betaFeedback

WorksEconomics, Econometrics and Finance

Statistical moments, risk weighting and percentiles of a loss distribution of a finite homogeneous credit portfolio; extending the Vasicek distribution

Abstract

The Vasicek distribution assumes an infinite portfolio of identical obligors and forms the basis of the Basel III/IV Risk Weighted Assets (RWA) calculation. We derive the central moments of the loss distribution for a finite number (n) of obligors. We use these moments to calculate percentiles of the fraction (x) of defaulted obligors, via the Cornish-Fisher expansion and via a weighting method. Both methods are based on scaling the standard deviation in x by a capital multiple. The weighting method assumes that the capital multiple is linear in the central moments and leads to an Extended Vasicek distribution. We apply this distribution in calculating an adjusted RWA for small portfolios. Simulations show that this adjustment is only relevant for small portfolios of less than 100 obligors and that the Extended Vasicek distribution allows analytical risk capital calculations for portfolios with different rating grades, even when the number of obligors per rating grade is small.

MSC classes: 91G40, 62P05

Introduction

In his seminal paper, Vasicek introduced a closed-form distribution of loan losses for credit portfolios with systemic risk expressed as a common risk factor (O. Vasicek, 1991) (O. Vasicek, 1987). This distribution, applicable for a large homogeneous credit portfolio, has been frequently applied, for example in pricing CDOs [8], importance sampling (P. Glasserman, 2003). Some authors use the Vasicek distribution as a control variate in Monte-Carlo simulations (V.Tchistiakov, 2004). The Vasicek distribution has also been extended in a multiple factor approach [8].

The Basel Committee on Banking Supervision (BCBS) also applied the Vasicek distribution in the Risk Weighted Assets (RWA) methodology. This RWA methodology first appeared in the revised Basel II framework in 2006 [2]. The latest final version with new requirements for the quality and quantity of regulatory capital appeared in 2011 in response to the financial crisis in 2007 – 2009 [3]. The Vasicek distribution is also relevant for calculating the Solvency Capital Requirement (SCR) for Type II Counterparty Default Risk in insurance [5].

The Vasicek distribution assumes a credit portfolio of an infinite number of obligors. The credit portfolio is also assumed homogeneous, with equal weights for all exposures. Some authors extended the Vasicek model for inhomogeneous portfolios based on the Herfindahl index (M. Osadchiy, 2025). The new Basel IV proposals account for the assumptions of the Vasicek distribution by introducing a floor on the RWA. This output floor ties the output of the bank’s internal risk calculation to the standardised risk approach. This prevents the bank’s own internal measurement of its risk exposure from yielding less than 72.5% of the standardised approach. Basel IV is the latest step in regulatory development since the 1980s.

The aim of this paper is two-fold: first, we introduce closed-form equations for statistical moments and percentiles of the loss distribution for a portfolio with a finite number of obligors. Second, we use these results to calculate RWAs. The next section reviews the Vasicek Distribution and its properties. Section 3 introduces a theorem to calculate central moments of the loss distribution for a finite homogeneous portfolio. This theorem is applied in deriving percentiles of the loss distribution, using two methods: the Cornish-Fisher expansion and a weighting approach. Section 4 compares these methods with simulations, revealing that they result in an appropriate estimation of the 99.9% percentile. This leads to a finite portfolio adjustment of the RWA. Section 5 shows these RWA adjustments have only substantial impact when number of obligors is less than 100. The last section concludes that the current RWA with the standardized floor is sufficiently prudent, based on comparing with the finite portfolio adjustment.

Review of the Vasicek Distribution

Here we briefly review the Vasicek distribution (O. Vasicek, 1991). This distribution associates each obligor ii in the portfolio with a latent variable ziz_i that depends on a single systematic risk factor ff and an obligor-specific factor εi\varepsilon_i:

zi=ρf+1−ρεif∼Φ[0,1]εi∼Φ[0,1](1)z_i = \sqrt{\rho}f + \sqrt{1-\rho}\varepsilon_i \quad f\sim\Phi[0,1]\quad\varepsilon_i\sim\Phi[0,1] \tag*{(1)}

The symbol Φ[]\Phi[] denotes the cumulative standard normal distribution. Since E[zi]=0\mathbb{E}[z_i] = 0, E[zi2]=1\mathbb{E}[z_i^2] = 1 and E[zizj]=ρ\mathbb{E}[z_i z_j] = \rho, Equation (1) results in a homogeneous n×nn \times n correlation matrix CnC_n for a credit portfolio of nn identical obligors:

Cn=[1ρρ⋯ρ1ρ⋯ρρ1⋯⋮⋮⋮⋱](2)C_n = \begin{bmatrix} 1 & \rho& \rho& \cdots\\ \rho& 1 & \rho& \cdots\\ \rho& \rho& 1 & \cdots\\ \vdots& \vdots& \vdots& \ddots \end{bmatrix} \tag*{(2)}

When the latent variable ziz_i is below a threshold τ=Φ−1[P0]\tau= \Phi^{-1}[P_0] that is related to the unconditional Probability of Default (P0P_0), obligor ii defaults on his debt obligations. Equation (1) shows that ziz_i is standard normally distributed, leading to the following default probability based on condition zi≤τz_i \leq\tau:

p(f)=Φ[τ−ρf1−ρ](3)p(f) = \Phi\left[\frac{\tau-\sqrt{\rho}f}{\sqrt{1-\rho}}\right] \tag*{(3)}

Then Equation (3) is applied as a conditional default probability in the cumulative binomial distribution Bn,p(f)(d)B_{n,p(f)}(d) of dd defaults for a portfolio of nn obligors. The portfolio is infinite and homogeneous, i.e. the exposure and default probability of every obligor is the same. The Vasicek distribution follows from integrating the binomial distribution Bn,p(f)(d)B_{n,p(f)}(d) over the systematic factor ff in the limit n→∞n \to\infty, which we call hereafter the Vasicek limit:

FV(x)=lim⁡n→∞P[kn≤x]=lim⁡n→∞∫01Bn,p(f)(k) dΦ(f)=Φ[1−ρ Φ−1[x]−τρ](4)F_V(x) = \lim_{n\to\infty} P\left[\frac{k}{n}\leq x\right] = \lim_{n\to\infty}\int_0^1 B_{n,p(f)}(k)\,\mathrm{d}\Phi(f) = \Phi\left[\frac{\sqrt{1-\rho}\,\Phi^{-1}[x]-\tau}{\sqrt{\rho}}\right] \tag*{(4)}

with x=dnx = \frac{d}{n} as the fraction of defaulted obligors.

We summarize here some properties of the Vasicek distribution: statistical moments, variance and percentiles, starting with the following theorem:

Theorem 1. The non-central moments of the Vasicek distribution with P0P_0 as the unconditional probability of default and with correlation ρ\rho follows from:

E[xm]=1(2π)m∣Cm∣∫−∞τ⋯∫−∞τexp⁡(−uˉTCm−1uˉ2) du1⋯dum=Φm[τ…τ⏟m times;ρ](5)\mathbb{E}[x^m] = \frac{1}{\sqrt{(2\pi)^m\lvert C_m\rvert}} \int_{-\infty}^{\tau} \cdots\int_{-\infty}^{\tau} \exp\left(\frac{-\bar{u}^{\mathsf{T}}C_m^{-1}\bar{u}}{2}\right) \, \mathrm{d}u_1 \cdots\mathrm{d}u_m = \Phi_m\left[\underbrace{\tau\ldots\tau}_{m\ \text{times}};\rho\right] \tag*{(5)}

with τ=Φ−1[P0]\tau= \Phi^{-1}[P_0] and matrix Cm−1C_m^{-1} as the inverse of the m×mm \times m homogeneous correlation matrix.

To prove Theorem 1, we transform xx in uu by FV(x)=Φ(u)F_V(x) = \Phi(u), giving:

E[xm]=∫01xm dFV(x)=∫01(Φ[τ−ρu1−ρ])m dΦ(u)(6)\mathbb{E}[x^m] = \int_0^1 x^m \,\mathrm{d}F_V(x) = \int_0^1 \left(\Phi\left[\frac{\tau-\sqrt{\rho}u}{\sqrt{1-\rho}}\right]\right)^m \,\mathrm{d}\Phi(u) \tag*{(6)}

Equation (29) in Appendix A shows that ∫01(Φ[a+by])m dΦ(y)=Φm[a1+b2…a1+b2⏟m times;b21+b2]\int_0^1(\Phi[a+by])^m \,\mathrm{d}\Phi(y) = \Phi_m\left[\underbrace{\frac{a}{\sqrt{1+b^2}}\ldots\frac{a}{\sqrt{1+b^2}}}_{m\ \text{times}};\frac{b^2}{1+b^2}\right].

Substituting a=τ1−ρa = \frac{\tau}{\sqrt{1-\rho}} and b=−ρ1−ρb = \frac{-\sqrt{\rho}}{\sqrt{1-\rho}} gives Theorem 1. □\square

Theorem 1 gives the expectation E[x]=Φ1[τ;ρ]=P0\mathbb{E}[x] = \Phi_1[\tau;\rho] = P_0 for m=1m=1 and the variance of the Vasicek distribution when combined with m=2m=2 [4]:

σx,∞2=E[x2]−(E[x])2=Φ2[τ,τ;ρ]−P02(7)\sigma_{x,\infty}^{2} = \mathbb{E}[x^2] - \left(\mathbb{E}[x]\right)^2 = \Phi_2[\tau,\tau;\rho] - P_0^2 \tag*{(7)}

with Φ2[τ,τ;ρ]=∫−∞τ∫−∞τ12π1−ρ2exp⁡{−x2−2ρxy+y22(1−ρ2)} dy dx\Phi_2[\tau,\tau;\rho] = \int_{-\infty}^{\tau}\int_{-\infty}^{\tau}\frac{1}{2\pi\sqrt{1-\rho^2}}\exp\left\{-\frac{x^2-2\rho xy+y^2}{2(1-\rho^2)}\right\}\,\mathrm{d}y\,\mathrm{d}x as the cumulative bivariate standard normal distribution. Table 1 shows how the Vasicek moments in Theorem 1 decrease with increasing mm.

