
Equipment
DiscoveryCommunications NASD DISCA Broadcasting
EZCORPInc. NASD EZPW Consumer
Financial Services
K-SwissInc. NASD KSWS Footwear
M.D.C.Holdings NYSE MDC Homebuilding
NewsCorp. NASD NWSA Media Diversifed
SandRidgeEnergyInc. NYSE SD Oil and Gas
Exploration/Production
ScientifcGamesCorp. NASD SGMS Casinos / Gaming
Corp., Enron Corp., Tyco Inter-
national and Worldcom, followed
by ESG breakdowns at American
International Group, BP, Massey
Energy Co. and others, have
be dominated by insiders, while
accounting indicated rapidly in-
creasing leverage, low cash levels
and concerns about inadequate
expense recognition. Yet another
Source: GMI, October 2011
UPFRONT
Company Rating
enchmarks, metrics
and other mathemat-
ics are hardly new to
risk managers or to
the discipline of risk manage-
ment. If anything, quantitative
analysis, and too much faith in it,
threw fnancial risk management
off course and helped precipi-
tate the credit market debacles
of 2007 and 2008.
To risk management prac-
titioners and experts like Cory
Gunderson, head of consulting
frm Protivitis U.S. fnancial ser-
vices practice, the industry has
gotten a better handle on risk in
the aftermath of the crisis, aided
by a new generation of sophisti-
cated analytical tools and report-
ing mechanisms. At the same
time, fnancial institution man-
agements and boards of direc-
tors are inundated by data that
taxes the capacity of the systems
that they depend on for high-lev-
el risk assessments, oversight and
strategic guidance.
An avalanche of data and
not a lot of analysis is how
Gunderson describes the situa-
tion. For that complex and hard-
to-manage problem, Gunderson
is developing a solution, de-
scribed as a risk index. It boils
hundreds of pages of risk data
that board members receive on a
monthly or quarterly basis down
to a single number.
Protiviti is not out to create
a global indicator of risk levels,
as has been done, for example,
by GARP, explains Gunder-
son. This is a tool to produce a
single-number snapshot of the
state of risk at a given fnancial
institution. The index is cus-
tomized, scalable to any size of
frm, and designed to enable an
interpretation of how it is do-
ing with regard to its risks at any
given moment, Gunderson tells
Risk Professional.
Deliberate Rollout
It is about a year since Gunder-
son began rolling the concept
out, establishing the beginnings
of a track record with selected,
undisclosed client frms. Protiviti
has not made an offcial release
of the product or of the meth-
odology underlying it, though
it published a fairly thorough,
four-page overview in its FS In-
sights newsletter.
B
By Jeffrey Kutler
It revealed, for
instance, the steps
taken to aggregate
and link an orga-
nizations strate-
gic objectives to
quantifable risk
measures and an
algorithm that ul-
timately calculates
the index. An in-
dex can be broken
down into different component
parts, allowing for drill-down
capability and a focus on core
issues that drive an entitys risk
profle, the article said.
Gunderson says a more de-
tailed white paper is forthcoming
in early 2012.
The newsletter noted that
two basic yet crucial questions
remain diffcult for board mem-
bers and management to answer
easily: Is our organization riskier
today than it was yesterday? Is
our organization likely to be-
come riskier tomorrow than it is
today?
The index promises to give
boards an accurate picture of
the organizations current and
future risk profle in a single, dis-
crete number. The tool would
similarly be applied by manag-
ers enterprise-wide and at senior,
business-line and departmental
levels, as well as geographically.
Whats more, Executives and
boards can use the
information . . .
to take action and
run what-if sce-
narios.
Wider
Applicability
Gunderson, a
University of Chi-
cago MBA who
is based in New
York and has been with Protiviti
since 2002 after 11 years with
Arthur Andersen, says, We be-
lieve we are onto a cutting-edge
idea. The index has been posi-
tively received where it has been
implemented and has attracted
a ton of interest, not just from
the fnancial industry. Finance is
my background, and an area
that lends itself to early adop-
tion, notes Gunderson, but
over time the index can work in
other sectors too.
We have not solved risk
management, he adds. We
created a tool a very effcient
tool to boil down where risks
are coming from. He stresses
that rather than replacing or
replicating any existing risk
management measures, the in-
dex is another tool in the tool-
kit.
Although Gunderson and
Protiviti are not ready to take
the wraps off, he says that mul-
Metrics International for its ESG
metrics and analysis, and Audit
Integrity for forensic accounting
analysis and risk models.
Future Risk List offerings will
examine non-U.S. frms and in-
dustrial sectors where high risk
factors are in play for all partici-
pants. GMI aims to publish its
lists twice a year, Zwingli said,
adding that GMI is looking to
expand beyond 10 companies, to
possibly as many as 50.
Zwingli stresses GMIs inde-
pendent status and the fact that it
does not provide consulting ser-
vices to frms seeking to improve
their ESG ratings, which differ-
entiates it from other research
and ratings frms. GMI competes
in some areas with the likes of
Glass, Lewis & Co. and Institu-
tional Shareholder Services, a
unit of MSCI.
We have been looking at risk
measures for several decades,
says Zwingli, and have been
able to identify those metrics and
patterns of behavior most pre-
dictive of negative events. We be-
lieve risk managers need to make
sure they are aware of the risks
that are most critical at a given
company and make them a part
of their everyday risk assessment
process.
Cory Gunderson
UPFRONT
tiple risk measures and time se-
ries go into the index calculation;
and that it has a quantitative
component while also having to
be responsive to the culture of
a company so that it resonates
with, and has value to, the peo-
ple within it.
As stated in the Protiviti news-
letter, By equipping companies
with a more comprehensive,
comprehensible and actionable
snapshot of their organizations
risk management progress, risk
indices can help risk offcers,
board members and sharehold-
ers become more confdent that
they understand the businesss
risk profle and whether risk
levels are rising or falling.
For a Protiviti commentary
on boards and risk commit-
tees, click here.
Slow Road to the Top
arren Buffett is
known as the sage
of Omaha as much
for his straight talk
and common-sense approach as
for his legendary investment suc-
cesses. A case in point was what
he said in a CNBC interview two
years ago: The CEO has to be
the chief risk offcer for a bank.
As much as he may have been
on point, the billionaire did not
mean that the CEO must liter-
ally or simultaneously wear the
offcial hat of CRO. But his re-
mark does prompt a question:
Why have there not been more
CROs or others with risk man-
agement backgrounds rising, as
has been anticipated for the last
few years, into CEO positions at
major banks?
The February 2009 Preview
Edition of this magazine head-
lined on its cover, The Rise of
the Risk Executive, alongside
a photo of Matthew Feldman,
president and CEO of the Feder-
al Home Loan Bank of Chicago.
Feldman, a former commercial
banker who had earlier been se-
nior vice president of risk man-
agement, remains one of the few
to climb from that rank to the top
of a fnancial institution of any
W
The rise of the CRO is taking time
By l. a. Wi noKur
prominence.
Also spotlighted in that early
issue of Risk Professional was Mau-
reen Miskovic, a career risk man-
ager who made the rare move
from board of directors to CRO
at Boston-based State Street
Corp. in 2008. In early 2011, she
became group CRO of Switzer-
land-based UBS, which was sub-
sequently hit by a costly rogue
trading scandal. UBSs previous
CRO, Philip Lofts, is now CEO
of UBS Group Americas.
Why havent more CROs fol-
lowed that path?
According to Kevin Buehler,
who co-founded McKinsey &
Co.s global risk management
practice and co-leads its cor-
porate and investment banking
practice, the answer is that it will
just take time. One factor work-
ing against CROs is their lack of
corporate longevity. Whenever
unexpected losses occur at large
fnancial institutions, the CRO is
often in the line of fre, Buehler
explains. This results in tremen-
dous turnover in the CRO posi-
tion, particularly at institutions
suffering a high level of losses at
a time of fnancial turmoil.
Hollywood portrayed this very
trend in Margin Call, a recently
released, criti-
cally praised
flm by J.C.
Chandor (see
the October
Risk Professional,
page 8). When
a fnancial cri-
sis threatened
the collapse of
a Wall Street
securities frm,
senior risk
managers took
the fall.
A n o t h e r
impediment has to do with the
role CROs play in the C-suite. In
some organizations, says Buehler,
it is viewed principally as a con-
trol function not unlike that of
the head of internal auditing
which is not an easy launching
pad for a prospective CRO.
Becoming Strategic
On the other hand, where the
CRO plays more of a strategic
role for instance, helping to
balance risk and return similar
to what a chief fnancial offcer
would do, or heading what is
regarded as a key business unit
there is a greater likelihood
of moving up the ladder and
contending for CEO. This is
especially true, says Buehler, for
candidates who have a proven
track record in both good times
and bad.
One more piece to the puzzle:
A greater risk consciousness at
the board level
of large fnan-
cial services
c ompa ni e s ,
which results
from having
more direc-
tors with risk
and regulatory
backgrounds.
Co r p o r a t e
boards pri-
marily care
about strat-
egy; compen-
sation and
succession planning for the man-
agement team; and risk issues
that could undermine the com-
pany, Buehler says.
Because of the fnancial crisis
and the regulatory response to it,
senior bank executives have been
spending increasing amounts of
time interacting with regulators,
the McKinsey director contin-
ues. In those instances where a
board enjoys a close relationship
with its CRO, he says, it would
increasingly value the skill set
that a CRO brings to the table.
Buehler points out that CROs
who play a central, increasingly
more strategic role working with
the board and on senior execu-
tive teams will be one step clos-
er to the CEO spot.
L.A. Winokur, a veteran business jour-
nalist based in the San Francisco area,
wrote the February 2009 article, The
Rise of the Risk Executive.
Special preview edition
A d v a n c i n g t h e R i s k P r o f e s s i o n
PRofessionAl
the riSe
of the riSk
executive
Risk managers gain prestige and
prominence, and take more heat,
as market pressures mount
Matthew FeldMan, CeO,
Federal hOMe lOan Bank OF ChiCagO
State Streets Top
Management Brings Its
CRO Into The Inner Circle
Will Too Many Controls
Kill The Credit Default Swap?
A New Technical Standard
Proposed For Risk Reporting
FeBRuARy 2009
pluS
UPFRONT
A Basket in Reserve
DX Organisation
Ltd. in London has a
new managing direc-
tor, and he has high
aspirations for what WDX is sell-
ing: the World Currency Unit,
aka Wocu. Ian Hillier-Brook, a
veteran fnancial industry and
technology executive who suc-
ceeded WDX founder Michael
King this summer, says he is aim-
ing by mid-2012 to have a dozen
exchanges around the world
actively using the Wocu and for
global banks to be posting Wocu
rates and prices.
Were attending conferences,
writing articles in the U.K. for
the Association of Corporate
Treasurers and working with a
number of global banks, Hill-
ier-Brook, who had previously
served as a director and consul-
tant to WDX, tells Risk Profes-
sional.
It has been as much an educa-
tional pursuit as it is a commer-
cial enterprise. A separate WDX
Institute promotes the intellec-
tual case for a currency unit.
The Wocu was conceived in
1996 as a weighted basket of
currency values the top 20 cur-
rencies based on gross domestic
product are in the basket that
would lessen the volatility associ-
ated with conventional trading in
currency pairs.
Wocu calculations have been
issued continuously since 2000,
and the commercial launch on
January 1, 2010, coincided with
concerns in emerging markets
and among some economists
about the continued, long-term
W
By Juli ette fai rley
viability of the U.S. dollar as a
reserve currency.
Aid in Trade
When his appointment was an-
nounced in September, Hilli-
er-Brook stated that there is
growing consensus among initial
target users that the Wocu will
be an invaluable component of
future international trade.
He also said, A balanced bas-
ket is the answer to dampening
volatility, while including the U.S.
dollar and the currencies of fast-
growing developing countries re-
fects the real world. WDX also
pointedly noted that the Wocu
would continue to function ef-
fectively, and uninterrupted, if
the euro were to break up.
Says Hillier-Brook: Corpo-
rate treasurers are saying they
will happily use the Wocu if their
banks are offering Wocu rates
and prices. The Warsaw Com-
modities Exchange in Poland an-
nounced in June that it intended
to use Wocu values in commod-
ity trading for exchange-rate risk
reduction.
Hurdles to Adoption
Market observers say the con-
cept has merits, but achieving
wide-scale adoption will be a
challenge.
I like this idea, but if you
look at competitors, such as a
stable currency index, theyve
struggled to keep traction. The
Wocu is more complex because
it is more diversifed, says Axel
Merk, president and chief invest-
ment offcer of Palo Alto, Cali-
fornia, currency fund manager
Merk Investments.
Having raised private capital
to commercialize its product,
WDX spent the last two years
building infrastructure, getting
technology in place and licensing
the Wocu, says Hillier-Brook.
It is not the only potential
alternative for the dollar-domi-
nated currency reserve system.
One proposal, articulated by
Chinas central bank chief in
2009, among others, was for the
International Monetary Fund
synthetic measure for reserve as-
sets, Special Drawing Rights, to
become the calculation mecha-
nism. Hillier-Brook sees disad-
vantages in the SDR: It is ex-
changeable with only four major
currencies and is updated every
fve years, versus the Wocus
twice-yearly rebalancing.
Volatility of currencies is a
major concern at the moment,
but simply dumping the U.S.
dollar does not make sense,
says Hillier-Brook. We have the
countries whose economies are
strongest in the basket, including
the U.S. dollar. A balanced bas-
ket reduces volatility and helps to
manage the problem, rather than
solve it by creating a derivative.
Persistent Dollar
Despite the knocks that the dollar
has taken, currency traders still
have a high opinion of it, asserts
Dev Kar, a former IMF senior
economist who is lead economist
of the Washington, D.C.-based
research and advocacy organiza-
tion Global Financial Integrity.
The dollar has the dominant
position in the invoicing of im-
ports and exports because it is
used as a third-party currency
between countries, such as India
and China, says Kar. To sup-
plant the dollar is a diffcult task.
All I can say is good luck.
Merk adds that the Wocu will
have a hard time rivaling the ma-
jor currencies liquidity. There
are non-deliverable currencies in
the basket, he points out. You
have to come up with a deliver-
able currency. The drawback is
that with emerging market cur-
rencies, there is often no full con-
vertibility.
WDX believes its time will
come and its formula is tested.
As its Web site (www.wocu.
com) explains, modern risk
theory and mathematics have
gone into the Wocu, a step-
changing currency instrument,
free of political infuence. Un-
like the outdated SDR, the
Wocu is distributed from the
WDX algorithm and modern,
real-time technology infrastruc-
ture. The Wocu truly refects
exchange rates and economic
power, calculated on a sub-sec-
ond basis. [It] will herald lower
risk, lower volatility and an an-
swer to the allocation puzzle of
reserve currencies.
Wind Shift
ho has seen the
wind? Neither I nor
you, wrote the 19th
century poet Chris-
tina Rossetti. Nor did wind farm
operators of RES Americas in
Texas during the winter of 2009-
10 when energy production
sagged.
But the weather changes.
This year saw the windiest July
in Texas in a long time, says
Michael Grundmeyer, head of
global business development
strategy for Seattle-based 3Tier,
which quantifes and manages
exposure to renewable energy
risks.
As wind power, increasing
26% annually, becomes a grow-
ing source of electricity genera-
tion, so does the demand for this
form of weather risk manage-
ment and weather derivatives for
hedging.
The natural variability of
wind led to conservative fnanc-
ing and a relatively high cost of
capital, says Martin Malinow,
CEO of Galileo Weather Risk
Management Advisors in New
York, which introduced Wind-
Lock in May using 3Tiers data.
It is a tool to manage the uncer-
tainty of wind-driven earnings.
Models Advance
Weather modeling generates his-
torical time series 3Tier has 30
years of data resulting in global
benchmarks for wind variability
risk assessment and mitigation,
and offering energy traders and
W
Weather derivatives make a mark
By Jani ce fi oravante
grid operators tools for scenario
analysis.
Weather prediction modeling
has evolved from the frst weath-
er derivatives of the late 1990s
that addressed temperature dif-
ferences in summer and winter
for energy companies. These
are still traded, still structured
today, Grundmeyer notes.
Standardized contracts trade
on the Chicago Mercantile Ex-
change, customized ones over-
the-counter. OTC deals are gen-
erally larger in dollar terms, up
to $150 million or $200 million
notional. There is an increased
balance between the exchange-
traded market and OTC mar-
kets, says Bill Windle, managing
director of RenRe Energy Advi-
sors in The Woodlands, Tex.
and president of the Weather
Risk Management Association
(WRMA). The product class
has grown to include hurricane,
rainfall and snowfall futures and
options.
For the wind segment, us-
ing multiple clusters of high-
powered computers for weather
prediction models, 3Tier aims at
making data digestible for those
with a fnancial stake in the num-
bers, explains Grundmeyer.
Key questions for wind farm op-
erators and developers concern
variability of wind speeds at a
given site, on what time scale and
with how much predictability.
We were involved in fve
projects in Texas that suffered
from the El Nio weather pat-
tern, and they couldnt get
enough wind, recalls Grund-
meyer. They have exposure to
a specifc geographic area and
want to lay off that risk from
their balance sheets. The owners
of wind power projects need to
know the points where there are
deviations from the norm that
affect the larger issues of grid
operability scheduling.
Enter Speculators
Operators want to avoid the im-
pact of intermittent wind and
having to buy replacement pow-
er at high prices on short notice.
The conventional swap ap-
proach would involve fnding a
counterparty to take the other
side of the risk trade, perhaps a
thermal power plant operator.
But maybe it doesnt have to
be a two-sided market, says Jeff
Hodgson, president of Chicago
Weather Brokerage, who thinks
speculators will make the mar-
ket.
With weather, your model
tells you something is going to
happen, says Hodgson, a for-
mer Merrill Lynch wealth man-
agement adviser who started
CWB in 2009 when he was ex-
cited by the CMEs introduction
of snowfall-related products.
If you think its going to snow
over 60 inches in Chicago, you
look for someone to take the
other side of the trade it is two
speculators betting against each
other. Someone will be right and
someone will be wrong, a zero-
sum game.
Hodgson is usually the expert-
educator on the subject, but he
tells of attending a conference
of snow-removal businesses and
hearing how they understood de-
rivatives as a way to off-load risk.
Theyd come up with ideas for
using futures and options that I
hadnt considered yet.
In only the second year of
trading, volumes in snowfall fu-
tures and options at the CME
have increased more than 400%.
We anticipate exponential
growth again next season, says
Hodgson.
These contracts benefts, he
says, include their non-correla-
tion with stock and bond mar-
kets.
Global Interest
With a tradable marketplace,
adds Hodgson, the farmer has
the ability to trade out of a por-
tion, or all, of the position prior
to expiration. This is something
that is unique to this market.
More than 1.4 million weath-
er derivative contracts, a record,
were written globally in the 12
months through March 2011,
according to the Washington-
based WRMA. The aggregate
notional value of OTC contracts
was $2.4 billion, out of a total
market of $11.8 billion.
Temperature contracts are the
most traded customized weather
hedge. Growth was also seen in
rainfall, snow, hurricane and
wind contracts, refecting end-
user participation from a wider
variety of industries such as ag-
riculture, construction and trans-
portation.
Weather risk is taking its place
alongside fnancial and valuation
risks as a factor in business rev-
enues and bottom lines. Recent
guidance from the Securities and
Exchange Commission focuses
on disclosure of climate-change
impacts. Investors may not want
to hear companies citing weather
as an excuse for missed earnings.
UPFRONT
If you manage risk, money or investments,
you recognize how critical it is to stay on
top of global industry developments
and to understand what they mean for
your organization.
GARP provides its Members with the expert risk information
and the global professional connections required to excel.
From the latest research and premium risk content, to
Members-only education and networking opportunities,
you can benefit from the depth and breadth of GARPs
global community.
Become a Member of GARP today and maintain your
professional advantage.
TM
AGLOBAL
NETWORK.
The ERM
Difference
Principal Financial relied on enterprise risk management
in an aggressive global expansion begun in the wake of the
dot-com crash and the post-9/11 downturn.
By Mi chael Shari
C R O - C E O D I A L O G U E
he Principal Financial Group was ahead of its
time in relying on risk management as a strate-
gic guide in insurance and retirement businesses
launched more than three decades ago. Risk man-
agement also provided underpinnings of decisions
such as the companys demutualization and initial public offer-
ing on October 26, 2001, just six weeks after the September 11
terrorist attacks.
While the U.S. was preparing for war amid continuing after-
shocks from the previous years dot-com bust, Principal set itself
apart from its peers by making acquisitions in Asia, Australia,
Europe and Latin America. Still aggressive on that front, the
company announced three acquisitions this year: global equity
investment frm Origin Asset Management of London in July;
pension manager HSBC AFORE of Mexico City in April; and
emerging markets fund manager Finisterre Capital of London,
also in April.
Founded in 1897, Des Moines, Iowa-based Principal is the
T
11th largest life insurer in the U.S., as well as the biggest man-
ager of bundled 401(k) retirement plans. It also owns Principal
Bank, which opened in 1998 after being one of the frst online
banks approved by its then-regulator, the Offce of Thrift Su-
pervision.
Because of its focus on small and medium-size companies,
which tend to outperform larger ones in an anemic economy,
Principals retirement business has largely averted the risk of
shrinking payrolls at a time of steadily high unemployment.
Whats more, senior vice president and chief risk offcer Greg
Elming sees a natural hedge in the fact that Principals $54
billion general account is divided into more than 10 distinct seg-
ments to match distinct businesses and the geographic diversity
of those businesses. Principal reported $9.16 billion in revenue
and $717 million in net income last year. It has offces in 15
countries, 16.5 million customers and $336 billion in assets un-
der management.
Elming and CEO Larry Zimpleman are both Iowa natives
P
h
o
t
o
g
r
a
p
h
s

