concurrent session 7: emerging risk, operational …...2015/12/07 · ex-post case study selection...
TRANSCRIPT
Concurrent Session 7: Emerging Risk, Operational Risk and Risk Culture
Joshua Corrigan, Milliman Principal and Convenor of the Actuaries Institute Risk
Management Practice Committee
Dunedin, 21 November 2014
Agenda • Introduction • Operational Risk • Emerging Risk • Risk Culture
INTRODUCTION Section 1
Companies are Complex Adaptive Systems Risk is an undesirable outcome of a complex system
Traditional Risk Management Frameworks Statistical models, assuming constant drivers Registers assuming single characteristics Scenarios “imagined” Emerging risks by spotting events
Frameworks based on complex systems Descriptions of risk profile taken holistically Scenarios derived from risk profile Models integrate all types of information Emerging risks spotted early from system
Risk management can be hard if looked at it through the wrong lens
OPERATIONAL RISK Section 2
Milliman Research Report • Recently published global research report, authored by
myself and Paola Luraschi (Milan) with input from global consultants
• Available for download at http://au.milliman.com/perspective/operational-risk-modelling-framework.php
• All developed markets
• Current and emerging techniques
• Operational risk assessment is a hot topic in the finance
industry and coming under increasing stakeholder scrutiny
Operational Risk Capital • Graph shows aggregate
required risk capital of top 4 Aussie banks as at end-2012 (99.9% VaR in AUD Billions)
• Op risk capital approximately double the aggregate of interest rate and market risk
• Roughly, wealth management accounts for around 10% of this = $0.9 Bn
Nature of Operational Risk Events Distribution of Number of Events by Size (ORX) Distribution of Total Gross Loss by Size (ORX)
Financial and Physical Consequences Industry Low Severity
High Likelihood Medium Severity
Medium Likelihood High Severity
Low Likelihood
Banking ATM failures Online security breach
Rogue trader
Insurance Claims processing Regulatory compliance failure
Mis-selling Mis-pricing
Mining Transport service interruption
Environmental contamination
Mine collapse
Energy Meter reading errors
Environmental contamination
Oil spill Gas plant fire
It’s all about the loss generation mechanisms, which are highly heterogeneous. Is the system generating the LGM stable or dynamic?
Model Framework Choices Risk identification, assessment, monitoring, mitigation, appetite etc. all depend upon the perspective taken. Traditional and statistical frameworks focus mainly on above the water line items, appropriate for stable systems. New complex systems based frameworks focus on dynamic systems, below the line items, embracing: • Holism • System drivers and dynamics • Non-linearity • Human bias • Emergence
Basic Indicators
Standard Formulas
LDA
Causal Models
Scenario Analysis
Basic Indicator and Standard Formula Country / Sector Indicator Factor (indicative)
Global, Basle II Gross income 12% to 18%
EU, Solvency II BSCR, premiums, liabilities, expenses
Floored at 30% of BSCR + 25% UL expenses
Australia, LAGIC Premium, liabilities, claims
Varies for Life vs General
Japan, SSR “BSCR” 3% if P&L < 0 2% if P&L > 0
South Africa, SAaM
BSCR, premiums, liabilities, expenses
Varies for Life vs General; Floored at 30% of BSCR + 25% UL expenses
Taiwan, RBC Premiums, AUM 0.5% life, 1% annuity, 1.5% other, 0.25% AUM
USA, Europe ex EU, Other Asia, Russia, NZ
None!
Operational risk capital scales in line with broad business metrics such as: • Gross income • Premiums, claims, expenses • Liabilities, Assets / AUM • Capital Assumes stable loss generation mechanisms (LGM) Simple, transparent, cheap, but… main problem is that it isn’t linked to the LGM itself ! • Rough proxy only • No incentive to manage op risk • Enables gaming of the system
Quant Risk Assessment or Scenario Analysis 1. Hypothesize loss severity and likelihood of
possible scenarios 2. Generally assume scenario independence, use
generalized binomial distribution to estimate loss distribution and thus capital (VaR / CTE).
