Abstract
The focus of this chapter is on sector concentrations. This type of concentration risk can occur if there is more than one systematic risk factor that influences credit defaults. The main research questions that are analyzed in this chapter are:
-
How can existing approaches for measuring sector concentration risk be modified and adjusted to be consistent with the Basel framework? Is the risk measure Value at Risk problematic when dealing with sector concentration risk?
-
Which methods are capable of measuring concentration risk and how good do they perform in comparison? What are the advantages and disadvantages of these methods?
In order to deal with these questions, it is initially determined how a multi-factor model can be parameterized to obtain a capital requirement which is consistent with Basel II. Furthermore, the models of Pykhtin (Risk 17(3):85–90, 2004), Cespedes et al. (J Credit Risk, 2(3):57–85, 2006), and Düllmann (Measuring business sector concentration by an infection model. Discussion Paper, Series 2: Banking and Financial Studies, Deutsche Bundesbank, (3), 2006), which have been developed to approximate the risk in the presence of sector concentrations, are presented and modified. Then, the accuracy of these models concerning their ability to measure sector concentration risk is compared.
The main results of this section comply with Gürtler et al. (2010).
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
BCBS (2005a), § 773.
- 2.
Cf. Sect. 3.3.
- 3.
The multi-factor model of Cespedes et al. (2006) is also specified against the background of the regulatory capital formula. However, within the deriviation of their formulas, the authors assume the regulatory capital requirement to be the upper barrier of risk, which is not consistent with the view of supervisors that we presented in Sect. 3.3 and especially in Fig. 3.2. Cf. Sect. 5.2.3 for details regarding this issue.
- 4.
Cf. Sect. 2.2.3.
- 5.
As shown by Morinaga and Shiina (2005), an assignment of borrowers to the wrong sectors usually leads to a higher estimation error than a non-optimal sector definition.
- 6.
In order to allow for negative intra-sector correlations, the factor loading could also be written as r i instead of \( \sqrt {{{\rho_{{\text{Intra}},i}}}} \). However, it is economically reasonable to assume that there is a positive relationship between the asset return of an obligor and the corresponding industry-sector. Thus, the chosen notation should be no practical limitation.
- 7.
Concretely, the independence of the risk factors is essential for the derivation of the Pykhtin-model in Sect. 5.2.2.
- 8.
This approach is a common mathematical method to generate correlated random variables and leads to the identical number of independent risk factors \( {\tilde{z}_k} \) and dependent sector factors \( {\tilde{x}_s} \), that is K equals S. Another common method to determine independent risk factors is the principal component analysis, which leads to a reduced number of risk factors.
- 9.
- 10.
The correlation structure based on the MSCI US is similar, see Düllmann and Masschelein (2007).
- 11.
Düllmann and Masschelein (2007) notice that the concentration is very similar to other countries like France, Belgium, and Spain.
- 12.
- 13.
See Fig. 4.7 for the portfolio characteristics.
- 14.
We tried several different functional forms but the formula above performed best. The multipliers 18.5% and 34% in function (5.8) were determined with a grid search using a reasonable parameter range, which is similar to the procedure of Lopez (2004) used for the single correlation parameter.
- 15.
E.g. Heitfield et al. (2006) determine the sector loadings, which equal \( \sqrt {{{\rho_{\text{Intra}}}}} \), for 50 industry sectors using KMV data on asset values. The resulting intra-sector correlation is on average 18.8% and the standard deviation is 8.3%. These inter-sectoral differences are not captured by the formula above.
- 16.
A correlation structure with one degree of freedom for every PD/sector-combination is practically unfeasible due to high data requirements.
- 17.
In our setting, the computation time could be reduced by more than 99.8%.
- 18.
The conditional PD stems from the Vasicek model, cf. Sect. 2.4 or 2.7.
- 19.
Cf. (5.9).
- 20.
For the determination of c i , we need both the intra- and inter-sector correlations, which can be taken from Sect. 5.2.1.
- 21.
This formula has already been derived for the granularity adjustment formula, cf. (4.18).
- 22.
Cf. Sect. 4.2.1.2.
- 23.
Cf. (5.2).
- 24.
See Appendix 5.5.2.
- 25.
The derivation of the variance decomposition can be found in Weiss (2005), p. 385 f.
- 26.
The quadratic computation effort is due to the determination of a double sum (see (5.22) and (5.24)).
- 27.
The results of the multi-factor adjustment do not differ whether different exposures with the same PD are aggregated or handled separately on borrower level. For details see Sect. 5.2.2.1 and Appendix 5.5.1.
- 28.
The computation time when calculating the multi-factor adjustment on bucket- instead on borrower-level can be reduced from 67 min to 5 s for a portfolio with 11 sectors, 7 PD-classes, and 5,000 creditors.
- 29.
In the strict sense, Cespedes et al. (2006) relate the multi-factor model to the economic capital in a single-factor model. But since they apply the regulatory capital formula and we require a relation to this formula, too, we use the term regulatory capital instead.
- 30.
The idea is related to Pykhtin (2004), who uses the VaR from the ASRF model as a weight when maximizing the correlation between the single factor of the comparable one-factor model and the sector factors; cf. (5.82)–(5.85).
- 31.
Cespedes et al. (2006) call this parameter the capital diversification index (CDI).
- 32.
This concentration measure corresponds to (2.87).
- 33.
The setting is similar to Cespedes et al. (2006). Until this point, the main difference is the definition of the intra- and inter-sector correlations.
- 34.
For the determination of the economic capital for one specific portfolio, the number of trials is slightly low but as we perform 25,000 simulations and the simulation noise of each simulation is unsystematic, the error terms should cancel out each other to a large extent.
- 35.
We have also tested the results when using the ES instead of the unexpected loss but the coefficient of determination is higher when subtracting the EL in the corresponding formulas when performing the simulations.
- 36.
To determine the Expected Shortfall with (4.59), a bivariate cumulative normal distribution has to be computed whereas the Value at Risk only makes use of univariate distributions.
- 37.
- 38.
The shape of the function is similar to Cespedes et al. (2006) but their range is from 0.1 to 1.0 whereas our function ranges from 0.2 to 1.5. In addition, they received a little higher R2 (99.4% instead of 95.5% or 97.9%) but this is mainly due to the different simulation setting. Cespedes et al. (2006) directly draw the parameter \( \bar{\beta } \) as an input parameter for each simulation, implying \( \bar{\beta } \) to fully define their correlation structure. We use a heterogeneous correlation structure instead and compute \( \bar{\beta } \) for the portfolios. Thus, in our setting \( \bar{\beta } \) does not reflect the complete correlation structure, which results in a lower R2 but does not imply a worse approximation.
- 39.
In comparison to the original procedure, the computation time could be reduced by almost 99.9% in our calculations.
- 40.
- 41.
See Appendix 5.5.5.
- 42.
See Appendix 5.5.5.
- 43.
See Appendix 5.5.6.
- 44.
Similar to the BET-model, the authors developed their model for CDOs but it can also be applied to standard credit portfolios.
- 45.
The expected number of defaults in the infectious defaults model is determined in Appendix 5.5.7.
- 46.
Cf. (5.61) for the corresponding expression without using the parameters of the BET-model.
- 47.
- 48.
See also definition (5.4) of Sect. 5.2.1.
- 49.
The portfolios used for calibration correspond to the setting of Düllmann (2006).
- 50.
Due to the characteristic of the correlation parameter (5.67), the parameter \( {\bar{r}_{\text{Inter}}} \) is always smaller than the parameter \( {\bar{r}_{\text{Intra}}} \).
- 51.
See (5.47).
- 52.
For the second fraction, a number of n elements has to be computed. Depending on the number of buckets or credits, the computation time can be longer than for the first term, but due to the linearity this term is virtually unproblematic.
- 53.
The computation time when calculating the infection model on bucket- instead on borrower-level can be reduced from 12 min to less than 1 s for a portfolio with 11 sectors, 7 PD-classes, and 5,000 creditors.
- 54.
The results refer to the total gross loss of a portfolio in terms of ES or VaR. To relate this to the unexpected net loss, the results have to be multiplied by the LGD and the EL has to be subtracted.
- 55.
The small mismatch is mainly due to keeping the ES-confidence level constant and not a result of the chosen intra-sector correlation function. If we directly compare the results from Monte Carlo simulations with the ES in the ASRF framework, the relative root mean squared error is reduced from 0.97% to 0.28%.
- 56.
In our analyses, the number of simulation runs is 500,000.
- 57.
If we consider all 25,000 simulated portfolios from Sect. 5.2.3, the lowest measured economic capital requirement was even 26% lower than the regulatory capital. This underlines the prospects of actively managing credit portfolios, e.g. with credit derivatives, but this is not in the scope of this thesis.
- 58.
CHKR I still corresponds to the DF-function based on Monte Carlo simulation and CHKR II on the Pykhtin formula.
- 59.
Acharya et al. (2006) examined credit portfolios of 105 Italian banks during the period 1993–1999. In this study, most bank portfolios had a HHI between 20% and 30%. However, it has to be considered that the number of different industry sectors was 23 whereas we use 11 different sectors. Thus, for a comparable degree of diversification their calculated HHI have to be slightly smaller than our HHIs.
- 60.
The runtimes refer to a quad-core PC with 2.66 GHz CPUs (calculated on one core).
- 61.
In contrast to this representation, Pykhtin (2004) applies these and the following formulas to n sector factors whereas we use K sector factors with \( \tilde{\bar{x}} \). This can lead to a significant reduction of the computation time as will be shown later on.
- 62.
This simplification of the Taylor series could already be used for the granularity adjustment in Sect. 4.2.1.1.
- 63.
Cf. Pykhtin (2004).
- 64.
The following calculations are based on Tasche (2006a), p. 41 ff.
- 65.
