Abstract
This chapter deals with the measurement of name concentrations. This type of concentration risk occurs if the weight of single credits in the portfolio does not converge to zero; thus, the individual risk component cannot be completely diversified. The main research questions on name concentrations that are considered in this chapter are:
-
In which cases are the assumptions of the ASRF framework critical concerning the credit portfolio size?
-
In which cases are currently discussed adjustments for the VaR-measurement able to overcome the shortcomings of the ASRF model?
Concerning the first question, it is analyzed how many credits are at least necessary implying the neglect of undiversified individual risk not to be problematic. Since there exist analytical formulas – the so-called granularity adjustment – which approximate these risks, it is further determined in which cases these formulas are able to lead to desired results.
This is a preview of subscription content, log in via an institution.
Buying options
Tax calculation will be finalised at checkout
Purchases are for personal use only
Learn about institutional subscriptionsNotes
- 1.
Another solution to the problem of the violation of assumption (A) or (B) might be to cancel risk quantification under the IRB Approach and use internal models. However, this solution is not designated in Basel II.
- 2.
Cf. Sect. 3.2.
- 3.
Gordy (2003) comes to the conclusion that the granularity adjustment works fine for risk buckets of more than 200 loans considering low credit quality buckets and for more than 1,000 loans for high credit quality buckets. However, he uses the CreditRisk+ framework from Credit Suisse Financial Products (1997) and not the Vasicek model that builds the basis of Basel II, and he does not analyze the effect of different correlation factors as they are assumed in Basel II.
- 4.
This question is also interesting when analyzing the Basel II formula because the designated add-on factor for the potential violation of assumption (A) was cancelled from the second consultative document to the third consultative document; see BCBS (2001a, 2003a). Thus, we only prove under which conditions the assumption (A) of the Vasicek model is fulfilled. Of course, this model may suffer from other assumptions like the distributional assumption of standardized returns. However, since we would only like to address the topic of concentration risk, our focus should be reasonable. Additionally, the distributional assumptions seem not to have a deep impact on the measured VaR; see Koyluoglu and Hickman (1998a, b), Gordy (2000) or Hamerle and Rösch (2005a, b, 2006).
- 5.
Wilde (2001) calls this “the granularity adjustment to first order in the unsystematic variance”.
- 6.
This procedure can be motivated by the fact that for market risk quantification of nonlinear exposures two factors of the Taylor series (fist and second order) are common to achieve a higher accuracy; see e.g. Crouhy et al. (2001) or Jorion (2001). This might be appropriate for credit risk as well. Furthermore, the higher order derivatives of VaR given by Wilde (2003) make it possible to systematically derive such a formula.
- 7.
The Basel Committee on Banking Supervision already stated that in principle the effect of portfolio size on credit risk is well understood but lacks practical analyses; see BCBS (2005b).
- 8.
Additionally, this study makes contribution to the ongoing research on analyzing differences between Basel II capital requirements and banks internal “true” risk capital measurement approaches. Since the harmonization of the regulatory capital requirements and the perceived risk capital of banks internal estimates for portfolio credit risk is often stated as the major benefit of Basel II, see e.g. Hahn (2005), p. 127, but often not observed in practice, this task might be of relevance in the future.
- 9.
See Appendix 4.5.2.
- 10.
This is valid because the added risk of the portfolio is unsystematic; see Martin and Wilde (2002) for further explanations.
- 11.
See Appendix 4.5.3.
- 12.
Cf. the identity 2.90.
- 13.
The notation n* refers to the effective number of credits as introduced in (2.87).
- 14.
The equivalent term for heterogeneous portfolios is \( O\left( {\sum\limits_{i = 1}^n {{w^3}} } \right) \).
- 15.
The mth moment of a random variable \( \tilde{X} \) about the mean \( {\eta_m}(\tilde{X}) \) is defined as \( {\eta_m}(\tilde{X}): = b{E}({[\tilde{X} - b{E}(\tilde{X})]^m}) \); cf. Abramowitz and Stegun (1972), 26.1.6.
- 16.
Cf. Appendix 4.5.4.
- 17.
This assumption can be critical for real-world portfolios. Especially, it is often assumed in ongoing research on credit portfolio modeling that the LGD is dependent on the systematic factor. However, the granularity adjustment formula would complicate significantly as neither the ELGD nor the VLGD could be treated as constant for the derivatives. Against this background, this assumption will be retained for the derivation.
- 18.
- 19.
Cf. Appendix 4.5.5.
- 20.
Gordy (2003) observes the concavity of the granularity add-on for a high-quality portfolio (A-rated) up to a portfolio size of 1,000 debtors.
- 21.
See Gordy (2004), p. 112, footnote 5, for a similar suggestion.
- 22.
See Appendix 4.5.8 for details regarding the order of these elements.
- 23.
Cf. (4.236) of Appendix 4.5.8.
- 24.
Precisely, the element \( {\eta_{3,c}} \) is the third conditional moment centered about the mean whereas the conditional skewness is the “normalized” third moment, defined as the third conditional moment about the mean divided by the conditional standard deviation to the power of three.
- 25.
Cf. (4.14).
- 26.
See Appendix 4.5.10.
- 27.
Cf. BCBS (2006), p. 10.
- 28.
The chosen portfolio exhibits high unsystematic risk and therefore serves as a good example in order to explain the differences of the four solutions. However, we evaluated several portfolios and basically, the results do not differ. Additionally, we claim that the general statements can also be applied to heterogeneous portfolios.
- 29.
See Rau-Bredow (2005) for a counter-example for very unusual parameter values. This problem can be addressed to the use of VaR as a measure of risk which does not guarantee sub-additivity; cf. Sect. 2.2.3.
- 30.
By contrast, we expected a significant enhancement by using the second order adjustment like mentioned in Gordy (2004), p. 112, footnote 5.
- 31.
To address to the minimum number after which the target tolerance will permanently hold, we have to add the notation “for all \( N \geq n \)” because the function of the coarse grained VaR exhibits jumps dependent on the number of credits.
- 32.
- 33.
See Sect. 2.7 for details. In both tables, (rounded) parameters ρ due to Basel II are marked.
- 34.
The case of heterogeneous portfolios will be analyzed in Sect. 4.2.2.5.
- 35.
Cf. Deutsche Bundesbank (2009).
- 36.
This is true not only for the first five derivatives but also for all following derivatives; see the general formula for all derivatives of VaR in (4.213).
- 37.
- 38.
- 39.
Due to the high number of trials, which corresponds to 3,000 hits in the tail for a confidence level of 0.999, the simulation noise should be negligible.
- 40.
Cf. Sect. 2.6.
- 41.
This is true for a violation of both the granularity and the single risk factor assumption.
- 42.
- 43.
See Albanese and Lawi (2004), p. 215, for this property of a reasonable risk measure.
- 44.
Of course the definition of the VaR does not allow a negative deviation and the VaR jumps to a higher value instead.
- 45.
See Appendix 4.5.11.
- 46.
See Appendix 4.5.12.
- 47.
As mentioned in Sect. 2.6, the VaR is exactly additive and therefore unproblematic in the context of the ASRF framework.
- 48.
We use the idealized default rates from Standard and Poors, see Brand and Bahar (2001), ranging from 0.01% to 18.27%, but the results do not differ widely for different values.
- 49.
The portfolios with high, average, low, and very low quality are taken from Gordy (2000). We added a portfolio with very high quality.
- 50.
The derivatives of ES are derived in Appendix 4.5.13 and 4.5.14.
- 51.
Cf. (4.8).
- 52.
The explanations regarding the order of the derivatives of VaR in Appendix 4.5.8 are valid for the derivatives of ES, too.
- 53.
See also Wilde (2003).
- 54.
See Appendix 4.5.14.
- 55.
Even if the calculations were based on the portfolio gross loss and thus on an LGD of 100%, the results remain identically for every constant LGD as the numerator and the denominator of the analyzed expressions are affected to the same degree.
- 56.
Cf. Schuermann (2005), p. 22, footnote 8.
- 57.
Cf. Schuermann (2005), p. 22, footnote 11.
- 58.
Probably, the data used to generate the figure did not include workout costs and therefore underestimate the true economic loss. Furthermore, the choice of the discount rate influences the effect of negative LGDs: If the recovery cash flows are discounted by the contractual rate, as required by IFRS and as proposed by the Basel II framework, a complete recovery without workout costs leads to a recovery rate of 100%, which shows that negative LGDs are not relevant at all.
- 59.
The issue of interconnections between LGDs and PDs via a systematic factor is not in the scope of this analysis.
- 60.
Cf. Altman et al. (2005), p. 46.
- 61.
Cf. Gupton et al. (1997), p. 80.
- 62.
See also Sect. 2.3.
- 63.
Cf. Bronshtein et al. (2007), p. 760, (16.80).
- 64.
Cf. Schönbucher (2003), p. 147 f.
- 65.
The aggregated data correspond to Fig. 4.8.
- 66.
The critical number of credits in a portfolio which leads to equality of the different parameter settings of the Basel consultative documents is not of interest in the subsequent analyses regarding the ES as both rely on the VaR.
- 67.
See Sect. 4.3.1.
- 68.
As the ASRF solution is constant and the coarse grained solution is monotonously decreasing in n for the ES (this is a result of the monotonicity of specific risk-property, cf. Sect. 4.3.1), the inequality also holds for every number above the first number that satisfies the inequality. Thus, the expression “for all \( N \geq n \)”, which had to be included in the corresponding analysis for the VaR, can be neglected.
- 69.
The corresponding value for deterministic LGDs is 91.64%.
- 70.
The omission of the zeroth-order terms could be foreseen as only the deviation from the systematic loss quantile is analyzed.
- 71.