b
y

2
0
1
1

S
c
o
t
t

S
i
n
k
l
i
e
r
Our net income held up better because of a
greater focus on enterprise risk management, and
that served us well during the fnancial crisis.
~ Larry Zimpleman, CEO, Principal Financial Group
who grew up within the organization.
Elming, a 29-year Principal veteran and
University of Iowa graduate who had
been comptroller for 10 years, became
CRO in April, succeeding Ellen Lama-
le, who had been with the insurer since
1976. Zimpleman, a 40-year veteran of
the company, a Drake University MBA
and now a trustee of the Des Moines
school, has been president since 2006,
CEO since 2008 and chairman since
2009.
Zimpleman, 60, and Elming, 51,
agree and are acutely aware that the
risks Principal faces are growing geo-
metrically as the company expands.
To manage those exposures, they rely
on a force of 300 risk professionals in
the CROs offce, in the business units
and even in shared services like human
resources. Their common touchstone is
a multi-page grid in which all of those
risks are painstakingly catalogued as f-
nancial or operational and then broken
down into several levels of specifcity,
from 30,000-foot overviews to granu-
lar close-ups. The executives candidly
walked Risk Professional through Princi-
pals risk management framework and
positioning relative to competitors in
this recent interview.
How did risk management devel-
op historically at Principal?
ZIMPLEMAN: In the late 1970s and
early 80s, we began to take some pret-
ty early and ground-breaking steps.
We had moved from the 1950s, when
interest rates were 3% to 4%, to the
late 1970s, when they were 20%, and
short rates were higher than long rates.
That had tremendous ramifcations for
how we thought about how to man-
age a multiline company like we were.
That ah-ha! moment told us a couple
things. You cant operate a single pool
of investments attached to multiple
businesses lines, because the character-
istics of life insurance are very different
from health insurance. We were one of
the very early insurance companies to
split up our general account into, Id
say, 10 to 12 investment pools. We be-
gan to adopt much more detailed stud-
ies of asset and liability management.
This was all going on in the early 1980s,
when that was just not the way that the
business was done.
What role did risk management
play in strategic moves such as
demutualization and expansion
in emerging markets?
ZIMPLEMAN: The reality of being a
mutual company was that the value of
the company was going to a small block
of individual policyholders. That was
not a fair or equitable sharing of the
value that was being put on the books.
Most of the customers were coming
into the retirement and mutual fund
spaces. They were not participating
policyholders. So we outgrew the mu-
tual form. Hand-in-hand with being a
public company was a need to focus on
an opportunity that would have a long
shelf life to it. We said one of the things
we do really well is that we have this
focus on enterprise risk management,
and we do really well in the U.S. retire-
ment space. So lets do a global scan for
countries that offer the best long-term
opportunity, particularly for small and
medium businesses. We started in the
early 2000s in greenfeld businesses
in markets like Mexico, Brazil, Hong
Kong, and Malaysia. We were 10 or 12
years ahead of our time in identifying
that opportunity and getting ourselves
into these markets.
Can you contrast Greg Elmings
approach as CRO with that of his
predecessor?
ZIMPLEMAN: When we started this
ERM effort in the 1980s, it was virtually
100% focused on mitigating fnancial
or economic risks. As we become more
global, we are getting into new levels of
currency risk and different levels of po-
litical risk. In the last three to fve years,
a whole new bucket of what I would call
business or operational risks has been
added to the suite of risks that the CRO
is responsible for -- reputational risks,
compliance risk, IT risk, cybersecurity
risk.
ELMING: It is not that Ellen [Lamale]
was not necessarily looking at [those
risks]. Its just perhaps placing a little
more emphasis on them, given the en-
vironment today. A lot of those things
were what I had to rock-and-roll on in
my comptroller role. But the vast major-
ity of my focus is to enhance and aug-
ment our ability around the operational
risk efforts at the specifc time that those
things are becoming probably more vo-
luminous, more complex, more involved
than maybe they have been in our his-
tory.
How would you describe the em-
powerment of the CRO?
ELMING: I dont want to imply too much
about the independence and oversight of
the job. But I have access to a department
that has all the professionals I need. Col-
laboration is huge here. We have no less
than 100 different risk committees within
the organization, dealing with all aspects
of managing risk. Plus I get the pleasure
of having quarterly discussions with our
board and very active subcommittees of
the board, whom I talk with frequently.
It cascades from Larry, from the board,
and is accepted as a good, solid business
practice by all the business leaders, which
I think is the ultimate blessing of empow-
erment, if you will.
Do you see yourself as a peer of
the CEO?
ELMING: I am not sure Id put myself
that high up, but I would consider my-
self as having the ear of anybody that I
need to have the ear of.
ZIMPLEMAN: Relative to our chief
information offcer, our chief invest-
ment offcer, our chief marketing offcer,
Gregs role, title and compensation are
absolutely on the same plane.
Is risk management closer to the
front offce at Principal than it
would be at other companies?
ELMING: We believe that risk man-
agement ought to be in three lines of
defense. The frst is the [300] business-
unit risk professionals in our fve key op-
erations as well as corporate functions
such as IT and human resources, which
span various business units. We feel very
strongly that the responsibility for iden-
tifying, prioritizing and mitigating all as-
pects of enterprise risk needs to sit with
the business units. We also need to know
we have oversight coordination, with the
greater good and consolidation perspec-
tive in mind. That is where we have the
CRO offce, which is our second line
of defense. I take an independent look
at things with a little bit of a skeptical
perspective. I report directly to Larry, so
there isnt anybody clouding my judg-
ment or tweaking my perspective. The
third line of defense is internal audit,
where we have 25 risk professionals from
a variety of backgrounds, be it opera-
tions or fnance or IT, who go in and ac-
tually kick the tires, independently assess
whether we are doing what we say we
are doing, and make sure that some oth-
er risk isnt there that we havent identi-
fed yet.
ZIMPLEMAN: Whats key is that the
CRO and the business units are in ac-
tive discussions. This is not someone
who is removed from those discussions
and is then asked later to provide some
sort of input or opinion about it. They
are sitting right there at the time when
discussions are going on, with an equal
voice, if you will, as to whether this is
something that the company does or
does not want to move forward on. Its
not someone being at a level higher than
someone else saying, I will pull the rank
card. It would be more around, Heres
a competitors product. If we are going
to offer something similar, what would
we do to make it commercially viable?
When that dialogue between risk pro-
fessionals and businesspeople would get
created, what would come back is, You
cant cover all those risks, so you will
excuse my English have to go naked on
those, which is not something you want
to do. Or you might be able to do it, but
the price is so outrageously high that the
product is not commercially viable.
What pitfalls have you avoided?
ZIMPLEMAN: The poster child is
variable annuities. I cant tell you how
many meetings I went to in 2004, 2005
and 2006, where we were continuously
asked, Why arent you selling variable
annuities in the wealth advisory chan-
nel? Our answer was, Well, we are
just not comfortable with the kind of
contract provisions, the pricing and the
subsequent hedging that we would have
to do. Its not as if we arent in those
products. For us, variable annuities with
living benefts is about $2 billion. At a
company like MetLife, it would be well
over $100 billion.
Any other areas avoided?
ZIMPLEMAN: If you look at our peers,
many would probably have 15% to 20%
of their general-account investment
portfolio in some sort of asset backed by
residential real estate. For us, you would
fnd it at more like 2%. We never invest-
ed much in those securities because we
were never that comfortable with the du-
rational and economic make-whole con-
ditions around them. And that is today
a huge, huge positive for us because of
the continuing trouble in residential real
estate, which we completely avoided.
Sell-side analysts have described
your exposure to commercial real
estate as a bit outsized.
ZIMPLEMAN: With any investment
portfolio, you have to decide where you
want to take your incremental oppor-
tunities. Commercial mortgages have
proven to be a much better source for
investments in insurance company ac-
counts than residential real estate. The
long-term historical loss ratios are quite
low. All of our commercial mortgages
are underwritten by our own people.
Very few companies have that level of
commitment. We would have exposure
in our general account in rough terms
today that is 20% to 25% in commercial
mortgages, as compared to peers at, lets
say, 10%.
To what extent is Principal safe
from experiencing what happened
at Lehman Brothers, Bear Stea-
rns, AIG and UBS?
ZIMPLEMAN: It is very hard for me
to relate to companies where internal
processes were such that CROs or risk
Collaboration is huge here.
We have 100 different risk committees.
~ Greg Elming, CRO, Principal Financial Group
professionals were considered second
class. We were as surprised and shocked
as most people were when we found out
how little attention was being paid to
risk. To some degree, that exists even in
the insurance industry. Manulife Finan-
cial is probably the best example of a
company that was held out from 2005
to 2007 as having very astute enterprise
risk management. Then you come to
fnd out that their variable annuity busi-
ness had absolutely no hedging and no
programs in place. It has cost them very
signifcantly.
How do you distinguish Princi-
pals risk management from that
of peers?
ZIMPLEMAN: Look at our net income,
which shows not only how operating earn-
ings performed, but also how the invest-
ment portfolio performed, from 2007 to
2009, from peak to trough. This is where
many of the ERM elements get embed-
ded into the portfolio performance. In
our case, net income dropped by roughly
50%. But for our peer group Pruden-
tial, Lincoln Financial, AXA and others
the entire net income was wiped out, on
average. We believe our net income held
up better because we have been operat-
ing with a greater focus on enterprise risk
management, and that it served us well
during the fnancial crisis.
How do you integrate risk man-
agement with portfolio manage-
ment?
ELMING: The way we view asset man-
agement risk, primarily as it relates to our
asset management operations, is mostly
around operational risk or business risk
the things associated with business
practice, people, process, systems and
non-economic factors that can infuence
us. Our asset management professionals,
including risk professionals, are closely
integrated with our general account, or
our Principal Life and investment man-
agement professionals, too, to make sure
we are managing things consistently. We
have both at the table every week when
we discuss the general account invest-
ment strategies.
Does the weekly meeting have a
name?
ELMING: It is the Principal Life In-
surance Investment Committee, Friday
mornings at 9:30. The primary mem-
bers are Julia Lawler, chief investment
offcer, who chairs the committee; the
CRO, CEO, and CFO [Terrance Lil-
lis]; two or three key business unit lead-
ers who are generating the liabilities that
we invest from; and James McCaughan,
president of Principal Global Investors.
Is risk management seen as hav-
ing helped mitigate losses, or as a
contributor to profts?
ELMING: Both. We view risk from both
lenses and gather data, analyze, track,
and set tolerances that are both lower-
bounded and upper-bounded. We are
not just looking to keep disasters from
happening. We also have measurements
to make sure we are taking enough risk,
and taking advantage of situations that
we need to take advantage of.
What is happening in the compli-
ance area?
ELMING: One of the key things con-
fronting the industry and me in my
job, personally is the food and furry
of regulatory, legislative, accounting and
compliance activity. It is pretty easy to
look at that stuff and get buried with,
Oh my gosh, Ive got to fnd a way to
comply with all this stuff and avert di-
saster. But we also look at that as, OK,
where are the opportunities in this?
What is your view on interest
rates?
ZIMPLEMAN: That is the type of
thing we have a very vigorous discus-
sion around at the investment commit-
tee meetings. Two to three years from
now, we are [still] going to be talking
about relatively low interest rates. There
is a view that, over time, there is some
chance that they would migrate higher,
primarily driven by infation, which is
coming out of the emerging economies
moreso than the developed economies.
ELMING: When managing interest
rate risk, you need to be pretty astute
and controlled in derivative utilization,
and that is something we monitor and
control very tightly. Currently, we utilize
over-the-counter derivatives. Of course,
youve got to manage exposures to spe-
cifc counterparties and make sure you
dont get outside your limits. Ellen and
Julia set up really good mechanisms on
that front.
This has been an unpredictable en-
vironment in which to deploy $700
million in capital, which Principal
had said it would. How are you
handling the risks involved?
ZIMPLEMAN: Since thats a fairly large
amount, you ought to expect us to be
active in all areas. We have done about
$350 million in M&A and about $450
million of share repurchases. And that
$700 million, by the way, became about
a $1 billion because of favorable events
during the year. That doesnt mean we
have to spend it. This is not the drunken
sailor syndrome. The board may simply
choose to see what happens over time.
C R E D I T R I S K
ver the last few years, and in particular in the light
of the fnancial crisis of 2007-2008, stress testing
has become a key tool in the risk management
arsenal of fnancial institutions and a crucial
input into the decision making process of a banks
board and top management. It is also a core component of the
dialogue between banks and regulators as part of the Supervi-
sory Review Process (Pillar 2 of Basel II).
In order to be a truly effective risk management tool, stress
and scenario analysis must make clear how a given scenario
impacts different portfolios, highlighting the assumptions on
which the analysis is based. Equipped with this information,
a frm will then be in a position to manage its portfolios pro-
actively, ensuring that the impact of a given stress is within the
frms tolerance to risk. Ultimately, robust stress testing is about
preparing for anything that might happen.
In this article, we specifcally focus on stress testing for banks
corporate portfolios and describe an approach that allows a
frm to assess in a transparent fashion how a given scenario
affects banks capital requirements for corporate portfolios, via
estimating stressed PDs and LGDs.
Our framework has three main advantages:
First, it links high-level scenarios to a more granular de-
scription of the economy, by estimating values for virtually any
Stress Testing: A Robust
End-to-End Approach
There is one approach for stress testing corporate
portfolios that can not only yield credible, transparent
results but also lead to improved accuracy of probability
of default and loss-given default forecasts.
By Bruno Eklund, Evgueni a Iankova, Alessandra Mongi ardi no,
Petroni lla Ni coletti , Andrea Serafi no and Zoran Stani savlj evi c
O
variable that can be considered to be a risk driver for a sector.
As such, it improves the accuracy of probability of default (PD)
and loss-given default (LGD) forecasts.
Second, by considering a large set of macroeconomic vari-
ables in a sound econometric model, it contributes to produce
credible and transparent results.
Third, it facilitates the identifcation of portfolios, or parts
of portfolios, that are particularly vulnerable to stressed condi-
tions. As such, it provides a sound basis for proactive risk and
portfolio management.
The main aim of stress testing is to evaluate the impact of se-
vere shocks on a bank and assess its ability to withstand shocks
and maintain a robust capital or liquidity position. If properly
conducted, it sheds light on vulnerabilities otherwise not identi-
fed, informs senior management in the decision-making pro-
cess, and underpins risk-mitigating actions to ensure the long
term viability of the frm.
When a scenario is set for example, either by a banks
management or by the regulator it is typically only articu-
lated on the basis of a few variables, such as GDP and infation.
However, the determinants of the solvency of frms, which are
likely to be sector specifc, may not be included among the sce-
nario variables. Trying to impose a relationship only between
the given scenario variables and the PD and LGD may not be
C R E D I T R I S K
appropriate, and may lead to biased and spurious results that
undermine the practical use of the stress analysis.
This article will describe a factor-model approach that al-
lows us to obtain any variables considered to be determinants
of the corporate PDs for each sector, based on broadly de-
fned scenarios. On the basis of this model, it is then possible
to estimate the impact of the scenario on portfolios PDs and
LGDs and, through these, on the banks capital requirements.
This feature underpins a transparent end-to-end approach
for stress testing of corporate portfolios that links the settings
of high-level scenarios all the way to the estimation of capi-
tal requirements under stress. It also provides a proactive risk
management tool as it helps to identify which portfolios are
most affected by a given stress and what mitigating actions
would be required, if any, to ensure a frms fnancial strength
under stress. The end-to-end process is summarized in the
diagram below.
Diagram: From Scenarios to Mitigating
Actions A Robust End-to-End Approach