It has to be noticed that the conditional correlation matrix is symmetric, so we have \( \begin{array} {c} \frac{{d\eta_{2,c}^\infty \left( {\bar{x}} \right)}}{{d\bar{x}}} = \sum\limits_{i = 1}^n {\sum\limits_{j = 1}^n {{w_i}} } {w_j}ELG{D_i}ELG{D_j} \\ \cdot \left( {\frac{{d{p_i}\left( {\bar{x}} \right)}}{{d\bar{x}}}\Phi \left( {\frac{{{\Phi^{ - 1}}\left[ {{p_j}\left( {\bar{x}} \right)} \right] - \rho_{ij}^{\bar{x}}{\Phi^{ - 1}}\left[ {{p_i}\left( {\bar{x}} \right)} \right]}}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}} \right)} \right. \\ + \frac{{d{p_j}\left( {\bar{x}} \right)}}{{d\bar{x}}}\Phi \left( {\frac{{{\Phi^{ - 1}}\left[ {{p_i}\left( {\bar{x}} \right)} \right] - \rho_{ij}^{\bar{x}}{\Phi^{ - 1}}\left[ {{p_j}\left( {\bar{x}} \right)} \right]}}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}} \right) \\ \left. { - \frac{{d{p_i}\left( {\bar{x}} \right)}}{{d\bar{x}}}{p_j}\left( {\bar{x}} \right) - \frac{{d{p_j}\left( {\bar{x}} \right)}}{{d\bar{x}}}{p_i}\left( {\bar{x}} \right)} \right). \\ \end{array} \) for all i, j.
References
Acharya VV, Hasan I, Saunders A (2006) Should banks be diversified? Evidence from individual bank loan portfolios. J Bus 79(3):1355–1412
Basel Committee on Banking Supervision (2005a) International convergence of capital measurement and capital standards – a revised framework, Updated November 2005, Bank for International Settlements, Basel
Basel Committee on Banking Supervision (2006) Studies on credit risk concentration: an overview of the issues and a synopsis of the results from the research task force project. BCBS Working Paper No. 15, Bank for International Settlements, Basel
Cespedes J, de Juan Herrero J, Kreinin A, Rosen D (2006) A simple multi-factor “factor adjustment” for the treatment of diversification in credit capital rules. J Credit Risk 2(3):57–85
Cifuentes A, O’Connor G (1996) The binomial expansion method applied to CBO/CLO analysis, Special report. Moody’s Investor Service, New York
Cifuentes A, Wilcox C (1998) The double binomial method and its application to a special case of CDO structures, Special report. Moody’s Investor Service, New York
Cifuentes A, Murphy E, O’Connor G (1996) Emerging market collateralized bond obligations: an overview, Special report. Moody’s Investor Service, New York
Davis M, Lo V (2001) Infectious defaults. Quant Fin 1(4):382–387
Düllmann K (2006). Measuring business sector concentration by an infection model. Discussion Paper, Series 2: Banking and financial studies. Deutsche Bundesbank (3)
Düllmann K (2007) Measuring concentration risk in credit portfolios. In: Christodoulakis G, Satchell S (eds) The analytics of risk model validation. Academic Press, Amsterdam, pp 59–78
Düllmann K, Masschelein N (2007) A tractable model to measure sector concentration risk in credit portfolios. J Fin Serv Res 32(1):55–79
Düllmann K, Küll J, Kunisch M (2008) Estimating asset correlations from stock prices or default rates – Which method is superior? Discussion Paper, Series 2: Banking and financial studies, Deutsche Bundesbank (4)
Gordy MB (2003) A risk-factor model foundation for rating-based capital rules. J Fin Intermediation 12(3):199–232
Grundke P (2008) Regulatory treatment of the double default effect under the New Basel accord: how conservative is it? Rev Manag Sci 2(1):37–59
Gürtler M, Hibbeln M, Vöhringer C (2010) Measuring concentration risk for regulatory purposes. J Risk 12(3):69–104
Heitfield E, Burton S, Chomsisengphet S (2006) Systematic and idiosyncratic risk in syndicated loan portfolios. J Credit Risk 2(3):3–31
Lopez J (2004) The empirical relationship between average asset correlation, firm probability of default, and asset size. J Fin Intermediation 13(2):265–283
Martin R, Wilde T (2002) Unsystematic credit risk. Risk 15(11):123–128
Morinaga S, Shiina Y (2005) Underestimation of sector concentration risk by mis-assignment of borrowers. Working Paper, Tokyo
Pykhtin M (2004) Multi-factor adjustment. Risk 17(3):85–90
Tasche D (2006a) Anwendungen der Stochastik in der Bankenaufsicht. Lecture Notes, Frankfurt University. http://www-m4.ma.tum.de/pers/tasche. Accessed 18 Aug 2009
Weiss NA (2005) A course in probability. Addison-Wesley, Boston
Wilde T (2001) Probing granularity. Risk 14(8):103–106
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
5.1.1 Optimal Choice of the Single Correlation Factor
To relate \( 1/11 = 9.1\% \) to \( \tilde{\bar{L}} \), it is assumed that the new systematic factor \( \tilde{L} \) has a linear dependence to the original sector factors:Footnote 61
Condition (5.74) satisfies that the new systematic factor still has a variance of 1. In order to specify the correlation factors c i and the coefficients b k , it will be required that the loss \( {\text{with}}\,\,\,\sum\limits_{k = 1}^K {b_k^2 = 1.} \) equals the conditional expectation of the “true” loss \( \tilde{\bar{L}} \). This assures that the first element of the subsequently performed Taylor series expansion vanishes.Footnote 62 To determine \( b{E}(\tilde{L}|\tilde{\bar{x}}) \), we first recall that the asset return of obligor i in sector s can be written as