For functions f, g with \( \mathop {{\lim }}\limits_{x \to {x_0}} f(x) = \mathop {{\lim }}\limits_{x \to {x_0}} g(x) = 0 \) or \( \mathop {{\lim }}\limits_{x \to {x_0}} f(x) = \mathop {{\lim }}\limits_{x \to {x_0}} g(x) = \infty \) it is true that\( \mathop {{\lim }}\limits_{x \to {x_0}} \frac{{f(x)}}{{g(x)}} = \mathop {{\lim }}\limits_{x \to {x_0}} \frac{{f'(x)}}{{g'(x)}} \) if \( \mathop {{\lim }}\limits_{x \to {x_0}} \frac{{f(x)}}{{g(x)}} \) exists; cf. Bronshtein et al. (2007), p. 54, (2.26).
- 72.
Cf. Wilde (2001).
- 73.
Cf. (2.14). The slightly different expressions compared to Rau-Bredow (2002) result from α instead of (1–α) representing the confidence level.
- 74.
Cf. Pitman (1999), p. 416.
- 75.
Cf. Rau-Bredow (2004), p. 66.
- 76.
Cf. Roussas (2007), p. 236.
- 77.
- 78.
Cf. Bronshtein et al. (2007), p. 710, (15.5).
- 79.
Cf. Bronshtein et al. (2007), p. 710, (15.8).
- 80.
Weisstein (2009a).
- 81.
Cf. Bronshtein et al. (2007), p. 672, Sect. 14.1.2.1.
- 82.
Cf. Bronshtein et al. (2007), p. 688, (14.41).
- 83.
Cf. Bronshtein et al. (2007), p. 691, (14.49).
- 84.
- 85.
Cf. Bronshtein et al. (2007), p. 692 f., Sect. 14.3.5.1.
- 86.
Cf. Bronshtein et al. (2007), p. 694, (14.56).
- 87.
Cf. Rowland and Weisstein (2009).
- 88.
Cf. Wilde (2003), p. 3 f.
- 89.
- 90.
Cf. Billingsley (1995), p. 146 ff., for details about moment generating functions.
- 91.
Cf. Miller and Childers (2004), p. 118.
- 92.
For ease of notation, the derivatives \( {{{\partial G}} \left/ {{\partial z}} \right.} \) and \( {{{\partial G}} \left/ {{\partial w}} \right.} \) will be abbreviated to \( {G_z} \) and \( {G_w} \), respectively. The function G is not associated with a random variable, so confusion should not arise with respect to the similar notation \( {F_{Y + \lambda Z}}(y) \), where the subscript of the distribution function F denotes the corresponding random variable.
- 93.
Cf. Wilde (2003), p. 7.
- 94.
See Abramowitz and Stegun (1972), Sect. 24.1.2(C). The notation \( p \prec m \) indicates that p is a partition of m, cf. Sect. 4.5.6.1.3.
- 95.
See Weisstein (2009b).
- 96.
Cf. Wilde (2003), p. 8.
- 97.
The relation between a partition \( u \) and \( \hat{u} \) is explained in Sect. 4.5.6.1.3.
- 98.
In order to demonstrate that the resulting formula is also valid for \( m = 1 \), the summand for partition \( \{ {1^1}\} \), which equals zero due to argument (4.216), is still considered.
- 99.
For ease of notation, the arguments \( \lambda = 0 \) of the left-hand as well as \( y = {q_\alpha }(\tilde{Y}) \) at the right-hand side are omitted.
- 100.
Cf. (4.213). The notation \( g \circ y \) means that a function g is composed with y.
- 101.
To illustrate that the first identity holds, an example will be demonstrated for m = 5: \( m = 5 \) Furthermore, see (4.9) for the switch between the systematic loss y and the systematic factor x.
- 102.
See (4.14).
- 103.
Cf. Wilde (2003), p. 11.
- 104.
Cf. Appendix 4.5.3.
References
Abramowitz M, Stegun I (1972) Handbook of mathematical functions: with formulas, graphs, and mathematical tables, 10th edn. Dover, New York
Albanese C, Lawi S (2004) Spectral risk measures for credit portfolios. In: Szegö G (ed) Risk measures for the 21st century. Wiley, Chichester, pp 209–226
Altman E, Resti R, Sironi A (2005) Loss given default: a review of the literature. In: Altman E, Resti R, Sironi A (eds) Recovery risk – the next challenge in risk management. Risk Books, London, pp 41–59
Basel Committee on Banking Supervision (2001a) Basel II: The New Basel capital accord, Second Consultative Paper. Bank for International Settlements, Basel
Basel Committee on Banking Supervision (2003a) Basel II: The New Basel capital accord, Third Consultative Paper. Bank for International Settlements, Basel
Basel Committee on Banking Supervision (2005b) Workshop ‘Concentration risk in credit portfolios’: background information. Bank for International Settlements, Basel. http://www.bis.org/bcbs/events/rtf05background.htm. Accessed 18 Aug 2009
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
Billingsley P (1995) Probability and measure, 3rd edn. Wiley, New York
Brand L, Bahar R (2001) Corporate defaults: will things get worse before they get better? S&P Special Report, pp 5–40
Bronshtein IN, Semendyayew KA, Musiol G, Muehlig H (2007) Handbook of mathematics, 5th edn. Springer, Heidelberg
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
Credit Suisse Financial Products (1997). CreditRisk+ – A credit risk management framework. Credit Suisse, London.
Crouhy M, Galai D, Mark R (2001) Risk management. McGraw-Hill, New York
Deutsche Bundesbank (2009) Credit register of loans of €1.5 million or more. http://www.bundesbank.de/bankenaufsicht/bankenaufsicht_kredit_evidenz.en.php. Accessed 11 Dec 2009
Düllmann K (2006). Measuring business sector concentration by an infection model. Discussion Paper, Series 2: Banking and financial studies. Deutsche Bundesbank (3)
Düllmann K, Erdelmeier M (2009) Stress testing german banks in a downturn in the automobile industry. Discussion Paper, Series 2: Banking and Financial Studies. Deutsche Bundesbank (2)
Düllmann K, Masschelein N (2007) A tractable model to measure sector concentration risk in credit portfolios. J Fin Serv Res 32(1):55–79
Emmer S, Tasche D (2005) Calculating credit risk capital charges with the one-factor model. J Risk 7(2):85–103
Frye J (2000) Depressing recoveries. Risk 13(11):108–111
Gordy MB (2000) A comparative anatomy of credit risk models. J Bank Fin 24(1–2):119–149
Gordy MB (2001) A risk-factor model foundation for rating-based capital rules. Working Paper. Board of Governors of the Federal Reserve System, Washington DC
Gordy MB (2003) A risk-factor model foundation for rating-based capital rules. J Fin Intermediation 12(3):199–232
Gordy MB (2004) Granularity adjustment in portfolio credit risk measurement. In: Szegö G (ed) Risk measures for the 21st century. Wiley, New York, pp 109–121
Gouriéroux C, Laurent J, Scaillet O (2000) Sensitivity analysis of values at risk. J Empir Fin 7(3–4):225–245
Gupton GM, Finger CC, Bhatia M (1997) CreditMetrics™ – Technical document. Morgan Guarantee Trust Co, New York
Gürtler M, Heithecker D, Hibbeln M (2008a) Concentration risk under Pillar 2: when are credit portfolios infinitely fine grained? Kredit Kapital 41(1):79–124
Gürtler M, Hibbeln M, Olboeter S (2008b) Design of collateralized debt obligations: the impact of target ratings on the first loss piece. In: Gregoriou GN, Ali P (eds) The credit derivatives handbook. McGraw-Hill, New York, pp 203–228
Hahn F (2005) The effect of bank capital on bank credit creation. Kredit Kapital 38(1):103–127
Hamerle A, Rösch D (2005a) Bankinterne Parametrisierung und empirischer Vergleich von Kreditrisikomodellen. Die Betriebswirtschaft 65(2):179–196
Hamerle A, Rösch D (2005b) Misspecified copulas in credit risk models: how good is Gaussian? J Risk 8(1):41–59
Hamerle A, Rösch D (2006) Parameterizing credit risk models. J Credit Risk 2(4):101–122
Heitfield E, Burton S, Chomsisengphet S (2006) Systematic and idiosyncratic risk in syndicated loan portfolios. J Credit Risk 2(3):3–31
Jorion P (2001) Value at risk, 2nd edn. McGraw-Hill, New York
Koyluoglu HU, Hickman A (1998a) A generalized framework for credit portfolio models. Working Paper, http://www.defaultrisk.com
Koyluoglu HU, Hickman A (1998b) Reconcilable differences. Risk 11(10):56–62
Martin R, Wilde T (2002) Unsystematic credit risk. Risk 15(11):123–128
Miller SL, Childers DG (2004) Probability and random processes: with applications to signal processing and communications, 2nd edn. Academic Press, Amsterdam
Overbeck L (2000) Allocation of economic capital in loan portfolios. In Frank WH, Stahl G (eds) Measuring risk in complex stochastic systems. Springer, New York, pp 1–17
Overbeck L, Stahl G (2003) Stochastic essentials for the risk management of credit portfolios. Kredit Kapital 36(1):52–81
Pitman J (1999) Probability. Springer, New York
Pykhtin M (2003) Unexpected recovery risk. Risk 16(8):74–78
Pykhtin M, Dev A (2002) Analytical approach to credit risk modelling. Risk 15(3):26–32
Rau-Bredow H (2002) Credit portfolio modelling, marginal risk contributions, and granularity adjustment. Working Paper, Würzburg
Rau-Bredow H (2004) Value at risk, expected shortfall, and marginal risk contribution. In: Szegö G (ed) Risk measures for the 21st century. Wiley, Chichester, pp 61–68
Rau-Bredow H (2005) Unsystematic credit risk and coherent risk measures. Working Paper, Würzburg
Roussas GG (2007) Introduction to probability. Academic Press, Amsterdam
Rowland T, Weisstein EW (2009) Complex residue, from mathworld – a Wolfram web resource. http://mathworld.wolfram.com/ComplexResidue.html. Accessed 18 Aug 2009
Schönbucher P (2003) Credit derivatives pricing models: models, pricing and implementation. Wiley, Chichester
Schuermann T (2005) What do we know about loss given default? In: Altman E, Sironi A (eds) Recovery risk – the next challenge in risk management. Risk Books, London, pp 3–24
Spiegel MR (1999) Schaum’s outline of theory and problems of complex variables: with an introduction to conformal mapping and its applications. McGraw-Hill, New York
Weisstein EW (2009a) Delta function, from MathWorld – a Wolfram Web resource. http://mathworld.wolfram.com/DeltaFunction.html. Accessed 18 Aug 2009
Weisstein EW (2009b) Leibniz identity, from MathWorld – a Wolfram web resource. http://mathworld.wolfram.com/LeibnizIdentity.html. Accessed 18 Aug 2009
Wilde T (2001) Probing granularity. Risk 14(8):103–106
Wilde T (2003) Derivatives of VaR and CVaR. Working Paper, CSFB
Gürtler M, Hibbeln M, Vöhringer C (2010) Measuring concentration risk for regulatory purposes. J Risk 12(3):69–104
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
4.1.1 Alternative Derivation of the First-Order Granularity Adjustment
With reference to Wilde (2001), the granularity adjustment will be derived as an approximation of the difference \( \Delta q \) between the true VaR of a granular portfolio \( {q^{(n)}} \) and the approximation \( {q^{(\infty )}} \) that results if infinite granularity is assumed to hold:
Instead of determining the add-on \( \Delta q \) directly, it will be analyzed how much the confidence level α will be overestimated or the probability \( p: = 1 - \alpha \) of exceeding the VaR will be underestimated if the portfolio is assumed to be infinitely granular. Thus, the probability
refers to the overestimation of the confidence level if only the systematic loss is considered. Here, α is the specified “target” confidence level, and by definition also the probability that the systematic loss will not exceed \( q_\alpha^{(\infty )} \):
By contrast, \( {\alpha^{(\infty )}} \) is the actual confidence level if the VaR is approximated by the ASRF model:
Subsequent to the derivation of \( \Delta p \), the result will be transformed into a shift of the loss quantile \( \Delta q \).