Our approach not only provides the tools to evaluate the
impact of a scenario on sector-specifc PDs, but is also less
likely to suffer from under-specifcation, a problem shared by
other macroeconomic models. In addition, unlike a purely
judgmental approach to stress testing, it makes every step in
the analysis transparent.
We should point out, though, that the model does not take
into consideration frm-specifc or idiosyncratic shocks, and is
used only to evaluate the impact of macroeconomic stresses
on PDs, LGDs and capital requirements. Furthermore, our
discussion focuses specifcally on assessing the impact of stress
and scenario analysis for banks corporate portfolios. The ex-
tension of this approach to other types of portfolios for
example, sovereign is an avenue for further work.
The article is divided in sections, loosely following the dia-
gram above. Section I explores the link between base case,
stress scenarios and losses. Section II discusses our analyti-
cal approach (the macroeconomic model, the construction
of factors and their use to replicate scenarios), while Section
III concentrates on modelling risk drivers to obtain scenario-
driven PDs and LGDs. Section IV explains the usefulness of
the approach as a basis not only for forecasting capital and
expected loss (EL), but also for productive discussions involv-
ing different stakeholders and risk mitigating actions. Conclu-
sions are found in Section V.
Section I: Why do Scenarios Matter? Base Case vs.
Stress
Banks proftability, liquidity and solvency depend on many
factors, including, crucially, the economic environment they
face. In the assessment of extreme scenarios, there are two
pressing questions for risk management: (1) What are the
consequences of the extreme scenario on a particular bank?
(2) Under that scenario, what can be done to improve the
banks resilience to the shock?
To address the frst question, the future impact of the
economy under normal conditions (base case) is compared
with that of the extreme scenario (stress). By affecting the
risk drivers, the stress shifts the loss distribution (see Figure 1,
pg. 21) and raises expected and unexpected losses and capi-
tal requirements. To understand the extent of the impact on
those variables, it is important to consider the link between
risk drivers and economic dynamics.
Our model does not take into
consideration frm-specifc or
idiosyncratic shocks, and is used
only to evaluate the impact of
macroeconomic stresses on PDs,
LGDs and capital requirements.
High level
scenarios
Full
macroeconomic
scenario - base
case and stress
Factor
analysis and
scenario
replication
Stakeholder
involvement and
mitigating actions
(if necessary)
Capital and
Expected Loss
forecasts - base
case and stress
Scenario driven
PDs and
LGDs
C R E D I T R I S K
Figure 1: Scenarios and Losses
For corporate credit risk, the approach described allows for
forecasting virtually any variable considered to be a risk driver
given any scenario, and thus provides the basis for a robust an-
swer to the frst question about assessing extreme scenarios. The
results are also useful in discussions on risk mitigating actions.
Section II: The Analytical Framework
The frst building block of our approach is a UK macroeco-
nomic model that uses 500+ UK and US quarterly macro and
fnancial time series from 1980-Q1. The large number of vari-
ables considered over a fairly long sample period, which covers a
few economic cycles, limits the risk of missing important drivers.
This model provides the background for stress and scenario
analysis. To use an analogy, suppose that we want to evaluate
the impact of a stone being dropped unexpectedly in the middle
of a lake. The task is to predict the number of boats sinking and
the number of fatalities. In this example, the probability of a
boat sinking can be viewed as a PD and the mortality rate as
a LGD.
The PDs and LGDs are affected by several factors in-
cluding, for example, the size of the stone; the size of the boat;
the distance between the boat and the stone; the experience of
the captain and crew; the availability of lifeboats; the strength of
the wind; and the proximity to the shore. Assume now that we
only use the frst two variables, which, in this analogy, represent
the set of scenario variables provided. On the basis of those two
variables only, we may draw substantially wrong conclusions
about the actual values of the PDs and LGDs.
On the other hand, the inclusion of too many variables can
make a model unmanageable. Our preferred solution is to build
a relatively small model, while still retaining most of the infor-
mation contained in the dataset by constructing principal com-
ponents.
The basic idea behind this method is that many economic
variables co-move, as if a small set of common underlying and
unobservable elements the principal components (also called
factors) is driving their dynamics. Under this assumption, a
small number of factors explain most of the variation in a large
data set, and thus those few factors can be used to predict the
variables in the dataset quite well.
1

We believe that building a vector autoregressive (VAR) model
based on the factors is a far better solution than the standard
approach of specifying a small VAR model using a subset of the
macro variables. A major drawback of a small VAR model is the
high chance of under-specifcation, which may lead to unreal-
istic estimates and therefore signifcantly limits its practical use.
For example, early VAR applications exhibited a price puzzle
where a positive monetary shock was followed by a counterin-
tuitive increase of the price level.
2,3
Furthermore, it would be
diffcult to construct a meaningful VAR model that also would
include the relevant determinants for the corporate sectors PDs.
Constructing the Factors
After collecting the economic and fnancial data, we use them to
derive factors, which are a parsimonious and manageable de-
scription of the state of the economy. All data are collected into
a single matrix, Xt, where each variable has been transformed
to be stationary and standardized to have a zero mean and unit
variance. The matrix Xt corresponds to the macro variables
box of historical data in Figure 2 below.
Figure 2: From Macro Variables to Factors, and
Vice Versa

Macro variables Factors
(3)
(1)
(2)
Historical
data
Forecast
horizon
C R E D I T R I S K
The principal components or factors can then be construct-
ed using the so-called singular value decomposition, which ex-
presses the matrix Xt as a product of three separate matrices:
where N is the number of observations and K is the number
of variables in the standardized data set X.
4
Using this decom-
position of the matrix X, the factors are constructed as
Note that in this step, we obtain as many factors as variables
included in the data set. Equation (2) corresponds to arrow (1)
in Figure 2 (see pg. 21), linking the macro data to the factors.
The data set can be retrieved by post-multiplying Ft with At,
arrow (2) in Figure 2, and thus mapping the factors back to the
macro data, as follows:
Since the general idea with using principal component fac-
tors is to reduce the number of variables that needs to be in-
cluded in the analysis, we only use the frst r < K factors. This
limited number of factors summarizes the information con-
tained in the underlying macroeconomic variables effciently.
5
When neglecting the remaining columns of Ft, a small er-
ror is normally made when transforming the factors back to
the macro data set. The re-constructed matrix Xt can thus be
expressed as a linear combination of the factors plus an error,
as follows:
where Ft and At contain the frst fve columns of Ft and At,
respectively, and t contains the estimated error made by not
using all factors. This relationship is used to retrieve forecasts
for the data matrix Xt from the factor forecasts over a given
horizon.
The factors can be compared to indices i.e., they can be
viewed as weighted averages constructed using a number of
different variables. For our UK macroeconomic model, four
out of the fve estimated factors have direct interpretations that
allow us to gain valuable insights into the fnal results. The frst
factor is highly correlated with GDP growth and related vari-
ables, and can be thought of as a proxy for economic activity.
Similarly, the second factor is related to asset prices, the third
to real interest rates and the fourth to productivity and employ-
ment costs.
Replicating Economic Scenarios
The framework previously described can be used to obtain
estimates of the risk drivers, even if they are not included in
the set of scenario variables provided, both under a base case
and under a stress scenario. Good replication of a scenario is
important for the overall performance of the stress testing pro-
cedure. Its purpose is to analyze how all other macro and fnan-
cial variables the potential risk drivers are affected by the
given stress scenario and, at a later stage, to analyze how this in
turn affects the estimates of the PDs and LGDs.
As mentioned in the introduction, an important advantage
of our framework is that it can be employed to forecast a wide
variety of economic variables that can be used to derive PDs
and LGDs. In our model, we can replicate a scenario by impos-
ing the dynamics of the scenario on factors and on all other
variables in the data set. The scenario usually consists of ex-
plicit values of the most common macro and fnancial variables
over a given forecast horizon.
The macroeconomic scenario can be represented in Figure
2 (pg. 21, see the vertical lines in the macro data set box). Each
column in this matrix is a separate variable. The lines are dot-
ted over the known sample period and solid over the forecast
horizon.
To be able to estimate the effect of the stressed macro vari-
ables on all other macro variables, we frst need to obtain good
estimates of the factor forecasts in box (3) of Figure 2. Once
these forecasts have been estimated, we can use the back trans-
formation, arrow (2) or equation 4, to re-construct the forecasts
of the remaining macro variables. This fnal step will fll in the
Figure 2: From macro variables to factors and vice versa
(3)
Historical
data
(2)
(1)
Forecast
horizon
The principal components or factors can then be constructed using the so called
singular value decomposition which expresses the matrix as a product of three
separate matrices:
t
X
(1)
T
t t t
,
t
A L U X
where N is the number of observations and K is the number of variables in the
standardized data set X.
4
Using this decomposition of the matrix X, the factors are
constructed as:
(2)
t t t
L U F .
Note that in this step we obtain as many factors as variables included in the data set.
Equation (2) corresponds to arrow (1) in the figure above linking the macro data to
the factors. The data set can be retrieved by post-multiplying with , arrow (2)
in Figure 2, thus mapping the factors back to the macro data:
t
F
T
t
A
(3)
t
T
t
.
t t
T
t t
X A L U A F
Since the general idea with using principal component factors is to reduce the number
of variables that needs to be included in the analysis, we only use the first r < K
factors. This limited number of factors summarizes the information contained in the
underlying macroeconomic variables efficiently.
5
4
U
t
and A
t
are (N N) and (K K) orthonormal matrices, U
t
T
U
t
=I
N
, A
t
T
A
t
=I
K
, L
t
is a (NK)
diagonal matrix with nonnegative elements in decreasing order.
5
The methodology developed by Bai and Ng (2002) has been adopted to determine the optimal number
of factors to use, in our case five, which explains a major proportion of the variance in .
t
X
6
(3)
Historical
data
(2)
(1)
Forecast
horizon
separate matrices:
t
X
(1)
T
t t t
,
t
A L U X
4
constructed as:
(2)
t t t
L U F .
t
F
T
t
A
(3)
t
T
t
.
t t
T
t t
X A L U A F
5
4
U
t
and A
t
t
T
U
t
=I
N
, A
t
T
A
t
=I
K
, L
t
is a (NK)
5
t
X
6
(1)
(2)
(3)
Historical
data
(2)
(1)
Forecast
horizon
separate matrices:
t
X
(1)
T
t t t
,
t
A L U X
4
constructed as:
(2)
t t t
L U F .
t
F
T
t
A
(3)
t
T
t
.
t t
T
t t
X A L U A F
5
4
U
t
and A
t
t
T
U
t
=I
N
, A
t
T
A
t
=I
K
, L
t
is a (NK)
5
t
X
6
When neglecting the remaining columns of a small error is normally made when
transforming the factors back to the macro data set. The re-constructed matrix can
thus be expressed as a linear combination of the factors plus an error:
t
F
t
X
(4)
t
T
t t t
A F X
~ ~
,
where
t
F
~
and
t
A
~
contain the first five columns of and respectively, and
t
F
t
A
t
contains the estimated error made by not using all factors. This relationship is used to
retrieve forecasts for the data matrix from the factor forecasts over a given
horizon.
t
X
The factors can be compared to indices, i.e. they can be viewed as weighted averages
constructed using a number of different variables. For our UK macroeconomic model,
four out of the five estimated factors have direct interpretations which allow us to gain
valuable insights into the final results. The first factor is highly correlated with GDP
growth and related variables and can be thought of as a proxy for economic activity.
Similarly, the second factor is related to asset prices, the third to real interest rates and
the fourth to productivity and employment costs.
II.3. Replicating economic scenarios
The framework above can be used to obtain estimates of the risk drivers even if they
are not included in the set of scenario variables provided, both under a base case and
under a stress scenario. Good replication of a scenario is important for the overall
performance of the stress testing procedure, since the purpose is to analyse how all
other macro and financial variables, the potential risk drivers, are affected by the
given stress scenario and, at a later stage, analyse how this in turn affects the
estimates of the PDs and LGDs.
As mentioned in the introduction, an important advantage of our framework is that it
can be used to forecast a wide variety of economic variables that can be used to derive
PDs and LGDs. In our model, we can replicate a scenario by imposing the dynamics
of the scenario on factors and on all other variables in the data set. The scenario
usually consists of explicit values of the most common macro and financial variables
over a given forecast horizon.
The macroeconomic scenario can be represented in Figure 2 above by the vertical
lines in the macro data set box. Each column in this matrix is a separate variable. The
lines are dotted over the known sample period and solid over the forecast horizon. To
be able to estimate the effect of the stressed macro variables on all other macro
variables, we first need to obtain good estimates of the factor forecasts in box (3).
Once these forecasts have been estimated we can use the back transformation, arrow
(2) or equation 4, to re-construct the forecasts of the remaining macro variables. This
final step will fill in the forecast horizon of all macro variables, not only for the
variables that were included in the scenario, and will thus allow us to use any macro
variable we desire in later steps of our stress testing process.
7
(3)
(4)
~
~
An important advantage of
our framework is that it can
be employed to forecast a wide
variety of economic variables
that can be used to derive PDs
and LGDs.
T
C R E D I T R I S K
forecast horizon of all macro variables (not just for the variables
that were included in the scenario), and will thus allow us to use
any macro variable we desire in later steps of our stress testing
process.
Because the replication of the scenario is subject to an ap-
proximation error, it is important that this process not only can
replicate the given scenario variables but also gives rise to realis-
tic forecasts of the other variables especially the possible risk
drivers. Below, we describe two alternative methods that can be
used to replicate scenarios, and the conditions under which one
is preferable to the other.
The frst method calls for the inclusion of the variables in the
stress scenario as exogenous variables in a VAR model (with
factors as endogenous variables), specifying a so called VARX
model. This can be depicted as follows:
where (L) represents a lag polynomial vector of fnite or-
der d, (L) is a lag polynomial of fnite order s, t is an error
term and the matrix Ht contains the variables whose fore-
casts are given in the scenario such as, for example, GDP
growth and CPI infation. The model should be specifed to
produce the best possible forecast performance and/or the
best ft. Since the stressed variables have known paths over
the forecast horizon, the factor forecasts (box (3) in Figure 2)
can easily be estimated.
6