The percentiles at confidence level α\alpha of the Vasicek distribution follow from inverting FV(xα,∞)=αF_V(x_{\alpha,\infty}) = \alpha using Equation (4):

xα,∞=Φ[τ+ρΦ−1[α]1−ρ](8)x_{\alpha,\infty}=\Phi\left[\frac{\tau+\sqrt{\rho}\Phi^{-1}[\alpha]}{\sqrt{1-\rho}}\right] \tag*{(8)}

This equation is the basis of the Basel RWA calculation [2]. The next section introduces a modified version of Equation (8) for portfolios with a small number of obligors.

Relaxing the Vasicek Limit

Central Moments of a loss distribution of a finite credit portfolio

Equation (7) and Equation (8) are based on approximating the cumulative binomial distribution in the Vasicek limit. Up to our knowledge, there is no closed-form equation for a finite portfolio with n<∞n < \infty. However, we can derive equations for the central moments of a binomial distribution with the conditional default probability as given in Equation (3). This gives an expectation E[x]=E[dn]=P0\mathbb{E}[x] = \mathbb{E}\left[\frac{d}{n}\right] = P_0 for a portfolio of nn obligors, of which dd obligors default on their loans. Then the central moments are defined as (μm)x,n=E[(x−P0)m](\mu_m)_{x,n} = \mathbb{E}[(x-P_0)^m]. By definition, the first central moment is zero, but the second central moment gives the variance in x=dnx=\frac{d}{n} for a portfolio with n<∞n < \infty obligors:

σx,n2=(μ2)x,n=∫01{np(f)+n(n−1)p(f)2n2} dΦ(f)−P02=(1−1n)(Φ2[τ,τ;ρ]−P02)+P0−P02n(9)\begin{aligned} \sigma_{x,n}^{2}=(\mu_2)_{x,n} &=\int_0^1\left\{\frac{np(f)+n(n-1)p(f)^2}{n^2}\right\}\,\mathrm{d}\Phi(f)-P_0^2\\ &=\left(1-\frac{1}{n}\right)\left(\Phi_2[\tau,\tau;\rho]-P_0^2\right)+\frac{P_0-P_0^2}{n} \tag*{(9)} \end{aligned}

in which we applied Theorem 1 in the second step. Using Equation (7), the variance σx,n2\sigma_{x,n}^2 can be written as a 1n\frac{1}{n}-weighted average of the variance σx,12=P0(1−P0)\sigma_{x,1}^2=P_0(1-P_0) of a one-obligor portfolio and the variance σx,∞2\sigma_{x,\infty}^2 in the Vasicek limit:

σx,n2=σx,12n+(1−1n)σx,∞2=P0(1−P0)n+(1−1n)σx,∞2(10)\sigma_{x,n}^2=\frac{\sigma_{x,1}^2}{n}+\left(1-\frac{1}{n}\right)\sigma_{x,\infty}^2=\frac{P_0(1-P_0)}{n}+\left(1-\frac{1}{n}\right)\sigma_{x,\infty}^2 \tag*{(10)}

Working out central moments of order m>2m>2 shows that this 1n\frac{1}{n}-weighting also holds approximately for higher central moments. We formalize this in the following Theorem:

Theorem 2. The central moments of a loss distribution with nn obligors with the same unconditional default probability P0P_0 and with mutual correlation ρ\rho, can be decomposed in the following:

(μm)x,n=1n(μm)x,1+(1−1n)(μm)x,∞+1n(1−1n)∑l=1mam,lΦl[τ…τ;ρ](11)(\mu_m)_{x,n} = \frac{1}{n}(\mu_m)_{x,1} + \left(1-\frac{1}{n}\right)(\mu_m)_{x,\infty} + \frac{1}{n}\left(1-\frac{1}{n}\right)\sum_{l=1}^{m}a_{m,l}\Phi_l[\tau\ldots\tau;\rho] \tag*{(11)}

with coefficients amla_{m l} summing up to zero: ∑l=1mam,l=0\sum_{l=1}^{m}a_{m,l}=0 and Φl[τ…τ;ρ]\Phi_l[\tau\ldots\tau;\rho] representing the non-central moments of the Vasicek distribution.

We prove Theorem 2 in Appendix B. Table 2 provides the coefficients amla_{ml} as calculated by Equation (42) for the first central moments. The coefficients are zero for m=2m=2 and add up to zero for m=3,4m=3,4. The table shows that ∑l=1mam,l=0\sum_{l=1}^{m}a_{m,l}=0 holds for m=2m=2, 3 and 4.

am,1a_{m,1}am,2a_{m,2}am,3a_{m,3}am,4a_{m,4}∑l=1mam,l\sum_{l=1}^{m}a_{m,l}
m=2m=2000
m=3m=3-13-20
m=4m=44P0−1−1n4P_0-1-\frac{1}{n}7n−12P0\frac{7}{n}-12P_06−12n+8P06-\frac{12}{n}+8P_06n−5\frac{6}{n}-50

Table 2. Coefficients in Equation (11) for the first three central moments m=2,3,4m=2,3,4.

Appendix B derives the general form of the coefficients am,la_{m,l}:

am,1=n∑k=2m(mk)(−P0)m−kn−(k−2)−1n−1a_{m,1}=n\sum_{k=2}^{m}\binom{m}{k}(-P_0)^{m-k}\frac{n^{-(k-2)}-1}{n-1}
am,l=n{∑k=lm(mk)(−P0)m−kbk,l−(ml)(−P0)m−l} for l>1(12)a_{m,l}=n\left\{\sum_{k=l}^{m}\binom{m}{k}(-P_0)^{m-k}b_{k,l}-\binom{m}{l}(-P_0)^{m-l}\right\}\text{ for }l>1 \tag*{(12)}

in which bk,l=Sk,ln(n−2)l−2nk−2b_{k,l}=S_{k,l}\frac{n^{(n-2)l-2}}{n^{k-2}} with Sk,lS_{k,l} as Stirling numbers of the 2nd2^{\mathrm{nd}} kind and nln^l as the falling power nl=n(n−1)(…)(n−l+1)n^l=n(n-1)(\ldots)(n-l+1) with n0=1n^0=1 and n1=nn^1=n. Appendix B proves that ∑l=1mam,l=0\sum_{l=1}^{m}a_{m,l}=0. Table 3 shows that coefficients bklb_{kl} are polynomials in 1n\frac{1}{n}.

bk,lb_{k,l} coefficients
k=2k=2b2,2=1b_{2,2}=1
k=3k=3b3,2=3nb_{3,2}=\frac{3}{n}, b3,3=1−2nb_{3,3}=1-\frac{2}{n}
k=4k=4b4,2=7n2b_{4,2}=\frac{7}{n^2}, b4,3=6n(1−2n)b_{4,3}=\frac{6}{n}\left(1-\frac{2}{n}\right), b4,4=(1−2n)(1−3n)b_{4,4}=\left(1-\frac{2}{n}\right)\left(1-\frac{3}{n}\right)
k=5k=5b5,2=15n3b_{5,2}=\frac{15}{n^3}, b5,3=25n2(1−2n)b_{5,3}=\frac{25}{n^2}\left(1-\frac{2}{n}\right), b5,4=10n(1−2n)(1−3n)b_{5,4}=\frac{10}{n}\left(1-\frac{2}{n}\right)\left(1-\frac{3}{n}\right), b5,5=(1−2n)(1−3n)(1−4n)b_{5,5}=\left(1-\frac{2}{n}\right)\left(1-\frac{3}{n}\right)\left(1-\frac{4}{n}\right)

Table 3. Coefficients bklb_{kl}, introduced in Equation (32).

Percentiles of a loss distribution for a finite portfolio

We may use Theorem 2 to calculate percentiles. Percentiles can be expressed in central moments like the variance, skewness and kurtosis, like the Cornish-Fisher expansion [6]. First, we introduce the capital multiple as a scaling factor to calculate percentiles [4]. The capital multiple cmncm_n expresses how much the percentile deviates from the expectation E[x]=P0\mathbb{E}[x] = P_0 in terms of standard deviations:

xα,n=P0+cmn⋅σx,n(13)x_{\alpha,n} = P_0 + cm_n \cdot\sigma_{x,n} \tag*{(13)}

The standard deviation σx,n\sigma_{x,n} can be calculated exactly for a portfolio of nn obligors as the square root of the variance in Equation (10).