Analogous to Appendix 2.8.3, the unconditional probability \( {p^{(\infty )}} \) can be expressed in terms of the conditional probability. Then, the substitution \( y: = q_\alpha^{(\infty )} + t \) is performed to center the integration at \( q_\alpha^{(\infty )} \):
with the shorter notation \( \tilde{Y}: = b{E}\left( {\tilde{L}|\tilde{x}} \right) \) for the conditional expectation. According to (4.106), the probability p can be written as
using the substitution \( y: = q_\alpha^{(\infty )} + t \) again, so that \( t(y = q_\alpha^{(\infty )}) = 0 \) and \( t(y = \infty ) = \infty \). Hence, (4.108) can be expressed as
The following transformations are performed for simplification of the integrand in order to solve the integral. A realization of the systematic loss implies a realization of the systematic factor. As the credit loss events are assumed to be independent for a realization of the systematic factor, the conditional credit losses follow a binomial distribution, which can be approximated by a normal distribution for a sufficient number of credits. This leads to
As \( b{E}(\tilde{L}) = b{E}(b{E}(\tilde{L}|\tilde{x})) = b{E}(\tilde{Y}) \), which is due to the law of iterated expectation, the conditional expectation of (4.111) equals
With the symmetry \( 1 - \Phi ( - x) = \Phi (x) \) and defining \( {\sigma^2}(y): = b{V}(\tilde{L}|\tilde{Y} = y) \), (4.111) results in
so that (4.110) can be written as
Subsequently, several linear approximations will be performed relying on the assumption that the loss quantile of the granular portfolio is close to the systematic loss quantile and the linearizations lead to minor errors. Linearizing the density function at \( q_\alpha^{(\infty )} \) leads to
The argument of the normal distribution can be approximated as
With the substitution \( y: = q_\alpha^{(\infty )} + t \), so \( dy/dt = 1 \) and \( y(t = 0) = q_\alpha^{(\infty )} \), the derivative of the conditional standard deviation can be rewritten as
Inserting (4.115)–(4.117) in (4.114) leads to
When the substitution \( t: = - t \) for the term \( \Delta {p_2} \) is performed and the symmetry of the normal distribution \( \Phi ( - x) - 1 = - \Phi (x) \) is used, both terms \( \Delta {p_1} \) and \( \Delta {p_2} \) are identical except for the algebraic signs:
A linearization of the normal distributions in \( \Delta {p_1} \) and \( \Delta {p_2} \) results in
Using this approximation, the terms \( \Delta {p_1} \) and \( \Delta {p_2} \) from (4.118) can be written as
The summands \( {\beta_0},\;{\gamma_0} \) are the points around which the linearizations have been performed. The summands \( {\beta_1},\;{\gamma_1} \) have resulted from the first-order approximations. Using this notation, the shift in probability \( \Delta p \) of (4.118) can notably be simplified to
Fortunately, both integrands are already first-order terms whereas the cross-terms \( {\beta_1} \cdot {\gamma_1} \) vanish.Footnote 70 Thus, there is no need for a further linearization. The remaining expression is
In order to solve the integrals, the substitution \( y: = t/\sigma (q_\alpha^{(\infty )}) \) is performed, with \( dy/dt = 1/\sigma (q_\alpha^{(\infty )}) \), \( y(t = - \infty ) = - \infty \) and \( y(t = 0) = 0 \):
For the second integral (**), it is used that the integrand is axially symmetric to the y-axis. Furthermore, the definition of the variance is utilized, considering that the standard normal distribution has mean \( {\mu_Y} = 0 \) and variance \( \sigma_Y^2 = 1 \):
The first integral (*) can be calculated with integration by parts:
For \( y = 0 \), the first term is zero but for \( y = - \infty \), the result is not obvious. Using l’Hôpital’s rule several times leads toFootnote 71
so that the first term of (4.126) vanishes. Using the result of the previous integration, (4.126) equals \( - {1/4} \). Hence, \( \Delta p \) from (4.124) is given as
Because of \( \sigma \frac{{d\sigma }}{{dy}} = \frac{1}{2}\frac{{d{\sigma^2}}}{{d\sigma }}\frac{{d\sigma }}{{dy}} = \frac{1}{2}\frac{{d{\sigma^2}}}{{dy}} \), (4.128) is equivalent to
This expression is the linearized deviation of the specified probability \( p = 1 - \alpha \) if only the systematic loss is considered for calculation of the loss quantile.
As initially noticed, the determined shift of the probability has to be transformed into a shift of the loss quantile (cf. Fig. 4.12). If the probability density function of the portfolio loss is assumed to be almost linear in a region around the quantile, the required transformation is
Two last first-order approximations lead to
Inserting (4.129) into (4.131) finally leads to
Using (4.8), this can be written as
which is identical to the first-order granularity adjustment of Sect. 4.2.1.1.Footnote 72
4.1.2 First and Second Derivative of VaR
The derivatives of VaR will be determined on the basis of Rau-Bredow (2002, 2004) in the following. Consider two continuous random variables \( \tilde{Y} \) and \( \tilde{Z} \) with joint probability density function \( f(y,z) \) and a variable \( \lambda \in b{R} \). The VaR (the quantile) \( q: = {q_\alpha }\left( {\tilde{L}} \right) \) of \( \tilde{L} = \tilde{Y} + \lambda \tilde{Z} \) can implicitly be defined asFootnote 73
Furthermore, the formula of the conditional density function will be used:Footnote 74
leading toFootnote 75
4.1.2.1 First Derivative
As the derivative of the constant α is zero, the derivative of (4.134) is
Performing the inner integration and the differentiation leads to
Using the formula for the conditional density function (4.135) and the integral representation of the conditional expectation, we get
This leads to the first derivative of VaR:
The first derivative at \( \lambda = 0 \) is
4.1.2.2 Second Derivative
Similar to (4.137), the second derivative of (4.134) is
With the first derivative of (4.138) and applying the product rule, this leads to
The derivative (*) can be determined with the chain rule:
Inserting (4.144) and the conditional density (4.136) into (4.143) results in
The first summand of (4.145) equals
In order to calculate the second summand of (4.145), the first derivative from (4.140) as well as the integral representation of the conditional variance is used:
With these summands, (4.145) can be written as
Thus, the second derivative of VaR is equal to
The second derivative at \( \lambda = 0 \) is
4.1.3 Probability Density Function of Transformed Random Variables
Let \( \tilde{X} \) be a random variable with density \( {f_X}(x) \) and let \( \tilde{Y} \) be a random variable with \( \tilde{Y} = g(\tilde{X}) \). If g is strictly monotonous and differentiable, the probability density function (PDF) of \( \tilde{Y} \) can be transformed using the inverse function theoremFootnote 76:
With \( {g^{ - 1}}(y) = x \), we obtain
which leads to
4.1.4 VaR-Based First-Order Granularity Adjustment for a Normally Distributed Systematic Factor
The granularity adjustment (4.10) can be expressed as
Because of
the granularity adjustment (4.154) can be written as
For the calculation of (4.156), the conditional expectation and variance have to be determined. Assuming stochastically independent LGDs and with ELGD and VLGD for the expectation and the variance of the LGD, respectively, the required moments are given asFootnote 77
4.1.5 VaR-Based First-Order Granularity Adjustment for Homogeneous Portfolios
For homogeneous portfolios, the granularity adjustment formula (4.28) can be simplified to
4.1.6 Arbitrary Derivatives of VaR
The following determination of all derivatives of VaR is based on Wilde (2003). The quantile q α of \( \tilde{L} = \tilde{Y} + \lambda \tilde{Z} \) can be written as \( q(\lambda ) \) to denote that the quantile depends on the parameter λ. Using this notation, the quantile can be defined implicitly as an argument of the distribution function F by \( F(q(\lambda ),\lambda ): = b{P}\left( {\tilde{Y} + \lambda \tilde{Z} \leq {q_\alpha }(\tilde{Y} + \lambda \tilde{Z})} \right) = \alpha \). In order to calculate the derivatives of q α , at first all derivatives of F are determined in Sect. 4.5.6.2.1. As the quantile is defined implicitly, the implicit derivatives of \( F(q(\lambda ),\lambda ) - \alpha = 0 \) have to be determined. This is done by application of the residue theorem in Sect. 4.5.6.2.2. As a next step, the result will be expressed in combinatorial form in Sect. 4.5.6.2.3. Using the results of the derivatives of the distribution function and the implicit derivatives, it is possible to determine all derivatives of VaR. This is performed in Sect. 4.5.6.2.4. As the resulting formula is quite complex, an expression for the first five derivatives of VaR is determined in Sect. 4.5.7. The mathematical basics to the Laplace transform, complex residues, and partitions, which are needed within the derivation, are presented in the following Sect. 4.5.6.1.