3. Or assume linear dependence, use correlations
Common method currently used Typical method used for Australian Superannuation entities (SPS 114) • ORFR must reflect the size, business
mix and complexity of the entity’s business operations
Forward looking and transparent, but suffer from: • selection bias • the when to stop problem • human bias (e.g. 1 in 1000 event?) • rubbery inter-relationship assumps • lack of uncertainty • allowance for complexity • no ability to use inference
Basle II allows for the use of an Advanced Measurement Approach (AMA) with regulatory approval. Current common practice in leading banks (including the big 4 in Aus). Distribution calibration leverages multiple data sources: • Internal loss data (ex-post) • External loss data (ex-post) • Scenario analysis (ex-ante) • Business environment and internal
control factors (ex-post, current, ex-ante)
Loss Distribution Approach (LDA)
LDA Distribution Choices Severity Distributions
• Continuous: Lognormal, Pareto, Gamma, Weibull
Frequency Distributions • Discrete: Poisson, Negative
Binomial
• Choice of prior distribution critical for low frequency events
Aggregation • Typically simulate the compounding effect of variation and uncertainty
through statistical models with dependency structures • Typical to assume independence between frequency and severity
𝜌 =1 ⋯ 𝜌1𝑛⋮ ⋱ ⋮𝜌𝑛1 ⋯ 1
Individual elements modelled on a marginal basis. Variability introduced using “random” behaviours which are assumed to follow particular shapes. Overall behaviour “reconstructed” through the use of dependency structures such as
correlation or copulae.
Issues • “Thing” being modelled is a complex adaptive system • These exhibit emergence and adaptation • Historical data therefore irrelevant for many behaviours
𝜌 =1 ⋯ 𝜌1𝑛⋮ ⋱ ⋮𝜌𝑛1 ⋯ 1
Models are not often used to understand “modal” behaviours…they are used to understand extremes. But the mechanisms of these behaviours are likely to be different to those seen often and are likely to adapt over time. Emergent behaviour requires us to
focus on interactions, but these modelling methods artificially set these.
An Anthropological Study of Op Risk
4. “The Business” gets on with life
2. “The Business” imparts wisdom
3. “The Business” is shown the model
1. Modeler meets “The Business”
Modelling Adaptive Systems x
x
Parametric shape assumed
x
x x x x x x x x Fit to historical data
0 1 2 3 4 5 6 7 8-6
-4
-2
0
2
4
6
8
10
12
x
y
Dependency locked in
Real distributions show far more variety of outcomes
Experts know that behaviours of this risk have two modes of operation
Experts know that this risk is nearly always benign but has a nasty sting in the tail
Structural / Causal Models 1. Elicit system structure
2. Identify critical drivers
3. Define driver states
4. Define inter-relationships
5. Aggregation and analysis
Loss outcomes are conditioned upon the underlying states of the drivers / risks constituting the LGM system “System” in the context of a complex adaptive system Designed to capture the important dynamics actually driving operational risk Incorporates and leverages the beneficial features of SA and LDA
Case Study: Rogue Trader Scenario
System Structure of Causal Drivers
Drivers of Rogue Trader Behaviour Willingness to Commit Fraud
Capture uncertainty in states of each driver
Capture non-linearities
Significant increase in the willingness to commit fraud driven by: – Increase in the view of potential profits
from committing fraud;
– Decrease in risk aversion; and
– Decrease in risk alignment of HR remuneration and benefits policy
Rogue Trader Scenario Modelling the Expected Loss
Model capture core drivers of exposure: – Likelihood of fraud
• Knowledge of Weaknesses • Willingness • Capability
– Severity of fraud
Total Expected Loss: – Mean: ~ 0.2 billion – 99.5%: ~ 2.9 billion
Sensitivity Analysis:
EMERGING RISK Section 3
Complexity / Connectivity / Emergence WEF Global Risks Map 2014