Now, each original sector factor \( {\tilde{a}_{s,i}} = \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {\tilde{x}_s} + \sqrt {{1 - {\rho_{{\text{Intra}},i}}}} \cdot {\tilde{\xi }_i}. \) is decomposed into a part that is related to the single-factor \( {\tilde{x}_s} \) and a part that is independent of this factor:
with \( {\tilde{x}_s} = {\bar{\rho }_s} \cdot \tilde{\bar{x}} + \sqrt {{1 - {{\bar{\rho }}_s}^2}} \cdot {\tilde{\eta }_s} \). Using (5.2), (5.73), and the independence of \( {\tilde{\eta }_s} \sim \mathcal{N}(0,1) \) if \( {\tilde{z}_i},\;{\tilde{z}_j} \), the correlation parameter \( i \ne j \) can be expressed as
Using this notation, the asset return (5.75) can now be written as
The independent standard normally distributed random variables \( \begin{array} {c} {{\tilde{a}}_{s,i}} = \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {{\tilde{x}}_s} + \sqrt {{1 - {\rho_{{\text{Intra}},i}}}} \cdot {{\tilde{\xi }}_i} \\ = \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot \left( {{{\bar{\rho }}_s} \cdot \tilde{\bar{x}} + \sqrt {{1 - {{\bar{\rho }}_s}^2}} \cdot {{\tilde{\eta }}_s}} \right) + \sqrt {{1 - {\rho_{{\text{Intra}},i}}}} \cdot {{\tilde{\xi }}_i} \\ = \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {{\bar{\rho }}_s} \cdot \tilde{\bar{x}} + \sqrt {{{\rho_{{\text{Intra}},i}} - {\rho_{{\text{Intra}},i}} \cdot {{\bar{\rho }}_s}^2}} \cdot {{\tilde{\eta }}_s} + \sqrt {{1 - {\rho_{{\text{Intra}},i}}}} \cdot {{\tilde{\xi }}_i}. \\ \end{array} \) and \( {\tilde{\eta }_s} \) can be combined into a new standard normally distributed random variable \( {\tilde{\xi }_i} \), leading to
with \( {\tilde{a}_{s,i}} = \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {\bar{\rho }_i} \cdot \tilde{\bar{x}} + \sqrt {{1 - {{\left( {\sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {{\bar{\rho }}_i}} \right)}^2}}} \cdot {\tilde{\zeta }_i}, \) for each obligor i in sector s. Since the variable \( {\bar{\rho }_i} = {\bar{\rho }_s} \) is independent of \( {\tilde{\zeta }_i} \), we can use the known formula of the single-factor model for the conditional expectation
The mentioned condition \( b{E}\left( {\tilde{L}|\tilde{\bar{x}}} \right) = \sum\limits_{i = 1}^n {{w_i} \cdot LG{D_i} \cdot \Phi \left[ {\frac{{{\Phi^{ - 1}}(P{D_i}) - \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {{\bar{\rho }}_i} \cdot {\Phi^{ - 1}}(\alpha )}}{{\sqrt {{1 - {{\left( {\sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {{\bar{\rho }}_i}} \right)}^2}}} }}} \right]} . \) leads to
using (5.9), (5.80), (5.77), and \( \begin{array} {lll} \tilde{\bar{L}} &= b{E}\left( {\tilde{L}|\tilde{\bar{x}}} \right) \\ & \Leftrightarrow \Phi \left[ {\frac{{{\Phi^{ - 1}}(P{D_i}) - {c_i} \cdot \tilde{\bar{x}}}}{{\sqrt {{1 - {c_i}^2}} }}} \right] = \Phi \left[ {\frac{{{\Phi^{ - 1}}(P{D_i}) - \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {{\bar{\rho }}_i} \cdot {\Phi^{ - 1}}(\alpha )}}{{\sqrt {{1 - {{\left( {\sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {{\bar{\rho }}_i}} \right)}^2}}} }}} \right] \\ & \Leftrightarrow {c_i} = \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {{\bar{\rho }}_i} = \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot \sum\limits_{k = 1}^K {{\alpha_{i,k}} \cdot {b_k}}, \\ \end{array} \) for each obligor i in sector s. While \( {\alpha_{i,k}} = {\alpha_{s,k}} \) and \( {\rho_{{\text{Intra}},i}} \) are known, the coefficients \( {\alpha_{i,k}} \) are unknown.
While (5.81) already satisfies that the first-order term of the Taylor series vanishes, the concrete choice of the parameter set {\( {b_k} \)} is critical concerning the distance between the zeroth-order term \( {b_k} \) and the unknown quantile \( {q_\alpha }(\tilde{\bar{L}}) \). Unfortunately, it is not obvious how this distance can be minimized. Thus, Pykhtin (2004) relies on the intuition that coefficients which maximize the (weighted) correlation between the single factor \( {q_\alpha }(\tilde{L}) \) and the sector factors {\( \tilde{\bar{x}} \)} should lead to good results. This leads to the following maximization problem:
subject to
The solution of this optimization problem isFootnote 63
where the positive constant Lagrange multiplier \( {b_k} = \sum\limits_{i = 1}^n {\frac{{{d_i} \cdot {\alpha_{ik}}}}{{2\tau }}}, \) is chosen in a way that {\( \tau \)} satisfies the constraint. As a final step, the weighting factor d i has to be chosen. After trying several specifications, Pykhtin (2004) uses
which is the VaR formula in a single-factor model. The intuition behind this choice is that obligors with a high exposure in terms of VaR should have a large weight in the maximization problem whereas obligors with a small VaR should have a minor impact. Summing up, the correlation parameter c i results from (5.81), where the coefficients b k are determined by (5.83)–(5.85).
5.1.2 Conditional Correlation
The correlation conditional on \( {d_i} = {w_i} \cdot LG{D_i} \cdot \Phi \left[ {\frac{{{\Phi^{ - 1}}(P{D_i}) + \sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {\Phi^{ - 1}}(\alpha )}}{{\sqrt {{1 - {\rho_{{\text{Intra}},i}}}} }}} \right], \) between the asset returns from (5.19) can be written as
using the independence of the factors \( \begin{array} {c} \rho_{ij}^{\bar{x}} = {\text{Corr}}\left( {{{\tilde{a}}_{s,i}},{{\tilde{a}}_{t,j}}|\tilde{\bar{x}}} \right) \\ = \frac{{{\text{Cov}}\left( {\sum\limits_{k = 1}^K {\left( {\sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {\alpha_{s,k}} - {c_i} \cdot {b_k}} \right)} \cdot {{\tilde{z}}_k},\sum\limits_{k = 1}^K {\left( {\sqrt {{{\rho_{{\text{Intra}},j}}}} \cdot {\alpha_{t,k}} - {c_j} \cdot {b_k}} \right)} \cdot {{\tilde{z}}_k}} \right)}}{{\sqrt {{b{V}\left( {{{\tilde{a}}_{s,i}}|\tilde{\bar{x}}} \right)}} \cdot \sqrt {{b{V}\left( {{{\tilde{a}}_{t,j}}|\tilde{\bar{x}}} \right)}} }} \\ = \frac{{\sum\limits_{k = 1}^K {\left( {\sqrt {{{\rho_{{\text{Intra}},i}}}} \cdot {\alpha_{s,k}} - {c_i} \cdot {b_k}} \right) \cdot \left( {\sqrt {{{\rho_{{\text{Intra}},j}}}} \cdot {\alpha_{t,k}} - {c_j} \cdot {b_k}} \right)} }}{{\sqrt {{1 - {c_i}^2}} \cdot \sqrt {{1 - {c_j}^2}} }}, \\ \end{array} \). The numerator can be simplified using \( {\tilde{z}_k} \) from (5.81) and \( \sum\limits_{k = 1}^K {{\alpha_{s,k}} \cdot {b_k}} = {{{c_i}} \mathord{\left/{\vphantom {{{c_i}} {\sqrt {{{\rho_{{\text{Intra}},i}}}} }}} \right.} {\sqrt {{{\rho_{{\text{Intra}},i}}}} }} \) from (5.74):
This leads to
5.1.3 Calculation of the Decomposed Variance
In order to determine the conditional variance, it is decomposed into the following terms:Footnote 64
For calculation of these terms, first the expressions (a) \( b{V}\left( {\tilde{L}|\tilde{\bar{x}} = \bar{x}} \right) = b{V}\left[ {b{E}\left( {\tilde{L}|\left\{ {{{\tilde{z}}_k}} \right\}} \right)|\tilde{\bar{x}} = \bar{x}} \right] + b{E}\left[ {b{V}\left( {\tilde{L}|\left\{ {{{\tilde{z}}_k}} \right\}} \right)|\tilde{\bar{x}} = \bar{x}} \right] \), (b) \( b{E}(\tilde{L}|\{ {\tilde{z}_k}\} ) \), and (c) \( b{E}({\tilde{L}^2}|\{ {\tilde{z}_k}\} ) \) will be calculated. The conditional loss is given as
and for stochastically independent LGDs this leads to
-
(a)
With \( \tilde{L}|\{ {\tilde{z}_k}\} = \sum\limits_i {{w_i}} \cdot {\widetilde{{LGD}}_i} \cdot \left( {{1_{\left\{ {{{\tilde{D}}_i}} \right\}}}|\{ {{\tilde{z}}_k}\} } \right). \) and \( b{E}\left( {{{\widetilde{{LGD}}}_i}} \right) = :ELG{D_i} \) we obtain:
-
(b)
Consider that \( b{E}\left( {\tilde{L}|\{ {{\tilde{z}}_k}\} } \right) = \sum\limits_i {{w_i}} \cdot ELG{D_i} \cdot {p_i}\left( {\{ {{\tilde{z}}_k}\} } \right). \), \( 1_{\left\{ {{{\tilde{D}}_i}} \right\}}^2 = {1_{\left\{ {{{\tilde{D}}_i}} \right\}}} \), and
as well as
Moreover, we have
Thus, we obtain:
-
(c)
The conditional variance \( \begin{array} {c} b{E}\left( {{{\tilde{L}}^2}|\{ {{\tilde{z}}_k}\} } \right) = b{E}\left[ {{{\sum\limits_i {\left( {{w_i}{{\widetilde{{LGD}}}_i}{1_{\left\{ {{{\tilde{D}}_i}} \right\}}}} \right)} }^2}|\{ {{\tilde{z}}_k}\} } \right] \\ = b{E}\left[ {\sum\limits_i {{w_i}^2} {{\widetilde{{LGD}}}_i}^21_{\left\{ {{{\tilde{D}}_i}} \right\}}^2|\{ {{\tilde{z}}_k}\} } \right] + b{E}\left[ {\sum\limits_i {\sum\limits_{j \ne i} {{w_i}{w_j}{{\widetilde{{LGD}}}_i}{{\widetilde{{LGD}}}_j}{1_{\left\{ {{{\tilde{D}}_i}} \right\}}}{1_{\left\{ {{{\tilde{D}}_j}} \right\}}}|\{ {{\tilde{z}}_k}\} } } } \right] \\ = \sum\limits_i {{w_i}^2} b{E}\left( {LG{D_i}^2} \right){p_i}\left( {\{ {{\tilde{z}}_k}\} } \right) \\ + \sum\limits_i {\sum\limits_{j \ne i} {{w_i}{w_j}b{E}\left( {LG{D_i}LG{D_j}} \right)b{E}\left( {{1_{\left\{ {{{\tilde{D}}_i}} \right\}}}{1_{\left\{ {{{\tilde{D}}_j}} \right\}}}} \right)} } \\ = \sum\limits_i {{w_i}^2} \left( {ELG{D_i}^2 + VLG{D_i}} \right){p_i}\left( {\{ {{\tilde{z}}_k}\} } \right) \\ + \sum\limits_i {\sum\limits_{j \ne i} {{w_i}{w_j}ELG{D_i}ELG{D_j}{p_i}\left( {\{ {{\tilde{z}}_k}\} } \right){p_j}\left( {\{ {{\tilde{z}}_k}\} } \right)} } \\ = \sum\limits_i {{w_i}^2} \left( {ELG{D_i}^2 + VLG{D_i}} \right){p_i}\left( {\{ {{\tilde{z}}_k}\} } \right) - \sum\limits_i {{w_i}^2ELG{D_i}^2{p_i}^2\left( {\{ {{\tilde{z}}_k}\} } \right)} \\ + \sum\limits_i {\sum\limits_j {{w_i}{w_j}ELG{D_i}ELG{D_j}{p_i}\left( {\{ {{\tilde{z}}_k}\} } \right){p_j}\left( {\{ {{\tilde{z}}_k}\} } \right)} } \\ = \sum\limits_i {{w_i}^2} \left( {ELG{D_i}^2\left[ {{p_i}\left( {\{ {{\tilde{z}}_k}\} } \right) - {p_i}^2\left( {\{ {{\tilde{z}}_k}\} } \right)} \right] + VLG{D_i}{p_i}\left( {\{ {{\tilde{z}}_k}\} } \right)} \right) \\ + {b{E}^2}\left( {\tilde{L}|\{ {{\tilde{z}}_k}\} } \right). \\ \end{array} \) is equal to
-
(d)
Using the law of iterated expectation, we have
Thus, with (5.98) the expectation of the conditional variance can be written as
For independent idiosyncratic factors \( \begin{array} {c} b{E}\left[ {b{V}\left( {\tilde{L}|\{ {{\tilde{z}}_k}\} } \right)|\tilde{\bar{x}} = \bar{x}} \right] = \sum\limits_i {{w_i}^2\left( {ELG{D_i}^2\left( {b{E}\left[ {{p_i}\left( {\{ {{\tilde{z}}_k}\} } \right)|\bar{x}} \right] - b{E}\left[ {{p_i}^2\left( {\{ {{\tilde{z}}_k}\} } \right)|\bar{x}} \right]} \right)} \right.} \\ \left. { + VLG{D_i}b{E}\left[ {{p_i}\left( {\{ {{\tilde{z}}_k}\} } \right)|\bar{x}} \right]} \right) \\ = \sum\limits_i {{w_i}^2\left( {ELG{D_i}^2\left( {{p_i}(\bar{x}) - b{P}\left[ {\left( {{1_{\left\{ {{{\tilde{D}}_i}} \right\}}} = 1} \right) \wedge \left( {{1_{\left\{ {{{\tilde{D}}_i}^\prime } \right\}}} = 1} \right)|\bar{x}} \right]} \right)} \right.} \\ \left. { + VLG{D_i}{p_i}(\bar{x})} \right). \\ \end{array} \) and with
we get
with the correlation conditional on \( \begin{array} {c} b{P}\left[ {\left( {{1_{\left\{ {{{\tilde{D}}_i}} \right\}}} = 1} \right) \wedge \left( {{1_{\left\{ {{{\tilde{D}}_i}^\prime } \right\}}} = 1} \right)|\bar{x}} \right] \\ = b{P}\left[ {{c_i} \cdot \tilde{\bar{x}} + \sqrt {{1 - {c_i}^2}} \cdot {{\tilde{\zeta }}_i} \leq {\Phi^{ - 1}}(P{D_i}),{c_i} \cdot \tilde{x} + \sqrt {{1 - {c_i}^2}} \cdot {{\tilde{\zeta }}_i}^\prime \leq {\Phi^{ - 1}}(P{D_i})|\bar{x}} \right] \\ = b{P}\left[ {{{\tilde{\zeta }}_i} \leq \frac{{{\Phi^{ - 1}}(P{D_i}) - {c_i} \cdot \bar{x}}}{{\sqrt {{1 - {c_i}^2}} }},{{\tilde{\zeta }}_i}^\prime \leq \frac{{{\Phi^{ - 1}}(P{D_i}) - {c_i} \cdot \bar{x}}}{{\sqrt {{1 - {c_i}^2}} }}} \right] \\ = {\Phi_2}\left( {{\Phi^{ - 1}}\left( {{p_i}(\bar{x})} \right),{\Phi^{ - 1}}\left( {{p_i}(\bar{x})} \right),\rho_{ii}^{\bar{x}}} \right), \\ \end{array} \) of (5.20). Hence, (5.100) results in
-
(e)
Using (5.92), the variance of the conditional expectation can be expressed as
leading to
Analogous to (5.102) and using the conditional correlation (5.20), this can be expressed as:
5.1.4 Derivatives of the Decomposed Variance Terms
As both conditional variance terms are linear in the bivariate normal distribution, the derivative of the bivariate normal distribution will be calculated subsequently. Then, the derivatives of \( \begin{array} {c} b{V}\left[ {b{E}\left( {\tilde{L}|\{ {{\tilde{z}}_k}\} } \right)|\tilde{\bar{x}} = \bar{x}} \right] = \sum\limits_i {\sum\limits_j {{w_i}{w_j}ELG{D_i}ELG{D_j}} } \\ \cdot \left[ {{\Phi_2}\left( {{\Phi^{ - 1}}\left( {{p_i}(\bar{x})} \right),{\Phi^{ - 1}}\left( {{p_j}(\bar{x})} \right),\rho_{ij}^{\bar{x}}} \right) - {p_i}(\bar{x}){p_j}(\bar{x})} \right]. \\ \end{array} \) and \( \eta_{2,c}^\infty (\bar{x}) \) will be computed.
Proposition
The derivative of the bivariate normal distribution can be written as:
Proof
Using the notation
and the chain rule, we get
For calculation of term (II) and (IV), we rewrite the bivariate normal distribution according to Appendix 2.8.6 as
Thus, we have
The term (*) is equivalent to
Hence, (5.111) can be written as
For solving the integral, we substitute \( \begin{array} {c} \frac{\partial }{{\partial {y_i}}}{\Phi_2}\left( {{y_i},{y_j},\rho_{ij}^{\bar{x}}} \right) = \frac{1}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}\int\limits_{z = - \infty }^{{y_i}} {\frac{1}{{2\pi }}\exp \left( { - \frac{1}{2}\left[ {{y_i}^2 + {{\left( {\frac{{z - \rho_{ij}^{\bar{x}}{y_i}}}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}} \right)}^2}} \right]} \right)dz} \\ = \varphi \left( {{y_i}} \right)\int\limits_{z = - \infty }^{{y_j}} {\frac{1}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}\varphi \left( {\frac{{z - \rho_{ij}^{\bar{x}}{y_i}}}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}} \right)dz} . \\ \end{array} \), and thus \( t: = \frac{{z - \rho_{ij}^{\bar{x}}{y_i}}}{{\sqrt {{1 - {{(\rho_{ij}^{\bar{x}})}^2}}} }} \). This leads to
Analogously, the term (IV) of (5.109) is equivalent to
The derivatives (I) and (III) of (5.109) are given as
Thus, inserting (5.114), (5.115), and (5.116) into (5.109), the derivative of the bivariate normal distribution finally results in
where the derivatives \( \begin{array} {c} \frac{d}{{d\bar{x}}}{\Phi_2}\left( {{y_i}\left( {\bar{x}} \right),{y_j}\left( {\bar{x}} \right),\rho_{ij}^{\bar{x}}} \right) = - \frac{{{c_i}}}{{\sqrt {{1 - {c_i}^2}} }}\varphi \left( {{y_i}} \right)\Phi \left( {\frac{{{y_j} - \rho_{ij}^{\bar{x}}{y_i}}}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}} \right) \\ - \frac{{{c_j}}}{{\sqrt {{1 - {c_j}^2}} }}\varphi \left( {{y_j}} \right)\Phi \left( {\frac{{{y_i} - \rho_{ij}^{\bar{x}}{y_j}}}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}} \right) \\ = \frac{{d{p_i}\left( {\bar{x}} \right)}}{{d\bar{x}}}\Phi \left( {\frac{{{\Phi^{ - 1}}\left[ {{p_j}\left( {\bar{x}} \right)} \right] - \rho_{ij}^{\bar{x}}{\Phi^{ - 1}}\left[ {{p_i}\left( {\bar{x}} \right)} \right]}}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}} \right) \\ + \frac{{d{p_j}\left( {\bar{x}} \right)}}{{d\bar{x}}}\Phi \left( {\frac{{{\Phi^{ - 1}}\left[ {{p_i}\left( {\bar{x}} \right)} \right] - \rho_{ij}^{\bar{x}}{\Phi^{ - 1}}\left[ {{p_j}\left( {\bar{x}} \right)} \right]}}{{\sqrt {{1 - {{\left( {\rho_{ij}^{\bar{x}}} \right)}^2}}} }}} \right), \\ \end{array} \) and \( \frac{{d{p_i}\left( {\bar{x}} \right)}}{{d\bar{x}}} \) are given by (5.16), which is equal to proposition (5.107).
As a next step, the derivatives of \( \frac{{d{p_j}\left( {\bar{x}} \right)}}{{d\bar{x}}} \) and \( \eta_{2,c}^\infty (\bar{x}) \) will be calculated. With
we get
Using the derivative of the bivariate normal distribution from (5.117) yields
Comparing the terms on the right-hand side, it can be found that the first and second summand as well as the third and fourth summand only differ concerning the indices i and j. Due to the double sum, each combination of i and j occurs twice. Thus, (5.120) can be simplified to:Footnote 65
Similarly, the derivative of
is given as
Inserting the derivative of the bivariate normal distribution (5.117) finally leads to
5.1.5 Moment Matching in the BET-Model
5.1.5.1 Matching the First Moment
The expected loss of the original portfolio can be calculated as
and the expected loss of the hypothetical portfolio as
with \( \begin{array} {c} b{E}\left( {{{\tilde{L}}^{\text{hyp}}}} \right) = \sum\limits_{i = 1}^D {\frac{1}{D} \cdot LGD \cdot b{E}\left( {{1_{\left\{ {{{\tilde{D}}_i}} \right\}}}} \right)} = \frac{1}{D} \cdot LGD \cdot \sum\limits_{i = 1}^D {\bar{p}} \\ = \frac{1}{D} \cdot LGD \cdot D \cdot \bar{p} = LGD \cdot \bar{p}, \\ \end{array} \) for all i. Thus, matching the expectation for both portfolios leads to
5.1.5.2 Matching the Second Moment
For the original portfolio, the variance can be calculated as
As the default variable is Bernoulli distributed, the variance terms equal
and we obtain
Due to the independence of the default events in the hypothetical portfolio, the variance of this portfolio is
Matching the variance terms (5.130) and (5.131) leads to
5.1.6 Interrelation of the Pairwise Default Correlation and the Asset Correlation
Using the standard calculus for the correlation and covariance as well as the variance of a Bernoulli distributed variable, the pairwise default correlation between borrower i in sector s and borrower j in sector t can be expressed as
The expectation values of the individual default events equal \( \begin{array} {c} {\text{Corr}}\left( {{1_{\left\{ {{{\tilde{D}}_{s,i}}} \right\}}},{1_{\left\{ {{{\tilde{D}}_{t,j}}} \right\}}}} \right) = \frac{{{\text{Cov}}\left( {{1_{\left\{ {{{\tilde{D}}_{s,i}}} \right\}}},{1_{\left\{ {{{\tilde{D}}_{t,j}}} \right\}}}} \right)}}{{\sqrt {{b{V}\left( {{1_{\left\{ {{{\tilde{D}}_{s,i}}} \right\}}}} \right)}} \cdot \sqrt {{b{V}\left( {{1_{\left\{ {{{\tilde{D}}_{t,j}}} \right\}}}} \right)}} }} \\ = \frac{{{\text{Cov}}\left( {{1_{\left\{ {{{\tilde{D}}_{s,i}}} \right\}}},{1_{\left\{ {{{\tilde{D}}_{t,j}}} \right\}}}} \right)}}{{\sqrt {{P{D_{s,i}} \cdot \left( {1 - P{D_{s,i}}} \right) \cdot P{D_{t,j}} \cdot \left( {1 - P{D_{t,j}}} \right)}} }} \\ = \frac{{b{E}\left( {{1_{\left\{ {{{\tilde{D}}_{s,i}}} \right\}}} \cdot {1_{\left\{ {{{\tilde{D}}_{t,j}}} \right\}}}} \right) - b{E}\left( {{1_{\left\{ {{{\tilde{D}}_{s,i}}} \right\}}}} \right) \cdot b{E}\left( {{1_{\left\{ {{{\tilde{D}}_{t,j}}} \right\}}}} \right)}}{{\sqrt {{P{D_{s,i}} \cdot \left( {1 - P{D_{s,i}}} \right) \cdot P{D_{t,j}} \cdot \left( {1 - P{D_{t,j}}} \right)}} }}. \\ \end{array} \) and \( P{D_{s,i}} \). Similar to (5.102), assuming a normally distributed asset return, the expectation value of a simultaneous default can be written as
Thus, we get
5.1.7 Expected Number of Defaults in the Infectious Defaults Model
Due to the homogeneity of the portfolio and the stochastic independence of all indicator variables, the expected number of defaults is
Rights and permissions
Copyright information
© 2010 Physica-Verlag HD
About this chapter
Cite this chapter
Hibbeln, M. (2010). Model-Based Measurement of Sector Concentration Risk in Credit Portfolios. In: Risk Management in Credit Portfolios. Contributions to Economics. Physica, Heidelberg. https://doi.org/10.1007/978-3-7908-2607-4_5
Download citation
DOI: https://doi.org/10.1007/978-3-7908-2607-4_5
Published:
Publisher Name: Physica, Heidelberg
Print ISBN: 978-3-7908-2606-7
Online ISBN: 978-3-7908-2607-4
eBook Packages: Business and EconomicsEconomics and Finance (R0)