Risk in Complex Systems: From No Data to Big Data

Risk in Complex Systems:
From No Data to Big Data
Patrick McSharry
Head of Catastrophe Risk Financing,
Smith School of Enterprise & the Environment,
University of Oxford Carnegie,
Mellon University in Rwanda
Risk in complex systems: from no data
to big data
Dr Patrick McSharry
Smith School of Enterprise and the Environment
Oxford MAN Institute of Quantitative Finance
University of Oxford
ICT Center of Excellence
Carnegie Mellon University
[email protected]
9th Oct 2014
Global Challenges Foundation & Future of Humanity Institute, Oxford University
Oxford, UK
Nuclear Risk
Forecast uncertainty
•  Honest forecasts are needed
•  Can society handle probabilistic forecasts?
•  Met Office says it will 'empower people to
make their own decisions'
•  “New Met Office forecast system likely to
mean 80% chance of confusion”
•  Tamara Cohen, Daily Mail, 10 November 2011.
Science and the media
•  Met Office and the BBC: “BBQ summer” and “Mild winter” •  Climategate (2009) •  L’Aquila earthquake: seven natural disaster experts sentenced to six years in jail for manslaughter •  Judge: “assessment of risks connected to the seismic acLvity in the area around L'Aquila turned out to be completely vague, generic and ineffecLve” Science and policymaking
Source: www.theage.com.au
Reputation: climategate
Human experts versus machines
•  Traditionally important decisions
have been made by human experts
•  Predictive analytics due to availability
of data, computational resources
and quantitative techniques
•  Accounting for uncertainty is key for
quantifying the confidence
underlying the decision-making
Big data
Large quantities of data
Structured and unstructured data
Temporal and spatial
Data on individuals, households, organisations
Transactional data (searches, clicks, tweets,
purchases, comments …)
•  Data about what people actually did and what
they are doing in real-time
Source: Mark Graham, OII
Risk and Reward
•  Big data involves considering potenLal rewards at the cost of privacy risk •  Data analyLcs will help organisaLons to beRer understand potenLal rewards as well as associated risks •  This will facilitate achieving an opLmal balance between future risk and reward Occam’s razor
William of Occam studied theology at the University of Oxford
from 1309 to 1321, but never completed his master's degree
•  Occam’s razor (principle of parsimony): seek the
simplest model that explains the data
•  How complicated should a model be?
•  How can we evaluate and compare models?
•  What are the appropriate benchmarks?
•  Examples:
–  Parsimonious risk indices (catastrophes, climate)
–  Diagnostics for decision-making
Global Trends
•  By 2050, world population will be 9 billion with
70% living in cities
•  Security of food, water and energy
•  Today, 925 million people experience hunger, with
food price spikes and volatility threatening the
sustainability of global food security
•  Interconnected risks: Rising food prices à
Tunisian street vendor, Mohamed Bouazizi, sets
himself on fire and dies à Ben Ali toppled from
from power à Arab Spring
Model Risk
•  Numerous major losses which were not adequately
–  Japan ($37bn); NZ quake ($12bn); Thai flood ($10bn)
•  Vendor model changes
–  Modelled versus real world & uncertainty
•  Regulation associated with Solvency II and increased
focus on 1 in 200 year events
–  Compliance issues relating to model uncertainty and
risk assessment
Systemic Risk
•  Exposure is more interconnected than ever
–  Sovereign risk, Arab Spring, Spatiotemporal extremes
•  Supply chain risk (outsourcing, lean
manufacturing, just in time inventory)
•  Local and regional models are inadequate for
assessing potential global losses
•  Thailand flooding was an unexpected loss
–  $10 bn (forecasted to reach $20 bn)
–  Contingent business interruption
–  Semiconductors, car manufacturing
Catastrophe Modelling
Risk = Hazard x Exposure x Vulnerability
Under certain assumptions,
extreme value theory or
simulation can be used to
perform extrapolation in order to
assess the probability of extreme
Loss from a 1 in 200 year event
is estimated at $152 billion. Loss
from Katrina was $90 billion.
Loss versus wind speed
Risk Forecasting
•  Quantitative predictive modelling informed by
qualitative analysis
•  Fusion of scientific knowledge and empirical
investigations based on data analysis
•  Provide prospective rather than retrospective
risk analysis
•  Bridge gap between state of the art modelling
and real-world decision making
Catastrophe risk
•  Member of the Willis Research Network working on
parsimonious models for catastrophe risk assessment
•  Assessment of economic losses arising from disasters
•  Demonstration of ability of atmospheric models to
generate synthetic hurricanes
•  Potential for global windstorm risk assessment
Loss estimates
from calibrated
HIGEM model
(solid) versus
commercial CAT
model (dashed).
Insufficient historical data
•  Ongoing collaboration with EU partners in the
SafeWind consortium
•  Assessment of risk of extreme wind speed to
measure resource
•  Wind farms sites only have a few years of data
•  Dynamical atmospheric models
•  Reanalysis data (~ 50 years)
Anastasiades & McSharry (2014). Wind Energy
Regulation: acceptable risk Q2
•  Pillar 1 of Solvency II will require (from 01Jan-2016) insurers operating in EU to use a
quantitative approach for calculating their
solvency capital ratio (SCR).
•  The SCR is the capital an insurer should hold to
meet its obligations over the next year with a
probability of at least 99.5%.
•  This corresponds to having sufficient capital to
withstand the loss from a 1 in 200 year event
Scenarios and model averaging
Hellman: range [2e-4, 5e-3]
Barrett: range [0.00001, 0.07]
President Obama & Gen.Powell: <0.001%
Focus group to assign weights to all three
models and combine this into a final
distribution for the estimate
•  Other scenarios should also be considered
with new countries having nuclear weapons
Dynamical systems
•  Intuitively, based on linear equilibrium systems, we
expect shocks to die away
•  Nonlinear dynamical systems (NDS) help us
understand the impacts of shocks in the real world
•  NDS tend to have multiple attractors (some safe
for humans and some not)
•  Moving from a safe attractor to an unsafe attractor
may be detrimental and difficult to reverse
Bifurcation diagram
Logistic map: xn+1 = axn(1-xn)
McSharry & Smith, Physical Review Letters, 1999.
Critical Transitions
Source: http://www.sparcs-center.org/key-concepts/tipping-points-and-resilience.html
•  Extreme events often cluster in time and space
•  Independent analysis of risks may be misleading
•  Using a Poisson assumption for a long memory
process will underestimate the risk
•  Correlated risks refer to a combination of
hazards such as wind and rain
•  Earthquake and tsunami, Japan (2011) cost $235
billion (World Bank) due to multi-hazards,
nuclear crisis and loss of economic activity
•  (1) PPPs to better understand risks by
constructing open-access models, improving data
quality and embracing forward-looking risk
forecasting techniques;
•  (2) Quantitative models can complement
qualitative reasoning to generate scenarios
•  (3) Parsimonious risk indices for communicating
to decision-makers and policy-makers
•  Email: [email protected]