Stochastic Reserving: Lessons Learned from General Insurance Matt Locke & Andrew D Smith Session F01 10:10-11:00, Friday 20 November 2015 Matt.Locke@UMACS.co.uk AndrewDSmith8@Deloitte.co.uk
Agenda Aims What is the bootstrap? Mortality example Back-test results Does the bootstrap underestimate extreme percentiles? Conclusions 2
Agenda Aims What is the bootstrap? Back-test results Does the bootstrap underestimate extreme percentiles? Conclusions 3
Aims of our investigation We are aiming to: Get a better understanding of the strengths and limitations of the over-dispersed Poisson bootstrap (ODPB) as described by England & Verrall (2002) Compare the predictive distribution from ODPB against the actual outcomes, using generated data Investigate the robustness of the ODP bootstrap s predictions when the model assumptions are violated 2 1.5 1 0.5 0 6 5 4 3 2 1 0 Low Medium High Extra High 10% rw vol 20% rw vol 4
Aims of our investigation We are not aiming to: Compare the performance of the ODP bootstrap with that of other mechanical or judgement based methods Bootstrap methods applied to paid claims; we do not consider incurred triangles or frequency/severity models here Promote or discourage the use of: The ODP bootstrap Other mechanical methods Judgement based methods 5
Agenda Aims What is the bootstrap? Mortality example Back-test results Does the bootstrap underestimate extreme percentiles? Conclusions 6
Bootstrap in Stochastic Reserving Delay to payment 1 2 3 4 5 6 Year of origin 2009 2010 2011 2012 2013 2014 Simulated runoff scenarios 7
Bootstrap Steps: Parameter Estimates Proportion of Claims Paid Parameters include: Development patterns Ultimate losses Variability Delay (years) 8
Bootstrap Steps: Forecasting Also called noise or process error Simulating one or more future claim scenarios based on estimated parameters (we use gamma distributions here) This is a familiar approach for many other risks besides reserving uncertainty 9
Bootstrap Steps: Back-Casting Re-creating hypothetical historical claim scenarios based on estimated parameters May use re-sampled residuals (non-parametric) or analytical distributions (parametric) 10
How the Steps Fit Together Input data triangle Back-cast by re-sampling residuals to capture parameter uncertainty Estimate runoff pattern and residuals Using basic chain ladder Re-estimate runoff pattern Stochastic forecasts incorporate both parameter and process error 11
Agenda Aims What is the bootstrap? Mortality example Back-test results Does the bootstrap underestimate extreme percentiles? Conclusions 12
Mortality Rates Example 1 1960 1970 1980 1990 2000 2010 0.1 m 80 m 70 m 60 0.01 m 50 m 40 m 30 0.001 m 20 Male population data.source: www.mortality.org 13
Model Uncertainty for Longevity Risk Probability Distributions for UK Male Annuity aged 65 (Based on Currie et al, 2013) 2.5 Probability Density 2 1.5 1 0.5 LC DDE LC(S) CBD Gompertz CBD Pspline APC 2DAP 0 11 12 13 Annuity Value 14 14
Fitting an AR2 Model to ln[m 50 ] 1 Coefficient of X t-2 Base case 25x bootstrap 0.5 0-1 -0.5 0 0.5 1 Coefficient of X t-1-0.5-1 15
Projections under the Base Model 1 1980 2000 2020 2040 2060 2080 2100 m 50 0.1 0.01 0.001 16
What Lies behind the Patterns? Although there is a general increasing trend in life expectancy, we know that some historic causes have had particularly large impact: Discovery of antibiotics (penicillin etc) Reduction in smoking Smoking in the UK has fallen from 50% to 20% in the last 40 years and it cannot feasibly fall to -10% over the next 40 years What future events are we aware of today whose occurrence is likely to be coupled with a significant impact on UK longevity? Introduction of plain cigarette packaging in the UK Use of novel diagnostic biomarkers KRAS targeted cancer treatment Genetic screening NHS Bowel Cancer Screening Programme Stem cell therapy and Parkinson's disease Polypill scenario Development of a universal influenza vaccine Source: Longevity Catalysts working party 17
Bootstrap Projections More Extremes 1 1960 1980 2000 2020 2040 2060 2080 2100 0.1 Orange = base Green = bootstrap 0.01 The bootstrap is blind to underlying causes. 0.001 18
Two Dimensional Examples We have shown one dimension of stochastic mortality models Two-dimensional (stochastic sheet) models have now become market standard. Several models proposed by Cairns-Blake-Down and by Plats Same principles apply for two-dimensional bootstraps (permuting residuals across calendar years and, within limits, across age at death) As in the univariate case, the bootstrap impact is to widen funnels of doubt, especially at extreme percentiles and over long time horizons. 19