Capital Multiple based on Cornish-Fisher expansion

The Cornish-Fisher expansion gives the percentile xαx_\alpha at confidence α\alpha (J.D. Lamb, 2016) [6]:

xα=μ+σ{uα+(uα2−1)s+(uα3−3uα)k−(2uα3−5uα)s2}(14)x_\alpha= \mu+ \sigma\{u_\alpha+ (u_\alpha^2 - 1)s + (u_\alpha^3 - 3u_\alpha)k - (2u_\alpha^3 - 5u_\alpha)s^2\} \tag*{(14)}

in which uα=Φ−1[α]u_\alpha= \Phi^{-1}[\alpha], μ=E[x]\mu= \mathbb{E}[x] is the expectation value and σ\sigma the standard deviation in xx. Comparing (14) and (13) gives a capital multiple,

cmn,CF=uα+(uα2−1)s+(uα3−3uα)k−(2uα3−5uα)s2(15)cm_{n,CF} = u_\alpha+ (u_\alpha^2 - 1)s + (u_\alpha^3 - 3u_\alpha)k - (2u_\alpha^3 - 5u_\alpha)s^2 \tag*{(15)}

Parameters ss and kk may be related to the skewness SxS_x and excess kurtosis KxK_x of xx by s≈Sx6s \approx\frac{S_x}{6} and k≈Kx24k \approx\frac{K_x}{24} respectively. However, (J.D. Lamb, 2016) and [6] have shown that these approximations only hold for small skewness and excess kurtosis. Generally, the skewness SxS_x and excess kurtosis KxK_x are also functions of parameters ss and kk. (J.D. Lamb, 2016) introduces a Newton-Raphson algorithm to retrieve parameters ss and kk from SxS_x and KxK_x. Since both skewness and excess kurtosis are ratios of central moments, a simpler approach is to use the central moments to find ss and kk. Theorem 2 gives central moments (μ3)x,n(\mu_3)_{x,n} and (μ4)x,n(\mu_4)_{x,n} and Equation (14) gives central moments μ3,CF(s,k)\mu_{3,CF}(s,k) and μ4,CF(s,k)\mu_{4,CF}(s,k), based on Cornish-Fisher expansion:

μm,CF(s,k)=E[(x−μ)m]=σmE[(z+(z2−1)s+(z3−3z)k−(2z3−5z)s2)m](16)\begin{aligned} \mu_{m,CF}(s,k) &= \mathbb{E}[(x-\mu)^m] \\ &= \sigma^m\mathbb{E}[(z+(z^2-1)s+(z^3-3z)k-(2z^3-5z)s^2)^m] \tag*{(16)} \end{aligned}

with zz as a standard normal distributed variable. The calculation of μ3,CF(s,k)\mu_{3,CF}(s,k) and μ4,CF(s,k)\mu_{4,CF}(s,k) using Equation (16) is cumbersome, but (J.D. Lamb, 2016) provides closed-form equations for these moments. Parameters ss and kk result from minimizing the following:

E={μ3,CF(s,k)−(μ3)x,n}2+{μ4,CF(s,k)−(μ4)x,n}2,(17)E=\left\{\mu_{3,CF}(s,k)-(\mu_{3})_{x,n}\right\}^{2}+\left\{\mu_{4,CF}(s,k)-(\mu_{4})_{x,n}\right\}^{2}, \tag*{(17)}

Furthermore, Equation (14) is only valid for limited values of ss and kk, since xαx_{\alpha} should be a monotonously increasing function of uαu_{\alpha} [6]. This requirement gives a specific domain RR in which parameters ss and kk are valid (J.D. Lamb, 2016):

R={(s,k)|0<s2<3−22, 1+11s2+s4−6s2+16<k<1+11s2+s4−6s2+16}(18)R=\left\{(s,k)\middle|0<s^{2}<3-2\sqrt{2},\ \frac{1+11s^{2}+\sqrt{s^{4}-6s^{2}+1}}{6}<k< \frac{1+11s^{2}+\sqrt{s^{4}-6s^{2}+1}}{6}\right\} \tag*{(18)}

The requirement (s,k)∈R(s,k)\in R guarantees that ∂xα∂uα>0\frac{\partial x_{\alpha}}{\partial u_{\alpha}}>0, and therefore that ∂cmn,CF∂α>0\frac{\partial c_{m n,CF}}{\partial\alpha}>0 by Equation (15). This imposes restrictions on applying the Cornish-Fisher approach.

Capital Multiple based on Weighting

This section provides an alternative to the Cornish-Fisher expansion to calculate percentiles xα,nx_{\alpha,n} for a portfolio of nn obligors. Theorem 2 shows that the central moments are approximately a 1n\frac{1}{n}-weighted average of the mthm^{\text{th}} central moment (μm)x,1(\mu_m)_{x,1} for a one-obligor portfolio and the mthm^{\text{th}} central moment (μm)x,∞(\mu_m)_{x,\infty} in the Vasicek limit. High-order terms consist of Vasicek moments Φl[τ…τ;ρ]\Phi_l[\tau\ldots\tau;\rho], which are small for l>1l>1 as shown by Table 1, and a pre-factor 1n(1−1n)\frac{1}{n}\left(1-\frac{1}{n}\right) for m>2m>2. This pre-factor is smaller than ¼ for n∈[1,∞)n\in[1,\infty). Assuming that the capital multiple is linear in the central moments, like Equation (14) with ignoring quadratic terms, we may approximate the capital multiple as a weighted average of a single-obligor capital multiple cm1cm_1 and the capital multiple cm∞cm_{\infty} in the Vasicek limit:

P0=2.00%P_0 = 2.00\% ρ=0.164\rho= 0.164P0=1.00%P_0 = 1.00\% ρ=0.193\rho= 0.193P0=0.10%P_0 = 0.10\% ρ=0.234\rho= 0.234P0=0.05%P_0 = 0.05\% ρ=0.237\rho= 0.237
Φ1[τ;ρ]\Phi_1[\tau;\rho]2.00⋅10−22.00\cdot10^{-2}1.00⋅10−21.00\cdot10^{-2}1.00⋅10−31.00\cdot10^{-3}5.00⋅10−45.00\cdot10^{-4}
Φ2[τ;ρ]\Phi_2[\tau;\rho]9.36⋅10−49.36\cdot10^{-4}3.26⋅10−43.26\cdot10^{-4}9.07⋅10−69.07\cdot10^{-6}2.98⋅10−62.98\cdot10^{-6}
Φ3[τ;ρ]\Phi_3[\tau;\rho]7.72⋅10−57.72\cdot10^{-5}2.21⋅10−52.21\cdot10^{-5}2.78⋅10−72.78\cdot10^{-7}6.88⋅10−86.88\cdot10^{-8}
Φ4[τ;ρ]\Phi_4[\tau;\rho]9.53⋅10−69.53\cdot10^{-6}2.46⋅10−62.46\cdot10^{-6}1.80⋅10−81.80\cdot10^{-8}3.60⋅10−93.60\cdot10^{-9}
Φ5[τ;ρ]\Phi_5[\tau;\rho]1.58⋅10−61.58\cdot10^{-6}3.87⋅10−73.87\cdot10^{-7}1.94⋅10−91.94\cdot10^{-9}3.35⋅10−103.35\cdot10^{-10}

Table 1. Moments Φm[τ…τ;ρ]\Phi_m[\tau\ldots\tau;\rho] of the Vasicek distribution for several values of P0P_0 and ρ\rho.

cmn,EV≈1ncm1+(1−1n)cm∞=1ncm1+(1−1n)xα,∞−P0σx,∞(19)cm_{n,EV}\approx\frac{1}{n}cm_1+\left(1-\frac{1}{n}\right)cm_{\infty} =\frac{1}{n}cm_1+\left(1-\frac{1}{n}\right)\frac{x_{\alpha,\infty}-P_0}{\sigma_{x,\infty}} \tag*{(19)}

with σx,∞\sigma_{x,\infty} and xα,∞x_{\alpha,\infty} as calculated from Equation (7) and (8). Equation (19) requires a capital multiple cm1cm_1 for a one-obligor portfolio. The default probability density for a one-obligor portfolio is P0x(1−P01−x)P_0^x(1-P_0^{1-x}) with discrete values x=0x=0 or x=1x=1. This gives an expectation E[x]=P0\mathbb{E}[x]=P_0 and a standard deviation P0(1−P0)\sqrt{P_0(1-P_0)}. Since the confidence level is high, α=99.9%\alpha=99.9\%, the obligor is certainly in default with a likelihood of 0.1%, giving xα,1=1x_{\alpha,1}=1. As such, the capital multiple for a one-obligor portfolio equals:

cm1=xα,1−P0σx,1=1−P0P0(1−P0)=1−P0P0(20)cm_{1} = \frac{x_{\alpha,1} - P_{0}}{\sigma_{x,1}} = \frac{1 - P_{0}}{\sqrt{P_{0}(1-P_{0})}} = \sqrt{\frac{1-P_{0}}{P_{0}}} \tag*{(20)}

Combining Equation (10), (13) and (19) gives the following equation for the percentile:

xα,n=P0+{1n1−P0P0+(1−1n)(xα,∞−P0σx,∞)}⋅P0(1−P0)n+(1−1n)σx,∞2(21)x_{\alpha,n} = P_{0} + \left\{\frac{1}{n}\sqrt{\frac{1-P_{0}}{P_{0}}} + \left(1-\frac{1}{n}\right)\left(\frac{x_{\alpha,\infty}-P_{0}}{\sigma_{x,\infty}}\right)\right\}\cdot\sqrt{\frac{P_{0}(1-P_{0})}{n}+\left(1-\frac{1}{n}\right)\sigma_{x,\infty}^{2}} \tag*{(21)}

with xα,∞=Φ[Φ−1[P0]+ρΦ−1[α]1−ρ]x_{\alpha,\infty} = \Phi\left[\frac{\Phi^{-1}[P_{0}]+\sqrt{\rho}\Phi^{-1}[\alpha]}{\sqrt{1-\rho}}\right] and σx,∞2=Φ2[τ,τ,ρ]−P02\sigma_{x,\infty}^{2} = \Phi_{2}[\tau,\tau,\rho]-P_{0}^{2} as in Equation (7). However, xα,n=0x_{\alpha,n}=0 when α\alpha is smaller than (1−P0)n(1-P_{0})^{n}. This occurs when nn is smaller than ln⁡(α)ln⁡(1−P0)\frac{\ln(\alpha)}{\ln(1-P_{0})}. Since α=0.995\alpha=0.995 in Basel III/IV, ln⁡(α)ln⁡(1−P0)≈0\frac{\ln(\alpha)}{\ln(1-P_{0})}\approx0 for most values of P0P_{0}. We will observe xα,n=0x_{\alpha,n}=0 for small nn only at low confidence levels like α=0.9\alpha=0.9.

We will test Equation (21) by comparing with simulations and refer to the underlying distribution FEV(x)F_{EV}(x) with FEV(xα,n)=αF_{EV}(x_{\alpha,n})=\alpha as the Extended Vasicek distribution (EV). We verify Equation (21) by simulation in § 4.

Testing the Vasicek Distribution and Extended Vasicek Distribution

Simulation set-up

We compared the Vasicek distribution and the Extended Vasicek distribution with the percentiles, based on 500,000 simulations. The objective is to see how close the percentiles of the Extended Vasicek distribution reproduce the percentiles of the simulated distribution.