Complex systems mean you can’t understand the whole by only studying the sum of the parts. It is the inherent and dynamic relationships between risks, causal drivers and outcomes that is key. Simple measures of dependency such as linear correlation are typically misleading Risks relating to complex adaptive systems exhibit emergent properties
Predicting Black Swans • Emerging risk events are simply new combinations
of known risk characteristics • We can analyse which risk characteristics exhibit
evolutionary change and hence are more likely to evolve into new emerging risk events
+ =
Paper presented in May 2013 at Actuaries Summit in Sydney
(a) paired fins, (b) jaws, (c) large dermal bones, (d) fin rays, (e) lungs, and (f) rasping tongue
Cladistics Technique - a Simple Example
Cladistic approach Scenario
Characteristics
1 2 3 4 5 6
A N N N N N N
B Y Y N N N Y
C Y N Y Y Y Y
D Y N Y N Y N
Most parsimonious solution (i.e. fewest changes)
Ex-post Case Study Selection of Large Derivative Trading Losses
Socitete GeneraleAmaranth Avisors
LongTerm Capital Management
Sumitomo Corportation
Aracruz CeluloseOrange CountyMetallgesellschaft
UBSBarings Bank
UBSNational Austrailia Bank0
1
2
3
4
5
6
7
8
9
1985 1990 1995 2000 2005 2010 2015
2011
Equ
ival
ent U
SD B
illio
ns
Socitete Generale
Amaranth Avisors
LongTerm Capital Management
Sumitomo Corportation
Aracruz Celulose
Orange County
Metallgesellschaft
Showa Shell Sekiyu
Kashima Oil
UBS
CITIC Pacific
Barings Bank
BAWAG
Daiwa Bank
Groupe Caisse d'Epargne
Sadia
Morgan Granfell & Co
Askin Capital Management
West LB
AIB Allfirst Financial
Bank of Monreal
China Aviation Oil
UBS
• Derivative losses seem to show no sign of abating in term of either frequency or severity
• How can we understand these events?
• Are they homogenous or heterogeneous?
• Are they relevant to my company?
• How can we understand the next emerging operational risk event?
Fraud clade
Normal trading activity gone wrong & primary activity financial / investing
Derivatives clade
1 Involving Fraud
2 Involving Fraudulent Trading
3 To Cover Up a problem
4 Normal trading activity gone wrong
5 Trading in Excess of limits
6 Primary Activity Financial or Investing
7 Failure to Segregate Functions
8 Lax Mgmt/control Problem
9 Long-term accumulated losses >3 years
10 Single Person
11 Physicals
12 Futures
13 Options
14 Derivatives
Cladogram of Drivers / Risks
RISK CULTURE Section 3
People are at the Heart of Risk Management Social sciences – risk is a social construct built on culture
Michael Thompson, IIASA, Dave Ingram, Willis Re
Anthropology
As scenarios are given detail, and cause and effect are described, it
is quite usual to find dis-agreement emerging – especially
with regard to uncertainties.
This is described by the Cultural Theory of Risk which asserts a
four-fold typology with respect to risk attitude – aligned with four types of human organisation.
Social science teaches us that
probability estimates are subjective – risk is socially
constructed.
Fatalist / Pragmatist Hierarchic / Manager
Individualist / Trader Egalitarian / Conservator
Fatalist / Pragmatist Hierarchic / Manager
Individualist / Trader Egalitarian/Conservator
Decision Making Under Contradictory Certainties Different views of risk exists within our stakeholder group. We need to find common ground.
Michael Thompson, IIASA, Dave Ingram, Willis Re
Typical of the sales director.
Risk isn’t a
concern
Typical of CRO / Actuary
Risks needs to be
managed DMUCC
Typical of compliance officer
auditor
See risk everywhere
Non-conformists
Outcomes are unpredictable and
uncontrollable
Diagnosing Risk Culture
• Simple comparative questions elicit opinion about risk processes
• Non-judgemental…diagnostic
Multi-national Insurer: Sub-Culture Insights Significant Differences Prevalent
Group CRO could see who to adapt framework for and in what areas
Multinational Insurer: Culture by Risk Process Significant Differences Prevalent
Questions / Comments?