Agenda Aims What is the bootstrap? Mortality example Back-test results Does the bootstrap underestimate extreme percentiles? Conclusions 20
Back-Testing Compare actual outcomes to a predictive distribution Actual outcome density 2 1 Predictive distribution ROC07 (High vol) 0 0% 50% 100% 150% 200% 21
Back-Testing the ODP Bootstrap Wait for the outcome Bootstrap scenarios x99 12 November 2015 22
Ranking the Outcomes Take 100 future claim scenarios 1 actual outcome and 99 from bootstraps Sort into increasing order of outstanding claims Divide into 10 buckets, each containing 10 observations aggregate outstanding claims Suppose the actual outcome and the bootstrap are independent samples from the same distribution Then there is 1-in-10 chance the red lies in each bucket 23
The Back-Test Bootstrap multiple historical claim triangles Multiple insurers Multiple projection years Or multiple random realistic triangles Aggregate the bucket counts across bootstraps Back-test passes if 10% of actual outcomes lie in each bucket Within random sampling tolerance 24
Back-test on Real Data: Leong et al (2012) 3 2 Commercial Auto Liability 3 Private Passenger Auto 3 2 Homeowners 1 2 1 0 1 0 3 2 Workers Compensation 0 3 Medical Professional Liability 3 2 Other Liability 1 2 1 0 1 0 0 3 2 Commerical Multi Peril Too many actual outcomes lie in the top and bottom bootstrap deciles, so bootstrap distribution too narrow 1 0 Different companies / years not independent so we do not know how significant this effect is 25
The Monte Carlo Back-Test (MCBT) Stage 1 Parametric back-cast Stage 2 One set of stage 2 parameters for each stage 1 back-cast triangle Non-parametric back-cast Stage 3 One set of stage 3 parameters for each stage 2 back-cast triangle 26
Example Output (Long dev, High vol) 1.6 1.4 1.2 Some previous studies focus on the worst 1% 1 0.8 0.6 0.4 0.2 With 150 000 Monte Carlo observations we can use 100 buckets rather than 10 We know the observations are independent, which implies bars below 0.95 or above 1.05 are significant at 95% confidence 0 1% 99% 27
MCBT Results: Proportion > 99%-ile Development pattern length Gamma Volatility Short Medium Long Extra Long Low 1.1% 0.7% 0.7% 0.6% Medium 1.5% 1.1% 0.8% 1.1% High 1.9% 1.5% 1.1% 1.5% Extra High 3.0% 2.7% 1.9% 2.7% 28
Experiment: Omitting the Back-Cast Stage 1 How much is the bootstrap adding? Stage 2 Perform the Monte Carlo Back Test using stochastic projection of the step 2 fitted parameters Therefore have no allowance for parameter uncertainty Stage 3 29
Unbiased Parameters fail Back-Test 9 8 7 It s not conventional to apply bootstrap methods to e.g. Catastrophe models 6 5 4 No bootstrap With bootstrap 3 2 1 0 30
Making Triangles More Realistic Scale parameter varies by delay Lumpy Distributions Change in development pattern by origin year Basic Gamma Model Payment acceleration / deceleration Superposed Inflation What else? 31
Origin Year and Calendar Year Effects 6 5 4 3 2 1 0 No effect 10% pa 20% pa Outcomes in worst 1% bootstrap bucket 6 5 4 3 2 1 0 10% rw vol 20% rw vol All Three Origin PH Calendar PH Inflation Base These figures relate to the long development pattern, and high gamma volatility 32
Origin Year and Calendar Year Effects 6 5 4 No effect 10% random walk vol pa. This is very volatile for inflation 20% random walk vol pa. This says claims inflation is as volatile as a stock market. Real data is unlikely to be this volatile but it shows how far the assumptions have to be violated before the bootstrap really breaks down 3 2 1 0 33
Agenda Aims What is the bootstrap? Back-test results Does the bootstrap underestimate extreme percentiles? Conclusions 34
Extreme Percentile Underestimation Given how simple the concept is, the bootstrap does well for most of the distribution We replicate results of ROC and others that bootstrap does not perfectly capture extreme tails In some instances the bootstrap distribution is less extreme than reality, and in others it is more extreme We would not expect perfection Bootstrap is remarkably robust to moderate assumption violations 35
Some alternatives to the Bootstrap Bootstrap is not the only way to address parameter uncertainty Classical methods estimate parameters by maximum likelihood and derive standard errors from the Fisher information matrix Bayesian methods (England & Cairns, 2009) Single (non-bootstrap) forecast, with the standard deviation multiplied by an adjustment factor Use the Monte Carlo Back Test to solve for the adjustment factor Retain unbiased point estimates as a control 36