We performed analyses for portfolios consisting of 2 to 2000 obligors. The correlation between obligors depends from the unconditional probability of default (P0P_{0}), according to BCBS [2] Ong (2005):

ρ=0.12(1−e−50⋅P01−e−50)+0.24(1−1−e−50⋅P01−e−50)(22)\rho= 0.12\left(\frac{1-e^{-50\cdot P_{0}}}{1-e^{-50}}\right)+0.24\left(1-\frac{1-e^{-50\cdot P_{0}}}{1-e^{-50}}\right) \tag*{(22)}

Equation (22) assumes that the average correlation decreases upon increasing P0P_{0}, lowering the effect of the single systematic risk factor on the obligor (N. Milonas, 2024). Our test is based on P0P_{0}, ranging from 0.0005 to 0.02, corresponding to correlations ranging from 0.24 to 0.16 respectively. The simulations are performed in RStudio v2024.12.0, using the following steps in each simulation run:

  1. A standard normal random number ziz_i is drawn for each obligor ii.

  1. The random numbers ziz_i are correlated via Cholesky decomposition of the correlation matrix as given by Equation (2).

  1. An obligor defaults when zi≤Φ−1[P0]z_i \le\Phi^{-1}[P_0], with P0P_0 as the unconditional probability of default of obligor i.

  1. The total defaults d and fraction of defaults x=dnx = \frac{d}{n} is calculated.

The important simulation outcome is the 99.9% percentile (x99.9%simx^{\mathrm{sim}}_{99.9\%}), based on the Harrell-Davis estimator [7]. Furthermore, we calculated the variance, 3rd and 4th central moments of the simulations and compared with the moments as calculated by Theorem 2. We present the results as a function of the logarithmic nn, with nn denoting the number of obligors.

Comparing the Extended Vasicek distribution with simulations

First, we compare the standard deviation σx,sim\sigma_{x,\mathrm{sim}} of the simulations with σx,∞\sigma_{x,\infty} of the Vasicek distribution as calculated by Equation (7) and with σx,n\sigma_{x,n} of the finite portfolio distribution as calculated by Equation (10). Figure 1 shows that all standard deviations converge to the same value for large nn, and Equation (10) gives the same values σx,n\sigma_{x,n} as σx,sim\sigma_{x,\mathrm{sim}} resulting from the simulations.

Eight-panel plots of standard deviation against number of obligors for different PD and ρ values

Figure 1. The standard deviation in xx (σx\sigma_x) of the Vasicek and finite portfolio distributions (nn obligors), compared with σx\sigma_x of simulations, as a function of the number of obligors in the credit portfolio.

Theorem 2 gives theoretical values for the third central moment (μ3)x,n(\mu_3)_{x,n} and fourth central moment (μ4)x,n(\mu_4)_{x,n}. Figure 2 and Figure 3 compare these theoretical central moments with the central moments based on the simulations, revealing that theoretical moments (μ3)x,n(\mu_3)_{x,n} and (μ4)x,n(\mu_4)_{x,n} agree with the corresponding simulation outcomes.

Eight-panel plots of the third central moment versus number of obligors for different probability of default values

Figure 2. The 3rd3^{\mathrm{rd}} central moment in xx of the finite portfolio distributions, compared with the 3rd3^{\mathrm{rd}} central moment of simulations (cmoment3), as a function of the number of obligors (nn) in the credit portfolio.

The 4rd central moment in x of the finite portfolio distributions compared with the 4rd central moment of simulations

Figure 3. The 4rd central moment in xx of the finite portfolio distributions, compared with the 4rd central moment of simulations (cmoment4), as a function of the number of obligors (nn) in the credit portfolio.

Capital Multiple based on the Cornish-Fisher expansion

We compared the Cornish-Fisher capital multiple cmn,CFcm_{n,CF} with the capital multiple cmn,simcm_{n,sim} based on simulations: cmn,simcm_{n,sim} follows from x99.9%simσx,n\frac{x_{99.9\%}^{sim}}{\sigma_{x,n}} and cmn,CFcm_{n,CF} is stepwise calculated:

  1. Theorem 2 gives the central moments (μ3)x,n(\mu_3)_{x,n} and (μ4)x,n(\mu_4)_{x,n}.

  1. Minimizing EE in Equation (17) gives ss and kk: a search on an (s,k)(s,k)-grid for minimum EE in the admissible domain RR in Equation (18) leads to large deviations between cmn,CFcm_{n,CF} and cmn,simcm_{n,sim}. Minimizing EE by the BFGS algorithm and allowing (s,k)(s,k) outside area RR leads to cmn,CF≈cmn,simcm_{n,CF} \approx cm_{n,sim} as shown in Figure 5.

Eight-panel plots of capital multiple versus number of obligors for different probability of default values

Figure 5. The capital multiple, based on simulations (green), on the Extended Vasicek distribution for finite portfolio (cmn,EV\mathrm{cm}_{n,\mathrm{EV}}, blue) and on the Cornish-Fisher expansion (cmn,CF\mathrm{cm}_{n,\mathrm{CF}}, orange) as a function of the number of obligors (nn) in the credit portfolio.

  1. Use ss and kk in Equation (15) to calculate cmn,CFcm_{n,CF}.

Since (s,k)(s,k) could be outside domain RR, Equation (15) will not always have a positive derivative, i.e. ∂cmn,CF∂α\frac{\partial cm_{n,CF}}{\partial\alpha} may be negative: cmn,CFcm_{n,CF} may not monotonously increase with increasing confidence level α\alpha. For example, minimizing Equation (17) gives (s=1.42,k=4.39)∉R(s = 1.42, k = 4.39) \notin R, leading to a minimum cmn,CFcm_{n,CF} at α=73.8%\alpha= 73.8\%. More general, Figure 4 shows cmn,CFcm_{n,CF} for (s,k)(s,k) values that are determined from minimizing Equation (17). The figure demonstrates that cmn,CFcm_{n,CF} only increases with increasing α\alpha for confidence levels that are larger than 90%. This works for risk capitalization at high confidence levels, but care should be taken when the Cornish-Fisher approach is applied for confidence levels lower than 90%.

Capital multiple based on Cornish-Fisher expansion as a function of confidence level

Figure 4. The capital multiple cmn,CFcm_{n,CF}, based on Cornish-Fisher expansion, as a function of confidence level α\alpha.

Figure 5 also shows that cmn,CFcm_{n,CF} deviates from cmn,simcm_{n,sim} for a small PD∼10−4PD \sim10^{-4}. Based on this observation, and that (s,k)∉R(s,k) \notin R for cmn,CF≈cmn,simcm_{n,CF} \approx cm_{n,sim}, we conclude that Cornish-Fisher expansion has limited applicability in calculating the capital multiple for RWA or risk capital calculations.

Capital Multiple based on weighting

The second method to calculate the capital multiple uses the approximation in Equation (19). Figure 5 also compares the capital multiple cmn,simcm_{n,sim} with the cmn,EVcm_{n,EV} as calculated theoretically by Equation (19). A first observation is the choppy behaviour of cmn,simcm_{n,sim} for small nn, caused by discreteness of the loss distribution as nn becomes small. The weighting of capital multiples in (19) smooths out these chops. Furthermore, both cmn,simcm_{n,sim} and cmn,EVcm_{n,EV} also depend on P0P_0:

  1. When P0<2%P_0 < 2\%, both cmn,simcm_{n,sim} and cmn,EVcm_{n,EV} decrease with increasing nn and finally converge to cm∞cm_{\infty} in (19). In this region, cm1cm_1 is larger than cm∞=xα,∞−P0σx,∞cm_{\infty} = \frac{x_{\alpha,\infty}-P_0}{\sigma_{x,\infty}}.

  1. A turning point exists for P0∼2%P_0 \sim2\%: cmn,EVcm_{n,EV} becomes flat as a function of nn since cm1=(1−P0)/P0cm_1 = \sqrt{(1-P_0)/P_0} approaches cm∞=xα,∞−P0σx,∞cm_{\infty} = \frac{x_{\alpha,\infty}-P_0}{\sigma_{x,\infty}}. The difference between cmn,EVcm_{n,EV} and cmn,simcm_{n,sim} relate to 1n(1−1n)\frac{1}{n}\left(1-\frac{1}{n}\right) and the fourth central moment. When cm1≈cm∞cm_1 \approx cm_{\infty}, high-order terms in (11) become more relevant than for cm1>cm∞cm_1 > cm_{\infty} or cm1<cm∞cm_1 < cm_{\infty}. However, this high-order effect disappears at large nn where cmn,1cm_{n,1} and cmn,simcm_{n,sim} become equal.

  1. When P0>2%P_0 > 2\%, both cmn,simcm_{n,sim} and cmn,EVcm_{n,EV} increase with increasing nn, converging to cm∞cm_{\infty} for high nn values. In this region, cm1cm_1 is smaller than cm∞=xα,∞−P0σx,∞cm_{\infty} = \frac{x_{\alpha,\infty}-P_0}{\sigma_{x,\infty}}.

To test (19), we regressed cmn,simcm_{n,sim} against 1n\frac{1}{n} and derived cm1cm_1 and cm∞cm_{\infty} from the regression coefficients. Table 4 compares the theoretical cm1cm_1 and cm∞cm_{\infty} from (19) with corresponding cm1,regcm_{1,reg} and cm∞,regcm_{\infty,reg}, resulting from regression. The table shows that cm∞≈cm∞,regcm_{\infty} \approx cm_{\infty,reg}, whereas cm1cm_1 and cm1,regcm_{1,reg} show substantial differences for P0<2%P_0 < 2\%. The slope (cm1,reg−cm∞,reg)(cm_{1,reg}-cm_{\infty,reg}) of the regression is insignificant in most cases. We conclude from Figure 5 and Table 4 that cm1cm_1 is hard to estimate from regression, due choppy behaviour of cmn,simcm_{n,sim} for small n. The 1n\frac{1}{n}-weighting smooths out this choppy behaviour in approximating the capital multiple.