The second method, an alternative to the VARX model,
is to model the factors individually, which corresponds to set-
ting the lag polynomial (L) equal to zero in equation (5).
In some cases, the second method is better at replicating
the scenario than the VARX, possibly because it is more par-
simonious. Including lagged dependent variables, as in the
VARX model, might put a lot of weight on the history of the
factors and could thus be worse at capturing the given stress
dynamics. Individual factor models will also guarantee that
any given shock to a macro variable will feed into the factor
forecasts directly.
However, this second method might miss key interactions
between the factors. The VARX also performs better at fore-
casting factors with low correlation to the set of scenario
variables given, and when the set of scenario variables avail-
able is very limited.
One fnal point to remember is that a successful scenario
replication requires an economic scenario derived using an
analytical approach, so that the included variables are eco-
nomically consistent with each other. This is especially im-
portant for stressed scenarios where variables might exhibit
unexpected relationships between each other. Any ad hoc
choices can severely affect any step of the stress testing pro-
cess, introducing spurious results and unrealistic effects on
macro and fnancial variables, and in turn on the forecasts
for PDs and LGDs.
Section III: Scenario-driven PDs and LGDs
Once a scenario has been replicated, forecasts of the associ-
ated risk drivers can be obtained. These estimates are then
used when modeling and deriving forecasts of the PDs and
LGDs, which will then drive the capital and EL estimation.
7

Our framework has two distinct advantages: (1) it produces
estimates of the risk drivers by linking them to the macroeco-
nomic environment; and (2) it takes into account the hetero-
geneity across business sectors.
Regarding the frst aspect, our approach is original because
it models the relationship between systematic factors, such as
GDP growth and interest rates, and PDs and LGDs, while
incorporating as much information as possible from the eco-
nomic environment. Several papers estimate links between
PDs and economic variables; however, our framework can
use both macro variables and factors as explanatory variables
for the PDs and LGDs.
As for the second aspect, our approach can differentiate
between sectors in order to identify the specifc vulnerabili-
ties of a particular scenario. Different sectors are sensitive to
sector-specifc drivers and respond differently to the system-
atic risk. Therefore, we model each sector PD separately and
use variables related to that particular sector e.g., house
prices for real estate in order to capture the sector-specifc
dynamics.
8
Section IV: Estimation of Stressed Expected Loss
and Capital, and Risk Mitigating Actions
After obtaining forecasts of PDs and LGDs, we can revert
to our original question of assessing the impact of the stress
on losses and capital requirements. The results of the stress
test, usually compared to the results under the base case, will
normally show an increase in EL and capital requirements.
Figure 3 (next page) presents some stress test results in which
the blue lines show the response to the base case and the red
lines to the stress.
Because the replication of the scenario is subject to an approximation error, it is
important that this process not only can replicate the given scenario variables, but also
gives rise to realistic forecasts of the other variables, especially the possible risk
drivers.
6
Below we describe two alternative methods that can be used to replicate
scenarios, and the conditions under which one is preferable to the other.
The first method is to include the variables in the stress scenario as exogenous
variables in a VAR model with factors as endogenous variables, specifying a so called
VARX model:
(5)
t t t t
H L F L F

) ( ) (
1
,
where represents a lag polynomial vector of finite order d, is a lag
polynomial of finite order s,
t
is an error term, and the matrix contains the
variables whose forecasts are given in the scenario such as, for example, GDP growth
and CPI inflation. The model should be specified to produce the best possible forecast
performance and/or the best goodness of fit. Since the stressed variables have known
paths over the forecast horizon the factor forecasts, box (3) in the figure, can easily be
estimated.
) (L ) (L
t
H
7
The second method, an alternative to the VARX model, is to model the factors
individually, which corresponds to setting the lag polynomial ) (L equal to zero in
equation (5).
In some cases, the second method is better at replicating the scenario than the VARX,
possibly because it is more parsimonious. Including lagged dependent variables, as in
the VARX model, might put a lot of weight on the history of the factors and could
thus be worse at capturing the given stress dynamics. Individual factor models will
also guarantee that any given shock to a macro variable directly will feed into the
factor forecasts. However, this second method might miss key interactions between
the factors. The VARX also performs better at forecasting factors with low correlation
to the set of scenario variables given, and when the set of scenario variables available
is very limited.
Finally, a successful scenario replication requires an economic scenario derived using
an analytical approach, so that the included variables are economically consistent with
each other. This is especially important for stressed scenarios where variables might
exhibit unexpected relationships between each other. Any ad hoc choices can severely
affect any step of the stress testing process, introducing spurious results and
unrealistic effects on macro and financial variables, and in turn on the forecasts for
PDs and LGDs.
Section III. Scenario-driven PDs and LGDs

7
A method to replicate scenarios exactly has been developed. However, when applying it, forecasts of
many of the non-scenario variables are unrealistic and meaningless. There is an apparent trade-off
between realistic results and the degree of accuracy of the scenario replication.
8
(5)
C R E D I T R I S K
Figure 3: An Example of Stress Test Results
The analysis is then followed by a comprehensive discussion
that challenges both assumptions and results by stakeholders in
different areas of the bank: e.g., risk and fnance. The results
emerging from this debate can then be used to inform risk- mit-
igating actions.
In particular, our framework allows for the evaluation of
corporate sector contributions to capital requirements and ex-
pected losses. If a sector under a certain stress, for example,
drives a rise in capital requirements and/or impairments above
the levels compatible with the banks risk appetite, actions can
be taken to limit the banks exposure to that sector.
In this process, the involvement of senior management and
the board of directors ensures that any decision taken is aligned
with the banks risk appetite and is effectively incorporated into
the wider portfolio and risk strategy.
Section V: Conclusions
The approach presented in this article shows how to establish
a clear link between a broadly defned scenario, often defned
only for a few macro variables, and a fully defned macroeco-
nomic scenario. It also demonstrates how to assess the credit
capital requirements for corporate portfolios via stressed PDs
and LGDs.
The merits of our framework are multiple. First, by estimat-
ing values for virtually any variable that can be considered to
be a risk driver for each corporate sector, it raises the accuracy
of PD and LGD forecasts. Second, by effciently considering
the information contained in a large set of macroeconomic
variables in a sound econometric model, it produces transpar-
ent results, which form the basis of discussion for proactive risk
management. Third, it can derive estimates of PD and LGD
determinants consistent with scenario variables produced by a
banks top management and regulators.
The approach is practical and transparent, and can be used
as a key input to assess the banks capital position at times of
stress, with respect to its own risk appetite as well as regulatory
requirements. Whenever mitigating actions need to be consid-
ered, the framework allows one to identify the specifc portfo-
lios toward which these actions should be targeted.
Of course, the output of the analytical framework should
not be used in a mechanistic way. Rather, it has to be subject to
a critical review based on sound judgment. The combination
of robust modeling and sound management is, we believe, the
basis for good risk management.
FOOTNOTES
1. See Jolliffe (2004) for details.
2. See Sims (1972) and Sims (1980).
3. For more technical information, please see Hamilton (1994), Lt-
kepohl (2005), Sims (1972, 1980), Stock and Watson (2002), and Ber-
nanke, Boivin and Eliasz (2005).
4. Ut and At are (N x N) and (K x K) orthonormal matrices; Ut Ut=IN,
At At=IK, Lt is a (N x K) diagonal matrix with nonnegative elements in
decreasing order.
5. The methodology developed by Bai and Ng (2002) has been ad-
opted to determine the optimal number of factors to use (in our case,
fve), which explains a major proportion of the variance in Xt.
6. A method to replicate scenarios exactly has been developed. How-
ever, when applying it, forecasts of many of the non-scenario vari-
ables are unrealistic and meaningless. There is an apparent trade-off
The approach presented in this
article shows how to establish a
clear link between a broadly defned
scenario, often defned only for a
few macro variables, and a fully
defned macroeconomic scenario.
T
T
Expected Loss
(% of base quarter), Total
160 -
140 -
120 -
100 -
80 -
60 -
Capital
(% of base quarter), Total
130 -
120 -
110 -
100 -
90 -
80 -
70 - - - - - - - - - - - - - - - - - - - - - - -
2
0
1
0

Q
2
2
0
1
0

Q
4
2
0
1
1

Q
2
2
0
1
1

Q
4
2
0
1
2

Q
2
2
0
1
2

Q
4
2
0
1
3

Q
2
2
0
1
3

Q
4
2
0
1
4

Q
2
2
0
1
4

Q
4
2
0
1
5

Q
2
- - - - - - - - - - - - - - - - - - - - - -
2
0
1
0

Q
2
2
0
1
0

Q
4
2
0
1
1

Q
2
2
0
1
1

Q
4
2
0
1
2

Q
2
2
0
1
2

Q
4
2
0
1
3

Q
2
2
0
1
3

Q
4
2
0
1
4

Q
2
2
0
1
4

Q
4
2
0
1
5

Q
2
C R E D I T R I S K
between realistic results and the degree of accuracy of the scenario
replication.
7. Here we focus on modeling probabilities of default. However, a
way forward would be to apply a similar methodology to LGDs.
8. A common problem is that internal time series of PDs normally
are too short, which means, in most cases, that the sample period
does not cover a full credit or business cycle. This is the case for us,
which is why we use MKMV Expected Default Frequencies (EDF)
as a proxy for the PDs.
REFERENCES
Bai, J. and S. Ng (2002). Determining the number of factors in ap-
proximate factor models, Econometrica, 70, 1, 191-221.
Bernanke, B.S., Boivin, J. and P. Eliasz (2005). Measuring the effects
of monetary policy: A factor-augmented vector autoregressive (FA-
VAR) approach, Quarterly Journal of Economics, 120, 1, 387-422.
Forni, M., Giannone, D., Lippi, M. and L. Reichlin (2007). Open-
ing the black box: Structural factor models with large cross-sections,
European Central Bank Working Paper, n.712.
Hamilton, J.D. (1994). Time Series Analysis, Princeton University
Press.
Jolliffe, I.T. (2004). Principal Component Analysis, Springer.
Ltkepohl, H. (2005). New Introduction to Multiple Time Series
Analysis, Springer.
Sims, C.A. (1972). Money, income and causality,American Eco-
nomic Review, 62, 540-552.
ibid. (1980). Macroeconomics and reality, Econometrica, 48, 1-48.
Stock, J.H. and M.W. Watson (2002). Macroeconomic forecasting
using diffusion indexes, Journal of Business & Economic Statistics,
147-162.
Alessandra Mongiardino (FRM) is the head of risk strategy and a member of the risk
management executive team for HSBC Bank (HBEU). She is responsible for develop-
ing and implementing the risk strategic framework, to ensure that risk management is
effectively embedded throughout the bank and across all types of risk.
Zoran Stanisavljevic is the head of wholesale risk analytics at HSBC. After working
at Barclays Capital as a credit risk data manager and a quantitative credit analyst, he
joined HSBC as a senior quantitative risk analyst in 2005. Shortly thereafter, he took
over the leadership on the banks stress testing and economic capital projects, and was
promoted to his current role in 2010.
Evguenia Iankova is a senior quantitative economist at HSBC, where she is currently
involved in several projects, such as scenario building and stressing risk drivers. Follow-
ing a stint as a quantitative economist in the economic research department at Natixis in
Paris, she joined HSBC in 2008 to work on macroeconomic stress testing.

