Integration & Aggregation in Risk Management: An Insurance Perspective Stephen Mildenhall Aon Re Services May 2, 2005
Overview Similarities and Differences Between Risks What is Risk? Source-Based vs. Characteristic-Based Classification Theoretical Tools Theoretical and Practical Challenges of Risk Integration Dependencies Modeling Philosophy & Guidelines Model Insights & Decision Making What Can We Expect From a Model? 2
What is Risk? Risk: The Possibility Actual Differs From Expected Balance Sheet Entries, Accruals, Valuations Inadequate or Redundant or Both Three Characteristics of Risk Severity Time Dependence Analysis/Synthesis Framework Analyze Severity & Time Components Separately Synthesis Requires Understanding of Dependence Between Risks 3
Classification of Risks Source-Based Classification (Practitioner) Underwriting, Credit, Market, Liquidity, Operational Developed Since 1990s in an Insurance Context Lowe, Standard Integrated DFA & Decision Support System, 1996 Catastrophe Models, Early 1990s Characteristic-Based Classification (Academic) Severity of Risk: Theory of Probability Distributions Developed Since 1700s Bernoulli, de Moivre, Laplace, Poisson, Gauss, Pareto Extreme Value Theory, Thick-Tailed, Sub-Exponential, Distributions Time Element: Stochastic Processes Developed Intensively Since 1930s Lévy, Khintchine, Kolmogorov, Doob, Meyer, Itô Brownian Motion, Markov Processes, Lévy Processes Critical to Development of Finance Dependence: Statistical Association, Copulas Newer Area of Research Since 1950s Fréchet, Sklar 4
Time Characteristics of Risk Static View of Risk P/C Actuaries Highly Trained in Static View of Risk What is Distribution of AY Ultimate Loss? Dynamic View of Risk ERM Requires Dynamic View of Risk How Will Booked AY Ultimate Evolve Over Time? Do Evaluations Between Statements Matter? (CP190, must at all times ) Theory of Stochastic Processes Highly Developed Cornerstone of Modern Finance Situation Vacant: Joint Stochastic Process Model (Paid Loss, Case Incurred, Bulk Reserve) t Bulk Reserve = f (Paid Loss, Case Reserve) Simulation of Ultimate Loss Must Be Expanded To Simulation of Evolution of Paid Loss, Reserve & Ultimate Loss Over Time Approach Crucial to Modeling Reserve Uncertainty 5
Time Characteristics of Risk 7,000 Losses Incurred During Exposure Period 6,000 5,000 4,000 3,000 2,000 1,000 Ultimate Distribution Specified 0 1/1/05 4/2/05 7/2/05 10/1/05 12/31/05 4/1/06 7/1/06 9/30/06 12/30/06 3/31/07 Expected UPR Cumul. Paid Case Incurred Ultimate+UPR 6
Time Characteristics of Risk Risk Can Evolve in Jumps or Continuously or Both Price Evolution of Contract to Pay A Portion of US Hurricane Losses in Sept. 2005 vs. US Earthquake Losses in Sept. 2005 Earthquake Contract With Event Hurricane Contract With Event Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Hurricane Contract, No Event Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 7
Time Characteristics of Risk Two Basic Processes Continuous Evolution: Brownian Motion Jump Evolution: Poisson Process Aggregate Loss Model Gives Jump Process A = X 1 + L+ X N Frequency N, E(N)=Expected Counts Per Unit Time N Often Poisson Severity X From Usual Suspects Generalizing Aggregate Loss Model To Poisson Process Define Frequency Density λ(t) Which Can Vary Over Time Expected Frequency Between 0 and t Given By Actuaries Well Placed to Analyze & Model Risk Evolution N( t) : = λ( t) dt t 0 8
The Challenge of Risk Integration Next Step In Analysis/Synthesis Framework: Risk Integration The Challenge: Dependence! Long Term Capital Management Marginals & Correlation Structure Do Not Determine Distribution Mean & Standard Deviation Do Not Determine Univariate Distribution Normal Copula t- Copula 9
The Challenge of Risk Integration Structural Economic-Scenario Based Models Correlations & Dependencies Among All Risk Sources, CAS Working Party Quasi-Structural Contagion Models (Glenn Meyers) Bivariate Fourier Transform (David Homer) Iman-Conover Method (SM) Copulas Reproduce Qualitative Behavior Useful When Aggregate All That Matters Use FFTs to Add Zero Mean White Noise 10
Iman-Conover Method Iman Conover (IC) Method Given Input Sample from Desired Marginal Distributions Re-order Sample to Have Same RANK ORDER as a Reference Multivariate Distribution With Desired Linear Correlation Method Effective Because Rank and Linear Correlation Close Easy to Produce Reference Multivariate Distributions IC Used By @Risk Software IC Algorithm, Inputs Sample (n x r matrix) From Marginal Distributions E.g. n ~ 10,000, r=2 for Bivariate Distribution Correlation Matrix (r x r matrix) IC Algorithm, Output Sample Re-ordered With Desired Correlation Reference Distributions Generated Using Choleski Trick Elliptically Contoured Distributions (Normal, t, Laplace) 11
