Capital requirements and portfolio optimization under solvency constraints: a dynamical approach

Transcription

1 Capital requirements and portfolio optimization under solvency constraints: a dynamical approach S. Asanga 1, A. Asimit 2, A. Badescu 1 S. Haberman 2 1 Department of Mathematics and Statistics, University of Calgary, Canada 2 Cass Business School, City University London, UK ASTIN Colloquium, The Hague May 21-24, 2013

2 Outline 1 Preliminaries and solvency constrained optimization 2 3 Empirical analysis 4

3 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint The setup A discrete-time framework with the set of trading dates T = {t t = 0,..., T }. A portfolio of n assets with the gross return process over the period [t, t + 1] defined by R t+1 = (R 1,t+1,..., R n,t+1 ) T. F t = σ(r 1,..., R t ) and we use the notation E[ F t ] = E t [ ]. Let x t = (x 1,t,..., x n,t ) T be the portfolio weights satisfying n the budget constraint, x i,t = 1, and the no short sales i=1 constraint, x i,t 0, i = 1,..., n. We introduce one-period optimization problems for a non-life insurance company over [t, t + τ], where τ is the solvency horizon. No rebalancing is allowed during the solvency period.

4 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint p t - the aggregate premium available for investment at time t. c t - the regulatory initial capital provided by the shareholders. No other premiums are collected and no capital is issued or retired between t and t + τ. The insurer s liability is modelled by a univariate random variable Y t+τ. It represents the aggregate claim amount over the solvency horizon which is assumed to be paid at time t + τ. We define the insurer s net loss: L t,t+τ := L(c t, x t ) = Y t+τ (p t + c t )R T t+τ x t. No other sources of risk other than the ones modelled through Y and R.

5 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint Each optimization problems characterized by minimizing the capital requirement c t, subject to two key constraints: (i) Solvency constraint: capital requirement imposed by the insurer s regulator and based on one of the following criteria: Ruin Probability (RP), Conditional Value-at-Risk (CVaR) and Expected Policyholder Deficit (EPD). (ii) Portfolio performance constraint: target return on capital (ROC) provided: ROC t,t+τ = L t,t+τ c t. Santos, Nogales, Ruiz and Van Dijk (2012). Mankai and Bruneau (2012). Cummins and Phillips (2009).

6 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint Optimization with RP constraint Motivated by Solvency II Regime which applies to EU based insurance companies. Identify the capital required to maintain a target level for the ruin probability over a specified period of time. min c t,x t c t s.t. E t î 1{Lt,t+τ >0}ó 1 α, E t î ROCt,t+τ ó ROC α, 1 T x t = 1, x t 0, c t 0. α represents the specified solvency level (e.g. α = 99.5%). ROC α is the lower bound for the expected return on capital.

7 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint The solvency chance constraint can be reformulated as a Value-at-Risk constraint, where the VaR of a loss r.v. Z at α is defined: VaR α (Z) := inf{z R : Pr(Z z) α}. The constraint becomes: VaR α t (L t,t+τ ) 0. There are two streams of literature dealing with solving chance constrained optimization. 1. Monte-Carlo estimators for the conditional expectation and perform further appropriate approximations. 2. Solve the chance constraint using the VaR representation.

8 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint 1. - Boyd and Vandenbergue (2004) - convex approximations by eliminating the indicator function. - Caliafore and Campi (2005), Nemirovski and Shapiro (2005) - scenario based approximations. - Nemirovski and Shapiro (2006) - convex approximation based on Bernstein scheme. - Luedtke and Ahmed (2008) - sample average approximation method and non-convex mixed-integer programming (MIP) reformulation, among others Larsen, Mausser and Uryasev (2002) - algorithms based on iterative CVaR optimizations. - Gaivoronski and Pflug (2004) - scenario-based methods. - Wozabal, Hochreiter and Pflug (2008) - difference of convex functions reformulation.

9 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint The Monte Carlo approximation of the solvency constraint: 1 m m j=1 1 {Yt+τ (j) (p t+c t)r T t+τ (j)xt>0} 1 α. This can be reformulated as a MIP problem. However, implementation becomes less efficient when m is large. Our approach: Semiparametric method; 1 m m j=1 E (j) î ó t 1{Yt+τ (p t+c t)r T 1 α. t+τ (j)xt>0} We used: E î F t {Rt+τ = R t+τ (j)} ó = E (j) î ó t. A sufficient condition for convexity of our problem: Y t+τ has a conditional convex survival function.

10 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint Optimization with CVaR constraint Rockafellar and Uryasev (2000) - alternative coherent risk measure to VaR; quantifies the loss severity in case of default. Defined as a weighted average of the corresponding VaR and conditional expected losses which strictly exceed VaR. CVaR coincides with the Expected Shortfall (ES) for continuous distributions. ES constitutes the basis for the target capital according to the Swiss Solvency Test (SST) (EIOPA, 2011). We take β = 99%. s.t. min c t c t,x t CVaRt β Ä ä Lt,t+τ î ó 0, E t ROCt,t+τ ROC β, 1 T x t = 1, x t 0, c t 0.

11 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint Rockafellar and Uryasev (2000) define CVaR: CVaR β (Z) = inf {s + 1 s R 1 β E î ó} (Z s) +. The optimization problem becomes (only solvency constraint): min s,c t,x t c t s.t. s β E tî (Lt,t+τ s) + ó 0. The traditional approach: use MC estimator and reformulate as a Linear Programming (LP) problem. Less efficient when m is large. Our approach: Semiparametric method: s + 1 m m(1 β) E (j) î t (Yt+τ (p t + c t )R T ó t+τ (j)x t s) + 0. j=1

12 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint Optimization with EPD constraint Introduced by Butsic (1994) as an alternative method to the ruin probability for measuring insolvency risk. Constitutes a useful tool in establishing the US Risk Based Capital (RBC) regulatory system (e.g. see NAIC, 2009). Defined as the expected loss in the event of insolvency: EPD(L t,t+τ ) = E t î (Yt+τ (p t + c t )R T t+τ x t ) + ó. Solvency criteria based on a target level of a deficit ratio: EPD Ä L t,t+τ ä E t î Yt+τ ó f. f is the maximum level for the EPD ratio with 0 f < 1. We arbitrarily take f = 0.25%.

13 The setup Optimization with RP constraint Optimization with CVaR constraint Optimization with EPD constraint A similar LP reformulations is available. For consistency, we consider our semiparametric representation: min 1 m s.t. m c t,x t c t j=1 E (j) t î (Yt+τ (p t + c t )R T t+τ (j)x t ) + ó f Et î Yt+τ ó, E t î ROCt,t+τ ó ROC f, 1 T x t = 1, x t 0, c t 0. A sufficient î condition for convexity: (Yt+τ (p t + c t )R T ó t+τ (j)x t ) + is a convex function in ct E (j) t and x t. This depends on the conditional distribution of Y t+τ.

14 Assets follow the DCC - GARCH model of Engle (2002) log R t+1 = m t+1 + ε t+1, ε t+1 F t MVN(0, H t+1 ), H t+1 = D 1/2 t+1 Σ t+1d 1/2 t+1, D t+1 = diag(h 1,t+1,..., h n,t+1 ), Σ t+1 = diag(q 1/2 11,t+1,..., q 1/2 nn,t+1 )Q t+1diag(q 1/2 11,t+1,..., q 1/2 nn,t+1 ), Q t+1 = (1 θ 1 θ 2 ) Q + θ 1 u t u T t + θ 2 Q t. D t+1 is the n n diagonal matrix formed with the univariate conditional variances GARCH(1,1). Σ t+1 is the time-varying conditional correlation matrix of R t+1. Liabilities are LogNormal distributed: Y t+τ LGN(µ t+τ, σ t+τ ). Y t+τ is independent of the enlarged filtration F t σ(rt+τ ).

15 Efficient frontiers and capital allocation Assets: 3-asset portfolios formed with T-Bills, NASDAQ and NYSE. Table: Descriptive statistics for NASDAQ and NYSE log-returns from January 3, July 29, 2011 for a total of 1,656 observations. Index Min Max Mean Std Skewness Kurtosis NASDAQ NYSE Liabilities: aggregate monthly claim amounts on property insurance for the same period used in the assets case. Table: Descriptive statistics for monthly claim amounts from January 3, July 29, 2011 for a total of 79 observations (figures are in thousands Euros). Min Max Mean StDev Skewness Kurtosis

16 Efficient frontiers and capital allocation 2.5 x (a) Conditional variance for NASDAQ 3 x (b) Conditional variance for NYSE (c) Conditional correlation

17 Efficient frontiers and capital allocation 1.04 UNI GARCH UNI GARCH UNI GARCH Expected Return on Capital Expected Return on Capital Expected Return on Capital Optimal Capital (a) RP Optimal Capital (b) CVaR Optimal Capital (c) EPD Figure: Efficient frontiers for DCC, CCC and UNI-GARCH models under the RP, CVaR and EPF -constrained problems.

18 Efficient frontiers and capital allocation Optimal Asset Allocation Nasdaq UNI GARCH Optimal Asset Allocation Nasdaq UNI GARCH Optimal Asset Allocation Nasdaq UNI GARCH Optimal Capital Optimal Capital Optimal Capital (a) RP (b) CVaR Figure: Optimal allocation in Nasdaq. (c) EPD

19 Efficient frontiers and capital allocation Optimal Asset Allocation NYSE UNI GARCH Optimal Asset Allocation NYSE UNI GARCH Optimal Asset Allocation NYSE UNI GARCH Optimal Capital Optimal Capital Optimal Capital (a) RP (b) CVaR Figure: Optimal allocation in NYSE. (c) EPD

20 Efficient frontiers and capital allocation UNI GARCH UNI GARCH UNI GARCH Optimal Asset Allocation T Bill Optimal Asset Allocation T Bill Optimal Asset Allocation T Bill Optimal Capital (a) RP Optimal Capital (b) CVaR Optimal Capital (c) EPD Figure: Optimal allocation in T-Bills.

21 Efficient frontiers and capital allocation Assets (sampled daily): Sample A: January 3, December 32, 2009 (l A = 1, 259) and Sample B: January 1, July 29, 2011 (l B = 397). Liabilities (sampled monthly): Sample A : January 3, December 32, 2009 (l A = 60) and Sample B : January 1, July 29, 2011 (l B = 19). Given the estimates based on Samples A and A, find the optimal solution (c t, x t ). Single rolling window: drop the first τ obs. from Sample A and replace them with the first τ observations from Sample B. Double rolling window: additionally, drop the first obs. from Sample A and replace it with the first obs. from B. Repeat the estimation and optimization steps until the length of the out-of-sample data set is reached (i.e. l B = 19 times).

22 Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Efficient frontiers and capital allocation Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Optimal Assets Optimal Assets Month of Investment (a) RP and single rolling window Month of Investment (b) RP and double rolling window Optimal Assets Optimal Assets Month of Investment Month of Investment (c) CVaR and single rolling window (d) CVaR and double rolling window Optimal Assets Optimal Assets Month of Investment Month of Investment (e) EPD and single rolling window (f) EPD and double rolling window

23 Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Efficient frontiers and capital allocation Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Univariate GARCH Jan 10 Apr Jul Oct Jan 11 Apr Jul Optimal Asset Allocation Nasdaq Optimal Asset Allocation Nasdaq Month of Investment (a) RP and single rolling window Month of Investment (b) RP and double rolling window Optimal Asset Allocation Nasdaq Optimal Asset Allocation Nasdaq Month of Investment (c) CVaR and single rolling window Month of Investment (d) CVaR and double rolling window Optimal Asset Allocation Nasdaq Optimal Asset Allocation Nasdaq Month of Investment Month of Investment (e) EPD and single rolling window (f) EPD and double rolling window

24 Efficient frontiers and capital allocation Using optimal solutions and time series of observed returns over the out-of-sample period, we test the performance of the DCC, CCC and UNI MGARCH models relative to two criteria: 1. Solvency performance: Average total assets invested (premium + optimal capital). Average solvency value. Maximum solvency value. 2. Portfolio performance: Average adjusted ROC. Standard deviation adjusted ROC. Sharpe Ratio. Portfolio Turnover.

25 Efficient frontiers and capital allocation Average solvency values: RP = ĈVaR = EPD = 1 l B l B ) Φ (d t+kτ, k=1 l B 1 ( E[Yt+kτ ] ( ) ) l B 1 β Φ σ t+kτ Φ 1 (β) R T t+kτ z t+(k 1)τ, k=1 1 l B l B k=1 î ( ) )ó E[Y t+kτ ]Φ d t+kτ + σ 2 t+kτ R T t+kτ z t+(k 1)τ (d Φ t+kτ. z t+(k 1)τ = (p t+(k 1)τ + c t+(k 1)τ )x t+(k 1)τ, d t+kτ = log R T t+kτ z t+(k 1)τ + µ t+kτ σ t+kτ,

26 Efficient frontiers and capital allocation Table: Out-of-sample solvency performance under the single rolling window exercise. Solvency Performance Avg. Assets Avg. Solvency Max. Solvency invested Value Value Problem 1. Avg. Max. Ruin Constraint Ruin Probability (%) Ruin Prob (%) Covariance Model DCC CCC UNI Problem 2. Avg. Max. CVaR Constraint CVaR CVaR Covariance Model DCC CCC UNI Problem 3. Avg. Max. EPD Constraint EPD Ratio (%) EPD Ratio (%) Covariance Model DCC CCC UNI

27 Efficient frontiers and capital allocation Table: Out-of-sample solvency performance under the double rolling window exercise. Solvency Performance Avg. Assets Avg. Solvency Max. Solvency invested Value Value Problem 1. Avg. Max. Ruin Constraint Ruin Probability (%) Ruin Prob (%) Covariance Model DCC CCC UNI Problem 2. Avg. Max. CVaR Constraint CVaR CVaR Covariance Model DCC CCC UNI Problem 3. Avg. Max. EPD Constraint EPD Ratio (%) EPD Ratio (%) Covariance Model DCC CCC UNI

28 Efficient frontiers and capital allocation Portfolio performance indicators: ˆµ AROC = ˆσ AROC = SR AROC = Turnover = l B 1 AROC t+(k 1)τ,t+kτ, l B k=1 à l B 1 (AROC t+(k 1)τ,t+kτ ˆµ AROC ) 2, l B k=1 ˆµ AROC ˆσ AROC, 1 l B 1 l B 1 k=1 n i=1 x i,t+kτ x i,t+(k 1)τ. These quantities are computed for the adjusted return on capital (AROC): AROC t,t+τ = (pt + c t )RT t+τ x t E[Y t+τ ] c t 1.

29 Efficient frontiers and capital allocation Table: Out-of-sample portfolio performance under the single rolling window exercise. Portfolio Performance Avg. Std. Sharpe Turnover AROC (%) AROC Ratio Problem 1. Ruin Constraint Covariance Model DCC CCC UNI Problem 2. CVaR Constraint Covariance Model DCC CCC UNI Problem 3. EPD Constraint Covariance Model DCC CCC UNI

30 Table: Out-of-sample portfolio performance under the double rolling window exercise. Portfolio Performance Avg. Std. Sharpe Turnover AROC (%) AROC Ratio Problem 1. Ruin Constraint Covariance Model DCC CCC UNI Problem 2. CVaR Constraint Covariance Model DCC CCC UNI Problem 3.. EPD Constraint Covariance Model DCC CCC UNI

31 We propose three problems to jointly solve for the optimal capital requirement and its optimal portfolio allocation. We provide a novel semiparametric approach for solving these problems. Asset correlation plays an important role in the behaviour of the optimal capital required and the portfolio structure. Optimal required capital is very stable when the liability parameters remain constant over the rolling horizon; however, the variation is substantial when the liability is re-estimated at each step. The differences between the optimal portfolio weights are not as pronounced for the single versus double rolling exercises. The out-of-sample exercise indicates that the DCC model outperforms the CCC and No-Correlation model relative to both the solvency and portfolio performance criteria.