Bruno Eklund is a senior quantitative analyst at HSBC. Before joining HSBC, he
worked at Bank of England, where he built models for stress testing for the UK bank-
ing system. He also worked previously as a researcher at the Central Bank of Iceland,
developing econometric models for forecasting the Icelandic business cycle.
Petronilla Nicoletti is a senior quantitative economist at HSBC, where she focuses on
macroeconomic scenario development, replication for stress testing and the impact of
shocks to risk drivers. Before joining HSBC, she worked as a senior risk specialist at the
Financial Services Authority, the UKs fnancial regulator.
Andrea Serafno is a senior econometrician at the FSA. Prior to joining the regulator in
2010, he worked at HSBC, specializing in macroeconomic stress testing. He has also
served as a consultant for the SAS Institute in Milan.
This article refects the work of HSBC employees at the time of their
employment at HSBC, but does not represent the views of either HSBC
or the FSA.
TrusT The leader.
TrusT Kaplan schweser.
schweser materials and instruction are the best quality and
best value in FrM
exam preparation, allowing you to study

anytime, anywhere with flexible, efficient and effective
resources that ensure success on exam day:
The best materials
schweser study notes are concise, easy
to understand, and truly the most effective
tool for exam preparation offered by
any prep provider. They will save you
countless hours of study.
Our practice exams are highly rigorous
assessment and learning tools designed
to evaluate your exam readiness.
The best instruction
schwesers online classes provide
a truly cost-effective and efficient
way to master the FrM curriculum.
Our instructors all have great
experience in preparing
candidates for the exam
and receive very high
candidate evaluations
year after year.
* Offer ends February 10, 2012
ENROLL TODAY!
www.schweser.com/frm
877-641-3916 608-779-5599
preMiuM
sTudy
pacKages
May 2012 FrM
early seasOn
discOunT
20
%
OFF*
FRM
Exam Prep
C O U N T E R P A R T Y R I S K
CVA desks have been
developed in response to
crisis-driven regulations for
improved counterparty risk
management. How do these
centralized groups differ
from traditional approaches
to manage counterparty
risk, and what types of data
and analytical challenges do
they face?
By Davi d Kelly
How the Credit
Crisis has Changed
Counterparty Risk
Management
he credit crisis and regulatory responses have
forced banks to update their counterparty risk
management processes substantially. New regu-
lations in the form of Basel III, the Dodd-Frank
Act in the U.S. and European Market Infrastruc-
ture Regulation (EMIR) have dramatically increased capital
requirements for counterparty credit risk. In addition to im-
plementing new regulatory requirements, banks are making
signifcant changes to internal counterparty risk management
practices.
There are three main themes inherent in these changes.
First, better frmwide consolidated risk reporting has become
a top priority. Second, centralized counterparty risk manage-
ment groups (CVA desks) are being created to more active-
ly monitor and hedge credit risk. Third, banks are making
signifcant investments in technology to better support the
frmwide risk reporting and CVA desk initiatives.
This article will explore some of the key changes to inter-
nal counterparty risk management processes by tracing typi-
cal workfows within banks before and after CVA desks, as
well as how increased clearing due to regulatory mandates
affects these workfows. Since CVA pricing and counterparty
risk management workfows require extensive amounts of
data, as well as a scalable, high-performance technology, it is
important to understand the data management and analytical
challenges involved.
Before CVA Desks
CVA desks, or specialized risk control groups tasked with more
actively managing counterparty risk, are becoming more
prevalent, partly because banks that had them generally fared
better during the crisis. To establish a basis for comparison,
it is important to review counterparty credit risk pricing and
post-trade risk management before the advent of CVA desks.
The case where a corporate end-user hedges a business risk
through a derivative transaction provides a useful example.
The corporate treasurer may want to hedge receivables in a
foreign currency or lock in the forward price of a commodity
input. Another common transaction is an interest rate swap
used to convert fxed rate bonds issued by the corporation into
foaters to mitigate interest rate risk.
In all of these cases, the corporate treasurer explains the
hedging objective to the banks derivatives salesperson, who
structures an appropriate transaction. The salesperson re-
quests a price for the transaction from the relevant trading
desk, which provides a competitive market price with credit
and capital add-ons to cover the potential loss if the coun-
terparty were to default prior to maturity of the contract. The
credit and capital charges are based on an internal qualitative
and quantitative assessment of the credit quality of the coun-
terparty by the banks credit offcers.
The credit portion of the charge is for the expected loss
i.e., the cost of doing business with risky counterparties. (It is
analogous to CVA, except that the CVA is based on market
implied inputs, including credit spreads, instead of histori-
cal loss norms and qualitative analysis). The capital portion
is for the potential unexpected loss. This is also referred to
as economic capital, which traditionally has been based on
historical experience but is increasingly being calculated with
market implied inputs. These charges go into a reserve fund
used to reimburse trading desks for counterparty credit losses
and generally ensure the solvency of the bank.
The trader on the desk works with the risk control group,
which may deny the transaction if the exposure limit with
that particular counterparty is hit. Otherwise, it provides the
credit and capital charges directly or the tools to allow the
trader to calculate them. The risk control group also provides
exposure reports and required capital metrics to the banks
regulator, which approves the capital calculation methodol-
ogy, audits its risk management processes and monitors its
ongoing exposure and capital reserves according to the Basel
guidelines.
Figure 1: Counterparty Risk Workfow Before
CVA Desks
T
Corporate Treasury
Bank Treasury (funding)
Collateral Management
Exchanges and Dealers
Regulatory Capital &
Reporting
Credit & Capital Reserves
Derivatives Trading Desk
Risk Control
Derivative Pricing
Bank Regulator
Bank Derivatives Sales
Derivative Transaction
Funding
Collateral Risk
Market Risk
While counterparty risk is managed through reserves in this
example, the market risk of the transaction is fully hedged by
the desk on exchanges or with other dealers. If the transac-
tion is uncollateralized, the banks treasury provides funding
for what is effectively a loan to the customer in the form of a
derivative line of credit. Some portion of the exposure may
also be collateralized, in which case there are additional op-
erational workfows around collateral management.
CVA Desks
One of the drivers for CVA desks was the need to reduce
credit risk i.e., to free capacity and release reserves, so
banks could do more business. In the late 1990s and early
2000s, in the wake of the Asian fnancial crisis, banks found
themselves near capacity and were looking for ways to reduce
counterparty risk. Some banks attempted to securitize and
redistribute it, as in JPMorgans Bistro transaction. Another
approach was to hedge counterparty risk using the relatively
new CDS market.
The recent international (2005) and U.S. (2007) account-
ing rules mandating the inclusion of CVA in the mark-to-
market valuations of derivatives positions provided additional
impetus for more precisely quantifying counterparty risk.
Some banks actively attempted to manage counterparty risk
like market risk, hedging all the underlying risk factors of
the CVA. Given the complexity of CVA pricing and hedg-
ing, these responsibilities have increasingly been consolidated
within specialized CVA desks. Now, with the increased capital
charges for counterparty default risk under Basel III and the
new CVA VaR charges, there is even more incentive to imple-
ment CVA desks.
With a CVA desk, most of the trading workfow is the same,
except the CVA desk is inserted between the derivatives trad-
ing desks and the risk control group. Instead of going to the
risk control group for the credit and capital charges, the trad-
er requests a marginal CVA price from the CVA desk, which
basically amounts to the CVA desk selling credit protection on
that counterparty to the derivatives desk. This is an internal
transaction between the two desks, which can be in the form
of a contingent credit default swap (CCDS).
The CVA charge is passed on to the external corporate
customer by adding a spread to the receive leg of the trade.
Unless the CVA is fully hedged, which is unlikely, the spread
charged to the customer should also include some portion of
economic capital to account for CVA VaR and potential un-
expected loss from default.
Since credit risk spans virtually all asset classes, the CVA
desk may deal with multiple internal trading desks. The risk
control group treats the CVA desk like other trading desks,
imposing trading limits and monitoring market risks using
traditional sensitivity metrics and VaR. The CVA desk ex-
ecutes credit and market risk hedges on exchanges or with
other dealers and relies on the banks internal treasury to fund
positions. To the extent the CVA desk attempts to replicate
(hedge) CVA fully, while the derivatives desks also hedge the
underlying market risks, there may be some ineffciencies due
to overlapping hedges.
Figure 2: Counterparty Risk Workfow with
CVA Desk
The key innovation
introduced by CVA desks
is quantifying and manag-
ing counterparty credit
risk like other market risks,
instead of relying solely
on reserves.
Corporate Treasury
Marginal CVA Price
CVA Desk
Derivative Pricing
Funding
Collateral Risk
Market Risk
Funding Credit Risk
Bank Treasury
Exposure Limits
Reporting
Risk Control
Bank Regulator
The key innovation introduced by CVA desks is quantifying
and managing counterparty credit risk like other market risks,
instead of relying solely on reserves. The main challenge is
that not all CVA risk can be hedged due to insuffcient CDS
liquidity and unhedgeable correlation and basis risks. There-
fore, some reserve-based risk management is unavoidable.
Based on trends among the top-tier banks, the optimal solu-
tion involves hedging as much of the risk as possible, consid-
ering the cost of rebalancing hedges in a highly competitive
CVA pricing context, and then relying on experienced trad-
ers well-versed in structured credit problems to manage the
residual exposures.
Clearing
The predominant issues with CVA desks are that they insert
another operational layer into the workfow and add substan-
tial analytical complexity. CVA models have to incorporate
all the market risk factors of the underlying derivative plus
the counterparty credit risk. Even the CVA on a plain vanilla
FX forward is complex, because the counterparty effectively
holds an American-style option to default. In addition to the
option risk profle, the CVA trader must also consider the cor-
relation between the counterpartys default probability and
the underlying market risk factors i.e., wrong-way risk.
Margining formulas are much simpler than CVA and eco-
nomic capital models, and new regulations are either mandat-
ing or heavily incentivizing banks to clear or fully collateralize
derivative transactions. In cleared transactions, the clear-
inghouse effectively replaces the CVA desk in the workfow.
Transactions are assigned or novated to the clearinghouse,
which becomes the counterparty to both sides of the trade.
Counterparty risk is virtually eliminated because the exposure
is fully collateralized and ultimately backed by the clearing-
house and its members.
There is a remote risk that the clearinghouse fails, but the
risk control group can focus on relatively simpler issues like
collateral management, liquidity and residual market risks.
Since cleared transactions are fully margined, the CVA charge
is replaced by the collateral funding cost, which is typically
based on an overnight rate.
Figure 3: Counterparty Risk Workfow
with Clearing

Not all trades can or will be cleared. Clearinghouses will
only be able to handle standardized contracts; moreover, the
Dodd-Frank and EMIR regulations specifcally exempt cor-
porate end-user hedge transactions from mandatory clearing,
so that banks can continue to provide credit lines and risk
transfer services through derivatives. It is estimated that at
least a quarter of derivatives transactions will remain OTC,
which means banks will need to maintain both CVA and
clearing workfows.
Data & Technology Challenges
Reliable data feeds that facilitate regular updates and data
integrity checks are absolutely fundamental to effective coun-
terparty risk management. Banks typically maintain separate
systems for their main trading desks, roughly aligned by the
major asset classes interest rates and foreign exchange,
credit, commodities and equities. Since counterparty risk
spans all asset classes, the counterparty risk system must ex-
Since counterparty risk
spans all asset classes, the
counterparty risk system
must exctract transactions,
market data and reference
data from the various
trading systems.
Corporate Treasury
Bank Treasury (funding)
Reporting
Collateral Funding Cost
Risk Control
Derivative Pricing
Bank Regulator
Funding
Collateral
Market Risk
Credit Risk Clearinghouse
tract transactions, market data and reference data from all
the various trading systems. In addition, supplemental refer-
ence data on legal entities and master agreements may need
to come from other databases. These systems typically use dif-
ferent data formats, symbologies and naming conventions, as
well as proprietary interfaces, adding signifcant complexity.
The next set of challenges involves balancing performance
and scalability with analytical robustness. The simulation en-
gine may have to value on the order of one million transac-
tions over a hundred future dates and several thousand market
scenarios.
Depending on the size of the portfolio and number of risk
factors, some shortcuts on the modeling side may be necessary.
However, given their role in the crisis, there are critical ana-
lytical aspects (such as wrong-way risk) that cannot be assumed
away. Regulators are continuously raising the bar in terms
of modeling every risk factor, such as potential shifts in basis
spreads, volatilities and correlations. New regulatory require-
ments for back-testing and stress testing are designed to ensure
the validity of the model.
Outputs of the simulation engine include current, expected
and potential future exposures, CVA and economic capital by
counterparty. The system should also capture counterparty
exposure limits and highlight breaches. Given the complexity
of the inputs and outputs, a robust reporting system is criti-
cal. The system should allow aggregation of exposure metrics
along a variety of dimensions, including industry and legal
jurisdiction. The system should also allow drilling down into
results by counterparty netting set and individual transactions.
By extension, users should have an effcient means to diagnose
unexpected results.
There may also be a separate system for marginal pricing
of new trades and active management of CVA. The marginal
pricing tools need to access results of the portfolio simulation,
since the price of each new trade is a function of the aggregate
exposure with that counterparty. Because of this, calculating
marginal CVA in a reasonable time frame can be a signifcant
challenge.
The CVA risk management system provides CVA sensitivi-
ties for hedging purposes. Hedges booked by the CVA desk
should fow through the simulation engine, so that they are re-
fected in the exposure and capital metrics. Whereas the CVA
desk may hedge credit and market risks, only approved credit
hedges including CDS, CCDS and, to some extent, credit
indices may be included for regulatory capital calculations.
Cleared transactions must be fed to the clearinghouses sys-
tems and trade repositories. Since clearing involves daily (or
more frequent) margining, the banks collateral management
system should be integrated with the clearinghouse. It should
also be integrated with the counterparty risk system. Ideally,
the simulation engine would upload current collateral positions
and revalue them for each market scenario to determine net
exposure. The simulation engine should also incorporate the
margin period of risk, i.e., the risk of losses from non-de-
livery of collateral.
Figure 4: Counterparty Risk Data Flows &
Technology Infrastructure
Even with the crisis-induced wave of improvements, the
picture remains very complex. The most sophisticated global
banks have targeted the CVA and clearing workfows and sup-
porting technology infrastructure described in this article. It is
expected that regional banks will follow suit over the next few
years in order to optimize capital and manage risk more effec-
tively, as well as comply with new regulations.
David Kelly is the director of credit products at Quantif, a software provider special-
izing in analytics, trading and risk management solutions. He has extensive knowl-
edge of derivatives trading, quantitative research and technology, and currently over-
sees the companys credit solutions, including counterparty credit risk. Prior to joining
Quantif in 2009, he held senior positions at Citigroup, JPMorgan Chase, AIG and
CSFB, among other fnancial institutions. He holds a B.A. in economics and math-
ematics from Colgate University and has completed graduate work in mathematics at
Columbia University and Carnegie Mellon.
Credit Commodities
Trading System(s)
Counterparty Exposure
Simulation Engine
Reporting
Clearing Systems &
Trade Repos
System
CVA Risk Management
System
Marginal CVA Pricing
Tools
Transaction
Market Date
Reference Date
Counterparties
Legal Agreements
IR/FX Equities
Hedges
THE QUANT CLASSROOM BY ATTI LI O MEUCCI
Mixing Probabilities,
Priors and Kernels
via Entropy Pooling
How to emphasize certain historical scenarios for risk
and portfolio management, according to their similarity
with the current market conditions.
he Fully Flexible Probabilities
framework discussed in Meucci (2010) represents the multi-
variate distribution f of an arbitrary set of risk drivers, X
(X
1
,...,X
N
), non-parametrically in terms of scenario-proba-
bility pairs, as follows:
where the joint scenarios can be
historical realizations or the outcome of simulations. The use
of Fully Flexible Probabilities permits all sorts of manipula-
tions of distributions essential for risk and portfolio manage-
ment, such as pricing and aggregation (see Meucci, 2011) and
the estimate of portfolio risk from these distributions.
The probabilities in the Fully Flexible Probabilities frame-
work (1) can be set by crisp conditioning, kernel smoothing, ex-
ponential time decay, etc. (see Figure 1, right).
(1)
Figure 1: Fully Flexible Probabilities
Specifcation via Entropy Pooling
Another approach to set the probabilities in (1) is based on
the Entropy Pooling technique by Meucci (2008). Entropy
Pooling is a generalized Bayesian approach to process views
on the market. It starts from two inputs, a prior market distri-
bution, f
0
, and a set of generalized views or stress tests, , and
yields a posterior distribution f that is close to the prior, but
incorporates the views.
Entropy Pooling can be used in the non-parametric sce-
T
Mixing Probabilities, Priors and Kernels
via Entropy Pooling
Attilio Meucci
1
1 Introduction
The Fully Flexible Probabilities framework discussed in Meucci (2010) repre-
sents the multivariate distribution of an arbitrary set of risk drivers X
(
1

)
0
non-parametrically in terms of scenario-probability pairs,
{x
}
=1
, (1)
where the joint scenarios x
(
1

)
0
can be historical realizations or
the outcome of simulations. The use of Fully Flexible Probabilities permits all
sorts of manipulations of distributions essential for risk and portfolio manage-
ment, such as pricing and aggregation, see Meucci (2011), and the estimate of
portfolio risk from these distributions.
The probabilities in the Fully Flexible Probabilities framework (1) can be
set by crisp conditioning, kernel smoothing, exponential time decay, etc., refer
to Figure 1.
Another approach to set the probabilities in (1) is based on the Entropy
Pooling technique by Meucci (2008). Entropy Pooling is a generalized Bayesian
approach to process views on the market. Entropy Pooling starts from two
inputs, a prior market distribution
0
and a set of generalized views or stress-
tests V, and yields a posterior distribution that is close to the prior, but
incorporates the views. Entropy Pooling can be used in the non-parametric
scenario-probability representation of the Fully Flexible Probabilities framework
(1), in which case it provides an optimal way to specify the probabilities p
(
1

)
0
of the scenarios. Alternatively, Entropy Pooling can be used with
parametric distributions
that are fully specied by a set of parameters ,

such as the normal distribution.
1
The author is grateful to Garli Beibi and David Elliott
1
via Entropy Pooling
Attilio Meucci
1
1 Introduction
(
1

)
0
{x
}
=1
, (1)
(
1

)
0
to Figure 1.
0
(
1

)
0

1
1
via Entropy Pooling
Attilio Meucci
1
1 Introduction
(
1

)
0
{x
}
=1
, (1)
(
1

)
0
to Figure 1.
0
(
1

)
0

1
1
nario-probability representation of the Fully Flexible Prob-
abilities framework (1), in which case it provides an optimal
way to specify the probabilities p (p1,...,p
J
)of the scenarios.
Alternatively, it can be used with parametric distributions f

that are fully specifed by a set of parameters , such as the
normal distribution.
In this article, we show that Entropy Pooling represents the
most general approach to optimally specify the probabilities
of the scenarios (1) and includes common approaches (such as
kernel smoothing) as special cases, when the prior distribution
f
0
contains no information. We will also demonstrate how to
use Entropy Pooling to overlay different kernels/signals to an
informative prior in a statistically sound way.
The remainder of the article will proceed as follows. In
Section 2, we review common approaches to assign prob-
abilities in (1) and we generalize such approaches using fuzzy
membership functions.
In Section 3, we review the non-parametric implementa-
tion of Entropy Pooling. Furthermore, we show how Entropy
Pooling includes fuzzy membership probabilities. Finally, we
discuss how to leverage Entropy Pooling to mix the above ap-
proaches and overlay prior information to different estima-
tion techniques.
In Section 4, we present a risk management case study,
where we model the probabilities of the historical simulations
of a portfolio P&L. Using Entropy Pooling we overlay to an
exponentially time-decayed prior a kernel that modulates the
probabilities according to the current state of implied volatil-
ity and interest rates.
Time/State Conditioning and Fuzzy Membership
Here we review a few popular methods to specify the prob-
abilities in the Fully Flexible Probabilities framework (1) ex-
ogenously. For applications of these methods to risk manage-
ment, refer to Meucci (2010).
If the scenarios are historical and we are at time T, a simple
approach to specify probabilities exogenously is to discard old
data and rely only on the most recent window of time . This
entails setting the probabilities in (1), as follows:
An approach to assign weights to historical scenarios dif-
ferent from the rolling window (2) is exponential smoothing
where > 0 is a given half-life for the decay.
An approach to assigning probabilities related to the rolling
window (2) that does not depend on time, but rather on state,
is crisp conditioning: the probabilities are set as non-null, and
all equal, as long as the scenarios of the market drivers xj lie in
a given domain
. Using the indicator function 1 (x), which is

1 if x X and 0 otherwise, we obtain
An enhanced version of crisp conditioning for assigning
probabilities, related to the rich literature on machine learn-
ing, is kernel smoothing. First, we recall that a kernel k (x)
is defned by a positive non-increasing generator function
k(d) [0.1], a target , a distance function d (x,y) and a radius,
or bandwidth, , as follows:
For example, the Gaussian kernel reads
where the additional parameter is a symmetric, positive def-
inite matrix. The Gaussian kernel (6) is in the form (5), where
d is the Mahalanobis distance
Using a kernel we can condition the market variables X
smoothly, by setting the probability of each scenario as pro-
portional to the kernel evaluated on that scenario, as follows:
As we show in Appendix A.1, available at http://symmys.
com/node/353, the crisp state conditioning (4) includes the
rolling window (2) as a special case, and kernel smoothing (7)
includes the time-decayed exponential smoothing (3) as a spe-
cial case (see Figure 1).
We can generalize further the concepts of crisp condition-
ing and kernel smoothing by means of fuzzy membership
functions. Fuzzy membership to a given set
is defned in
terms of a "membership" function m(x) with values in the
range [0,1], which describes to what extent x belongs to a
2 Time/state conditioning and fuzzy member-
ship
Here we review a few, popular methods to exogenously specify the probabilities
in the Fully Flexible Probabilities framework (1). For applications of these
methods to risk management, refer to Meucci (2010).
If the scenarios are historical and we are at time T, a simple approach to
exogenously specify probabilities is to discard old data and rely only on the
most recent window of time t. This entails setting the probabilities in (1) as
follows
j
1 if t
T t
0 otherwise.
. (2)
An approach to assign weights to historical scenarios dierent from the
rolling window (2) is exponential smoothing
j
ln 2
||
, (3)
where t 0 is a given half-life for the decay.
An approach to assigning probabilities related to the rolling window (2)
that does not depend on time, but rather on state, is crisp conditioning: the
probabilities are set as non-null, and all equal, as long as the scenarios of the
market drivers x
lie in a given domain X. Using the indicator function 1

X
(x),
which is 1 if x X and 0 otherwise, we obtain
j
1
X
(x
) (4)
An enhanced version of crisp conditioning for assigning probabilities, related
to the rich literature on machine learning, is kernel smoothing. First, we recall
that a kernel /
(x) is dened by a positive non-increasing generator function

/ (d) [0, 1], a target , a distance function d (x, y), and a radius, or bandwidth,
c as follows
/
(x) /(
d (x, )
c
). (5)
/
(x) c
1
2
2
(x)
0
1
(x)
, (6)
where the additional parameter is a symmetric, positive denite matrix. The
Gaussian kernel (6) is in the form (5), where d is the Mahalanobis distance
d
2
(x, ) (x )
0
1
(x ).
Using a kernel we can condition the market variables X smoothly, by setting
the probability of each scenario as proportional to the kernel evaluated on that
scenario
j
(x
). (7)
3
(2)
ship
follows
j
1 if t
T t
0 otherwise.
. (2)
j
ln 2
||
, (3)
market drivers x

X
(x),
j
1
X
(x
) (4)
that a kernel /

c as follows
/
(x) /(
d (x, )
c
). (5)
/
(x) c
1
2
2
(x)
0
1
(x)
, (6)
d
2
(x, ) (x )
0
1
(x ).
scenario
j
(x
). (7)
3
ship
follows
j
1 if t
T t
0 otherwise.
. (2)
j
ln 2
||
, (3)
market drivers x

X
(x),
j
1
X
(x
) (4)
that a kernel /

c as follows
/
(x) /(
d (x, )
c
). (5)
/
(x) c
1
2
2
(x)
0
1
(x)
, (6)
d
2
(x, ) (x )
0
1
(x ).
scenario
j
(x
). (7)
3
ship
follows
j
1 if t
T t
0 otherwise.
. (2)
j
ln 2
||
, (3)
market drivers x

X
(x),
j
1
X
(x
) (4)
that a kernel /

c as follows
/
(x) /(
d (x, )
c
). (5)
/
(x) c
1
2
2
(x)
0
1
(x)
, (6)
d
2
(x, ) (x )
0
1
(x ).
scenario
j
(x
). (7)
3
ship
follows
j
1 if t
T t
0 otherwise.
. (2)
j
ln 2
||
, (3)
market drivers x

X
(x),
j
1
X
(x
) (4)
that a kernel /

c as follows
/
(x) /(
d (x, )
c
). (5)
/
(x) c
1
2
2
(x)
0
1
(x)
, (6)
d
2
(x, ) (x )
0
1
(x ).
scenario
j
(x
). (7)
3
ship
follows
j
1 if t
T t
0 otherwise.
. (2)
j
ln 2
||
, (3)
market drivers x

X
(x),
j
1
X
(x
) (4)
that a kernel /

c as follows
/
(x) /(
d (x, )
c
). (5)
/
(x) c
1
2
2
(x)
0
1
(x)
, (6)
d
2
(x, ) (x )
0
1
(x ).
scenario
j
(x
). (7)
3
ship
follows
j
1 if t
T t
0 otherwise.
. (2)
j
ln 2
||
, (3)
market drivers x

X
(x),
j
1
X
(x
) (4)
that a kernel /

c as follows
/
(x) /(
d (x, )
c
). (5)
/
(x) c
1
2
2
(x)
0
1
(x)
, (6)
d
2
(x, ) (x )
0
1
(x ).
scenario
j
(x
). (7)
3
(4)
(5)
(6)
(7)
(3)
given set
.
Using a fuzzy membership mx for a given set
of potential
market outcomes for the market
, we can set the probability

of each scenario as proportional to the degree of membership
of that scenario to the set of outcomes
A trivial example of membership function is the following
indicator function, which defnes crisp conditioning (4):
With the indicator function, membership is either maximal
and equal to 1, if x belongs to
, or minimal and equal to 0,

if x does not belong to
.
A second example of membership function is the kernel (5),
which is a membership function for the singleton
The membership of x to is maximal when x is . The larger

the distance d of x from , the less x "belongs" to , and thus
the closer to 0 the membership of x.
Entropy Pooling
The probability specifcation (8) assumes no prior knowledge
of the distribution of the market risk drivers X. In this case,
fuzzy membership is the most general approach to assign
probabilities to the scenarios in (1). Here, we show how non-
parametric Entropy Pooling further generalizes fuzzy mem-
bership specifcations to the case when prior information is
available, or when we must blend together more than one
membership specifcation.
First, we review the non-parametric Entropy Pooling imple-
mentation. (Please refer to the original article Meucci (2008)
for more details and more generality, as well as for the code.)
The starting point for non-parametric Entropy Pooling is
a prior distribution f
(0)
for the risk drivers X, represented as
in (1) by a set of scenarios and associated probabilities, as fol-
lows:
The second input is a view on the market X, or a stress-
test. Thus, a generalized view on X is a statement on the
yet-to-be-defned distribution defned on the same scenarios
A large class of such views can be
characterized as expressions on the expectations of arbitrary
functions of the market v(X), as follows:
where v* is a threshold value that determines the intensity of
the view.
To illustrate a typical view, consider the standard views, a
la Black and Litterman (1990), on the expected value
aX
of
select portfolios returns aX, where X represents the returns
of N securities, and a is a K x N matrix, whose each row are
the weights of a different portfolio. Such view can be written
as in (12), where
Our ultimate goal is to compute a posterior distribution
f that departs from the prior to incorporate the views. The
posterior distribution f is specifed by new probabilities p on
the same scenarios (11). To this purpose, we measure the "dis-
tance" between two sets of probabilities p and p
0
by the rela-
tive entropy
The relative entropy is a "distance" in that (13) is zero only
if p = p
0
, and it becomes larger as p diverges away from p
0
.
We then defne the posterior as the closest distribution to the
prior, as measured by (13), which satisfes the views (12), as
follows:
where the notation p means that p satisfes the view (12).
Applications of Entropy Pooling to the probabilities in the
Fully Flexible Probabilities framework are manifold. For in-
stance, with Entropy Pooling, we can compute exponentially
decayed covariances where the correlations and the variances
are decayed at different rates. Other applications include con-
ditioning the posterior according to expectations on a market
panic indicator. (For more details see Meucci, 2010.)
As highlighted in Figure 1, the Entropy Pooling posterior
(14) also includes as special cases the probabilities defned in
terms of fuzzy membership functions (8). Indeed, let us as-
sume that in the Entropy Pooling optimization (14), the prior
As we show in Appendix A.1, available at http://symmys.com/node/353,
the crisp state conditioning (4) includes the rolling window (2) as a special case,
and kernel smoothing (7) includes the time decayed exponential smoothing (3)
as a special case, see Figure 1.
We can generalize further the concepts of crisp conditioning and kernel
smoothing by means of fuzzy membership functions. Fuzzy membership to a
given set X is dened in terms of a "membership" function
X
(x) with values
in the range [0 1], which describes to what extent x belongs to a given set X.
Using a fuzzy membership
X
for a given set X of potential market outcomes
for the market X, we can set the probability of each scenario as proportional to
the degree of membership of that scenario to the set of outcomes

X
(x
). (8)
A trivial example of membership function is the indicator function that
denes crisp conditioning (4)
X
(x) 1
X
(x) . (9)
With the indicator function, membership is either maximal, and equal to 1, if
x belongs to X, or minimal, and equal to 0, if x does not belong to X.
A second example of membership function is the kernel (5), which is a mem-
bership function for the singleton X
(x)
(x) . (10)
The membership of x to is maximal when x is . The larger the distance
of x from , the less x "belongs" to and thus the closer to 0 the membership
of x.
3 Entropy Pooling
The probability specication (8) assumes no prior knowledge of the distribution
of the market risk drivers X. In this case, fuzzy membership is the most general
approach to assign probabilities to the scenarios in (1). Here we show how non-
parametric Entropy Pooling further generalizes fuzzy membership specications
to the case when prior information is available, or when we must blend together
more than one membership specication.
First, we review the non-parametric Entropy Pooling implementation. Please
refer to the original article Meucci (2008) for more details, more generality, and
for the code.
The starting point for non-parametric Entropy Pooling is a prior distribution
(0)
for the risk drivers X, represented as in (1) by a set of scenarios and
associated probabilities
0
{x

(0)
}
=1
. (11)
4
X
(x) with values
X

X
(x
). (8)
X
(x) 1
X
(x) . (9)
(x)
(x) . (10)
of x.
3 Entropy Pooling
for the code.
(0)
0
{x

(0)
}
=1
. (11)
4
X
(x) with values
X

X
(x
). (8)
X
(x) 1
X
(x) . (9)
(x)
(x) . (10)
of x.
3 Entropy Pooling
for the code.
(0)
0
{x

(0)
}
=1
. (11)
4
X
(x) with values
X

X
(x
). (8)
X
(x) 1
X
(x) . (9)
(x)
(x) . (10)
of x.
3 Entropy Pooling
for the code.
(0)
0
{x

(0)
}
=1
. (11)
4
The second input is a view on the market X, or a stress-test. Thus a gen-
eralized view on X is a statement on the yet-to-be dened distribution dened
on the same scenarios {x
}
=1
. A large class of such views can
be characterized as expressions on the expectations of arbitrary functions of the
market (X).
V : Ep
{ (X)}
, (12)
where
is a threshold value that determines the intensity of the view.

To illustrate a typical view, consider the standard views a-la Black and
Litterman (1990) on the expected value
aX
of select portfolios returns aX,
where X represents the returns of securities, and a is a matrix, whose
each row are the weights of a dierent portfolio. Such view can be written as in
(12), where (X) (a
0
a
0
)
0
and
(
aX

0
aX
)
0
.
Our ultimate goal is to compute a posterior distribution which departs
from the prior to incorporate the views. The posterior distribution is specied
by new probabilities p on the same scenarios (11). To this purpose, we measure
the "distance" between two sets of probabilities p and p
0
by the relative entropy
E (p p
0
) p
0
(lnp lnp
0
) . (13)
The relative entropy is a "distance" in that (13) is zero only if p = p
0
and it
becomes larger as p diverges away from p
0
.
Then we dene the posterior as the closest distribution to the prior, as
measured by (13), which satises the views (12)
p argmin
qV
E (q p
0
) , (14)
where the notation p V means that p satises the view (12).
Applications of Entropy Pooling to the probabilities in the Fully Flexible
Probabilities framework are manifold. For instance, with Entropy Pooling, we
can compute exponentially decayed covariances where the correlations and the
variances are decayed at dierent rates. Other applications include conditioning
the posterior according to expectations on a market panic indicator. For more
details see Meucci (2010).
As highlighted in Figure 1, the Entropy Pooling posterior (14) also includes
as special cases the probabilities dened in terms of fuzzy membership functions
(8). Indeed, let us assume that in the Entropy Pooling optimization (14) the
prior is non-informative, i.e.
p
0
1. (15)
Furthermore, let us assume that in the Entropy Pooling optimization (14) we
express the view (12) on the logarithm of a fuzzy membership function
X
(x)
[0 1]
(x) ln(
X
(x)) . (16)
Finally, let us set the view intensity in (12) as
=1
X(x) ln X(x)
P
=1
X(x)
. (17)
5
}
=1
market (X).
V : Ep
{ (X)}
, (12)
where

aX
(12), where (X) (a
0
a
0
)
0
and
(
aX

0
aX
)
0
.
0
E (p p
0
) p
0
(lnp lnp
0
) . (13)
0
and it
0
.
p argmin
qV
E (q p
0
) , (14)
p
0
1. (15)
X
(x)
[0 1]
(x) ln(
X
(x)) . (16)
=1
X(x) ln X(x)
P
=1
X(x)
. (17)
5
}
=1
market (X).
V : Ep
{ (X)}
, (12)
where

aX
(12), where (X) (a
0
a
0
)
0
and
(
aX

0
aX
)
0
.
0
E (p p
0
) p
0
(lnp lnp
0
) . (13)
0
and it
0
.
p argmin
qV
E (q p
0
) , (14)
p
0
1. (15)
X
(x)
[0 1]
(x) ln(
X
(x)) . (16)
=1
X(x) ln X(x)
P
=1
X(x)
. (17)
5
(12)
}
=1
market (X).
V : Ep
{ (X)}
, (12)
where

aX
(12), where (X) (a
0
a
0
)
0
and
(
aX

0
aX
)
0
.
0
E (p p
0
) p
0
(lnp lnp
0
) . (13)
0
and it
0
.
p argmin
qV
E (q p
0
) , (14)
p
0
1. (15)
X
(x)
[0 1]
(x) ln(
X
(x)) . (16)
=1
X(x) ln X(x)
P
=1
X(x)
. (17)
5
}
=1
market (X).
V : Ep
{ (X)}
, (12)
where

aX
(12), where (X) (a
0
a
0
)
0
and
(
aX

0
aX
)
0
.
0
E (p p
0
) p
0
(lnp lnp
0
) . (13)
0
and it
0
.
p argmin
qV
E (q p
0
) , (14)
p
0
1. (15)
X
(x)
[0 1]
(x) ln(
X
(x)) . (16)
=1
X(x) ln X(x)
P
=1
X(x)
. (17)
5
(9)
(8)
(10)
(11)
(13)
(14)
is non-informative -- i.e.,
Furthermore, let us assume that in the Entropy Pooling opti-
mization (14) we express the view (12) on the logarithm of a
fuzzy membership function mx (x) [0,1] as
Then, as we show in Appendix A.2, available at http://sym-
mys.com/node/353, the Entropy Pooling posterior (14) reads
This means that, without a prior, the Entropy Pooling poste-
rior is the same as the fuzzy membership function, which in
turn is a general case of kernel smoothing and crisp condi-
tioning (see also the examples in Appendix A.3, available at
http://symmys.com/node/353).
Not only does Entropy Pooling generalize fuzzy member-
ship, it also allows us to blend multiple views with a prior.
Indeed, let us suppose that, unlike in (15), we have an infor-
mative prior p
(0)
on the market X such as, for instance,
the exponential time decay predicate (3) that recent infor-
mation is more reliable than old information. Suppose also
that we would like to condition our distribution of the market
based on the state of the market using a membership function
as in (18) such as, for instance, a Gaussian kernel. How do
we mix these two conficting pieces of information?
Distributions can be blended in a variety of ad-hoc ways.
Entropy Pooling provides a statistically sound answer: we sim-
ply replace the non-informative prior (15) with our informa-
tive prior p
(0)
in the Entropy Pooling optimization (14) driven
by the view (12) on the log-membership function (16) with
intensity (17). In summary, the optimal blend reads
More generally, we can add to the view in (19) other views
on different features of the market distribution (as in Meucci,
2008).
It is worth emphasizing that all the steps of the above process
are computationally trivial, from setting the prior p
(0)
to setting
the constraint on the expectation, to computing the posterior
(19). Thus the Entropy Pooling mixture is also practical.
Case Study: Conditional Risk Estimates
Here we use Entropy Pooling to mix information on the dis-
tribution of a portfolios historical simulations. A standard
approach to risk management relies on so-called historical
simulations for the portfolio P&L: current positions are evalu-
ated under past realizations {xt}t=1,...,T of the risk drivers X,
giving rise to a history of {xt}t=1,...,T. To estimate the risk in
the portfolio, one can assign equal weight to all the realiza-
tions. The more versatile Fully Flexible Probability approach
(1) allows for arbitrary probability weights
Refer to Meucci, 2010, for more details.
In our case study we consider a portfolio of options, whose
historically simulated daily P&L distribution over a period of
ten years is highly skewed and kurtotic, and defnitely non-
normal. Using Fully Flexible Probabilities, we model the ex-
ponential decay prior that recent observations are more rel-
evant for risk estimation purposes
where, is a half-life of 120 days. We plot these probabilities
in the top portion of Figure 2 (below).
Figure 2: Mixing Distributions via Entropy Pooling
}
=1
market (X).
V : Ep
{ (X)}
, (12)
where

aX
(12), where (X) (a
0
a
0
)
0
and
(
aX

0
aX
)
0
.
0
E (p p
0
) p
0
(lnp lnp
0
) . (13)
0
and it
0
.
p argmin
qV
E (q p
0
) , (14)
p
0
1. (15)
X
(x)
[0 1]
(x) ln(
X
(x)) . (16)
=1
X(x) ln X(x)
P
=1
X(x)
. (17)
5
}
=1
market (X).
V : Ep
{ (X)}
, (12)
where

aX
(12), where (X) (a
0
a
0
)
0
and
(
aX

0
aX
)
0
.
0
E (p p
0
) p
0
(lnp lnp
0
) . (13)
0
and it
0
.
p argmin
qV
E (q p
0
) , (14)
p
0
1. (15)
X
(x)
[0 1]
(x) ln(
X
(x)) . (16)
=1
X(x) ln X(x)
P
=1
X(x)
. (17)
5
}
=1
market (X).
V : Ep
{ (X)}
, (12)
where

aX
(12), where (X) (a
0
a
0
)
0
and
(
aX

0
aX
)
0
.
0
E (p p
0
) p
0
(lnp lnp
0
) . (13)
0
and it
0
.
p argmin
qV
E (q p
0
) , (14)
p
0
1. (15)
X
(x)
[0 1]
(x) ln(
X
(x)) . (16)
=1
X(x) ln X(x)
P
=1
X(x)
. (17)
5
(15)
(16)
(17)
(18)
Then, as we show in Appendix A.2, available at http://symmys.com/node/353,
the Entropy Pooling posterior (14) reads

X
(x
). (18)
This means that, without a prior, the Entropy Pooling posterior is the same as
the fuzzy membership function, which in turn is a general case of kernel smooth-
ing and crisp conditioning, see also the examples in Appendix A.3, available at
http://symmys.com/node/353.
Not only does Entropy Pooling generalize fuzzy membership, it also allows
us to blend multiple views with a prior. Indeed, let us suppose that, unlike in
(15), we have an informative prior p
(0)
on the market X, such as for instance
the exponential time decay predicate (3) that recent information is more reli-
able than old information. Suppose also that we would like to condition our
distribution of the market based on the state of the market using a membership
function as in (18), such as for instance a Gaussian kernel. How do we mix these
two conicting pieces of information?
Distributions can be blended in a variety of ad-hoc ways. Entropy Pooling
provides a statistically sound answer: we simply replace the non-informative
prior (15) with our informative prior p
(0)
in the Entropy Pooling optimization
(14) driven by the view (12) on the log-membership function (16) with intensity
(17). In summary, the optimal blend reads
p argmin
q
E(q p
(0)
). (19)
where Eq
{ln(
X
(X))}
.
More in general, we can add to the view in (19) other views on dierent features
of the market distribution, as in Meucci (2008).
It is worth emphasizing that all the steps of the above process are com-
putationally trivial, from setting the prior p
(0)
to setting the constraint on the
expectation, to computing the posterior (19). Thus the Entropy Pooling mixture
is also practical.
4 Case study: conditional risk estimates
Here we use Entropy Pooling to mix information on the distribution of a port-
folios historical simulations. A standard approach to risk management relies
on so-called historical simulations for the portfolio P&L: current positions are
evaluated under past realizations {x
}
=1
of the risk drivers X, giving rise
to a history of P&Ls {
}
=1
. To estimate the risk in the portfolio, one
can assign equal weight to all the realizations. The more versatile Fully Flexible
Probability approach (1) allows for arbitrary probability weights
{(
}
=1
, (20)
refer to Meucci (2010) for more details.
6

X
(x
). (18)
(0)
(0)
p argmin
q
E(q p
(0)
). (19)
where Eq
{ln(
X
(X))}
.
(0)
is also practical.
}
=1
}
=1
{(
}
=1
, (20)
6