Copulas Copula: A Multivariate Distribution With Uniform Marginals Sklar s Theorem: Copulas Determine Multivariate Dependencies Pr( X1 < x1, K, X n < xn ) = C( F1( x1 ), K, Fn ( xn )) Copulas Generate Many Different Dependency Structures Simulating From Copulas Can Be Difficult Archimedean Copulas Easy To Simulate From Cook FGM Venter HRT 12
Modeling Philosophy & Guidelines Avoid Sweeping Generalizations Begin With The End In Mind Understand Process Then Model Model Insights: Reasonable & Unreasonable Expectations 13
Avoid Sweeping Generalizations For Every Rule About Risk There Is A Counter-Example Pathological Examples 99 th Percentile As Risk Adjusted Value Any Percentile Can Be Less Than The Mean Implies Negative Risk Load Standard Deviation as Risk Measure Pareto Can Have Same Mean & Lower SD Than a Uniform Uncorrelated But Dependent t-copula vs. Normal Copula Be Aware of Limitations of Assumptions Intellectually Rigorous Framework Desirable Coherent Measures of Risk 14
Begin With the End in Mind Building An ERM Model Like Building A Car Both Require Goal-Driven Design Objectives ERM Goals Include Reinsurance Decisions Capital Determination Capital Allocation Set BU Profit Targets General Business Planning Investment Opportunities Acquisitions Growth Strategy Investment vs. UW Risk Reserving & Capital 15
Understand Process Then Model Don t Let Modeling Expediencies Drive Model Process Workers Compensation Claim Payment Process Driven By Mortality & Medical Cost Escalation Assumptions Not Modeled Well Using Traditional P/C Actuarial Methods Triangle Methods Ignore Changing Claimant Demographics Premium Correlation vs. Loss Correlation Dependence in Results Driven By Premium Dependence Catastrophe Losses Exhibit Quantifiable Loss Correlation Minimum Pension Liability Difference of Asset & Liability Under Statutory Accounting Very Sensitive To Investment Return Assumptions Example: Stock Price Returns 16
Example: Stock Prices Density Log Density -0.010-0.008-0.006-0.004-0.002 0 0.002 0.004 0.006 0.008 0.010 One Minute Return -0.050-0.040-0.030-0.020-0.010 0 0.010 0.020 0.030 0.040 0.050 One Minute Return Empirical Normal Fit Empirical Normal Fit Density of 1 Minute Returns Not Normally Distributed Largest Observed Changes ±4% Most Big Moves Occurred Late In Trading Day, Between 15:10 and 15:20 For Normal Model ± 4% is a 1 in 10 233 Event Actually Occurred Twice in 19,000 Observations Is Difference in Distribution Important? Perhaps! 17
2.0E-04 Mean Example: Stock Prices 3.0 Skewness 1.5E-04 2.0 1.0E-04 5.0E-05 0.0E+00-5.0E-05 1.0 0.0-1.0-2.0-3.0-1.0E-04-4.0-1.5E-04 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000-5.0 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 1.4E-03 Standard Deviation 300.0 Kurtosis 1.2E-03 250.0 1.0E-03 8.0E-04 6.0E-04 200.0 150.0 4.0E-04 100.0 2.0E-04 50.0 0.0E+00 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 0.0 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 Sequentially Computed Moments of 1 Minute Returns, Mandelbrot Converging Moment Test F. Longin, Asymptotic Distribution of Extreme Stock Market Returns, J. of Bus., 1996 69(3) Concluded First Two Moments Exist From 29,000 Daily Observations 18
Example: Stock Prices Bivariate Distribution of 1 Minute Returns For Two Large Stock Companies, Feb-Apr 2005 SD 1 =0.075%, SD 2 =0.103% Correlation 18.34% 0.5% 0.4% 0.3% 0.2% 0.1% 0.0% -0.5% -0.4% -0.3% -0.2% -0.1% 0.0% 0.1% 0.2% 0.3% 0.4% 0.5% -0.1% -0.2% -0.3% -0.4% -0.5% 19
Example: Stock Prices, IC Method Actual Marginals, Normal Copula Actual Marginals, t-copula, 5 DoF Actual Marginals, t-copula, 1 DoF Simulated Marginals, Normal Copula Simulated Marginals, t-copula, 5 DoF Simulated Marginals, t-copula, 1 DoF 20
Use of Model Results What Can We Expect From Models? Model Output Always Reflects Model Assumptions Management Reaction To Events & Feedback Loops Impossible to Model Reasonable Expectations Reinsurance Adequacy & Effectiveness Capital Determination & Allocation Detailed Short-Term Calculations Cash-Flow Projections RBC, BCAR Projections Growth Strategy Adequate Income & Capital to Support Business Plan? Stochastic Analysis of Static Plans Weed Out Bad Strategic Options Unreasonable Expectations Optimize Management Role To Decide Between Efficient Choices No Universal Evaluation Criteria Model Can Provide Guidance Investment Decisions Parrot Assumptions Assumptions Article of Faith Tony Day, Financial Economics & Actuarial Practice, NAAJ 8(3) 21
Summary Actuarial Analysis of Severity Well Developed Theory of Time Evolution of Risk Available & Readily Comprehensible to Actuaries Theory of Risk Dependencies Still Under Development Model With Goal in Mind Question Model Insights; Apply With Caution 22