Theoretical, Equation (19)Regression
PDcm1cm_1cm∞cm_{\infty}cm1,regcm_{1,reg}cm∞,regcm_{\infty,reg}
0.05%44.712.136.610.7
0.20%22.311.117.19.6
0.50%14.19.910.08.8
1.00%9.98.75.88.2
2.00%7.07.43.57.0
5.00%4.45.86.15.6
10.00%3.04.93.24.8
20.00%2.04.02.04.0

Table 4. Comparison of cmncm_n, based on Equation (19) and based on linear regression of simulated cmncm_n, versus 1/n1/n.

Figure 6 compares the percentile x99.9%simx^{\mathrm{sim}}_{99.9\%} of the simulations with the following:

  • The 99.9% percentile x99.9%EVx^{\mathrm{EV}}_{99.9\%} of the Extended Vasicek distribution, based on Equation (21).

  • The 99.9% percentile x99.9Vx_{99.9}^{V} of the original Vasicek distribution, based on Equation (8).

  • The 99.9% percentile x99.9CFx_{99.9}^{CF}, using the Cornish-Fisher expansion in Equation (15) as capital multiple.

The figure shows that the percentiles x99.9%EVx_{99.9\%}^{EV} of the Extended Vasicek distribution are considerably closer to the x99.9%simx_{99.9\%}^{sim} than the percentiles x99.9%Vx_{99.9\%}^{V} of the original Vasicek distribution for a small number of obligors. All percentiles converge to the same limit for large nn, as expected. Overall, the percentiles based on the Extended Vasicek distribution in Equation (21) agree fairly well with the simulations. We conclude that the Extended Vasicek distribution is a good approximation for credit portfolios with a finite number of obligors.

Figure 6 also compares the percentiles x99.9%simx_{99.9\%}^{sim} with x99.9%CFx_{99.9\%}^{CF}, based on the Cornish-Fisher expansion. The percentiles x99.9%CFx_{99.9\%}^{CF} align with x99.9%simx_{99.9\%}^{sim} of the simulation, except for small PD<2⋅10−3PD < 2 \cdot10^{-3}. Furthermore, the parameters (s,k)(s,k) of the Cornish-Fisher expansion are outside the admissible domain RR.

The 99.9% percentiles, based on simulations compared with those resulting from the Vasicek distribution, the Extended Vasicek distribution and the Cornish-Fisher expansion, as a function of the number of obligors in the credit portfolio

Figure 6. The 99.9% percentiles, based on simulations (green), compared with those resulting from the Vasicek (red) distribution, the Extended Vasicek distribution (blue) and the Cornish-Fisher expansion (orange), as a function of the number of obligors (n) in the credit portfolio.

Comparing RWAs of Vasicek and Extended Vasicek Distributions with simulations

Equation (21) may serve as a “RWA” for a finite credit portfolio for α=99.9%\alpha= 99.9\%:

RWAα,n=12.5{[1n1−P0P0+(1−1n)(xα,∞−P0)]P0(1−P0)n⋅σx,∞2+1−1n}⋅LGD⋅EAD(23)RWA_{\alpha,n} = 12.5 \left\{\left[\frac{1}{n}\sqrt{\frac{1-P_0}{P_0}}+\left(1-\frac{1}{n}\right)(x_{\alpha,\infty}-P_0)\right]\sqrt{\frac{P_0(1-P_0)}{n\cdot\sigma_{x,\infty}^{2}}+1-\frac{1}{n}}\right\}\cdot LGD\cdot EAD \tag*{(23)}

with xα,∞=Φ[Φ−1[P0]+ρΦ−1[α]1−ρ]x_{\alpha,\infty}=\Phi\left[\frac{\Phi^{-1}[P_0]+\sqrt{\rho}\Phi^{-1}[\alpha]}{\sqrt{1-\rho}}\right] and σx,∞2=Φ2[τ,τ,ρ]−P02\sigma_{x,\infty}^{2}=\Phi_2[\tau,\tau,\rho]-P_0^2. This RWA factor converges to the current Basel RWA factor in the Vasicek limit n→∞n\to\infty: RWAα,n=12.5⋅{xα,∞−P0}⋅LGD⋅EADRWA_{\alpha,n}=12.5\cdot\{x_{\alpha,\infty}-P_0\}\cdot LGD\cdot EAD. This section compares the RWA in Equation (23) with the Basel III/IV RWA calculation.

Figure 7 shows the RWA, based on Equation (23) and the Basel III/IV RWA, floored with 72.5% of the RWA standardised approach, as a function of the number of obligors (nn). We assumed an LGD of 0.4 and an EAD of 100. We consider two cases:

Basel III/IV RWA based on the Extended Vasicek distribution for A- and BB-rated obligors

Figure 7. Basel III/IV RWA, based on the Extended Vasicek distribution, with an LGD of 0.4, and EAD of 100 and an A and BB rated obligor.

  • An obligor with an S&P rating A, corresponding to a PD of 0.04% and a risk weight of 50% under Basel III Standardised Approach. This gives floor of 72.5% x 50% x 100 = 36.25.

  • An obligor with an S&P rating BB, corresponding to a PD of 0.64% and a risk weight of 50% under Basel III Standardised Approach. This gives floor of 72.5% x 100% x 100 = 72.50.

PDs are based on the latest observed default rates of S&P (N. Serino, 2025). The RWA based on Equation (23) converges to the Basel III/IV RWA, based on the Vasicek distribution, when nn increases beyond 400. For small nn, Equation (23) leads to a higher RWA than the Basel III/IV RWA, but the proposed 72.5% floor makes the Basel III/IV prudent, unless the portfolio has less than 20 obligors in case a) or less than 50 BB-rated obligors in case b).

We performed a second study for a credit portfolio, based on 4 ratings I∈{A,B,C,D}I \in\{A, B, C, D\}, with A representing low credit risk and D representing the highest credit risk. Obligors within the same rating category have the same PD as shown in Table 5. All obligors have the same LGD and EAD. The correlations between different rating grades result in a positive definite correlation matrix.

RatingPDEADLGDABCD
A0.0011000.410.250.150.10
B0.0021000.40.2510.170.19
C0.0151000.40.150.1710.22
D0.0201000.40.100.190.2201

Table 5. Credit portfolio with 4 rating grades, their PD, EAD (in € million) and LGD and their correlations.

We calculated RWAs for different portfolios Π1…Π6\Pi_1 \ldots\Pi_6, as presented in Table 6. The first four rows present the number of obligors per rating category. Portfolio Π1\Pi_1 has the same number of obligors in each grade, but Portfolio Π6\Pi_6 is concentrated in rating A. The inhomogeneity is reflected by the Herfindahl–Hirschman index HHI=∑I∈{A,B,C,D}fI2HHI = \sum_{I \in\{A,B,C,D\}} f_I^2 with fIf_I as the fraction of obligors in rating category II. Row nTotaln_{Total} gives the total number of obligors in the portfolio. The last three rows in Table 6 compare different RWAs for α=99.9%\alpha= 99.9\%:

Rating:Π1\Pi_1Π2\Pi_2Π3\Pi_3Π4\Pi_4Π5\Pi_5Π6\Pi_6
A (low risk)1065100200260
B106581010
C102158826
D (high risk)10725844
nTotaln_{Total}404040124222300
HHI0.250.350.440.660.820.76
RWAα,BIII/IVRWA_{\alpha,BIII/IV}205224672728316845416918
RWAα,simRWA_{\alpha,sim}281132663703363247926638
RWAα,EVRWA_{\alpha,EV}277932683649346447456334

Table 6. Credit portfolios (Πi\Pi_i) with 4 rating grades and different rating distributions, reflected in the HHI.

  • RWAα,BIII∣IVRWA_{\alpha,BIII|IV} is the sum of the Basel III/IV RWA per obligor. As such, this aggregation assumes perfect correlation between different rating categories. The standardized floor is not applied here.

  • RWAα,simRWA_{\alpha,sim} is calculated from 500,000 simulations as 12.5⋅nTotal(xα,sim−⟨x⟩sim)⋅LGD⋅EAD12.5 \cdot n_{Total}(x_{\alpha,sim}-\langle x\rangle_{sim}) \cdot LGD \cdot EAD with ⟨x⟩sim\langle x\rangle_{sim} as the averaged xx and xα,simx_{\alpha,sim} as the percentile at α=99.9%\alpha=99.9\% resulting from the simulations.

  • An analytical RWAα,EVRWA_{\alpha,EV} results from aggregating RWAα,nI,IRWA_{\alpha,nI,I} in rating grade II, which is based on the Extended Vasicek distribution in Equation (23). Since this aggregation requires a covariance σI,J2\sigma_{I,J}^{2} between the different rating categories II and JJ, we use the following identity to convert the correlations ρI,J\rho_{I,J} in Table 5 into default correlations ρI,Jdef\rho_{I,J}^{def} [4]:

σI,J2=Φ2[Φ−1[PI],Φ−1[PJ];ρI,J]−PIPJ=ρI,JdefPI(1−PI)PJ(1−PJ)=ρI,JdefσIσJ(24)\sigma_{I,J}^{2}=\Phi_{2}\left[\Phi^{-1}[P_I],\Phi^{-1}[P_J];\rho_{I,J}\right]-P_IP_J=\rho_{I,J}^{def}\sqrt{P_I(1-P_I)P_J(1-P_J)} \\ =\rho_{I,J}^{def}\sigma_I\sigma_J \tag*{(24)}

After multiplying each σI\sigma_I with the capital multiple, EAD and LGD, this gives RWAα,EV=nTotal∑I∈{A,B,C,D}∑J∈{A,B,C,D}ρI,JdefRIRJRWA_{\alpha,EV}=n_{Total}\sqrt{\sum_{I\in\{A,B,C,D\}}\sum_{J\in\{A,B,C,D\}}\rho_{I,J}^{def}R_IR_J} with RI=fI⋅RWAα,nI,IR_I=f_I\cdot RWA_{\alpha,nI,I}.