X
(x
). (18)
(0)
(0)
p argmin
q
E(q p
(0)
). (19)
where Eq
{ln(
X
(X))}
.
(0)
is also practical.
}
=1
}
=1
{(
}
=1
, (20)
6
In our case study we consider a portfolio of options, whose historically simu-
lated daily P&L distribution over a period of ten years is highly skewed and kur-
totic, and denitely non-normal. Using Fully Flexible Probabilities, we model
the exponential decay prior that recent observations are more relevant for risk
estimation purposes
(0)
ln 2
()
, (21)
where, is a half-life of 120 days. We plot these probabilities in the top portion
of Figure 2.
Prior(exponentialtimedecay)
Stateindicator (2yrswaprate currentlevel)
Membership(Gaussiankernel)
Posterior(EntropyPoolingmixture)
time
Stateindicator (1>5yr ATMimpliedswaption vol currentlevel)
Figure 2: Mixing distributions via Entropy Pooling
Then we condition the market on two variables: the ve-year swap rate,
which we denote by
1
, and the one-into-ve swaption implied at-the-money
volatility, which we denote by
2
. We estimate the 22 covariance matrix of
these variables, and we construct a quasi-Gaussian kernel, similar to (6), setting
as target the current values x
of the conditioning variables
exp(
1
2
2
(x
)
0
1
(x
). (22)
In this expression the bandwidth is 10 and 04 is a power for the
Mahalanobis distance, which allows for a smoother conditioning than = 2.
If we used directly the membership levels (22) as probabilities
,
we would disregard the prior information (21) that more recent data is more
valuable for our analysis. If we used only the exponentially decayed prior (21),
7
(21)
Prior (exponential time decay)
State indicator (2yr swap rate - current level)
State indicator (1>5yr ATM implied swaption vol - current level)
Membership (Gaussian kernel)
Posterior (Entropy Pooling mixture)
time
(20)
(19)
We condition the market on two variables: the fve-year swap
rate, which we denote by X1, and the one-into-fve swaption
implied at-the-money volatility, which we denote by X2. We
estimate the 2 x 2 covariance matrix of these variables, and
we construct a quasi-Gaussian kernel, similar to (6), setting as
target the current values x
T
In this expression, the bandwidth 10 and 0.4 is a power
for the Mahalanobis distance, which allows for a smoother
conditioning than =2.
If we used directly the membership levels (22) as probabili-
ties pt mt, we would disregard the prior information (21) that
more recent data is more valuable for our analysis. If we used
only the exponentially decayed prior (21), we would disregard
all the information conditional on the market state (22). To
overlay the conditional information to the prior, we compute
the Entropy Pooling posterior (19), which we write here
Notice that for each specifcation of the kernel bandwidth
and radius depth in (22) we obtain a different posterior.
Hence, a further refnement of the proposed approach lets
the data determine the optimal bandwidth, by minimizing the
relative entropy in (23) as a function of q as well as (,). We
leave this step to the reader.
The kernel-based posterior (23) can be compared with al-
ternative uses of Entropy Pooling to overlay a prior with par-
tial information. For instance, Meucci (2010) obtains the pos-
terior by imposing that the expected value of the conditioning
variables be the same as the current value, i.e. Eq{X} =x
T
.
This approach is reasonable in a univariate context. How-
ever, when the number of conditioning variables X is larger
than one, due to the curse of dimensionality, we can obtain
the undesirable result that the posterior probabilities are
highly concentrated in a few extreme scenarios. This does not
happen if we condition through a pseudo-Gaussian kernel as
in (22)-(23)
In Figure 2 (see pg. 34), we plot the Entropy Pooling poste-
rior probabilities (23) in our example. We can appreciate the
hybrid nature of these probabilities, which share similarities
with both the prior (21) and the conditioning kernel (22).
Using the Entropy Pooling posterior probabilities (23),
we can perform all sorts of risk computations, as in Meucci
(2010). In (24) we present a few signifcant statistics for our
historical portfolio (20), and we compare such statistics with
those stemming from the standard exponential decay (21) and
the standard conditioning kernel (22). For more details, we
refer the reader to the code available at http://symmys.com/
node/353.
On the last row of (24) we also report the effective num-
ber of scenarios, a practical measure of the predictive power
of the above choices of probabilities, discussed in detail in
Meucci (2012).
References
Black, F. and R. Litterman, 1990. "Asset allocation: Combining
Investor Views with Market Equilibrium," Goldman Sachs Fixed
Income Research.
Meucci, A., 2008. "Fully Flexible Views: Theory and Practice,"
Risk 21, 97102. Article and code available at http://symmys.
com/node/158.
ibid, 2010. "Historical scenarios with Fully Flexible Probabilities,"
Risk Professional December, 4043. Article and code available at
ibid, 2011. "The Prayer: Ten-step Checklist for Advanced Risk
and Portfolio Management," Risk Professional April/June, 54
60/5559. Available at http://symmys.com/node/63.
ibid, 2012. "Effective Number of Scenarios with Fully Flexible
Probabilities," Working Paper article and code available at http://
symmys.com/node/362.
Attilio Meucci is the chief risk offcer at Kepos Capital LP. He runs the 6-day "Ad-
vanced Risk and Portfolio Management Bootcamp," see www.symmys.com. He is
grateful to Garli Beibi and David Elliott.
In our case study we consider a portfolio of options, whose historically simu-
lated daily P&L distribution over a period of ten years is highly skewed and kur-
totic, and denitely non-normal. Using Fully Flexible Probabilities, we model
the exponential decay prior that recent observations are more relevant for risk
estimation purposes
(0)
ln 2
()
, (21)
where, is a half-life of 120 days. We plot these probabilities in the top portion
of Figure 2.
Prior(exponentialtimedecay)
Stateindicator (2yrswaprate currentlevel)
Membership(Gaussiankernel)
Posterior(EntropyPoolingmixture)
time
Stateindicator (1>5yr ATMimpliedswaption vol currentlevel)
Figure 2: Mixing distributions via Entropy Pooling
Then we condition the market on two variables: the ve-year swap rate,
which we denote by
1
, and the one-into-ve swaption implied at-the-money
volatility, which we denote by
2
. We estimate the 22 covariance matrix of
these variables, and we construct a quasi-Gaussian kernel, similar to (6), setting
as target the current values x
exp(
1
2
2
(x
)
0
1
(x
). (22)
In this expression the bandwidth is 10 and 04 is a power for the
Mahalanobis distance, which allows for a smoother conditioning than = 2.
If we used directly the membership levels (22) as probabilities
,
we would disregard the prior information (21) that more recent data is more
valuable for our analysis. If we used only the exponentially decayed prior (21),
7
(22)
we would disregard all the information conditional on the market state (22).
To overlay the conditional information to the prior, we compute the Entropy
Pooling posterior (19), which we write here
p argmin
q
0
ln m
E(q p
(0)
). (23)
Notice that for each specication of the kernel bandwidth and radius depth in
(22) we obtain a dierent posterior. Hence, a further renement of the proposed
approach lets the data determine the optimal bandwidth, by minimizing the
relative entropy in (23) as a function of q as well as ( ). We leave this step to
the reader.
The kernel-based posterior (23) can be compared with alternative uses of En-
tropy Pooling to overlay a prior with partial information. For instance, Meucci
(2010) obtains the posterior by imposing that the expected value of the con-
ditioning variables be the same as the current value, i.e. Eq
{X} = x
. This
approach is reasonable in a univariate context. However, when the number of
conditioning variables X is larger than one, due to the curse of dimensionality,
we can obtain the undesirable result that the posterior probabilities are highly
concentrated in a few extreme scenarios. This does not happen if we condition
through a pseudo-Gaussian kernel as in (22)-(23)
In Figure 2 we plot the Entropy Pooling posterior probabilities (23) in our
example. We can appreciate the hybrid nature of these probabilities, which
share similarities with both the prior (21) and the conditioning kernel (22).
Time decay Kernel Entropy Pooling
Exp. value 0.12 0.05 0.10
St. dev. 1.18 1.30 1.26
Skew. -2.65 -2.56 -2.76
Kurt. 12.58 11.76 13.39
VaR 99% -4.62 -5.53 -5.11
Cond. VaR 99% -6.16 -6.70 -6.85
Eective scenarios 471 1,644 897
(24)
Using the Entropy Pooling posterior probabilities (23) we can perform all sorts
of risk computations, as in Meucci (2010). In (24) we present a few signicant
statistics for our historical portfolio (20), and we compare such statistics with
those stemming from the standard exponential decay (21) and the standard
conditioning kernel (22). For more details, we refer the reader to the code
available at http://symmys.com/node/353.
On the last row of (24) we also report the eective number of scenarios, a
practical measure of the predictive power of the above choices of probabilities,
discussed in detail in Meucci (2012).
References
Black, F., and R. Litterman, 1990, Asset allocation: combining investor views
with market equilibrium, Goldman Sachs Fixed Income Research.
8
(23)
(24)
Timedecay Kernel EntropyPooling
Exp. value 0.12 0.05 0.10
St. dev. 1.18 1.30 1.26
Skew. -2.65 -2.56 -2.76
Kurt. 12.58 11.76 13.39
VaR 99% -4.62 -5.53 -5.11
Cond. VaR 99% -6.16 -6.70 -6.85
Effectivescenarios 47 11,644 897
GARP Webcasts bring the risk professions top thought leaders and innovators direct to your desktop,
so you can learn from their insights on the critical issues facing risk professionals today. GARP produces
between fifteen and twenty live Webcasts each year. As a benefit of Membership, the US$49 fee per paid Webcast is
waived for GARP Individual and Student Members, who also receive unlimited access to the Webcast archive.
Upcoming Webcasts
w
Risk Appetite, Governance and Corporate StrategyPresented by SAS | December 6 >Register
w
The Ramifications of Dodd-Frank: Consumer Protection | December 13 >Register
Featured On-Demand Webcasts
w
The Ramifications of Dodd-Frank: The Volcker Rule >View
w
The Ramifications of Dodd-Frank: Securitization >View
w
The Ramifications of Dodd-Frank: Derivatives >View
w
Using Operational Risk to Gain a Competitive Edge >View
A full selection of GARP Webcasts can be found at www.garp.org/webcasts
TM
The worlds leading risk
practitioners can now be found
on your desktop.
A Fresh
Assessment
Recently refned risk management approaches could be
more realistically aligned with strategy and governance.
By Gaurav Kapoor and Conni e Valenci a
W
R I S K S T R A T E G Y
hen tightrope walkers perform, they never look
down. Even when their rope sways dangerously,
they keep their eyes fxed on their destination. If
they dont, they could lose their balance and fall.
Managing a business is a lot like walking a
tightrope. The risks below are plenty. But if you focus only on
the risks, you could lose sight of the goal and fall.
Historically, few businesses focused on managing their risks.
Most indulged in high-risk ventures without thinking of the
consequences. Then a series of accounting scandals, security
breaches and fnancial crises forced businesses to push risk man-
agement up front and center. In response to the public outcry, the
auditing profession adopted a risk-based approach. Risks were
assessed, categorized and prioritized based on their likelihood
of occurrence and impact. As a result, maximum resources and
attention became focused on high-impact, high-frequency risks.
The risk-based approach, though well accepted, has room
for improvement. For example, current guidance is designed to
analyze risks in isolation, when, in reality, risks usually occur in
a series or in combination. In addition, and most importantly,
the risk-based approach categorizes low-frequency, not likely
to occur risks as low-priority. In other words, current guidance
doesnt provide analysis on so-called black swan events or unex-
pected/unknown risks. However, over the last decade, headlines
were flled with unexpected or unknown risk events such as the
BP oil spill, the fnancial meltdown, the Toyota safety recall and
the 2011 tsunami in Japan. These indicate that black swans oc-
cur, and with alarming frequency.
An old saying is appropriate: Dont fx your fat tire if your
goal is to stay home. Risk management is no different. Assess-
ing risks without understanding the impact each risk has on the
companys goals is not very meaningful. The shotgun approach
of analyzing and tracking every possible what could go wrong
risk may lead an organization down the path of focusing on risks
that may not have a signifcant impact on the company and its
objectives.
There is a new sound to risk management that begins with
identifying the companys goals. Ask the executive team, what
are we trying to achieve (as opposed to, what keeps you up at
night)? Once the goals and objectives have been clearly de-
fned, the next step is to focus only on the risks that threaten the
achievement of the goals.
Taking a strategy-based approach to risk management, like
the tightrope walker, puts the focus is on the goals and objec-
tives. The strategy-based approach enables the risk manage-
ment analysis to concentrate on all risks, including potential
black swans, that threaten the achievement of the corporate
goals; to analyze and measure the companys capacity to achieve
the goals; and to approach risk mitigation efforts in a balanced,
focused and more cost-effective way.
The Right Questions
Asking the right questions takes as much skill as getting the right
answers. Doing it right creates a healthy dialogue among the
executive and risk management teams to understand the rela-
tionships among, and pervasiveness of, the companys goals and
objectives. Certain goals may confict with others. Consider, for
example, a cruise lines desires to enhance passenger ticket rev-
enue, increase customer service and decrease costs. They can be
in inherent confict with each other.
Management might want to ask, how can we increase ticket-
ing prices while maintaining customer satisfaction and reducing
costs? Such thought-provoking questions can facilitate a healthy
debate and, in turn, an exploration of relevant risks. In the cruise
line example, executives might decide not only to raise passenger
ticket prices, but also to improve customer service without incur-
ring excessive costs, perhaps through incentives or streamlining
staff.
Once the goals and related risks have been identifed, the next
step is a survey of management. Why a survey? It allows for any
number of participants to weigh in, the data collection is auto-
mated and effcient, and the anonymity fosters open and honest
communication regarding risks.
Management is best positioned to assess these impacts because
it is responsible for managing the day-to-day risks of operations.
Because senior executives set strategy and business-line manag-
ers have a handle on the risks, both constituencies are invested in
the process and the outcome.
A true enterprisewide risk assessment solicits the feedback of
those responsible for managing risks across the organization.
Participants for the survey should be chosen wisely. They should
have a macro view of the company and understand the intrica-
cies of each goal, as well as the relationships of risks to each goal.
For a survey to have maximum effectiveness, the key is to strike
the right balance in portraying the seeming conficts among the
goals and identifying the impact of each risk to each goal. De-
signing a right ft survey is part of the art of an effective risk
management assessment. Again, using the example of the cruise
line company, the survey question might read as follows:
Analyzing Results
Once the responses are collected, the next challenge is in ana-
lyzing the results. This is where the strategy-based approach
suggests applying some science and technology to the tradi-
tional performance of a risk assessment. Technology does not
predict the future; only humans can speculate on future events.
However, technology can analyze data trends and patterns,
which are useful in studying the assessment results.
Any mathematical model can determine the correlation (the
relevancy) between the risks identifed and the impact each the
risk has on each goal. In addition, most models can also ana-
lyze the compounding impact of several risks occurring at the
same time. Analyzing the compounding risks as they occur si-
multaneously is important, as most risks do not occur in isola-
tion. In fact, most emergency response professionals will agree
that an accident is the result of a combination of multiple risks
occurring at the same time.
We believe that corporate entities are no different. Most or-
ganizations can handle isolated occurrences such as a fnancial
restatement. However, a combination of risks, triggering simul-
taneously within the same area, may have a catastrophic impact.
Think of a company fling a fnancial restatement due to sus-
pected foul play by its executives, compounded by a revised, re-
ported loss during a down economy. This situation is dramatic,
but not uncommon. Analyzing risk patterns is the most effective
method for projecting the compounding impact of risks.
Managers Involved
Running the analysis of risk patterns based on the survey re-
sponses is the easy part. The next step requires the involvement
of managers enterprisewide to defne and implement an appro-
priate risk mitigation strategy and various control measures.
Back to the cruise line example: The risk of overbooking is
found to have a low impact versus the risk of not enhancing
revenue. Alternatively, overbooking cabins has a 20% impact
on threatening guest satisfaction. However, the combined im-
pact of overbooking cabins when on a cruise that has to tender
to port (due to limited port docking capacity) exponentially
increases the impact of guest dissatisfaction by over 50%. A
proper risk mitigation strategy based on this impact analysis
would be for management to enhance the sophistication of the
online booking system or to implement more accurate mea-
sures to forecast demand. In addition, management should
actively focus on ensuring as many port-docking spaces as pos-
sible for its ships itineraries.
To align the risk mitigation strategy to the companys risk
0 1 2 3 4 5
Limited Port Capacity
Rising Fuel Costs
Global Terrorism
Overbooking Cabins
Down Economy
3
2
3
4
5
How does increasing passenger ticket sales impact our
capacity to increase guest satisfaction when consider-
ing the following risks (rate impact on scale of 1-5)?
patterns properly, the following must be defned:
The relationship between the companys risks and goals.
Risks that could potentially threaten the companys goals must
be correctly identifed. In other words, did we ask the right
what could go wrong questions? It is also important to un-
derstand the strength of the correlation (or relationship) be-
tween the goals set by the executive team and the risks identi-
fed and assessed by management.
The pervasive impact each risk has to your corporate strat-
egy. How many goals are impacted by each risk, and how great
is the impact of each risk on each goal?
Risks and controls should be monitored continually, as the
impact of each risk on each goal varies over time. For instance,
during peak season, the risk of overbooking a cabin may be
greater than at any other time of the year. The ease of the
survey approach allows for real-time updates as frequently as
the organization demands.
Predictive Analytics and Technology
A beneft of analyzing risk patterns over a period of time is the
emergence of predictive analytics. Mathematical models that
analyze and track risk patterns can also identify recurring pat-
terns. The more such patterns there are for the model to ana-
lyze, the stronger the correlation between risks and goals. The
stronger the correlation, the greater the level of confdence in
the model to be predictive or to detect black swans.
To have an effective risk pattern analysis over time, the goals
of the organization must be consistent. Any newly set strategic
goal will change the risk landscape and require a new assess-
ment.
Together with the mathematical model, a well designed
technology framework enables a systematic and focused ap-
proach to risk and optimizes the deployment of resources. As
a result, enterprise risk-resilience can be strengthened, custom-
ers and stakeholders can be protected and proftability can be
improved.
The key capabilities of technology that enable effective
strategy-based risk management include:
Integration. A centralized risk framework integrates risk
management processes across the enterprise by providing a
single point of reference to identify, assess and profle risks.
This simplifes the mapping of risks to goals. Similarly, an inte-
grated repository of risk assessments, controls, key risk indica-
tors, loss events and other risk-related data enables managers
to understand quickly the relationships among various risks,
as well their relationships to goals. The survey management
itself benefts from an integrated approach. Platform-based
technology can facilitate a streamlined, systematic and eff-
cient process of survey design, distribution, implementation
and response collection across departments, business units and
geographic locations.
Automation. Automating certain critical processes, such as
the scoring and tabulation of risk assessments, can accelerate
risk-goal mapping and save precious time, resources and ef-
fort. It also eliminates the need for manual and cumbersome
paper-based processes. Alerts can be confgured to be sent out
automatically to the relevant users when a risk threshold has
been breached or when a control assessment is not up-to-date.
Visibility. Displaying the quantifed risks in color-coded
charts or heat maps enables risk managers to understand the
impact of risks on goals easily and quickly. If the maps can be
drilled down into, managers can study each goal in relation
to its corresponding risks. They can also help ascertain how
each unit or department is managing risks. Technology can
thus bring transparency to risks. It also simplifes the creation
of risk reports, automatically generates trend analyses and en-
ables managers to track risk management processes across the
enterprise in real time.
Sustainability. A technology-enabled framework can drive
a proactive and sustainable approach to risk management by
continually monitoring risks and their relationship to goals.
The framework can also establish clear links between each
goal, risk and control, thereby simplifying control assessments
and monitoring.
If equipped with loss-event and incident management ca-
pabilities, the technology framework can automatically track
near misses and other threat issues, along with their root causes
and owners. Subsequently, the issues can be routed through a
systematic process of investigation and resolution.
As we have seen, risk management in isolation is not refec-
tive of real-world risk environments. Aligning risk manage-
ment to strategic goals yields more accurate and focused re-
sults. Integrating a strategy-based risk model with compliance,
audit and governance processes, and confdently walking the
tightrope of risks, will build resilience and keep stakeholders
and customers happy.
Gaurav Kapoor is chief operating offcer of MetricStream (www.metricstream.com), a
Palo Alto, California-based provider of governance, risk and compliance solutions to a
broad range of industries. Connie Valencia, a consultant who specializes in process im-
provement and internal controls, is principal of Miami-based Elevate Consulting (www.
elevateconsulting.com).

Professional: On The Same Page

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Professional: On The Same Page

Uploaded by

Copyright:

Available Formats

PROFESSIONAL

A GARP Publication DECEMBER 2011

exam preparation, allowing you to study

that are fully specied by a set of parameters ,

that are fully specied by a set of parameters ,

that are fully specied by a set of parameters ,

. Using the indicator function 1 (x), which is

lie in a given domain X. Using the indicator function 1

(x) is dened by a positive non-increasing generator function

lie in a given domain X. Using the indicator function 1

(x) is dened by a positive non-increasing generator function

lie in a given domain X. Using the indicator function 1

(x) is dened by a positive non-increasing generator function

lie in a given domain X. Using the indicator function 1

(x) is dened by a positive non-increasing generator function

lie in a given domain X. Using the indicator function 1

(x) is dened by a positive non-increasing generator function

lie in a given domain X. Using the indicator function 1

(x) is dened by a positive non-increasing generator function

lie in a given domain X. Using the indicator function 1

(x) is dened by a positive non-increasing generator function

, we can set the probability

, or minimal and equal to 0,

The membership of x to is maximal when x is . The larger

is a threshold value that determines the intensity of the view.

is a threshold value that determines the intensity of the view.

is a threshold value that determines the intensity of the view.

is a threshold value that determines the intensity of the view.

is a threshold value that determines the intensity of the view.

is a threshold value that determines the intensity of the view.

is a threshold value that determines the intensity of the view.

is a threshold value that determines the intensity of the view.

of the conditioning variables

of the conditioning variables

You might also like