Quantitative Risk Management:

Concepts, Techniques, Tools

Paul Embrechts
Department of Mathematics
Director of RiskLab, ETH Zurich
Senior SFI Chair
www.math.ethz.ch/~embrechts

Five QRM-examples with as

ingredients:

Risk and Uncertainty

Climate, Meteorology, Geology
Extremes and Correlation
Quantitative Modelling
Societal Consequences
Insurance and Finance

Ex.1: 31. Jan. 1953 1. Feb. 1953*

1836 people killed

72000 people evacuated
49000 houses and farms
floaded
201000 cattle drowned
500 km coastal defenses
destroyed; more than 400
breaches of dykes
200000 ha land floaded
*Antwerp (Schoten):
3rd February, 1953

The Delta-Project

Coastal fload-protection
Requested dyke height at l: hd(l)

Safety margin at l: MYSS(l) =

Maximal Yearly Sea Surge at l:
Probability(MYSS(l) > hd(l))
should be small, whereby small is defined as (Risk):
1 / 10000 in the Randstad
1 / 250 in the Deltaregion to the North
Similar requirements for rivers, but with 1/10 1/100
For the Randstad (Amsterdam-Roterdam):
Dyke height = Normal-level (= NAP) + 5.14 m

PE

Laurens de Haan

Guus Balkema

Ex.2: The 1994 Northridge earthquake

Ex.3: February 1995

The Great Hanshin (Kobe) earthquake of

January 17, 1995

Prime example for Operational Risk,

external event (on top of all else)

How Kobe earthquake and a straddle position finally

broke the back of Barings bank
Straddle = Short Call and Short Put on Common Strike

Volume of Nikkei Futures

Ex.4: The 1986 Challenger explosion

Escaping fuel!

A brief discussion on the immediate cause:

O-Ring

Richard Feynman

Some key modelling input:

Logistic regression: theory exists!

Rare event prediction (31 degrees F.)
Model Uncertainty
Statistical analysis: data matters!
Statistical estimation of this uncertainty
(95% confidence intervals)
These intervals are typically very wide for
the estimation of rare events

Ex.5: The financial tsunami (2007/9)

Recipe for Disaster: The Formula That Killed Wall Street

By Felix Salmon 23 February, 2009
Wired Magazine

Error, )

Standard - model
(3)

Stress - model
(12)

And of course:

RiskLab
Founded October 1994, mainly for research on
QRM for Insurance and Banking
Modelling of Extreme Events more broadly
Understanding dependence (beyond correlation)
Examples of current applied research related to
the environment include:
- Modelling Flood Risk for an atomic reactor
- Modeling the North Atlantic cyclone
frequency, jointly with NCAR, Boulder, Co

RiskLab QRM Research - Examples

(1997, 2005, 2007)
1997

2005

2007

 (2005) contains

Chapter on Extreme Value Theory

life beyond Normality

Chapter on Dependence Modelling

life beyond Linear Correlation

and much more

P(X>x), power law!

20+% drop!

Some isues:
RM too often frequence oriented ...
- every so often (rare event)
- return period, 1 in x-year event
- Value-at-Risk (VaR)
... rather than more relevant severity orientation
- what if
- loss size given the occurence of a rare event
- Expected Shortfall E[X I X > VaR]
This is not just about theory but a RM attitude!

The Peaks Over Threshold (POT) Method

Crucial point!

Mean Excess Function: e(u) = E(X-u I X > u)

99%-quantile with 95% aCI (Profile Likelihood):

27.3 (23.3, 33.1)
99% Conditional Excess: E( X I X > 27.3) with aCI

99%-quantile

99%-conditional excess
27.3

Several extensions exist:

Non-stationarity
Co-variable modelling within POT
Beware of discrete data, non-standard theory
Multivariate extremes: definitions
Several, question dependent approaches exist
Dynamic, stochastic process models
Diagnostic and graphical tools
Important: Communicating extreme events!!!!

A note on risk measures

Axiomatics ---> coherent/convex risk measures
Example: q(,X) as a quantile risk measure or
return period, P(X > q(,X))=1 - (100%-VaR)
- estimation for close to 1 ---> EVT
- nice properties for elliptical models (MVN)
- cases which are problematic wrt non-convexity,
q(, X+Y) > q(, X) + q(,Y),
concern very skew, or very heavy-tailed risks, or
risks with special dependence ---> research!
(next speaker)

A note on linear correlation

Question from practice: simulate two risk
factors each having a lognormal distribution,
one with parameters (0,1), the other with
parameters (0,9), and this with given linear
correlation coefficient in [-1,+1].
Any reaction, intuition?
Answer: One can either have no-, 1-, or
infinitely many solutions, it all depends on !

Explanation: Frchet-Hoeffding
(A fallacy!}

Explanation: the lognormal example

=3

For more information on

correlation pitfalls, see:
Embrechts, P., McNeil, A., Straumann, D.:
Correlation and dependence in risk management:
properties and pitfalls
In: Risk Management: Value at Risk and Beyond, ed.
M.A.H. Dempster, Cambridge University Press,
Cambridge, 2002, pp. 176-22
(or the QRM book)

Some points to take forward:

Research
Teaching
Interdisciplinarity
Communication
Prevention! : RM is most effective at
prevention. Failing at prevention results in
damage control, which is often expensive and
ineffective.

Thank you!