Table 6 shows that RWAα,EV≈RWAα,simRWA_{\alpha,EV}\approx RWA_{\alpha,sim}. When nTotaln_{Total} is small, RWAα,BIII∣IVRWA_{\alpha,BIII|IV} is smaller than RWAα,simRWA_{\alpha,sim}. This aligns with expectations, as RWAα,BIII∣IVRWA_{\alpha,BIII|IV} is based on the Vasicek distribution, assuming large portfolios. When nTotaln_{Total} increases to 200, the Vasicek distribution becomes valid and RWAα,BIII∣IVRWA_{\alpha,BIII|IV} approaches RWAα,simRWA_{\alpha,sim}. Portfolio Π6\Pi_6 shows that RWAα,BIII∣IVRWA_{\alpha,BIII|IV} exceeds both RWAα,simRWA_{\alpha,sim} and RWAα,EVRWA_{\alpha,EV}, since RWAα,BIII∣IVRWA_{\alpha,BIII|IV} is based on adding RWAs per rating grade and does not include diversification between rating categories as reflected in Table 5.

The effect of concentration is further explored in Figure 8. The figure compares RWAα,simRWA_{\alpha,sim} with RWAα,BIII/IVRWA_{\alpha,\mathrm{BIII/IV}} and RWAα,EVRWA_{\alpha,\mathrm{EV}} for portfolios with different HHI and the same nTotal=40n_{Total}=40. A high HHI indicates that obligors are concentrated in a rating category. RWAα,simRWA_{\alpha,sim} and RWAα,EVRWA_{\alpha,\mathrm{EV}} are similar, but RWAα,BIII/IVRWA_{\alpha,\mathrm{BIII/IV}} is lower than RWAα,simRWA_{\alpha,sim} since nTotaln_{Total} is small and the Vasicek limit does not hold. Nonetheless, all three RWAs increase with HHI, meaning that they account for concentration in a specific rating grade.

RWA as a function of the Herfindahl-Hirschman index

Figure 8. RWA, resulting from simulations, Basel III/IV RWA and the RWA based on Equation (21) as a function of the HHI, assuming a portfolio of 40 obligors, distributed over four different rating grades.

Conclusion

We introduced Theorem 2 to calculate central moments of a loss distribution of a finite credit portfolio. These central moments may be used to calculate percentiles of the loss distribution via scaling of the standard deviation with a capital multiple. We proposed two methods to achieve this: either via the Cornish-Fisher expansion or a 1n\frac{1}{n}-weighting of capital multiple and variance. We refer to the second method, giving Equation (21) for percentiles, as the Extended Vasicek distribution.

Comparison with percentiles of 500,000 simulations show that both methods work in general, but they both have drawbacks in calculating percentiles. The Cornish-Fisher method gives too low percentiles and capital multiples for small PDs. Another drawback is that the (s,k)(s,k)-parameterization of the Cornish-Fisher expansion is outside the admissible region. As such, a capital multiple that increases monotonously with increasing confidence level is not guaranteed.

The second method gives percentiles that better align with the simulations than the Cornish-Fisher method. A drawback is the assumption that the capital multiple is linear in the central moments, so that this scaling is a 1n\frac{1}{n} – weighting of the capital multiple of a single obligor portfolio and infinite portfolio. Nonetheless, (21) leads to a close approximation of the 99.9% percentiles that result from the simulations.

We deem that our analyses lead to two important conclusions. First, comparing the Basel III/IV RWA calculation with our alternative RWA, based on the Extended Vasicek distribution, shows that the Basel III/IV calculation includes enough prudency for finite portfolios, unless the portfolios are unrealistically small (n≤20n \le20). When n>100n > 100, the Extended Vasicek distribution approaches the original Vasicek distribution, valid for n→∞n \to\infty. Then the original Vasicek distribution is more convenient to apply than the Extended Vasicek distribution.

A second conclusion is that the percentiles, based on the Extended Vasicek distribution, could be used as control variate in Monte-Carlo simulations of credit portfolios. The control variate may substantially reduce variance in the Monte-Carlo calculations of percentiles (V.Tchistiakov, 2004). Simulations show that the Extended Vasicek distribution can also be applied for a portfolio with different rating grades, even when the number of obligors per rating grade is small.

References

Appendix A: Calculation of integrals over the Cumulative Gaussian Distribution

This appendix evaluates the integrals over cumulative Gaussian functions. We start with the following integral:

∫01Φ[a+by] dΦ(y)=∫−∞∞∫−∞a12πexp⁡{−y22−(x−by)22} dxdy=∫−∞∞∫−∞∞12πexp⁡{−12[(1+b2)(y−bx1+b2)2]−12(x21+b2)} dxdy(25)\begin{aligned} \int_{0}^{1} \Phi[a+by]\,\mathrm{d}\Phi(y) &= \int_{-\infty}^{\infty}\int_{-\infty}^{a} \frac{1}{2\pi}\exp\left\{-\frac{y^{2}}{2}-\frac{(x-by)^{2}}{2}\right\}\,\mathrm{d}x\mathrm{d}y \\ &= \int_{-\infty}^{\infty}\int_{-\infty}^{\infty} \frac{1}{2\pi}\exp\left\{-\frac{1}{2}\left[(1+b^{2})\left(y-\frac{bx}{1+b^{2}}\right)^{2}\right]-\frac{1}{2}\left(\frac{x^{2}}{1+b^{2}}\right)\right\}\,\mathrm{d}x\mathrm{d}y \tag*{(25)} \end{aligned}

in which we used completing the square in the exponential in the second step. Integration over yy results in:

∫01Φ[a+by] dΦ(y)=∫−∞aexp⁡{−12(x21+b2)}2π(1+b2) dx=Φ[a1+b2](26)\int_{0}^{1} \Phi[a+by]\,\mathrm{d}\Phi(y) = \int_{-\infty}^{a} \frac{\exp\left\{-\frac{1}{2}\left(\frac{x^{2}}{1+b^{2}}\right)\right\}}{\sqrt{2\pi(1+b^{2})}}\,\mathrm{d}x = \Phi\left[\frac{a}{\sqrt{1+b^{2}}}\right] \tag*{(26)}

The same approach may be applied in integration over Φ(  )\Phi(\ \ ) with power mm:

∫01(Φ[a+by])m dΦ(y)=∫−∞∞∫−∞a⋯∫−∞a⏟m times1(2π)(m+1)/2exp⁡{−12(y2+∑i=1m(xi−by)2)}∏i=1mdxi dy(27)\int_{0}^{1}\left(\Phi[a+by]\right)^{m}\,\mathrm{d}\Phi(y) = \int_{-\infty}^{\infty}\underbrace{\int_{-\infty}^{a}\cdots\int_{-\infty}^{a}}_{m\ \text{times}} \frac{1}{(2\pi)^{(m+1)/2}} \exp\left\{-\frac{1}{2}\left(y^{2}+\sum_{i=1}^{m}(x_i-by)^{2}\right)\right\} \prod_{i=1}^{m}\mathrm{d}x_i\,\mathrm{d}y \tag*{(27)}

Completing the square in the exponential gives:

∫01(Φ[a+by])m dΦ(y)=∫−∞∞∫−∞a⋯∫−∞a⏟m times1(2π)(m+1)/2exp⁡{−12∑i=1m(1+mb2)(y−b∑i=1mxi1+mb2)2+12(b∑i=1mxi1+mb2)2−∑i=1mxi2} dx1⋯dxm dy(28)\begin{aligned} \int_{0}^{1}\left(\Phi[a+by]\right)^{m}\,\mathrm{d}\Phi(y) ={}&\int_{-\infty}^{\infty}\underbrace{\int_{-\infty}^{a}\cdots\int_{-\infty}^{a}}_{m\ \text{times}} \frac{1}{(2\pi)^{(m+1)/2}} \exp\left\{-\frac{1}{2}\sum_{i=1}^{m}(1+mb^{2})\left(y-\frac{b\sum_{i=1}^{m}x_i}{1+mb^{2}}\right)^{2}\right.\\ &\left.\qquad+\frac{1}{2}\left(\frac{b\sum_{i=1}^{m}x_i}{1+mb^{2}}\right)^{2}-\sum_{i=1}^{m}x_i^{2}\right\} \,\mathrm{d}x_1\cdots\mathrm{d}x_m\,\mathrm{d}y \tag*{(28)} \end{aligned}

Integration over yy and collecting terms with xi2x_i^{2} and with xixjx_i x_j gives:

∫01(Φ[a+by])m dΦ(y)=(1+b2)m(2π)m(1+mb2) ∫−∞a/1+b2⋯∫−∞a/1+b2⏟m timesexp⁡{−12(∑i=1mhi,ixi2+2hi,j∑i=1m∑j>imxixj)} dx1⋯dxm=Φm[a1+b2,…,a1+b2;b21+b2](29)\begin{aligned} \int_{0}^{1}\left(\Phi[a+by]\right)^{m}\,\mathrm{d}\Phi(y) ={}&\sqrt{\frac{(1+b^{2})^{m}}{(2\pi)^{m}(1+mb^{2})}}\, \underbrace{\int_{-\infty}^{a/\sqrt{1+b^{2}}}\cdots\int_{-\infty}^{a/\sqrt{1+b^{2}}}}_{m\ \text{times}} \exp\left\{-\frac{1}{2}\left(\sum_{i=1}^{m}h_{i,i}x_i^{2} +2h_{i,j}\sum_{i=1}^{m}\sum_{j>i}^{m}x_i x_j\right)\right\} \,\mathrm{d}x_1\cdots\mathrm{d}x_m\\ ={}&\Phi_m\left[\frac{a}{\sqrt{1+b^{2}}},\ldots,\frac{a}{\sqrt{1+b^{2}}};\frac{b^{2}}{1+b^{2}}\right] \tag*{(29)} \end{aligned}

with hi,i=(1+b2)(1+(m−1)b2)1+mb2h_{i,i}=\frac{(1+b^{2})(1+(m-1)b^{2})}{1+mb^{2}} and hi,j=−b2(1+b2)1+mb2h_{i,j}=-\frac{b^{2}(1+b^{2})}{1+mb^{2}}. In the second step, we changed the integration interval from (−∞,a)(-\infty,a) into (−∞,a1+b2)\left(-\infty,\frac{a}{\sqrt{1+b^{2}}}\right).