4.1.6.1 Mathematical Basics
4.1.6.1.1 Laplace Transform and Dirac’s Delta Function
The Laplace transform \( L \) of a function \( f(t) \) with \( t \in {b{R}^{+} } \) is given asFootnote 78
with \( s = c + i\omega \,\, \in \,\,b{C} \), where \( b{C} \) denotes the set of all complex numbers. The inverse Laplace transform \( {L^{ - 1}} \) can be represented asFootnote 79
Dirac’s delta function \( \delta (x) \) can be defined asFootnote 80
A more illustrative, heuristic definition of \( \delta (x) \) is given by
Using the definition of the Laplace transform and the inverse Laplace transform, Dirac’s delta function can be written as
4.1.6.1.2 Laurent Series, Singularities, and Complex Residues
If \( f(z) \) is differentiable in all points of an open subset of the complex plane \( H \subset b{C} \), then we call \( f(z) \) holomorphic on H.Footnote 81 For a function \( f(z) \), which is holomorphic in a simply connected region H, according to the Cauchy integral theorem we haveFootnote 82
with C being a closed path in H. If a function \( f(z) \) is holomorphic in z 0 and in a circular region around z 0, we can perform a Taylor series expansion, which is analogous to the real plane:Footnote 83
However, if a function \( f(z) \) is only holomorphic inside the annulus between two concentric circles with center z 0 and radii r 1 and r 2, which is the region \( H = \left\{ {z|0 \leq {r_1} < \left| {z - {z_0}} \right| < {r_2}} \right\} \), the function \( f(z) \) can be expressed as a generalized power series, the so-called Laurent series:Footnote 84
Thus, the function has to be holomorphic only inside the annulus and not inside the inner circle or outside the outer circle.
If a function \( f(z) \) is holomorphic in a neighborhood of z 0 but not in the point z 0, then z 0 is called an isolated singularity of the function \( f(z) \). The concrete type of a singularity can be classified according to the analytic part of the Laurent series:Footnote 85
-
The point z 0 is a removable singularity if \( {a_n} = 0\,\,\forall n < 0 \). In this case, the Laurent series is identical to the Taylor series above.
-
The point z 0 is a pole of order m if the principal part consists of a finite number of terms with \( {a_m} \ne 0 \) and \( {a_n} = 0 \) for \( n < m < 0 \).
-
The point z 0 is an essential singularity if the principal part consists of an infinite number of terms.
The coefficient \( {a_{ - 1}} \) of the Laurent series (4.167) around an isolated singularity z 0 is the residue of \( f(z) \) in z 0. This will subsequently be denoted by \( {\text{Re}}{{\text{s}}_{{z_0}}}(f) \). The residue can also be defined as
where C is a contour with winding number 1 in a holomorphic region H around an isolated singularity in z 0. If the contour C encloses a finite number of isolated singularities \( {z_1},{z_2},...,{z_m} \) with corresponding residues \( {a_{ - 1}}({z_\mu }) \) \( (\mu = 1,...,m) \), we have
which is the residue theorem.Footnote 86
The residue \( {\text{Re}}{{\text{s}}_{{z_0}}}(f) \) with z 0 being a pole of order m can be calculated asFootnote 87
For a function \( f = {{{g(z)}} \left/ {{h(z)}} \right.} \), where h has a simple zero in z 0, the residue can be determined with
4.1.6.1.3 Partitions
A partition p of a positive integer m is a way to express m as a sum of positive integers in non-decreasing order. A partition p of m will be denoted by \( p \prec m \). A partition p can be indicated by \( p = {1^{{e_1}}},{2^{{e_2}}},...,{m^{{e_m}}} \), where \( {e_i} \) is the frequency of the number i in the partition. The number of summands of p is expresses by \( \left| p \right| \), which is the sum \( \left| p \right| = {e_1} + {e_2} + ... + {e_m} \). The notation \( \hat{p} \) indicates the partition which results if each summand of a partition p is increased by 1. This means that for \( p \prec m \) the partition \( \hat{p} \) refers to a specific partition of \( m + \left| p \right| \).Footnote 88
Example
-
For \( m = 5 \), there exist seven partitions \( p \prec m \): \( p \prec m = \left\{ {1 + 1 + 1 + 1 + 1,\;1 + 1 + 1 + 2,\;1 + 2 + 2,\;1 + 1 + 3,\;2 + 3,\;1 + 4,\;5} \right\}. \) Thus, a concrete partition for \( m = 5 \) is \( p = 3 + 1 + 1 \).
-
This partition can also be denoted by \( p = {1^{{e_1}}}\,{2^{{e_2}}}\,...\,{m^{{e_m}}} = {1^2}{3^1} \), leading to \( {e_1} = 2, \) \( {e_2} = 0, \) \( {e_3} = 1, \) \( {e_4} = 0 \), and \( {e_5} = 0 \). Thus, the number m results from: \( m = 1 \cdot {e_1} + 2 \cdot {e_2} + ... + m \cdot {e_m} = 1 \cdot 2 + 3 \cdot 1 = 5 \).
-
The number of summands of this partition is \( \left| {p = {1^2}{3^1}} \right| = {e_1} + {e_2} + ... + {e_m} = 2 + 1 = 3 \).
-
The partition \( \hat{p} \) appendant to the partition \( p = 3 + 1 + 1 \) is \( \hat{p} = 4 + 2 + 2 \), which is a specific partition of \( m + \left| p \right| = 5 + 3 = 8 \).
4.1.6.2 Determination of the Derivatives
4.1.6.2.1 Derivatives of the Distribution Function
Proposition
The derivatives of the distribution function of losses \( {F_{Y + \lambda Z}}(y) = b{P}(\tilde{Y} + \lambda \tilde{Z} \leq y) \) at \( \lambda = 0 \) are given asFootnote 89
Proof
Using the definition of the Laplace transform (4.160) and recognizing that the loss \( \tilde{L} = \tilde{Y} + \lambda \tilde{Z} \) cannot go below zero so that the probability density function is \( {f_{Y + \lambda Z}}(y) = 0 \) for all \( y < 0 \), we get for the Laplace transform of \( {f_{Y + \lambda Z}}(y) \)
With the definition of the expectation operator
(4.173) is equivalent to
Applying the definition of the inverse Laplace transform (4.161) and using the moment generating function M of \( \tilde{Y} + \lambda \tilde{Z} \), which is defined asFootnote 90
the probability density function equalsFootnote 91
Thus, the derivatives of the probability density function at \( \lambda = 0 \) can be determined using the approach
Applying definition (4.176), we obtain for the derivatives of M
With (4.179) and \( {s^m}{{\text{e}}^{s\left( {\tilde{Y} - y} \right)}} = {( - 1)^m}\frac{{{\partial^m}}}{{\partial {y^m}}}{{\text{e}}^{s\left( {\tilde{Y} - y} \right)}} \), (4.178) is equivalent to
According to (4.164), Dirac’s delta function can be written as
which leads to
for \( t = \tilde{Y} - y \). Hence, (4.180) is equivalent to
With \( b{E}[{\tilde{Z}^m}\delta (\tilde{Y} - y)] = b{E}[{\tilde{Z}^m}|\tilde{Y} = y] \cdot {f_Y}(y) \), the derivatives of the distribution function result after integration of (4.183):
which is proposition (4.172). In order to determine the derivatives of the quantile \( {{{{d^m}q}} \left/ {{d{\lambda^m}}} \right.} \), the implicit derivatives of \( F(q(\lambda ),\lambda ) - \alpha = 0 \) with \( F(q(\lambda ),\lambda ): = {F_{\tilde{Y} + \lambda \tilde{Z}}}({q_\alpha }(\tilde{Y} + \lambda \tilde{Z})) \) \( = b{P}\left( {\tilde{Y} + \lambda \tilde{Z} \leq {q_\alpha }(\tilde{Y} + \lambda \tilde{Z})} \right) \) will be calculated in the following.
4.1.6.2.2 Implicit Derivatives: Complex Residue Form
Consider a function \( G(z,w) \) of two variables \( z,w \in b{C} \). Suppose there exists an analytic function \( w = w(z) \) in a region around a pole \( z = {z_0} \), such that \( G(z,w(z)) = 0 \). The first derivative \( {{{dw}} \left/ {{dz}} \right.} \) can be determined as follows:Footnote 92
Proposition
For \( {G_w}({z_0},{w_0}) \ne 0 \), the derivatives \( {{{{d^m}w}} \left/ {{d{z^m}}} \right.} \) are given as
Proof
According to (4.186), the first derivative is
As z 0 is a pole of G and \( G({z_0},w) = 0 \), an application of (4.171) leads to
which is equal to (4.185). This shows that the formula is correct for \( m = 1 \).
Applying the residue theorem (4.169)
and recognizing that there is only a singularity at \( z = {z_0} \) leads to
Differentiating and applying the residue theorem again results in
which is the proposition presented in (4.186). This result is a generalization of the Lagrange inversion theorem.Footnote 93
4.1.6.2.3 Implicit Derivatives: Combinatorial Form
In order to express the implicit derivatives (4.191) in combinatorial form, Faà di Bruno’s formula will be used. According to this formula, the following equation holds for a function \( g = g(y) \) with \( y = y(x) \):Footnote 94
with \( {\alpha_p} = \frac{{m!}}{{{{(1!)}^{{e_1}}} \cdot {e_1}!\, \cdot ... \cdot {{(m!)}^{{e_m}}} \cdot {e_m}!}} \), \( \frac{{{d^{\left| p \right|}}g}}{{d{y^{\left| p \right|}}}} \) as ordinary \( {\left| p \right|^{\text{th}}} \) derivative, and
Proposition
Equation (4.191) is equivalent to
Proof
For ease of notation, it will be assumed that \( {z_0} = {w_0} = 0 \), so that \( G(0,0) = 0 \). With \( {{{\partial \ln G}} \left/ {{\partial z}} \right.} = {{{{G_z}}} \left/ {G} \right.} \), (4.191) is equivalent to
The mth derivative of \( \ln G \) can be calculated using Faà di Bruno’s formula:
with \( {{{{\partial^p}G}} \left/ {{\partial {z^p}}} \right.} = :{G_{z,p}} \). This leads to
According to (4.170), the residue of a function h(w) in w 0, with w 0 being a pole of order r, can be calculated as
With \( r = \left| p \right| \), we obtain for the derivative (4.197)
Using the Leibniz identity for arbitrary-order derivatives of products of functions, we get:Footnote 95
As a next step, the derivative \( \frac{{{\partial^s}}}{{\partial {w^s}}}{\left( {\frac{{G(0,w)}}{w}} \right)^{ - \left| p \right|}} \) contained in (4.200) will be calculated. Performing a Taylor series expansion of \( G(0,w) \) at \( w = 0 \), we have
Thus, for \( {{{G(0,w)}} \left/ {w} \right.} \), we obtain
with \( {\varphi_r} = \frac{1}{{r + 1}} \cdot \frac{{{{{{\partial^{r + 1}}}} \left/ {{\partial {w^{r + 1}}G(0,0)}} \right.}}}{{{{\partial } \left/ {{\partial w}} \right.}G(0,0)}}. \) Another application of Faà di Bruno’s formula results in:Footnote 96
withFootnote 97
Applying (4.203) and (4.204) to (4.200) leads to
Summarizing the sums, using \( ( - 1) \cdot {( - 1)^{\left| p \right| - 1}} \cdot {( - 1)^{\left| u \right|}} = {( - 1)^{\left| p \right| + \left| u \right|}} \), and
(4.205) can be simplified to
which concludes the proof.
4.1.6.2.4 Completion of the Derivation
Application of (4.207) can be used to determine the derivatives of a quantile, which will be calculated subsequently. With \( F(q(\lambda ),\lambda ) - \alpha = 0 = G(w(z),z) \), the derivatives are given as
where the right-hand side can be determined with (4.207). The derivatives of G contained in (4.207) can be calculated with (4.172):
where we define \( {\mu_{s,c}}: = b{E}({\tilde{Z}^s}|\tilde{Y} = y) \) and \( f: = {f_Y}(y) \) for convenience. Using definition (4.193) for the pth derivative with \( p \prec m \), this leads to
Similarly the \( {\hat{u}^{\text{th}}} \) derivative can be determined with \( u \prec s \). It has to be considered that for each partition u the elements of the corresponding partition \( \hat{u} \) are increased by 1. Thus, the smallest number is 2 and the largest is \( s + 1 \). Hence, we obtain
Furthermore, we have \( {G_w} = {{{dF}} \left/ {{dy}} \right.} = f \) and \( {( - 1)^{\left| p \right| + \left| u \right|}} \cdot {f^{\left| p \right| + \left| u \right|}} = {( - f)^{\left| p \right| + \left| u \right|}} \). Using these formulas, we finally get for (4.207) or (4.208):
which is the formula for arbitrary derivatives of VaR. Written without abbreviations this is
with \( {\alpha_p} = \frac{{m!}}{{{{(1!)}^{{e_p}_1}}{e_p}_{,1}!\, \cdot ... \cdot {{(m!)}^{{e_{p,m}}}}{e_{p,m}}!}} \).
4.1.7 Determination of the First Five Derivatives of VaR
The general form of the mth derivative of VaR is given by (4.213). Subsequently, the first five derivatives will be determined with this formula. For each derivative, we have summands for all partitions \( p \prec m \) and \( u \prec s \leq \left| p \right| - 1 \). For the considered cases \( 1 \leq m \leq 5 \), the following partitions \( p \prec m \) exist:
By construction, the expectation of the unsystematic loss is zero:
which is called the “granularity adjustment condition”. Consequently, for all partitions with \( {e_{p1}} \ne 0 \), the summands of (4.213) are zero, too:
Hence, the only relevant partitions \( p \prec m \) of (4.214) with non-zero terms and the corresponding numbers \( \left| p \right| \) are given asFootnote 98
For the associated terms
we obtain
According to (4.217), we only have \( \left| p \right| = 1 \) and \( \left| p \right| = 2 \), leading to the following partitions \( u \prec s \leq \left| p \right| - 1 \):
As we have one summand for each \( p \prec m \) and \( u \prec s \leq (|p| - 1) \), we obtain one summand for \( m = 1,\;2,\;3 \) and three summands for \( m = 4,\;5 \):
where the summands are determined with the following variables:
The first summand (I), with \( p = {m^1},\;\left| p \right| = 1,\;s = 0,\;u = 0,\left| u \right| = 0,\;\hat{u} = {1^1} \), \( {e_{pm}} = 1 \), and \( {e_{pi}} = 0 \) for all \( i \ne m \), equals:Footnote 99
For \( m = 4 \), the second summand (II.[4]), with values \( p = {2^2},\;\left| p \right| = 2,\;s = 0,\;u = 0,\left| u \right| = 0,\;\hat{u} = {1^1} \), \( {e_{p2}} = 2 \), and \( {e_{pi}} = 0 \) for all \( i \ne 2 \), is equivalent to
For \( m = 5 \), we have \( p = {2^1}{3^1},\;\left| p \right| = 2,\;s = 0,\;u = 0,\left| u \right| = 0,\;\hat{u} = {1^1},\;{e_{p2}} = 1,\;{e_{p3}} = 1, \) and \( {e_{pi}} = 0 \) for all \( i \ne 2,3 \), leading to
The third summand for \( m = 4 \) (III.[4]), with \( p = {2^2},\;\left| p \right| = 2,\;s = 1,\;u = {1^1},\;\left| u \right| = 1,\;\hat{u} = {2^1},\;{e_{p2}} = 2, \) \( {e_{pi}} = 0 \) for all \( i \ne 2 \), and \( {e_{u1}} = 1 \) equals
For \( m = 5 \), we have \( p = {2^1}{3^1},\;\left| p \right| = 2,\;s = 1,\;u = {1^1},\;\left| u \right| = 1,\;\hat{u} = {2^1},\;{e_{p2}} = 1,\;{e_{p3}} = 1, \) \( {e_{pi}} = 0 \) for all \( i \ne 2,3 \), and \( {e_{u1}} = 1 \). Hence, we get
Summing up the relevant elements from (4.223) to (4.227) and multiplying by \( {( - 1)^m} \) leads to
and
Comparing these terms, we find that the derivatives for \( m = 1,\;...,\;5 \) can be written as
or without abbreviations as
with \( \kappa (1) = \kappa (2) = 0,\,\,\kappa (3) = 1,\,\,\kappa (4) = 3 \), and \( \kappa (5) = 10 \), which is the result of Wilde (2003).
4.1.8 Order of the Derivatives of VaR
For any \( m \in b{N} \), the (m+1)th element of the Taylor series can be written asFootnote 100
with g being a function that is independent of the number of credits n. With \( {\mu_i} \) as the ith moment about the origin and \( {\eta_i} \) as the ith moment about the mean, it is possible to writeFootnote 101
for each m. Thus, the derivatives are given as
Due toFootnote 102
with \( {\eta_i}\left( {\tilde{L}|\tilde{x} = x} \right) = {\eta_i}^*(x) \cdot \sum\limits_{j = 1}^n {{w_j}^i} \leq {\eta_i}^*(x) \cdot {\left( {\frac{b}{a}} \right)^i} \cdot \frac{1}{{{n^{i - 1}}}} = O\left( {\frac{1}{{{n^{i - 1}}}}} \right), \) for all i, and revisiting (4.235) and (4.236), it is straightforward to see that only for \( 0 < a \leq EA{D_i} \leq b \) and \( m = 3 \) there exist terms which are at maximum of order O(1/n 2):
All terms with higher derivatives of VaR are at least of Order O(1/n 3).
4.1.9 VaR-Based Second-Order Granularity Adjustment for a Normally Distributed Systematic Factor
For convenience, the summands of the second-order granularity add-on \( \begin{array} {c} \sum\limits_{p \prec 3} {\prod\limits_{i = 1}^3 {{{\left( {{\eta_i}\left[ {\tilde{L}|\tilde{Y} = y} \right]} \right)}^{{e_{pi}}}}} } = {\eta_3}\left[ {\tilde{L}|\tilde{Y} = y} \right] = O\left( {\frac{1}{{{n^2}}}} \right), \\ \sum\limits_{p \prec 4} {\prod\limits_{i = 1}^4 {{{\left( {{\eta_i}\left[ {\tilde{L}|\tilde{Y} = y} \right]} \right)}^{{e_{pi}}}}} } = {\eta_4}\left[ {\tilde{L}|\tilde{Y} = y} \right] + {\left( {{\eta_2}\left[ {\tilde{L}|\tilde{Y} = y} \right]} \right)^2} = O\left( {\frac{1}{{{n^3}}}} \right) + O\left( {\frac{1}{{{n^2}}}} \right). \\ \end{array} \) will be calculated separately:
The term \( \begin{array} {c} \Delta {l_2} = \frac{1}{{6\varphi }}\frac{d}{{dx}}\left( {\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}\frac{d}{{dx}}\left[ {\frac{{{\eta_{3,c}}\varphi }}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right]} \right) \\ + \frac{1}{{8\varphi }}\frac{d}{{dx}}{\left. {\left[ {\frac{1}{\varphi }\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}{{\left( {\frac{d}{{dx}}\left[ {\frac{{{\eta_{2,c}}\varphi }}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right]} \right)}^2}} \right]} \right|_{x = {\Phi^{ - 1}}(1 - \alpha )}} \\ = :{\left. {\Delta {l_{2,1}} + \Delta {l_{2,2}}} \right|_{x = {\Phi^{ - 1}}(1 - \alpha )}}. \\ \end{array} \) equals
For the calculation, we need the first and second derivative of the density function \( \begin{array} {c} \Delta {l_{2,1}} = \frac{1}{6}\left[ {\frac{d}{{dx}}\left( {\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)\frac{1}{\varphi }\frac{d}{{dx}}\left( {\frac{{{\eta_{3,c}}\varphi }}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right) + \frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}\frac{1}{\varphi }\frac{{{d^2}}}{{d{x^2}}}\left( {\frac{{{\eta_{3,c}}\varphi }}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)} \right] \\ = \frac{1}{6}\left[ {\frac{d}{{dx}}\left( {\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)\left( {\underbrace {\frac{1}{\varphi }\frac{d}{{dx}}\left( {{\eta_{3,c}}\varphi } \right)}_{ = :A}\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + {\eta_{3,c}}\frac{d}{{dx}}\left( {\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)} \right)} \right. \\ + \frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}\frac{1}{\varphi }\frac{d}{{dx}}\left[ {\underbrace {\frac{d}{{dx}}\left( {{\eta_{3,c}}\varphi } \right)\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}}_{ = :B} + \underbrace {{\eta_{3,c}}\varphi \frac{d}{{dx}}\left( {\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)}_{ = :C}} \right]. \\ \end{array} \). As the systematic factor is assumed to be normally distributed, we have
Furthermore, we need the derivative
Herewith, the term A form (4.240) can easily be calculated:
Furthermore, \( A = \frac{1}{\varphi }\frac{d}{{dx}}\left( {{\eta_{3,c}}\varphi } \right) = \frac{{d{\eta_{3,c}}}}{{dx}} + \frac{{{\eta_{3,c}}}}{\varphi }\frac{{d\varphi }}{{dx}} = \frac{{d{\eta_{3,c}}}}{{dx}} - {\eta_{3,c}}x. \) is equal to
Similarly, \( \begin{array} {c} \frac{{dB}}{{dx}} = \frac{d}{{dx}}\left( {\frac{d}{{dx}}\left( {{\eta_{3,c}}\varphi } \right)\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right) \\ = \frac{{{d^2}}}{{d{x^2}}}\left( {{\eta_{3,c}}\varphi } \right)\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + \frac{d}{{dx}}\left( {{\eta_{3,c}}\varphi } \right)\frac{d}{{dx}}\left( {\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right) \\ = \frac{d}{{dx}}\left( {\frac{{d{\eta_{3,c}}}}{{dx}}\varphi + {\eta_{3,c}}\frac{{d\varphi }}{{dx}}} \right)\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + \left( {\frac{{d{\eta_{3,c}}}}{{dx}}\varphi + {\eta_{3,c}}\frac{{d\varphi }}{{dx}}} \right)\left( { - \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right) \\ = \left( {\frac{{{d^2}{\eta_{3,c}}}}{{d{x^2}}}\varphi + 2\frac{{d{\eta_{3,c}}}}{{dx}}\frac{{d\varphi }}{{dx}} + {\eta_{3,c}}\frac{{{d^2}\varphi }}{{d{x^2}}}} \right)\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} \\ - \left( {\frac{{d{\eta_{3,c}}}}{{dx}}\varphi + {\eta_{3,c}}\frac{{d\varphi }}{{dx}}} \right)\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}. \\ \end{array} \) is equivalent to
Using these terms, \( \begin{array} {c} \frac{{dC}}{{dx}} = \frac{d}{{dx}}\left( {{\eta_{3,c}}\varphi \left( { - \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right)} \right) \\ = - \frac{d}{{dx}}\left( {{\eta_{3,c}}\varphi } \right)\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}} - {\eta_{3,c}}\varphi \frac{d}{{dx}}\left( {\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right) \\ = \left( { - \frac{{d{\eta_{3,c}}}}{{dx}}\varphi - {\eta_{3,c}}\frac{{d\varphi }}{{dx}}} \right)\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}} \\ - {\eta_{3,c}}\varphi \left( {\frac{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}\left( {{{{{d^3}{\mu_{1,c}}}} \left/ {{d{x^3}}} \right.}} \right) - 2\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right){{\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}^2}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^4}}}} \right). \\ \end{array} \) results in
Applying the derivatives of \( \begin{array} {c} \Delta {l_{2,1}} = \frac{1}{6}\left[ { - \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}\left( {\frac{{{{{d{\eta_{3,c}}}} \left/ {{dx}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} - \frac{{{\eta_{3,c}}x}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} - {\eta_{3,c}}\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right)} \right. \\ + \frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}\frac{1}{\varphi }\left[ {\left( {\frac{{{d^2}{\eta_{3,c}}}}{{d{x^2}}}\varphi + 2\frac{{d{\eta_{3,c}}}}{{dx}}\frac{{d\varphi }}{{dx}} + {\eta_{3,c}}\frac{{{d^2}\varphi }}{{d{x^2}}}} \right)\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right. \\ - 2\left( {\frac{{d{\eta_{3,c}}}}{{dx}}\varphi + {\eta_{3,c}}\frac{{d\varphi }}{{dx}}} \right)\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}} \\ \left. { - {\eta_{3,c}}\varphi \left( {\frac{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}\left( {{{{{d^3}{\mu_{1,c}}}} \left/ {{d{x^3}}} \right.}} \right) - 2\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right){{\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}^2}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^4}}}} \right)} \right]. \\ \end{array} \) from (4.242) and (4.243) leads to
Henceforward, the summand \( \begin{array} {c} \Delta {l_{2,1}} = \frac{1}{6}\left[ { - 3\frac{{\left( {{{{d{\eta_{3,c}}}} \left/ {{dx}} \right.}} \right)\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^3}}} + 3\frac{{{\eta_{3,c}}x\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^3}}} + 3{\eta_{3,c}}\frac{{{{\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}^2}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^4}}}} \right. \\ + \frac{{{{{{d^2}{\eta_{3,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}} - 2x\frac{{{{{d{\eta_{3,c}}}} \left/ {{dx}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}} + \frac{{{\eta_{3,c}}\left( {{x^2} - 1} \right)}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}\left. { - {\eta_{3,c}}\frac{{{{{{d^3}{\mu_{1,c}}}} \left/ {{d{x^3}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^3}}}} \right] \\ = \frac{1}{{6{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}\left[ {{\eta_{3,c}}\left( {{x^2} - 1 - \frac{{{{{{{{d^3}{\mu_{1,c}}}} \left/ {{dx}} \right.}}^3}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + \frac{{3x\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + \frac{{3{{\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}^2}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right)} \right. \\ \left. { + \frac{{d{\eta_{3,c}}}}{{dx}}\left( { - 2x - \frac{{3\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right) + \frac{{{d^2}{\eta_{3,c}}}}{{d{x^2}}}} \right]. \\ \end{array} \) will be simplified:
The term (*) is the negative twice of the first-order granularity adjustment, so that we can use the resulting equation (4.18). This leads to
Using the derivative of a normal distribution \( \begin{array} {c} \Delta {l_{2,2}} = \frac{1}{{8\varphi }}\frac{d}{{dx}}\left( {\frac{\varphi }{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}{{\left[ { - \frac{{x\,{\eta_{2,c}}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + \frac{{{{{d{\eta_{2,c}}}} \left/ {{dx}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} - \frac{{{\eta_{2,c}}{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right]}^2}} \right) \\ = \frac{1}{8}\left[ {\underbrace {\frac{1}{\varphi }\frac{d}{{dx}}\left( {\frac{\varphi }{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^3}}}} \right)}_{ = :(I)}{{\left( { - x\,{\eta_{2,c}} + \frac{{d{\eta_{2,c}}}}{{dx}} - \frac{{{\eta_{2,c}}{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)}^2}} \right. \\ \left. { + \frac{1}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^3}}}\underbrace {\frac{d}{{dx}}\left( {{{\left[ { - x\,{\eta_{2,c}} + \frac{{d{\eta_{2,c}}}}{{dx}} - \frac{{{\eta_{2,c}}{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right]}^2}} \right)}_{ = :(II)}} \right]. \\ \end{array} \), the term (I) is equivalent to
Term (II) can be written as
Using these expressions, \( \begin{array} {c} (II) = \frac{d}{{dx}}\left( {{{\left[ { - x\,{\eta_{2,c}} + \frac{{d{\eta_{2,c}}}}{{dx}} - \frac{{{\eta_{2,c}}{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right]}^2}} \right) \\ = 2\left( { - x\,{\eta_{2,c}} + \frac{{d{\eta_{2,c}}}}{{dx}} - \frac{{{\eta_{2,c}}{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)\left( { - {\eta_{2,c}} - x\frac{{d{\eta_{2,c}}}}{{dx}} + \frac{{{d^2}{\eta_{2,c}}}}{{d{x^2}}}} \right. \\ \left. { - \frac{d}{{dx}}\left( {{\eta_{2,c}}\frac{{{d^2}{\mu_{1,c}}}}{{d{x^2}}}} \right)\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} - {\eta_{2,c}}\frac{{{d^2}{\mu_{1,c}}}}{{d{x^2}}}\frac{d}{{dx}}\left( {\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)} \right) \\ = 2\left( { - x{\eta_{2,c}} + \frac{{d{\eta_{2,c}}}}{{dx}} - \frac{{{\eta_{2,c}}{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)\left( { - {\eta_{2,c}} - x\frac{{d{\eta_{2,c}}}}{{dx}} + \frac{{{d^2}{\eta_{2,c}}}}{{d{x^2}}}} \right. \\ \left. { - \frac{{d{\eta_{2,c}}}}{{dx}}\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} - {\eta_{2,c}}\frac{{{{{{d^3}{\mu_{1,c}}}} \left/ {{d{x^3}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + {\eta_{2,c}}\frac{{{d^2}{\mu_{1,c}}}}{{d{x^2}}}\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right). \\ \end{array} \) from (4.251) is equal to
which leads to
Adding the terms \( \begin{array} {c} \Delta {l_{2,2}} = \frac{1}{{8{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^3}}}{\left[ {\left( { - x - 3\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)\left( {{\eta_{2,c}}\left[ { - x - \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right] + \frac{{d{\eta_{2,c}}}}{{dx}}} \right)} \right.^2} \\ + 2\left( {{\eta_{2,c}}\left[ {x + \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right] - \frac{{d{\eta_{2,c}}}}{{dx}}} \right)\left( {{\eta_{2,c}}\left[ {1 + \frac{{{{{{d^3}{\mu_{1,c}}}} \left/ {{d{x^3}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} - \frac{{{{\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}^2}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right]} \right. \\ \left. {\left. { + \frac{{d{\eta_{2,c}}}}{{dx}}\left[ {x + \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right] - \frac{{{d^2}{\eta_{2,c}}}}{{d{x^2}}}} \right)} \right]. \\ \end{array} \) and \( \Delta {l_{2,1}} \) together results in
4.1.10 Third Conditional Moment of Losses
Subsequently, the third conditional moment of the portfolios loss about the mean, \( \begin{array} {c} \Delta {l_2} = \frac{1}{{6{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}\left[ {{\eta_{3,c}}\left( {{x^2} - 1 - \frac{{{{{{{{d^3}{\mu_{1,c}}}} \left/ {{dx}} \right.}}^3}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + \frac{{3x\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + \frac{{3{{\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}^2}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right)} \right. \\ \left. { + \frac{{d{\eta_{3,c}}}}{{dx}}\left( { - 2x - \frac{{3\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right) + \frac{{{d^2}{\eta_{3,c}}}}{{d{x^2}}}} \right] \\ + \frac{1}{{8{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^3}}}{\left[ {\left( { - x - 3\frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)\left( {{\eta_{2,c}}\left[ { - x - \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right] + \frac{{d{\eta_{2,c}}}}{{dx}}} \right)} \right.^2} \\ + 2\left( {{\eta_{2,c}}\left[ {x + \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right] - \frac{{d{\eta_{2,c}}}}{{dx}}} \right)\left( {{\eta_{2,c}}\left[ {1 + \frac{{{{{{d^3}{\mu_{1,c}}}} \left/ {{d{x^3}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} - \frac{{{{\left( {{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}} \right)}^2}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right]} \right. \\ {\left. {\left. {\left. { + \frac{{d{\eta_{2,c}}}}{{dx}}\left[ {x + \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right] - \frac{{{d^2}{\eta_{2,c}}}}{{d{x^2}}}} \right)} \right]} \right|_{x = {\Phi^{ - 1}}\left( {1 - \alpha } \right)}}. \\ \end{array} \), shall be expressed in terms of the moments of separated factors \( {\eta_{3,c}} = {\eta_3}(\tilde{L}|\tilde{x} = x) \) and \( \widetilde{{LG{D_i}}} \). With
which is due to the conditional independence property, we need to determine \( \begin{array} {c} {\eta_{3,c}} = {\eta_3}\left( {\tilde{L}|\tilde{x} = x} \right) \\ = {\eta_3}\left( {\sum\limits_{i = 1}^n {{w_i} \cdot \widetilde{{LG{D_i}}} \cdot {1_{\left\{ {{{\tilde{D}}_i}} \right\}}}} |\tilde{x} = x} \right) \\ = \sum\limits_{i = 1}^n {{w_i}^3 \cdot {\eta_3}\left( {\widetilde{{LG{D_i}}} \cdot {1_{\left\{ {{{\tilde{D}}_i}} \right\}}}|\tilde{x} = x} \right)}, \\ \end{array} \). In general, the third moment about the mean is equal to
Thus, the conditional moment \( \begin{array} {c} {\eta_3}\left( {\tilde{X}} \right) = b{E}\left( {{{\left[ {\tilde{X} - b{E}\left( {\tilde{X}} \right)} \right]}^3}} \right) \\ = b{E}\left[ {{{\tilde{X}}^3} - 3{{\tilde{X}}^2}b{E}\left( {\tilde{X}} \right) + 3\tilde{X}{b{E}^2}\left( {\tilde{X}} \right) - {b{E}^3}\left( {\tilde{X}} \right)} \right] \\ = b{E}\left( {{{\tilde{X}}^3}} \right) - 3b{E}\left( {{{\tilde{X}}^2}} \right)b{E}\left( {\tilde{X}} \right) + 3b{E}\left( {\tilde{X}} \right){b{E}^2}\left( {\tilde{X}} \right) - {b{E}^3}\left( {\tilde{X}} \right) \\ = b{E}\left( {{{\tilde{X}}^3}} \right) - 3b{E}\left( {{{\tilde{X}}^2}} \right)b{E}\left( {\tilde{X}} \right) + 2{b{E}^3}\left( {\tilde{X}} \right). \\ \end{array} \) can be written as
Using the conditional independence property again, considering that the LGDs are assumed to be stochastically independent of each other, and with \( \begin{array} {c} {\eta_3}\left( {\widetilde{{LG{D_i}}} \cdot {1_{\left\{ {{{\tilde{D}}_i}} \right\}}}|\tilde{x}} \right) = b{E}\left( {{{\left[ {\widetilde{{LGD}} \cdot {1_{\left\{ {{{\tilde{D}}_i}} \right\}}}|\tilde{x}} \right]}^3}} \right) - 3\,b{E}\left( {{{\left[ {\widetilde{{LGD}} \cdot {1_{\left\{ {{{\tilde{D}}_i}} \right\}}}|\tilde{x}} \right]}^2}} \right) \cdot b{E}\left( {\widetilde{{LG{D_i}}} \cdot {1_{\left\{ {{{\tilde{D}}_i}} \right\}}}|\tilde{x}} \right) \\ + 2\,{b{E}^3}\left( {\widetilde{{LG{D_i}}} \cdot {1_{\left\{ {{{\tilde{D}}_i}} \right\}}}|\tilde{x}} \right). \\ \end{array} \), we have