Agenda Aims What is the bootstrap? Mortality example Back-test results Does the bootstrap underestimate extreme percentiles? Conclusions 37
The Importance of Quantitative Testing Rhetorical Quantitative Mechanical Methods Subjective methods Bootstrap takes account of parameter error Assumptions do not hold in practice Frequentist peg in a Bayesian hole Use underwriting knowledge, common sense, relevant for the board, practical decisions Telling management what they want to hear, profit smoothing. Prob{outcome > 99%-ile} Robustness to model mis-specification Unbiasedness / efficiency? Wanted: outcome-based tests for subjective methods 38
Points for Discussion (1) What to do about known changes affecting past mortality? Examples: medical breakthroughs, smoking, alcohol consumption, cohort effects? Should we strip them out of the data and put them back into the forecast? Or is that part of the noise we re trying to measure and extrapolate? The bootstrap allows for these mechanically, but only to the extent that these fluctuations affect the past and the future 39
Points for Discussion (2) Parameter and model errors pervade many risk models: market risks, longevity / mortality and, in general insurance, premium risk, cat risk, reserving risk, credit risk, market risk etc. For reserving risk we have the bootstrap. It s not perfect but we d give it 8/10 for capturing model and parameter uncertainty For the other risks, we probably ignore model and parameter error, scoring 0/10 We can try to perfect the bootstrap, but should we prioritise other risks? 40
Acknowledgements We are grateful for the support from the Managing Uncertainty Qualitatively working party, the Managing Uncertainty with Professionalism working party and from our employers Special thanks to Sarah MacDonnell, Tom Wright and Peter England for detailed comments on earlier drafts All views expressed and any remaining errors are ours alone 41
Further Reading Cairns M and England P D (2009) Are the upper tails of predictive distributions of outstanding liabilities underestimated when using bootstrapping? General Insurance Convention (presentation only) http://www.actuaries.org.uk/sites/all/files/documents/pdf/a09england.pdf England, P.D. and Verrall, R.J. (2002) Stochastic Claims Reserving in General Insurance (with discussion). British Actuarial Journal, 8, pp 443-544 http://www.actuaries.org.uk/sites/all/files/documents/pdf/sm0201.pdf. Note that this link is to the originally distributed sessional meeting paper. The is a crucial typographical error for the residual adjustment in Appendix 3 which is corrected in the paper finally published in the BAJ, and also in this paper. Efron B & Tibshirani, R J (1993). An introduction to the bootstrap. Chapman and Hall. General Insurance Reserving Oversight Committee Working Party (Chair: Lis Gibson, 2007) Best Estimates and Reserving Uncertainty. General Insurance Reserving Oversight Committee Working Party (Chair: Neil Bruce, 2008) Reserving Uncertainty. Previous two papers available here: http://www.actuaries.org.uk/practice-areas/pages/general-insurance-reservingoversight-committee-gi-roc-0 Leong J, Wang S and Chen H. (2012) Back-Testing the ODP Bootstrap of the Paid Chain-Ladder Model with Actual Historical Claims Data. Casualty Actuarial Society E-forum. http://www.casact.org/pubs/forum/12sumforum/leong_wang_chen.pdf Pinheiro, P J. R; João Manuel Andrade e Silva and Maria de Lourdes Centeno. (2003). Bootstrap Methodology in Claim Reserving. The Journal of Risk and Insurance, 70: 701-714. Shapland M R and Leong, J. (2010) Bootstrap Modeling: Beyond the Basics. Casualty Actuarial Society E-forum. 42
Questions Comments The views expressed in this presentation are those of invited contributors and not necessarily those of the IFoA. The IFoA do not endorse any of the views stated, nor any claims or representations made in this presentation and accept no responsibility or liability to any person for loss or damage suffered as a consequence of their placing reliance upon any view, claim or representation made in this presentation. The information and expressions of opinion contained in this publication are not intended to be a comprehensive study, nor to provide actuarial advice or advice of any nature and should not be treated as a substitute for specific advice concerning individual situations. Unpaid volunteers have produced this presentation to promote discussion in the public interest. You may copy, distribute, display and perform the work and make derivative works based on it provided you acknowledge the original authors. 43