Equation (29) results in Equation (5) when a=τ1−ρa = \frac{\tau}{\sqrt{1-\rho}} and b=ρ1−ρb = \frac{\sqrt{\rho}}{\sqrt{1-\rho}}:

  • The diagonal elements and off-diagonal elements become hi,i=1+(m−2)ρ(1−ρ)[1+(m−1)ρ]h_{i,i} = \frac{1+(m-2)\rho}{(1-\rho)[1+(m-1)\rho]} and hi,j=−ρ(1−ρ)[1+(m−1)ρ]h_{i,j} = \frac{-\rho}{(1-\rho)[1+(m-1)\rho]}, corresponding to the elements of the inverse correlation matrix by Cm−1C_m^{-1} as shown in Appendix C.

  • The normalization factor becomes (1+b2)m(2π)m(1+mb2)={[1+(m−1)ρ](1−ρ)m−1}−1/2\sqrt{\frac{(1+b^2)^m}{(2\pi)^m(1+mb^2)}} = \{[1+(m-1)\rho](1-\rho)^{m-1}\}^{-1/2}, corresponding to the inverse square root of the determinant ∣Cm∣\lvert C_m\rvert of the homogeneous correlation matrix, also shown in Appendix C.

Appendix B: Proof of Theorem 2

To prove Theorem 2, we start with a general equation for the central moments, with E[x]=P0\mathbb{E}[x] = P_0:

(μm)x,n=E[(x−P0)m]=cm+∑k=2mcm,kE[xk](30)(\mu_m)_{x,n} = \mathbb{E}[(x-P_0)^m] = c_m + \sum_{k=2}^{m} c_{m,k}\mathbb{E}[x^k] \tag*{(30)}

with cm=(−1)m(1−m)P0mc_m = (-1)^m(1-m)P_0^m and cm,k=(mk)(−P0)m−kc_{m,k} = \binom{m}{k}(-P_0)^{m-k}. The non-central moments E[xk]\mathbb{E}[x^k] in Equation (30) follow from E[dk]=∑l=0knlSk,lpl\mathbb{E}[d^k] = \sum_{l=0}^{k} n^l S_{k,l}p^l for a binomial distribution of dd defaults, with the following components:

  1. Sk,lS_{k,l} are Stirling numbers of the second kind, with Sk,0=0S_{k,0}=0, Sk,1=Sk,k=1S_{k,1}=S_{k,k}=1.

  2. The falling power is defined as nl=n(n−1)(…)(n−l+1)n^l = n(n-1)(\ldots)(n-l+1) with n0=1n^0=1 and n1=nn^1=n.

We derive central moments in terms of the fraction x=dnx = \frac{d}{n} of defaulted obligors. Since the probability p=p(f)p=p(f) is conditional on one systematic factor ff, we use Theorem 1 for p=p(f)p=p(f) and integrate over ff:

E[xk]=E[(dn)k]=E[dk]nk=∑l=0knlnkSk,l∫01p(f)l dΦ(f)=∑l=0knlnkSk,lΦl[τ…τ;ρ]=nnkΦ1[τ;ρ]+∑l=2knlnkSk,lΦl[τ…τ;ρ](31)\begin{aligned} \mathbb{E}[x^k] &= \mathbb{E}\left[\left(\frac{d}{n}\right)^k\right] = \frac{\mathbb{E}[d^k]}{n^k} = \sum_{l=0}^{k}\frac{n^l}{n^k}S_{k,l}\int_0^1p(f)^l\,d\Phi(f) = \sum_{l=0}^{k}\frac{n^l}{n^k}S_{k,l}\Phi_l[\tau\ldots\tau;\rho] \\ &= \frac{n}{n^k}\Phi_1[\tau;\rho] + \sum_{l=2}^{k}\frac{n^l}{n^k}S_{k,l}\Phi_l[\tau\ldots\tau;\rho] \tag*{(31)} \end{aligned}

For a one-obligor portfolio, n=1n=1 and the non-central moments equal E[xk]=Φ1[τ;ρ]=P0\mathbb{E}[x^k]=\Phi_1[\tau;\rho]=P_0. In the Vasicek limit, the non-central moments equal E[xk]=Φk[τ…τ;ρ]\mathbb{E}[x^k]=\Phi_k[\tau\ldots\tau;\rho]. For the intermediate case, 1<n<∞1<n<\infty, we use nl=n(n−1)⋅(n−2)l−2n^l=n(n-1)\cdot(n-2)^{l-2} and Equation (31) may be written as:

E[Xk]=1nk−1Φ1[τ;ρ]+(1−1n)∑l=2kbk,lΦl[τ…τ;ρ](32)\mathbb{E}[X^{k}] = \frac{1}{n^{k-1}}\Phi_{1}[\tau;\rho] + \left(1-\frac{1}{n}\right)\sum_{l=2}^{k} b_{k,l}\Phi_{l}[\tau\ldots\tau;\rho] \tag*{(32)}

with coefficients bk,l=Sk,l(n−2)l−2nk−2b_{k,l}=S_{k,l}\frac{(n-2)^{l-2}}{n^{k-2}}. Before continuing the proof, we give some properties of coefficients and bk,lb_{k,l} that will be used. Coefficients bk,kb_{k,k} equal one in the Vasicek limit:

lim⁡n→∞bk,k=lim⁡n→∞Sk,k(n−2)k−2nk−2=lim⁡n→∞(1−2n)(1−3n)…(1−k−1n)=1k≥3(33)\lim_{n\to\infty}b_{k,k}=\lim_{n\to\infty}S_{k,k}\frac{(n-2)^{k-2}}{n^{k-2}}=\lim_{n\to\infty}\left(1-\frac{2}{n}\right)\left(1-\frac{3}{n}\right)\ldots\left(1-\frac{k-1}{n}\right)=1\quad k\geq3 \tag*{(33)}

and specifically, b2,2=1b_{2,2}=1 for every n. Since ∑l=0kSk,lnl=nk\sum_{l=0}^{k}S_{k,l}n^{l}=n^{k} holds for Stirling numbers Sk,lS_{k,l}, the coefficients bk,lb_{k,l} add up as follows:

∑l=2kbk,l=∑l=2kSk,l(n−2)l−2nk−2=∑l=2kSk,ln(n−1)(n−2)l−2n(n−1)nk−2=∑l=2kSk,lnlnk−1(n−1)=∑l=0kSk,lnl−Sk,1n−Sk0nk−1(n−1)=nk−n−0nk−1(n−1)=n−n−(k−2)n−1(34)\begin{aligned} \sum_{l=2}^{k}b_{k,l}=\sum_{l=2}^{k}S_{k,l}\frac{(n-2)^{l-2}}{n^{k-2}}=\sum_{l=2}^{k}S_{k,l}\frac{n(n-1)(n-2)^{l-2}}{n(n-1)n^{k-2}}=\sum_{l=2}^{k}\frac{S_{k,l}n^{l}}{n^{k-1}(n-1)} \\ &=\frac{\sum_{l=0}^{k}S_{k,l}n^{l}-S_{k,1}n-S_{k0}}{n^{k-1}(n-1)}=\frac{n^{k}-n-0}{n^{k-1}(n-1)}=\frac{n-n^{-(k-2)}}{n-1} \tag*{(34)} \end{aligned}

As shown in Table 3, the coefficients bklb_{kl} are polynomials in 1n\frac{1}{n}. We substitute Equation (32) in Equation (30):

(μm)x,n=cm+∑k=2mcm,k{1nk−1Φ1[τ;ρ]+(1−1n)∑l=2kbk,lΦl[τ…τ;ρ]}(35)(\mu_m)_{x,n}=c_m+\sum_{k=2}^{m}c_{m,k}\left\{\frac{1}{n^{k-1}}\Phi_1[\tau;\rho]+\left(1-\frac{1}{n}\right)\sum_{l=2}^{k}b_{k,l}\Phi_l[\tau\ldots\tau;\rho]\right\} \tag*{(35)}

Equation (35) gives the following values for an one-obligor portfolio (n=1n=1) and the Vasicek limit n→∞n\to\infty:

(μm)x,1=lim⁡n→1(μm)x,n=cm+∑k=2mcm,kΦ1[τ;ρ](36)(\mu_m)_{x,1}=\lim_{n\to1}(\mu_m)_{x,n}=c_m+\sum_{k=2}^{m}c_{m,k}\Phi_1[\tau;\rho] \tag*{(36)}
(μm)x,∞=lim⁡n→∞(μm)x,n=cm+∑k=2mcm,kΦk[τ…τ;ρ](37)(\mu_m)_{x,\infty}=\lim_{n\to\infty}(\mu_m)_{x,n}=c_m+\sum_{k=2}^{m}c_{m,k}\Phi_k[\tau\ldots\tau;\rho] \tag*{(37)}

in which we used Equation (33) in Equation (37). Using Equation (36) and (37) in Equation (35) gives:

(μm)x,n=1n(μm)x,1+(1−1n)(μm)x,∞+Rm(38)(\mu_m)_{x,n}=\frac{1}{n}(\mu_m)_{x,1}+\left(1-\frac{1}{n}\right)(\mu_m)_{x,\infty}+R_m \tag*{(38)}

with a residual term RmR_m calculated by:

Rm=1n(1−1n){n∑k=2mcm,kn−(k−2)−1n−1Φ1[τ;ρ]+n∑k=2m∑l=2kcm,kbk,lΦl[τ…τ;ρ]−n∑l=2mcm,lΦl[τ…τ;ρ]}(39)R_m = \frac{1}{n}\left(1-\frac{1}{n}\right)\left\{n\sum_{k=2}^{m}c_{m,k}\frac{n^{-(k-2)}-1}{n-1}\Phi_1[\tau;\rho]+n\sum_{k=2}^{m}\sum_{l=2}^{k}c_{m,k}b_{k,l}\Phi_l[\tau\ldots\tau;\rho]-n\sum_{l=2}^{m}c_{m,l}\Phi_l[\tau\ldots\tau;\rho]\right\} \tag*{(39)}

Equation (39) is a linear combination of Vasicek moments Φk[τ…τ;ρ]\Phi_k[\tau\ldots\tau;\rho] as calculated by Equation (6). Using the identity ∑k=2m∑l=2kxk,l=∑l=2m∑k=lmxk,l\sum_{k=2}^{m}\sum_{l=2}^{k}x_{k,l}=\sum_{l=2}^{m}\sum_{k=l}^{m}x_{k,l} gives:

Rm=1n(1−1n){n∑k=2mcm,kn−(k−2)−1n−1Φ1[τ;ρ]+n∑l=2m∑k=lm(cm,kbk,l−cm,l)Φl[τ…τ;ρ]}(40)R_m = \frac{1}{n}\left(1-\frac{1}{n}\right)\left\{n\sum_{k=2}^{m}c_{m,k}\frac{n^{-(k-2)}-1}{n-1}\Phi_1[\tau;\rho]+n\sum_{l=2}^{m}\sum_{k=l}^{m}(c_{m,k}b_{k,l}-c_{m,l})\Phi_l[\tau\ldots\tau;\rho]\right\} \tag*{(40)}

So we may write the residual as:

Rm=1n(1−1n)∑l=1mam,lΦl[τ…τ;ρ](41)R_m=\frac{1}{n}\left(1-\frac{1}{n}\right)\sum_{l=1}^{m}a_{m,l}\Phi_l[\tau\ldots\tau;\rho] \tag*{(41)}

with coefficients depending on bklb_{kl} as polynomials in 1n\frac{1}{n}:

am,1=n∑k=2mcm,kn−(k−2)−1n−1a_{m,1}=n\sum_{k=2}^{m}c_{m,k}\frac{n^{-(k-2)}-1}{n-1}
am,l=n{∑k=lmcm,kbk,l−cm,l} for l>1(42)a_{m,l}=n\left\{\sum_{k=l}^{m}c_{m,k}b_{k,l}-c_{m,l}\right\}\text{ for }l>1 \tag*{(42)}

To complete the proof, we have to show that ∑l=1mam,l=0\sum_{l=1}^{m}a_{m,l}=0 holds for every mm.

1n∑l=1mam,l=1nam,1+1n∑l=2mam,l=∑k=2mcm,kn−(k−2)−1n−1+∑l=2m∑k=lm(cm,kbk,l−cm,l)(43)\frac{1}{n}\sum_{l=1}^{m}a_{m,l}=\frac{1}{n}a_{m,1}+\frac{1}{n}\sum_{l=2}^{m}a_{m,l}=\sum_{k=2}^{m}c_{m,k}\frac{n^{-(k-2)}-1}{n-1}+\sum_{l=2}^{m}\sum_{k=l}^{m}(c_{m,k}b_{k,l}-c_{m,l}) \tag*{(43)}

which is Equation (39) for RmR_m, divided by (1−1n)\left(1-\frac{1}{n}\right) and with Φl[τ;ρ]\Phi_l[\tau;\rho] replaced by 1. Again, switching indices in the last term gives:

1n∑l=1mam,l=∑k=2mcm,kn−(k−2)−1n−1+∑k=2m∑l=2mcm,kbk,l−∑l=2mcm,l=∑k=2mcm,kn−(k−2)−1n−1+∑k=2m∑l=2mcm,kbk,l−∑k=2mcm,k=∑k=2mcm,k(n−(k−2)−1n−1+∑l=2mbk,l−1)=∑k=2mcm,k(n−(k−2)−1n−1+n−n−(k−2)n−1−1)=0(44)\begin{aligned} \frac{1}{n}\sum_{l=1}^{m}a_{m,l} &= \sum_{k=2}^{m}c_{m,k}\frac{n^{-(k-2)}-1}{n-1}+\sum_{k=2}^{m}\sum_{l=2}^{m}c_{m,k}b_{k,l}-\sum_{l=2}^{m}c_{m,l}\\ &=\sum_{k=2}^{m}c_{m,k}\frac{n^{-(k-2)}-1}{n-1}+\sum_{k=2}^{m}\sum_{l=2}^{m}c_{m,k}b_{k,l}-\sum_{k=2}^{m}c_{m,k}\\ &=\sum_{k=2}^{m}c_{m,k}\left(\frac{n^{-(k-2)}-1}{n-1}+\sum_{l=2}^{m}b_{k,l}-1\right)\\ &=\sum_{k=2}^{m}c_{m,k}\left(\frac{n^{-(k-2)}-1}{n-1}+\frac{n-n^{-(k-2)}}{n-1}-1\right)=0 \tag*{(44)} \end{aligned}

The last step uses Equation (34). This completes the proof of the Theorem 2. □

Inverse of the homogeneous correlation matrix C_n

The correlation matrix CnC_n in Equation (2) has two eigenvalues:

  1. The largest eigenvalue 1+(n−1)ρ1+(n-1)\rho corresponds to one eigenvector vˉ1=[1,1,…,1]T\bar{v}_1=[1,1,\ldots,1]^T.

  2. The eigenvalue 1−ρ1-\rho corresponds to multiple eigenvectors vˉ2=[1,−1,0,0,0,…]T\bar{v}_2=[1,-1,0,0,0,\ldots]^T, vˉ3=[0,1,−1,0,0,…]T\bar{v}_3=[0,1,-1,0,0,\ldots]^T up to vˉn=[…,0,0,0,1,−1]T\bar{v}_n=[\ldots,0,0,0,1,-1]^T. As such, the geometric multiplicity of this eigenvalue is n−1n-1.

Knowing the eigenvalues and eigenvectors of CnC_n, its determinant follows from the product of the eigenvalues: ∣Cn∣=[1+(n−1)ρ](1−ρ)n−1|C_n|=[1+(n-1)\rho](1-\rho)^{n-1} and the diagonal elements of inverse Cn−1C_n^{-1} follow from the minors Mi,i=∣Cn−1∣M_{i,i}=|C_{n-1}|:

(Cn−1)i,i=(−1)i+iMi,i∣Cn∣=∣Cn−1∣∣Cn∣=1+(n−2)ρ(1−ρ)[1+(n−1)ρ](45)(C_n^{-1})_{i,i}=(-1)^{i+i}\frac{M_{i,i}}{|C_n|}=\frac{|C_{n-1}|}{|C_n|}=\frac{1+(n-2)\rho}{(1-\rho)[1+(n-1)\rho]} \tag*{(45)}

Since vˉ1=[1,1,1,…,1]T\bar{v}_1=[1,1,1,\ldots,1]^T is an eigenvector of CnC_n with eigenvalue 1+(n−1)ρ1+(n-1)\rho, it is also an eigenvector of Cn−1C_n^{-1} with eigenvalue (1+(n−1)ρ)−1(1+(n-1)\rho)^{-1}. Therefore, the sum of each row ii of Cn−1C_n^{-1} should equal:

(Cn−1)i,i+(n−1)(Cn−1)i≠j=11+(n−1)ρ(46)(C_n^{-1})_{i,i}+(n-1)(C_n^{-1})_{i\ne j}=\frac{1}{1+(n-1)\rho} \tag*{(46)}

Using Equation (45) for the diagonal elements (Cn−1)i,i(C_n^{-1})_{i,i} in Equation (46) gives:

1+(n−2)ρ(1−ρ)[1+(n−1)ρ]+(n−1)(Cn−1)i≠j=11+(n−1)ρ(47)\frac{1+(n-2)\rho}{(1-\rho)[1+(n-1)\rho]}+(n-1)(C_n^{-1})_{i\ne j}=\frac{1}{1+(n-1)\rho} \tag*{(47)}

giving the off-diagonal elements (Cn−1)i≠j=−ρ(1−ρ)[1+(n−1)ρ](C_n^{-1})_{i\ne j}=\frac{-\rho}{(1-\rho)[1+(n-1)\rho]}.

References

  1. [2]BCBS. (2006). International convergence of capital measurements and capital standards: a revised framework. Bank for International Settlements, Basel Committee on Banking Supervision.
  2. [3]BCBS. (2011). Basel III: A global regulatory framework for more resilient banks and banking systems - revised version. Basel Committee on Banking Supervision.
  3. [4]C. Bluhm, L. O. (2003). Credit Risk Modelling. Chapman & Hall/CRC.
  4. [5]CEIOPS. (2009). CEIOPS’ Advice for Level 2 Implementing Measures on Solvency II: SCR standard formula - Counterparty default risk module. Retrieved from https://register.eiopa.europa.eu/CEIOPS-Archive/Documents/Advices/CEIOPS-L2-Final-Advice-SCR-SF-Counterparty-default-risk.pdf
  5. [6]D. Maillard. (2020). A User's Guide to the Cornish Fisher Expansion. (hal-02987694). doi:http://dx.doi.org/10.2139/ssrn.1997178
  6. [7]F.E. Harrell, C. D. (1982). A new distribution-free quantile estimator. Biometrika, 69, 635-640. doi:https://doi.org/10.1093/biomet/69.3.635
  7. [8]J. Hull, A. W. (2004). Valuation of a CDO and an n to default CDS without Monte-Carlo simulations. Journal of Derivatives, 12(2), 8-23.DOI

Paper details

vlnc:W_sb6r5vkj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents