Natural Catastrophe Risk and the Changing Environment: Overview and Challenges Shree Khare, Senior...
-
Upload
benjamin-mackenzie -
Category
Documents
-
view
216 -
download
0
Transcript of Natural Catastrophe Risk and the Changing Environment: Overview and Challenges Shree Khare, Senior...
Natural Catastrophe Risk and the Changing Environment: Overview and Challenges
Shree Khare, Senior Catastrophe Risk Modeller, RMS Ltd., London
2Confidential
Talk Outline
Focus: Hurricane Risk Modelling
– Financial motivation
– Catastrophe modelling basics: Event set framework
– Components
– Climate change
– Model development project Mathematical and scientific challenges Opportunities for
collaboration
3Confidential
Introduction to RMS
Founded at Stanford University in 1988 Multi-disciplinary skills: Applied mathematics, statistics,
physical sciences and engineering applied to insurance Solely focused on risk management issues Independent and objective information source Global presence in major insurance markets
“ At RMS, our goal is to help clients manage catastrophe risk through the practical application of the most advanced quantitative risk assessment techniques available.”
- Hemant Shah, President & CEO
4Confidential
Top 10 Insured Cat Losses, 1990-2005
* Includes liability losses
CountryEventYearInsured Loss ($billions)
Indonesia, Thailand
U.S., Bahamas
France, Switzerland
France, U.K.
Japan
U.S., Caribbean
U.S., Caribbean
U.S.
U.S., Bahamas
U.S.
Earthquake & Tsunami
Hurricane Wilma
Winterstorm Lothar
Winterstorm Daria
Typhoon Mireille
Hurricane Charley
Hurricane Ivan
Northridge Earthquake
Hurricane Andrew
Terrorist Attack on WTC
2004
2005
1999
1990
1991
2004
2004
1994
1992
2001
5.0
6.5
6.6
7.8
8.0
11.0
17.8
21.5
31.7*
U.S.Hurricane Katrina200545.0
6.0-6.8
Swiss Re Sigma 2/2006; Triple I 1/2006
5Confidential
Framework: Event Based Modelling
Assess WindSpeed
- Peak gusts experienced at each location
Calculate Damage
- Varies by structure type
Define Hurricane
-Track-intensity
Quantify FinancialLoss
- Apply policy termsand Reinsurance
structures
90%
$ Loss$ Loss
Apply property exposure
Using physical and statistical modelling - simulate events in time and quantify financial loss for each event
Model components are consistent with observed data
6Confidential
Framework: Event Based Modelling
Assess WindSpeed
- Peak gusts experienced at each location
Calculate Damage
- Varies by structure type
Define Hurricane
-Track-intensity
Quantify FinancialLoss
- Apply policy termsand Reinsurance
structures
90%
$ Loss$ Loss
Apply property exposure
Simulation of hundreds of thousands of years can be used to quantify modelled probabilities of financial loss
7Confidential
Framework: Event Based Modelling
Assess WindSpeed
- Peak gusts experienced at each location
Calculate Damage
- Varies by structure type
Define Hurricane
-Track-intensity
Quantify FinancialLoss
- Apply policy termsand Reinsurance
structures
90%
$ Loss$ Loss
Apply property exposure
Model output is used to inform Enterprise Risk Management: Rate setting, capital allocation, securities …
8Confidential
Hurricane Risk Model Components
‘Rates’ (5-year view, long-term projections in a changing climate)
‘Track modelling’: Trajectories of tropical vortices in space/time
‘Windfield’ Surface roughness and topography Transitioning of tropical extra-tropical storms Vulnerability Exposure Financial Model On the horizon: Parametric and model-choice
uncertainty
9Confidential
Need to quantify expected number of landalling hurricanes: models are validated using historical data
Data source: NOAA NHC HURDAT “Best Track” 1950-2005: 597 time series for named North Atlantic TCs
Modelling Hurricane Rates
10Confidential
Insurance/Re-insurance industry typically interested in 5-year projections
Data source: NOAA NHC HURDAT “Best Track” 1950-2005: 597 time series for named North Atlantic TCs
Modelling Hurricane Rates
11Confidential
Modelling Hurricane Rates
Cat 1-5 Storms
Blue Basin NumbersRed Landfall Numbers
HURDAT data Jarvinen et al. (1984)
RMS has built an exhaustive collection of statistical models for predicting this non-stationary time series
Annually, we gather world-leading hurricane experts to give us their recommendations as to which of our models are best for predicting future rates (expert elicitation)
12Confidential
Modelling Hurricane Tracks
For most diagnostics in most regions (but not all) the historical TCs fall within the range of values in the synthetic TC set (Hall and Jewson, Tellus, 2007).
Evaluation criterion: historical TCs should be statistically indistinguishable from equal-sized samples of synthetic TC set.
On most coast regions track model’s landfall predictions “beat” predictions derived solely from local landfall events, based on out-of-sample likelihood analysis (Hall and Jewson, JAM, 2007).
HISTORICAL (1950-2005) SYNTHETIC (1000 YRS)
13Confidential
Long-Term Risk Management: Climate Change
14Confidential
Long-Term Risk Management: Climate Change
‘Natural’ forcing can not explain 20th century warming
15Confidential
Rates and Track Modelling in a Changing Climate
Clients are increasingly interested in quantifying hurricane risk in future climates
Given the changing climate, quantifying future risk is a significant challenge (more later …)
16Confidential
Model Development Example: Hurricane Winds
Natural catastrophe risk models are comprised of components (rates, track, winds, …)
Need to generate millions of simulations Need to explore efficient methods of generating windfields
along the modelled tracks Given some validation data set, can use cross-validation to
perform model selection Quick overview of hurricane vortex model comparison Apologies in advance for jargon …
17Confidential
Goal: To model maximum 1-minute/3-second winds over ocean and land (10 m height with roughness) for a large number of simulated events
Given spatial scales of hurricanes, full 3-dimensional numerical modelling can not feasibly be used to generate the full stochastic set
Model Development Example: Hurricane Winds
18Confidential
Wind Modelling Basics
We need some approximations: Steady Pressure Field Heating source ‘maintains’ a steady pressure gradient on time scales of
6 hours - also ignoring feedbacks, convection, vertical acceleration … Approximate pressure distribution as radially symmetric: p(r)
19Confidential
Wind Models: PBL + Linear Analytical
Our interest is 10m winds: Consider the atmospheric boundary layer
Surface layer is ‘turbulent’: Ultimately arising from surface friction – has effect of slowing down winds at surface
20Confidential
Wind Models: PBL + Linear Analytical
Space/time scales of turbulent motions can be extremely small, hence difficult to model
Attempt to model larger scale flow by ‘Reynolds Averaging’
',',' wwwvvvuuu
21Confidential
Wind Models: PBL + Linear Analytical
The (approximate) momentum equations (in translating system)
z
wvuf
r
vuv
r
v
r
vu
t
v
z
wu
r
pvf
r
vu
r
v
r
uu
t
u
o
''
''12
22Confidential
Wind Models: PBL + Linear Analytical
PBL (Chow, Vickery, Cardone, FHLC): Vertical mean – friction parameterization
H
vcFuf
r
vuv
r
v
r
vu
t
v
H
ucF
r
pvf
r
vu
r
v
r
uu
t
u
o
),(
),(12
23Confidential
Wind Models: PBL + Linear Analytical
For Gradient Wind let H ∞, and look at the steady state solution, which is the root (with the proper limiting property) of:
01
2
r
pvf
r
v
o
24Confidential
Linear Analytical Boundary Layer Model
Analytical theory developed in Kepert (2001) for 3-dimensional flow in a translating vortex for a prescribed pressure field Model has friction, vertical diffusion, ‘slip’ boundary condition at surface
2
2
2
221
z
vKuf
r
vu
z
vw
v
r
v
r
vu
z
uK
r
pvf
r
v
z
uw
u
r
v
r
uu
o
25Confidential
Linear Analytical Boundary Layer Model
Idea: Linearize equations about gradient wind, solve first order equations Efficient (free) to run, encapsulates physics causing asymmetries z, Cd and K can be optimized
uu
vVv g
~
~
26Confidential
Model Selection Study Using H*WIND H*WIND is consists of 10 m, 1-minute mean winds over ocean which summarizes
nearly all available data (surface obs, flight level …) Put together by researchers at Hurricane Research Division of NOAA in Miami We are the first group to perform such a thorough study …
27Confidential
Mathematical and Scientific Challenges: Collaboration
RMS is in a unique position, serving as an intermediary between academic/government research and the financial industry
Our models involve many components – some of which are developed through collaboration with the wider research community
This involves pure academic research and paid consultancies Example institutions: LSE, NASA, University of Miami,
National Center for Atmospheric Research, Oxford, … Collaboration often leads to peer-reviewed journal publications We work with PhD students, University Faculty, US
Government Researchers, Post-Docs, … We are very open to new collaboration …
28Confidential
Extreme Value Theory
EVT is not often used in catastrophe risk modelling With event based mathematical modelling, spatially correlated
extremes are naturally accounted for – a challenge in EVT Output from cat models may provide a rich ‘data’ set to ‘play’ with Can EVT be used to gain greater insight into cat model output? Can EVT be used to build better cat models?
29Confidential
Use of Climate Models in Catastrophe Risk General circulation models are used by research groups
to simulate the evolution of future climates Climate researchers and catastrophe risk modellers ask
related, yet unique questions It is challenging for catastrophe risk modellers to make
best use of climate simulations How we make best use of climate simulations will involve
extensive research and statistical analysis
30Confidential
Model Choice Uncertainty
Catastrophe models are made of components Components have parameters, which have been estimated
using observed data Financial loss can be sensitive to uncertain parameters – this
kind of information will be included in future cat models Financial loss is also sensitive to choice of model components
(track model A vs. track model B) How do we best quantify model choice sensitivity/uncertainty? How do we optimally use ensembles of models? Bayesian model averaging seems inadequate due to ‘double-
counting’ (e.g. Hoeting et al., 1999, Statistical Science) Cat modelling requires a proper statistical framework to
answer these questions
31Confidential
QUESTIONS?