2006 General Meeting Assemblée générale 2006 Chicago, Illinois 2006 General Meeting Assemblée générale 2006 Chicago, Illinois Canadian Institute of Actuaries.

Présentation au sujet: "2006 General Meeting Assemblée générale 2006 Chicago, Illinois 2006 General Meeting Assemblée générale 2006 Chicago, Illinois Canadian Institute of Actuaries."— Transcription de la présentation:

1 2006 General Meeting Assemblée générale 2006 Chicago, Illinois 2006 General Meeting Assemblée générale 2006 Chicago, Illinois Canadian Institute of Actuaries Canadian Institute of Actuaries L’Institut canadien des actuaires L’Institut canadien des actuaires

2 Validation of long term equity return models for equity-linked guarantees Mary Hardy, Keith Freeland and Matthew Till

3 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Outline The models Does it matter? Traditional model selection Bootstrap evidence Abusing the bootstrap

4 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Accumulation Factors Let Y t represent log return in t th month 1-year accumulation factors are exp(Y t +Y t+1 +…Y t+11 ) Similarly for 5-year and 10-year 40 years data  4 non-overlapping observations of 10-Year accumulation factor

5 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Canadian Calibration Table Accumulation Factor 2.5 %ile5 %ile10%ile 1-year0.760.820.90 5-year0.750.851.05 10-year0.851.051.35

6 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 US Calibration %ile1-Yr5-Yr10-Yr20-Yr 2.5%0.780.720.79N/A 5%0.840.810.941.51 10%0.900.941.162.10 90%1.282.173.639.02 95%1.352.454.3611.70 97.5%1.422.725.12N/A

7 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

8 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 S&P data not much auto-correlation but correlation is not always a good measure of independence notice bunching of poor returns (eg last 2 years) and association of high volatility with crashes –ie large movement down more than up

9 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Some equity models …. Regime Switching Log Normal (Hardy, 2001) GARCH(1,1) MARCH (Chan and Wong, 2005) ‘Stochastic Log Volatility’ (AAA C3-Phase 2) Regime Switching Draw Down (Panneton, 2003) Regime Switching GARCH (Gray 1996, JFE)

10 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 RSLN-2 REGIME 2  2 High Volatility  2 Low Mean  2 REGIME 1  1 Low Volatility  1 High Mean  1

11 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 11 The RSLN-2 Model The regime process is a hidden Markov process 2 Regimes are usually enough for monthly data. 2 Regime model has 6 parameters: –  ={  1,  2,  1,  2, p 12, p 21 } Regime 1: Low Vol, High Mean, High Persistance (small p 12 ) Regime 2: High Vol, Low Mean, Low Persistance (large p 21 )

12 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 GARCH(1,1) Where  t ~ N(0,1), iid Given F t-1,  t is the only stochastic element We generally require  1 + 

13 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 MARCH (2;0,0;2,0)

14 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 MARCH(2;0,0;2,0) MARCH(K; p 1 …,p K ;q 1,…q K ) is a mixture of K AR-ARCH models, p j and q j are the AR-order and ARCH-order of the j th mixture RV According to Chan and Wong, provides superior fit to 3 rd and 4 th moments of monthly log-return disn cf RSLN

15 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 SLV Z v,t and Z y,t are standard normal RVs, with correlation  The v t process is constrained by upper and lower bounds

16 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 SLV According to C3P2, SLV –“Captures the full benefits of stochastic volatility in an intuitive model suitable for real world projections” –Stoch vol models are widely used in capital markets to price derivatives… –Produces very “realistic” volatility paths

17 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Regime Switching Draw Down (RSDD)

18 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 RSDD 2 Regimes proposed by Panneton D t is the draw-down factor RSLN-2 is recovered when   =0, for  =1,2 Captures ‘tendency to recover’

19 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 RSGARCH Two GARCH regimes Markov switching After Gray (1995)

20 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Does it matter? 6 models, each being championed by someone. 2 RS, 2 conditional heteroscedatic, 1 ‘stochastic volatility’. Each fitted by MLE (-ish) to S&P500 data Does it make any difference to the results for Equity-Linked Capital Requirements?

21 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Two methods for Equity Linked Life Insurance Actuarial Approach: –Simulate liabilities, –apply risk measure, –discount at risk-free rate Determines the economic capital requirement to write the contract for a given solvency standard.

22 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Two methods for Equity Linked Life Insurance Dynamic Hedging Approach –Simulate hedge under real world measure –Estimate distribution of unhedged liability –Apply risk measure and discount at r-f rate –Add to cost of initial hedge

23 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Example Contract Single Premium GMAB, premium P Issue age 50, MER=3% p.y. Guarantee risk premium = 0.2% p.y. Deterministic mortality and lapses Assets in policyholder’s fund = F t at t F t follows model stock returns, less MER

24 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Example Contract Benefit on death or maturity is max(F T, G T ) Guarantee G t at t – G t =P for 0 < t  10 – G t =max(F 10,P) for 10 < t  20 Payable on death or maturity

25 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

26 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Does the model matter using the actuarial approach? Oh Yes !!!

27 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Using hedging? Straight Black-Scholes (LN) delta hedge –r=0.05;  =0.20 Simulate additional cost arising from –Discrete hedge –Model Error ie P-measure is GARCH; RSLN etc –Transactions costs

28 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

29 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Does the model matter using the hedging approach? Not so much….

30 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 But Few companies are hedging Pressure to adopt models giving lower capital requirements Can we use traditional methods to eliminate any of the models?

31 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Likelihood Comparison Model# parameters Max LL RSDD81047.1 MARCH71039.8 SLV71035.2* GARCH41030.1 RSLN61042.0 RSGARCH 81056.0

32 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

33 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

34 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

35 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

36 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

37 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

38 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 So far … Likelihood based selection doesn’t help much AIC is too simple, BIC depends on sample size, LRT has technical limitations Residuals can be useful, but are tricky in multifactor cases

39 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 so good… The regime switching models look good on likelihood and on residuals (all pass J-B test) But -- big difference in application between rsdd and rsln or rsgarch What causes the big difference? Which rs model should we believe?

40 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

41 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

42 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Bootstrapping time series The traditional bootstrap is applied to independent observations. Dependent time series require different treatment. Order matters.

43 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Time series bootstrap Bootstrap from original observations in blocks of b consecutive values. If the blocks are too small, lose dependence factor  results too thin tailed (if +vely autocorrelated) If blocks are too large lose data,  results too thin tailed (extreme results averaged out) Block size 3-6 months for stock data

44 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Bootstrap Quantile Estimates 1-Year Accumulation Model2.5%ile5%ile10%ile Bootstrap 90% CI 0.67  0.870.76  0.910.84  0.97 RSDD 0.7680.8310.901 RSLN 0.7640.8290.908 RSGARC H 0.7920.8470.910 This doesn’t help us much.

45 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 10-year accumulation factor We can do the same thing But the original data only has 4 non- overlapping observations minimum 10-year observed AF is estimate of 1/5=20%ile So we bootstrap B samples of 4 observations

46 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Bootstrap Quantile Estimates 10-Year Accumulation Model20 %ile Bootstrap 90% CI Mean 0.95  2.83 1.706 RSDD1.953 (1.92,1.97) RSLN1.773 RSGARCH1.660 (1.63,1.68) And this doesn’t help us much either.

47 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Oversampling Bootstrapping re-samples from original data  Four 10-year accumulation factors from 584 observations What happens if we break the rules and keep sampling? The ‘empirical distribution’

48 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Oversampling If data are independent or +vely autocorrelated then oversampling  thin tails – positive bias for low quantiles; negative for high quantiles –Bias should be small for large original sample

49 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Oversampling If data are +vely auto-correlated and block size is not large enough to capture long down or up periods  even thinner tails

50 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Oversampling If data are –vely auto-correlated –Oversampling with small block size will fatten tails –Overall effect depends on correlation But we are estimating AFs so we also look at these correlations.

51 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

52 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Back to the data No significant negative autocorrelations… So oversampling should over- estimate left tail quantiles (on average) And underestimate right tail quantiles

53 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Left tail, 10-Year AFs Model2.5%5%10% Bootstrap (sort of…) 1.041 (1.03, 1.06) 1.228 (1.20,1.25) 1.478 (1.47,1.49) RSDD1.2771.4391.653 SLV1.0821.2541.468 RSGARCH0.9051.0861.315 RSLN0.9141.1051.378

54 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006

55 2006 General Meeting Assemblée générale 2006 2006 General Meeting Assemblée générale 2006 Summing up We need to pay attention to model econometrics Huge financial implications – especially with traditional actuarial methods Abusing the bootstrap offers some info Multiple state models for equity returns.