Portfolio risk and dependence modeling: Application of factor copula models 1 The International Journal of Banking and Finance, Volume 9 (Number 3) 2012: pages 1-14 PORTFOLIO RISK AND DEPENDENCE MODELING: APPLICATION OF FACTOR AND COPULA MODELS Arsalan Azamighaimasi Wuhan University of technology, china ____________________________________________________________ Abstract We consider portfolio credit risk modeling with a focus on two approaches, the factor model, and the copula model. While other models have received greater scrutiny, both factor and cupola models have received little attention although these are appropriate for rating-based portfolio risk analysis. We review the two models with emphasis on the joint default probability. The copula function describes the dependence structure of a multivariate random variable. In this paper, it is used as a practical to simulation of generate portfolio with different copula, we only use Gaussian and t–copula case. And we generate portfolio default distributions and study the sensitivity of commonly used risk measures with respect to the approach in modeling the dependence structure of the portfolio. Key Words: Gaussian copula, Factor model, Copula model JEL Classification: C15, C38 _____________________________________________ 1. Introduction There is a need to understand components of portfolio risk and their interaction. The Basel Committee for Banking Supervision in its Basel proposed (BIS, 2001) to develop an appropriate framework for a global financial regulation system. Several portfolio credit risk models developed in the industry have been made public since then. Examples are: CreditMetrics (Gupton et al., 1997), CreditRisk+ (Credit Suisse Financial Products, 1997) and Credit Portfolio View (Wilson 1997a; 1997b). Others systems remain proprietary, such as KMV’s Portfolio Manager (Kealhofer, 1996). Although the models appear quite different on the surface, recent http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#basle01 2 theoretical work has shown an underlying mathematical equivalence among them (Gordy, 2000; and Koyluoglu and Hickman, 1998). The credit portfolio models to obtain portfolio loss distributions, which are statistical models, can be classified as based on credit rating systems; See Crouhy et al. (2001) for exact description and discussion of the various models. Frey and McNeil (2001) study the mathematical properties of the models and consider the modeling of dependent defaults in large credit portfolios using latent variable models and mixture models. Crouhy et al. (2000) compared and reviewed models on benchmark portfolio using credit migration approach, the structural approach, the actuarial approach, and McKinsey approach. However, few studies have attempted to investigate aspects of portfolio risk based on rating-based credit risk models. Gordy (2000) offered a comparative anatomy of two especially influential benchmarks for credit risk models, the Risk Metrics Group's Credit Metrics and Credit Suisse Financial Product’s . Kiesel et al. (1999) employ a mark-to-market model and stress the importance of stochastic changes in credit spreads associated with market values, an aspect also highlighted in Hirtle et al. (2001). The aim of this paper is to contribute to the understanding of the performance of rating- based credit portfolio models, long ignored in the field. We apply a default-mode model to assess the effect of changing dependence structure within the portfolio. First, in the ensuing section, we discuss about the copula model as one of the dependency approaches within the portfolio. Second, we describe a factor model by focusing on the effects of default dependence model within the portfolio. Finally, in the penultimate section, we simulated types of copula model with different degree of freedom within the portfolio. 2. Copula Modelling An overview of basic copula uses in structural systems and models is provided in this section. Copulas provide a natural way to study and measure dependence between random variables. Suppose we have specified a portfolio of obligors, with default times . The variable International Journal of Banking and Finance, Vol. 9, Iss. 3 [2012], Art. 1 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#croughy-galai-mark01 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#frey-mcneil01 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#CY-GI-MK99 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#KL-PN-TR99 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#hirtle01 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#hirtle01 3 of default of obligor , at time t, is donated as . The probability space is . This space has filtration : (1) For the joint default probability at time t, evaluated at time 0, as (2) The survival property as (3) We take for granted the copula definition as a joint distribution function with uniform margins, which implies that and take for granted the fundamental Sklar’s theorem, in terms of a copula and the marginal distribution functions : (4) The joint survival probability with survival copula, and the marginal survival functions : (5) Factor copula is, (6) In the credit risk case, since the variables are default time, the copula represents default dependence. It is donated as , (7) (8) 4 According to Merton model (1974) if default of firm occurs, the values of asset or values of shares cross from barrier line of outstanding debt at debt maturity. Default is occurred when the firm’s asset value falls to the liability one, , the time of default is: (9) The default probability at time is, (10) The marginal default probability can be easily computed to be (11) Then, (12) And is the instantaneous return on assets, which equates the riskless rate under the risk neutral measure. The joint default probability of assets is (13) Where, is the distribution function of a standard normal vector with correlation matrix R. the marginal default probabilities is follows ) (14) To study the effect of different copula on default correlation, we use the following examples of copula (further details on these copula can be found in Embrechts et al., 2001). (i) Gaussian copula: (15) Where, denotes the joint distribution function of the - variety normal with linear correlation matrix ,and the inverse of the distribution function of the univariate standard normal. International Journal of Banking and Finance, Vol. 9, Iss. 3 [2012], Art. 1 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#ES-LG-ML01 5 (ii) A student copula: (16) Where is the standardized multivariate Student’s distribution, with correlation matrix and degrees of freedom, While is the inverse of the corresponding margin. Gumbel copula: (17) Where . This class of copula is a sub-class of the class of Archimedean copula. According to the table[1], joint default probabilities of two obligors are represented through three types of obligors with individual default probabilities corresponding to rating classes.as you will see that and Gumbel copula have higher joint default probabilities than the Gaussian copula. The joint default probabilities of two Obligors are represented through three types of obligors with individual default probabilities corresponding to rating classes. Table 1: Copula and default probability copula Default probability Class A Class B Class C 6.89 3.38 52.45 46.55 7.88 71.03 134.80 15.35 97.96 Gumbel 57.20 14.84 144.56 Gumbel 270.60 41.84 283.67 6 3. Factor Modelling Another popular approach to default modeling allows us to switch to the so called product copula. The reduction technique, which is widely adopted for the evaluation of losses in high- dimensional portfolios, with hundreds of obligors (see for instance Laurent and Gregory (2003)), is the standard approach of (linear) factorization, or transformation into a Bernoulli factor model. In the typical portfolio analysis the vector is embedded in a factor model, which allows for easy analysis of correlation, the typical measure of dependence. We assume that the underlying variables are driven by a vector of common factors. (18) Where is dimensional normal vector, and is independent normally distributed random variables. Here is obligor to factor , i.e. the so-called factor loading and is volatility of the risk contribution. The default indicators of the obligor are independent Bernoulli variables, with probability: (19) Where is cut-off point for default obligor . The individual default probabilities are, (20) And the joint default probability is, (21) If we denote by the correlation of the underlying latent variables and by the default correlation of obligors and , then we obtain the default correlation formula (22) International Journal of Banking and Finance, Vol. 9, Iss. 3 [2012], Art. 1 7 Under assumption above, we obtain the joint default probability, (23) Where is bivariate normal density with correlation coefficient . 4. Simulation Results of Copula Model Here, we want to generate portfolios with given marginal and the above copula. we only use Gauss and copula case . We looking for random sample generation for this mean we obtain the generation of an -variety normal with liner correlation matrix ,to take realizations from a Gaussian copula we simply have to transform the marginal: • Set • To generate random varieties from the –copula we assume the random vector X act the stochastic process (24) With Where Z and Y are independent, and then X is distributed with mean and covariance matrix we assume , while the stochastic process is still valid the parameters has to change for . We will have algorithm (this is algorithm in Embrechts et al. (2001)): • Set • Set • . http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#ES-LG-ML01 8 We can replace the with in order to have multivariate distribution with –copula and normal marginal, to obtain the –copula . Figure 1 shows three simulation results with 1000, 500, and 50 observations from a multivariate normal distribution. As you see the represents tree types of observations from a multivariate normal distribution with mean vector mu and covariance matrix. The figure 2 shows to computes a scatterplot of a normal sample and in a second plot the contour ellipses for mu =# (3, 2) and sigma = # (1,-1.5) ~# (-1.5, 4) with different observations. Figure1: Simulation results from samples of 1,000, 500 and 50 observations A=1000 B=500 C=50 Figure 2: Scatter plots of normal sample and second plot of the contour ellipses A=1000 B=500 C=50 International Journal of Banking and Finance, Vol. 9, Iss. 3 [2012], Art. 1 9 Further analyses of the same data are plotted in Figure 2. These are scatterplots of a normal sample. In an adjacent plot next to each sample, we present a second plot as the contour ellipses for mu =# (3, 2) and sigma = # (1,-1.5) ~# (-1.5, 4) with different size. 4.1 Portfolio For our first simulation exercise, we assume that the underlying variables are normally distributed within a single factor framework, i.e. and in formula as follow: (25) They are constant and are chosen so that the correlation for the underlying latent variables is (Kiesel et al., 1999. Note that we use three rating classes, named A, B, and C with default probabilities 0.005, 0.05, and 0.15 roughly corresponding to default probabilities from standard rating classes (Ong, 1999). To generate different degrees of tail correlation, we link the individual assets together using a Gaussian, a and a -copula. The information in table 2, 3 and 4 represent the effect tail-dependence has on the high quintiles of highly-rated portfolios at different quintiles: Table 2 is for 99 percentile. Table2: Effect of normal copula with default probability set at 0.005 Portfolio Copula Mean variance A=1000 normal 0.115 0.13391 1 2 A=500 normal 0.106 0.119 1 1 A=50 normal 0.18 0.19143 1 2 B=1000 normal 0.99 1.8277 4 6 B=500 normal 1.038 1.8442 4 6 B=50 normal 1.18 2.3955 4 6 C=1000 normal 3.029 7.0953 8 11 C=500 normal 2.998 6.9078 8 11 C= 50 normal 3.1 7.3163 9 10 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#KL-PN-TR99 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode40.html#ONG99 10 The copula is more than three-times larger than the corresponding quintile for the Gaussian copula. The same effect can be observed for lower rated portfolios although not quite with a similar magnitude. Table3: effect of with default probability 0.05 Portfolio Copula Mean variance A=1000 0.101 0.26907 1 2 A=500 0.098 0.15671 1 2 A=50 0.14 0.36776 1 4 B=1000 0.963 2.38 4 6 B=500 0.994 2.1984 4 6 B=50 1.06 2.9147 4 9 C=1000 3.008 7.9799 9 11 C=500 3.05 7.9474 9 12 C=50 3.42 8.9016 9 11 We assume the second factor, i.e. in (4), for a sub-portfolio of 100 obligors increasing the correlation of the latent variables within the sub-portfolio to 0.5 Table4: effect of with default probability 0.15 Portfolio Copula Mean variance A=1000 0.088 0.39665 0 2 A=500 0.084 0.24543 0 2 A=50 0.22 2.42 0 11 B=1000 0.924 3.1454 5 9 B=500 1 3.0261 7 5 B=50 1.02 3.5302 4 11 C=1000 2.997 9.5860 10 12 C=500 3.028 9.0213 9 13 C=50 3.34 9.2086 9 12 International Journal of Banking and Finance, Vol. 9, Iss. 3 [2012], Art. 1 11 Table 5: the effect of correlation cluster with default probability 0.005 portfolio copula First subportfolio Second subportfolio mean variance A=1000 normal 100 150 1.237 6.8447 5 13 A=500 normal 50 75 0.6 1.6433 2 7 A=50 normal 20 30 0.24 0.47184 1 4 B=1000 normal 100 150 12.723 204.41 41 71 B=500 normal 50 75 6.198 47.951 20 33 B=50 normal 20 30 2.58 7.3506 10 11 C=1000 normal 100 150 37.972 871.43 96 132 C=500 normal 50 75 18.832 200.1 49 63 C=50 normal 20 30 7.74 30.36 20 23 Table 6: the effect of correlation cluster with default probability 0.05 portfolio copula First subportfolio Second subportfolio mean variance A=1000 100 150 1.451 27.335 7 28 A=500 50 75 0.644 6.7668 3 11 A=50 20 30 0.2 0.32653 1 3 B=1000 100 150 11.76 299.29 52 83 B=500 50 75 6.28 85.605 24 44 B=50 20 30 2.32 11.365 10 17 C=1000 100 150 38.24 1104.7 105 148 C=500 50 75 18.638 263.7 52 75 C=50 20 30 7.5 31.235 17 24 12 Table 7: the effect of correlation cluster with default probability 0.15 portfolio copula First subportfolio Second subportfolio mean variance A=1000 100 150 1.635 70.278 7 42 A=500 50 75 0.682 14.554 3 21 A=50 20 30 0.36 2.1943 1 10 B=1000 100 150 13.385 592.25 65 128 B=500 50 75 6.266 132.82 28 61 B=50 20 30 2.26 16.074 13 18 C=1000 100 150 38.465 1395 117 157 C=500 50 75 18.676 331.96 56 80 C=50 20 30 7.56 41.109 23 27 for this reasaning we want to shows the effects of increased correlation within parts of the portfolio; we change the factor loading within parts of our portfolio. These results are shown in tables 7, 9 and 10. As expected, the results in Tables 5, 6, 7 show increase in the quantiles due to the increased correlation within the portfolio. However, comparing the three tables we will see that the sensitivity of the portfolio loss quantiles is higher with regard to the underlying copula than to the correlation within the portfolio. 5. Conclusions To investigate the riskiness of credit-risky portfolios is one of the big challenging in financial mathematics. An important thing for a model of credit-risky portfolios is the dependence structure of the underlying obligors. We studied two approaches, a factor structure, and the direct specification of a copula. We generated portfolio default distributions and studied the sensitivity International Journal of Banking and Finance, Vol. 9, Iss. 3 [2012], Art. 1 http://sfb649.wiwi.hu-berlin.de/fedc_homepage/xplore/tutorials/xfghtmlnode39.html#cortab 13 of commonly used risk measures with respect to the approaches in modeling the dependence structure of the portfolio using as a rating-based approach using cupola mathematics. The simulation results indicate that the degree of tail dependence of the underlying copula plays a major role. That is identified as a credit risk. The copula modeling links the underlying variables together, which is of crucial importance especially for portfolios of highly-rated obligors. Author Information: Arsalan Azamighaimasi is a faculty member in the Department of Management, Wuhan University of Technology, China. He may be contacted at E-mail: Arsalan.azami2011@gmail.com. References BIS (2001), "Overview of the new Basel capital accord," Basel Committee on Banking Supervision. Caouette, J., Altman, E. and Narayanan, P., (1998). "Managing Credit Risk, The Next Great Financial Challenge," Wiley Frontiers in Finance,Vol. Wiley Frontiers in Finance, Wiley & Sons, Inc, New York. M. Carey, (2000). "Dimensions of credit risk and their relationship to economic capital requirements.," Federal Reserve Board. M. Gordy, (1997). "A comparative anatomy of credit risk models," Journal of Banking and Finance, vol. 24, pp. 119-149, 2000J. Morgan, "Creditmetrics," JP Morgan, New York. Crouhy, M., Galai, D., and Mark, R., (2001). Risk management, McGraw Hill. Crouhy, M., Galai, D. and Mark, R, (2000). "A comparative analysis of current credit risk models," ournal of Banking and Finance, Vols. 24(1-2), pp. 59-117. Embrechts, P., Lindskog, F. and McNeil, A, (2001). "Modelling dependence with copulas and applications to risk management," ETH Zürich. Frey, R. and McNeil, A, (2001). "Modelling dependent defaults," ETH Zürich. P. Jorion, (2000). Value at Risk, New York. : McGraw-Hill. Kiesel, R., Perraudin, W. and Taylor, A, (1999). "The structure of credit risk," Birkbeck College. Koyluoglu, H., and Hickmann, A., (1998). "A generalized framework for credit portfolio models," Wyman & Company, mailto:Arsalan.azami2011@gmail.com 14 Luciano, E., (2008)."Copula-based default dependence modelling: where do we stand?," in Nickell, P., Perraudin, W. and Varotto, S, "Ratings-versus equity-based credit risk models: An empirical investigation," unpublished Bank of England mimeo. M. Ong, (1999). Internal Credit Risk Models. Capital Allocation and Performance Measurement, London: Risk Books. International Journal of Banking and Finance, Vol. 9, Iss. 3 [2012], Art. 1