The Use of the Tukey s g h family of distributions to Calculate Value at Risk and Conditional Value at Risk

Transcription

1 The Use of the Tukey s g h family of distributions to Calculate Value at Risk and Conditional Value at Risk José Alfredo Jiménez and Viswanathan Arunachalam Journal of Risk, vol. 13, No. 4, summer, 2011 Taller de Matemáticas Financieras y Finanzas Computacionales Universidad de los Andes, Colombia Julio 26-28, 2011

2 Contents 1 Introduction 2 Preliminary 3 Estimation 4 Measure of Risk Traditional approaches 5 An Illustration 6 Conclusions 7 References

3

4 Introduction When calculating the Value at Risk (VaR) there is a little importance to the most extreme losses, since which is not adequately reflect the skewness and kurtosis of the distribution. Moreover, normality is assumption to overestimate the VaR values for upper percentiles, while it underestimates for lower percentiles of values which correspond to more extreme events. In this article we propose use of the Tukey s g h family of distributions relevant for the calculation of VaR and Conditional Value at Risk CVaR, as this distribution enables the skewness and kurtosis. We have also calculated explicit formula for CVaR using a Cornish-Fisher expansion to approximate the percentiles. An illustrative example is presented to compare with others models.

5

6 Tukey s g h family of distributions Let Z be a random variable with standard normal distribution and g and h are two constants (parameters). The random variable Y given by Y = T g,h (Z) = 1 g (exp(gz) 1) exp(hz 2 /2) with g 0, h R, (2.1) has Tukey s g h distribution.

7 Tukey s g h family of distributions Let Z be a random variable with standard normal distribution and g and h are two constants (parameters). The random variable Y given by Y = T g,h (Z) = 1 g (exp(gz) 1) exp(hz 2 /2) with g 0, h R, (2.1) has Tukey s g h distribution. Observations

8 Tukey s g h family of distributions Let Z be a random variable with standard normal distribution and g and h are two constants (parameters). The random variable Y given by Y = T g,h (Z) = 1 g (exp(gz) 1) exp(hz 2 /2) with g 0, h R, (2.1) has Tukey s g h distribution. Observations 1 When h = 0 the Tukey s g h distribution reduces to T g,0 (Z) = 1 (exp(gz) 1) (2.2) g which is Tukey s g distribution.

9 Tukey s g h family of distributions Let Z be a random variable with standard normal distribution and g and h are two constants (parameters). The random variable Y given by Y = T g,h (Z) = 1 g (exp(gz) 1) exp(hz 2 /2) with g 0, h R, (2.1) has Tukey s g h distribution. Observations 1 When h = 0 the Tukey s g h distribution reduces to T g,0 (Z) = 1 (exp(gz) 1) (2.2) g which is Tukey s g distribution. 2 Similarly, when g 0 the Tukey s g h distribution is given by T 0,h (Z) = Z exp{hz 2 /2} (2.3)

10 Ordinary moments Martínez and Iglewicz (1984) establishing the n th moments of Tukey s g h family distributions, when h < 1 n, as follows { 1 n ) ( ) } 2 exp g n 1 nh µ n = E (Y n ) = ( 2 n 2 n 2 k=0 ( 1) k( n k 1 2 n k 1 nh g g 0 0 for n odd n! )! (1 nh) n+1 for n even g = 0 (2.4) then the mean of the Tukey s g h distribution given by { [ { } ] 1 µ g,h = E [T g,h (Z)] = g 1 g exp 2 1 h 2 1 h 1 g 0, 0 h < 1, 0 g = 0. (2.5)

11 Table 1 shows the values of g and h that approximate a selected set of well known distributions Name of Parameters Estimates Values Distribution A B g h Cauchy µ, σ > 0 µ σ 0 1 Exponential λ > 0 1 λ ln 2 g λ 0,773 0,09445 Laplace α, β > 0 α β 0 0 Logistic α, β > 0 α β 0 1, Log-normal µ, σ 2, C > 0 C µ gc µ σ ln C 0 Normal µ, σ 2 µ σ 0 0 t 10 ν = , Cuadro: Values of g and h for some distributions

12 Method of quantiles We start by the method proposed in Hoaglin (1985) which takes into account the properties presented in Dutta and Babbel (2002) and is complemented by the proposal given in Jiménez (2004). 1 X is a strictly increasing transformation of Z. This is the transformation of a normal standard of g h is one to one. 2 The location parameter of the Tukey s g h distribution is estimated by the median of the data, ie A = x 0,5. 3 The estimate of the parameter that controls the skewness of the distribution (g), is estimated, usually by the median of the logarithms of the following expression: e gzp = UHS p LHS p, p > 0,5 (3.1) where UHS p = x p x 0,5 and LHS p = x 0,5 x 1 p, denote the p-th upper half-spread and lower half-spread, respectively, defined in Hoaglin et al. (1985).

13 Method of quantiles (4) If there is θ R, with θ x 0,5, such that x p θ = x 0,5 θ p > 0,5, (3.2) θ x 0,5 θ x 1 p then h = 0. In particular, the expression (3.2) is satisfied if θ = A B g. This constant is known as a threshold parameter and was obtained in Hoaglin (1985). (5) When g 0, the parameter that controls the elongation (or kurtosis) of the tails (h), can be estimated conditionally on this value of g ln (x 0,5 θ p ) = ln where θ p < x 0,5 for all p > 0,5, and ( ) B + h Z p 2 g 2, (3.3) θ p = x px 1 p x 2 0,5 UHS p LHS p for all p (0, 1), p 0,5, (3.4)

14 Method of moments The idea of method of moments proposed in Majumder and Ali (2008), to estimate the parameters is to get as many equations as the number of parameters. However, the location and scale parameters are not determined, for this we estimate the parameters A and B as follows A = µ X Bµ g,h, (3.6) where µ X is the mean of random variable X and µ g,h is given in (2.5) and the scale parameter is estimated by B = sgn (β 1 (X )) σ X σ Y. here sgn( ) denote the signum function and β 1 (X ) denote the coefficient of skewness from the variable we want to approximate.

15 Value at Risk (VaR) Considering a confidence level α and a time horizon of T days, the VaR can be calculated as follows: 1 α = H (S VaR α ) g (S) ds where g(s) is the density function of S and H (u) is the Heaviside step function.

16 Value at Risk (VaR) Considering a confidence level α and a time horizon of T days, the VaR can be calculated as follows: 1 α = H (S VaR α ) g (S) ds where g(s) is the density function of S and H (u) is the Heaviside step function. Conditional Value at Risk (CVaR) Conditional Value at Risk CVaR is defined as the expected loss given that is larger than or equal to VaR. The CVaR is the average losses over a Q % probability level to be identified by α, that is losses that can be expected with this probability. The CVaR of a sample is obtained as follows: CVaR α = 1 1 α 1 α VaR q dq. (4.1)

17

18 Measure of risk coherent To determine the efficiency of a good indicator of market risk, Artzner et al. (1997) derive four desirable properties that should comply with a measure of risk to be called coherent. A risk indicator ρ must satisfy

19 Measure of risk coherent To determine the efficiency of a good indicator of market risk, Artzner et al. (1997) derive four desirable properties that should comply with a measure of risk to be called coherent. A risk indicator ρ must satisfy 1 Positive homogeneity: ρ (λu) = λρ (u). Increasing the value of portfolio in λ, the risk must also increase λ.

20 Measure of risk coherent To determine the efficiency of a good indicator of market risk, Artzner et al. (1997) derive four desirable properties that should comply with a measure of risk to be called coherent. A risk indicator ρ must satisfy 1 Positive homogeneity: ρ (λu) = λρ (u). Increasing the value of portfolio in λ, the risk must also increase λ. 2 Monotonicity u v implies ρ (u) ρ (v). If the portfolio u has a consistently lower return than the portfolio v, then risk must be lower.

21 Measure of risk coherent To determine the efficiency of a good indicator of market risk, Artzner et al. (1997) derive four desirable properties that should comply with a measure of risk to be called coherent. A risk indicator ρ must satisfy 1 Positive homogeneity: ρ (λu) = λρ (u). Increasing the value of portfolio in λ, the risk must also increase λ. 2 Monotonicity u v implies ρ (u) ρ (v). If the portfolio u has a consistently lower return than the portfolio v, then risk must be lower. 3 Translation invariance: ρ (u + a) = ρ (u) + a. Add cash of an amount a then add the risk by a.

22 Measure of risk coherent To determine the efficiency of a good indicator of market risk, Artzner et al. (1997) derive four desirable properties that should comply with a measure of risk to be called coherent. A risk indicator ρ must satisfy 1 Positive homogeneity: ρ (λu) = λρ (u). Increasing the value of portfolio in λ, the risk must also increase λ. 2 Monotonicity u v implies ρ (u) ρ (v). If the portfolio u has a consistently lower return than the portfolio v, then risk must be lower. 3 Translation invariance: ρ (u + a) = ρ (u) + a. Add cash of an amount a then add the risk by a. 4 Subadditivity: ρ (u + v) ρ (u) + ρ (v). The portfolio composition should not increase the risk. If ρ satisfies these properties, then it is considered as a coherent risk measure.

23 Traditional approaches Delta-normal method for approaches to VaR This parametric method to calculate VaR was proposed by Baumol (1963) as a criterion of confidence limits expected gain. Assuming a confidence level fixed α (0, 1] and a time horizon of T days, the VaR can be easily calculated from σ, by the following expression: VaR α = µ S Φ 1 (α) σ S T, (4.2) where Φ(x) is the standard normal distribution function.

24 Traditional approaches Delta-normal method for approaches to VaR This parametric method to calculate VaR was proposed by Baumol (1963) as a criterion of confidence limits expected gain. Assuming a confidence level fixed α (0, 1] and a time horizon of T days, the VaR can be easily calculated from σ, by the following expression: VaR α = µ S Φ 1 (α) σ S T, (4.2) where Φ(x) is the standard normal distribution function. Delta-normal method for approaches to CVaR Under the assumption of normality, the parametric model that determines the CVaR of a position is as follows: CVaR α = µ S σ S T 1 α ϕ (z α), (4.3) where ϕ(x) is the standard normal density function. This last formula coincides with the expression given by Jondeau et al. (2009, page 335) and McNeil et al. (2005, page 45).

25 Traditional approaches Cornish-Fisher approximation Based on the Fisher and Cornish (1960) expansion, Zangari (1996) approximate the percentiles of the probability distribution of S and obtain the VaR for a confidence level α % and a horizon of T days as follows VaR α = E [S] ω α σ S T, (4.4) where ω α is defined (Abramowitz and Stegun (1965)) as follows: ω α =z α + 1 ( z 2 6 α 1 ) β 1 (S) + 1 ( ) z 3 24 α 3z α (β2 (S) 3) 1 ( ) 2z 3 36 α 5z α β 2 1 (S) 1 ( z 4 24 α 5zα ) β 1 (S) (β 2 (S) 3), (4.5) here β 1 (S), β 2 (S) denote the skewness and kurtosis from the distribution of S. Note that when the skewness coefficient β 1 (S) and excess kurtosis β 2 (S) are zero, we obtain the quantile of the variable N (0, 1).

26 Traditional approaches Cornish-Fisher approximation The CVaR can be approximated using the Cornish-Fisher expansion for a confidence level α % and a horizon of T days as follows CVaR α = E [S] 1 1 α ω ασ S T, (4.6) with ω α = 1 α { ω q dq = 1 + z α 6 β 1(S) + z2 α 1 24 z3 α 2z α β 1 (S) (β 2 (S) 3) 24 (β 2 (S) 3) 2z2 α 1 β (S) } ϕ (z α ), where ω q is given in (4.5). Note that when the skewness coefficient β 1 (S) and excess kurtosis β 2 (S) are zero, this expression reduces to (4.3).

27 Traditional approaches Approximation by the Tukey s g h distribution If S is approximated by S = A + BY, then we obtain the VaR for α > 0,5, VaR α =A + BT g,h (Z α ) (4.7) VaR 1 α =A B exp{ gz α }T g,h (Z α ). This expression coincides with the corresponding result presented in Nam and Gup (2003).

28 Traditional approaches Approximation by the Tukey s g h distribution If S is approximated by S = A + BY, then we obtain the VaR for α > 0,5, VaR α =A + BT g,h (Z α ) (4.7) VaR 1 α =A B exp{ gz α }T g,h (Z α ). This expression coincides with the corresponding result presented in Nam and Gup (2003). Approximation by the Tukey s g h distribution If S is approximated by S = A + BY, then we obtain the CVaR for α > 0,5, CVaR α =A + B [ µ g,h Φ (δ 2α ) + 1 ] Φ (δ 2α ) Φ (δ 1α ), (4.8) 1 α 1 h δ 2α δ 1α where µ g,h is the mean of the Tukey s g h distribution given in (2.5) and δ 1α = 1 h z α, δ 2α =δ 1α + g 1 h, (4.9)

29 Traditional approaches Special Cases 1 When h = 0, VaR α = A + B g ( e gz α 1 ) = θ + e µ+σzα, (4.10) where θ = A e µ and g = σ. Jiménez and Martínez (2006) propose that in this case, if S = ln (X θ) follows a normal law N ( µ, σ 2), we get, µ X =θ + exp {µ + 12 } σ2 and σ 2 X = (µ X θ) 2 [ exp { σ 2} 1 ], solving for µ, σ, and substituting this into (4.10) resulting VaR α =θ + µ } X θ exp {Z 1 + ρ 2 α ln (1 + ρ 2X ), (4.11) X σ X where ρ X = µ X θ, which coincides with the coefficient of variation of random variable X, when θ = 0.

30 Traditional approaches 1 Suppose that g = 0, we obtain VaR α = A + BZ α e 1 2 hz 2 α, when h=1 we have the cauchy distribution with parameters µ and σ ie VaR α = µ + σz α e 1 2 Z 2 α. 2 If g = h = 0, using the constants given for location and scale parameters in Jiménez (2004), we obtain VaR α = µ + σz α. Note that this last expression coincides with the classical formula of VaR (see Jorion (2007)).

31 Traditional approaches Special Cases 1 Supposing that h = 0, then CVaR α = θ + Φ (σ z α) 1 α e µ+σ2 /2, (4.12) where θ = A e µ. Following the approach used in Jiménez (2004), by solving µ, σ, to replace (4.12), it follows that CVaR α =θ + µ { [ ]} X θ Φ ln (1 + ρ 2 X 1 α ) Z α, (4.13) σ X where ρ X = µ X θ, which coincides with the coefficient of variation of random variable X, when θ = 0.

32 Traditional approaches 1 If g = 0, we can use the Mean Value Theorem, accordingly we obtain Φ (b) Φ (a) b a CVaR α = A + ϕ (c), where c (a, b), B 1 α ϕ ( 1 h Z α ) 1 h 2 When g = h = 0, using the constants given for location and scale parameters in Jiménez (2004), we have CVaR α = µ σ 1 α ϕ (Z α). Note that this last expression coincides with the formula for the CVaR given in (4.3)..

33 An Illustration Now we compare the procedure developed above with the classical method, historical simulation and the Cornish-Fisher approximation to estimate the VaR and CVaR. We consider a portfolio constructed with three largest market capitalization stocks in Spain: Banco Bilbao Vizcaya Argentaria (BBVA), Endesa (ELE) and Banco Santander (SAN). The data were obtained from and the sample covers 2, 081 trading days from 01/01/2003 to 17/01/ Get the arithmetic daily rate of return for each stock, ie R t = P t P t 1 P t 1 t =1, 2,..., T, (5.1) where P t denotes the price of the stock at time t. Asset Return BBVA ( %) ELE ( %) SAN ( %)

34 1 Covariance matrix of the portfolio Σ = 4, , , , , , , , , The Global Minimum Variance Portfolio (GMVP) Asset GMVP BBVA ELE SAN We assume an investment of V 0 = 1 million currency units and positions in this portfolio is Asset GMVP BBVA 255,744 ELE 672,138 SAN 72,117

35 Now µ V = w r = 352,58188 and σv 2 = 2, If X is approximated by X = A + BY, then X = 243, , 968, { } h g [exp{gz} 1] exp 2 Z 2 ; where g = 0,29507 and h = 0, As shown in figure 1, there is a difference between the empirical distribution of portfolio returns GMVP (represented by the histogram) and the normal distribution. Tukey s g h distribution better approximates the empirical. The distribution of portfolio returns as expected GMVP tends to be more leptokurtic than the normal distribution has heavier tails.

36 500 PORTFOLIO daily returns: histogram Frequency Daily returns (Change (%)) Figura: Portfolio vs. Normal Distribution and Tukey s g h distribution

37 The following table presents the statistics of portfolio returns GMVP. Statistics Values Mean Median Stan. Dev. 15, Minimum -174, Maximum 121, Skewness Kurtosis JB test 14, Cuadro: Descriptive Statistics Kurtosis, skewness and the test proposed by Jarque and Bera (1987), statistics reported in Table 2 indicate that the null hypothesis of normal distribution can be rejected for the variable under study. Figure 1 shows such a histogram, it is evident that the returns of the series have a slight degree of bias to the right, leptokurtic and do not follow the normal distribution.

38 In this case, the calculation of VaR, initially assumed to be normal in returns. Table 3 presents the loss of calculate VaR for the portfolio GMVP under the following confidence levels: 90 %, 95 %, 97,5 % and 99 %. VaR Confidence levels α = 0,90 α = 0,95 α = 0,975 α = 0,99 Classical 20,130 25,937 30,973 36,829 Historical 15,279 22,262 28,423 48,198 Cornish-Fisher 19,553 24,406 48,381 32,869 GH (A, B, g, h) 20,279 29,448 39,533 54,704 Cuadro: Comparison of VaR methodologies

39 As shown in Figure 2 there is a perceptible difference between the VaR methodologies 0 1 x 104 Comparison of methods for calculate of Value at Risk Historical Classical Cornish Fisher Tukey s g h 2 3 VaR Confidence Levels Figura: Comparison of VaR methodologies

40 Table 4 presents the loss of calculate CVaR for the portfolio GMVP under the following confidence levels: 90 %, 95 %, 97,5 % and 99 %. CVaR Confidence levels α = 0,90 α = 0,95 α = 0,975 α = 0,99 Classical 27,697 32,615 37,012 42,245 Cornish-Fisher 35,087 59,102 88, ,600 GH (A, B, g, h) 41,660 67, , ,015 Cuadro: Comparison of CVaR methodologies As can be noted CVAR losses for each of the methods are greater than the losses of VaR.

41 Figure 3 shows that there is a difference between the CVaR methodologies 0 x 105 Comparison of methods for calculate of Conditional Value at Risk CVaR Classical Cornish Fisher Tukey is g h Confidence Levels Figura: Comparison of CVaR methodologies

42 Conclusions This article presents an alternative methodology to establish the VaR and CVaR when the portfolio distribution has skewness and kurtosis. Whereas, normality assumption overestimates the VaR and CVaR for upper percentiles, while it underestimates for lower percentiles of values that correspond to extreme values. The formulas obtained to calculate VaR and CVaR are explicit and we get the classical model as a particular case when the parameters g and h are considered equal to zero. The losses obtained for CVaR for each of the methods used are greater than losses of VaR. Thus our model exhibiting many of the characteristics of the other models in the literature.

43 References I Abramowitz, M. and Stegun, I. (1965), Handbook of Mathematical Functions, with Formulas, Graphs and Mathematical Tables, Dover Publications, Inc., New York. Baumol, W. J. (1963), An expected gain confidence limit criterion for portfolio selection, Management Science 10(1), Fisher, R. A. and Cornish, E. A. (1960), The percentile points of distributions having known cumulants, Technometrics 2(2), Hoaglin, D. (1985), Summarizing shape numerically: the g-and-h distributions, In: Hoaglin, D.C., Mosteller, F., Tukey, J.W. (Eds.), Exploring Data Tables, Trends, and Shapes. Wiley (New York), Hoaglin, D. C., Mosteller, F. and Tukey, J. W. (1985), Exploring Data Tables, Trends, and Shapes, John Wiley & Sons.

44 References II Jarque, C. M. and Bera, A. K. (1987), A test for normality of observations and regression residuals, International Statistical Review 55(2), Jiménez, J. A. (2004), Aproximaciones de las funciones de riesgo del tiempo de sobrevivencia mediante la distribución g-h de tukey, Especialista en actuaría, Facultad de Ciencias. Departamento de Matemáticas. Universidad Nacional de Colombia. Sede Bogotá. Jiménez, J. A. and Martínez, J. (2006), Una estimación del parámetro de la distribución g de tukey, Revista Colombiana de Estadística 29(1), Jondeau, E., Poon, S. H. and Rockinger, M. (2009), Financial modeling under non-gaussian distributions, Springer Finance Series, Springer-Verlag New York, LLC. Jorion, P. (1996), Risk: Measuring the risk in value at risk, Financial Analysts Journal 52(6),

45 References III Majumder, M. M. A. and Ali, M. M. (2008), A comparison of methods of estimation of parameters of tukey s gh family of distributions, Pakistan Journal of Statistics 24(2), Martínez, J. and Iglewicz, B. (1984), Some properties of the tukey g and h family of distributions, Communications in Statistics - Theory and Methods 13(3), Zangari, P. (1996), A var methodology for portfolios that include options, RiskMetricsTM Monitor pp Morgan Guaranty Trust Company: New York.