With the abbreviations \( \begin{array} {c} {\eta_3}\left( {\widetilde{{LG{D_i}}} \cdot {1_{\left\{ {{{\tilde{D}}_i}} \right\}}}|\tilde{x}} \right) = b{E}\left( {{{\left[ {\widetilde{{LGD}}|\tilde{x}} \right]}^3}} \right)p\left( {\tilde{x}} \right) - 3\,b{E}\left( {{{\left[ {\widetilde{{LGD}}|\tilde{x}} \right]}^2}} \right)b{E}\left( {\widetilde{{LGD}}|\tilde{x}} \right){p^2}\left( {\tilde{x}} \right) \\ + 2\,{b{E}^3}\left( {\widetilde{{LGD}}|\tilde{x}} \right){p^3}\left( {\tilde{x}} \right) \\ = b{E}\left( {{{\widetilde{{LGD}}}^3}} \right)p\left( {\tilde{x}} \right) - 3\,b{E}\left( {{{\widetilde{{LGD}}}^2}} \right)b{E}\left( {\widetilde{{LGD}}} \right){p^2}\left( {\tilde{x}} \right) \\ + 2\,{b{E}^3}\left( {\widetilde{{LGD}}} \right){p^3}\left( {\tilde{x}} \right). \\ \end{array} \), \( ELGD = b{E}(\widetilde{{LGD}}) \) as well as \( VLGD = b{V}(\widetilde{{LGD}}) \) and using (4.258) again, we obtain
Consequently, (4.260) is equivalent to
Thus, the conditional moment of the portfolio loss (4.257) can finally be written as
4.1.11 Difference Between the VaR Definitions
For the case of homogeneous credits and with \( \begin{array} {c} {\eta_{3,c}} = \sum\limits_{i = 1}^n {{w_i}^3 \cdot {\eta_3}\left( {\widetilde{{LG{D_i}}} \cdot {1_{\left\{ {{{\tilde{D}}_i}} \right\}}}\left| {\widetilde{x} = x} \right.} \right)} \\ = \sum\limits_{i = 1}^n {{w_i}^3\left[ {\left( {ELG{D_i}^3 + 3 \cdot ELG{D_i} \cdot VLG{D_i} + SLG{D_i}} \right) \cdot {p_i}(x)} \right.} \\ \left. { - 3 \cdot \left( {ELG{D_i}^3 + ELG{D_i} \cdot VLG{D_i}} \right) \cdot {p_i}^2(x) + 2 \cdot ELG{D_i}^3 \cdot {p_i}^3(x)} \right]. \\ \end{array} \), the possible realizations of losses are
which implies
If we define \( b{P}\left[ {\tilde{L} \leq l} \right] = b{P}\left[ {\tilde{L} < \left( {l + {{1} \left/ {n} \right.}} \right)} \right] \), we get
4.1.12 Identity of ES Within the Basel Framework
Using the result of the ASRF framework (2.93), the definition of the ES (2.19), the integral representation of the conditional expectation, and the identity of the condition as in (4.9), the ES of the portfolio loss equals
With the conditional independence property as in (2.92), the conditional PD of the Vasicek model (2.66), the integral representation (2.126), and the symmetry of the normal distribution, the ES can be written as
4.1.13 Arbitrary Derivatives of ES
According to (2.20), the ES can be written as
Thus, for continuous distributions, all derivatives of ES can be expressed as
The derivative of VaR is a function of \( \frac{{{d^m}E{S_\alpha }}}{{d{\lambda^m}}} = \frac{{{d^m}}}{{d{\lambda^m}}}\left( {\frac{1}{{1 - \alpha }}\int\limits_\alpha^1 {{q_u}du} } \right) = \frac{1}{{1 - \alpha }}\int\limits_\alpha^1 {\frac{{{d^m}{q_u}}}{{d{\lambda^m}}}du} . \) and \( {f_Y}(y) \) evaluated at \( {\mu_{i,c}}(y) \). The substitution \( {q_u}(\tilde{Y}) \), so that \( u = {F_Y}(y) \), \( {{{du}} \left/ {{dy}} \right.} = {f_Y}(y) \), and \( y(u = \alpha ) = F_Y^{ - 1}(\alpha ) = {q_\alpha }(\tilde{Y}) \), leads to:Footnote 103
where the expression resulting from the derivative of VaR simply has to be evaluated at y since \( {\left. {\frac{{{d^m}E{S_\alpha }}}{{d{\lambda^m}}}} \right|_{\lambda = 0}} = \frac{1}{{1 - \alpha }}\int\limits_{u = \alpha }^1 {{{\left. {\frac{{{d^m}{q_u}}}{{d{\lambda^m}}}} \right|}_{\lambda = 0}}du} = \frac{1}{{1 - \alpha }}\int\limits_{y = {q_\alpha }\left( {\tilde{Y}} \right)}^\infty {{{\left. {\frac{{{d^m}{q_u}}}{{d{\lambda^m}}}} \right|}_{\lambda = 0}}{f_Y}dy}, \). Using the derivatives of VaR from (4.212), this leads to
with \(\begin{array}{ll} \left.\frac{{d^m} E{S_\alpha }}{{d{\lambda^m}}}\right|_{\lambda = 0} & = \frac{1}{{1 - \alpha }} \int\limits_{y = {q_\alpha }\left( {\tilde{Y}} \right)}^\infty{{( - 1)}^m} \left[\sum\limits_{ p \prec m, u \prec s \leq |p| - 1 } \right. {\frac{{{\alpha_p}{\alpha_{\hat{u}}}\left( {\left| p \right| + \left| u \right| - 1} \right)!}}{{\left( {s + \left| u \right|} \right)!\left( {\left| p \right| - 1 - s} \right)!}}} \\ & \cdot{{\left( { - f} \right)}^{ - \left| p \right| - \left| u \right|}} \cdot \left( {\prod\limits_{i = 1}^s {{{\left[ {\frac{{{d^i}f}}{{d{y^i}}}} \right]}^{{e_{ui}}}}} } \right)\left. { \cdot \frac{{{d^{\left| p \right| - 1 - s}}}}{{d{y^{\left| p \right| - 1 - s}}}}\left( {\prod\limits_{i = 1}^m {{{\left[ {\frac{{{d^{i - 1}}\left( {{\mu_{i,c}}f} \right)}}{{d{y^{i - 1}}}}} \right]}^{{e_{pi}}}}} } \right)} \right]f\,dy,\end{array}\).
4.1.14 Determination of the First Five Derivatives of ES
Instead of solving the integral (4.272) for each of the derivatives of VaR (4.228)–(4.232), we will directly evaluate the integral for the first five derivatives. Using the expression for the first five derivatives of VaR (4.233), we obtain
This term is equal to
or written without abbreviations as
with \( \begin{array} {c} {\left. {\frac{{{d^m}E{S_\alpha }\left( {\tilde{Y} + \lambda \tilde{Z}} \right)}}{{d{\lambda^m}}}} \right|_{\lambda = 0}} = {\left( { - 1} \right)^m} \cdot \frac{1}{{1 - \alpha }} \cdot \left( {\frac{{{d^{m - 2}}\left( {{\mu_m}\left( {\tilde{Z}|\tilde{Y} = y} \right){f_Y}(y)} \right)}}{{d{y^{m - 2}}}}} \right. \\ - \kappa (m) \cdot \left[ {\frac{1}{{{f_Y}(y)}}} \right.{\left. {\left. {\left. { \cdot \frac{{d\left( {{\mu_2}\left( {\tilde{Z}|\tilde{Y} = y} \right){f_Y}(y)} \right)}}{{dy}} \cdot \frac{{{d^{m - 3}}\left( {{\mu_{m - 2}}\left( {\tilde{Z}|\tilde{Y} = y} \right){f_Y}(y)} \right)}}{{d{y^{m - 3}}}}} \right]} \right)} \right|_{y = {q_\alpha }\left( {\tilde{Y}} \right)}}, \\ \end{array} \), and \( \kappa (1) = \kappa (2) = 0,\,\,\kappa (3) = 1,\,\,\kappa (4) = 3 \). This is the result of Wilde (2003), except that the algebraic signs of Wilde (2003) seem to be wrong.
4.1.15 ES-Based Second-Order Granularity Adjustment for a Normally Distributed Systematic Factor
The summands of the second-order granularity add-on \( \kappa (5) = 10 \) can be expressed as
Using the derivative of the normal distribution (4.242), the summand \( \begin{array} {c} \Delta {l_2} = \frac{1}{{6\left( {1 - \alpha } \right)}}\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}\frac{d}{{dx}}\left( {\frac{{{\eta_{3,c}}\varphi }}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right) \\ + \frac{1}{{8\left( {1 - \alpha } \right)}}\frac{1}{\varphi }\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}{\left. {{{\left[ {\frac{d}{{dx}}\left( {\frac{{{\eta_{2,c}}\varphi }}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)} \right]}^2}} \right|_{x = {\Phi^{ - 1}}(1 - \alpha )}} \\ = :{\left. {\Delta {l_{2,1}} + \Delta {l_{2,2}}} \right|_{x = {\Phi^{ - 1}}(1 - \alpha )}}. \\ \end{array} \) equals
Using the same transformations, the summand \( \begin{array} {c} \Delta {l_{2,1}} = \frac{1}{{6\left( {1 - \alpha } \right)}}\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}\frac{d}{{dx}}\left( {\frac{{{\eta_{3,c}}\varphi }}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right) \\ = \frac{1}{{6\left( {1 - \alpha } \right)}}\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}\left[ {\frac{d}{{dx}}\left( {{\eta_{3,c}}\varphi } \right)\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} + {\eta_{3,c}}\varphi \frac{d}{{dx}}\left( {\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)} \right] \\ = \frac{1}{{6\left( {1 - \alpha } \right)}}\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}\left[ {\left( {\frac{{d{\eta_{3,c}}}}{{dx}}\varphi + {\eta_{3,c}}\frac{{d\varphi }}{{dx}}} \right)\frac{1}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}} - {\eta_{3,c}}\varphi \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}} \right] \\ = \frac{1}{{6\left( {1 - \alpha } \right)}}\frac{\varphi }{{{{\left( {{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}} \right)}^2}}}\left[ {\frac{{d{\eta_{3,c}}}}{{dx}} - {\eta_{3,c}}\left( {x - \frac{{{{{{d^2}{\mu_{1,c}}}} \left/ {{d{x^2}}} \right.}}}{{{{{d{\mu_{1,c}}}} \left/ {{dx}} \right.}}}} \right)} \right]. \\ \end{array} \) is equivalent to
leading to a second-order adjustment of
4.1.16 Probability Density Function of the Logit-Normal Distribution
The derivation of the density function is based on the inverse function theoremFootnote 104
For the logit function \( {f_Y}(y) = {f_X}\left( {{g^{ - 1}}(y)} \right) \cdot \left| {\frac{{d{g^{ - 1}}(y)}}{{dy}}} \right|. \), we have
and
Using the density of a normal distribution (4.82) for \( \frac{{d{g^{ - 1}}(y)}}{{dy}} = \frac{d}{{dy}}\left( { - \ln \left( {\frac{1}{y} - 1} \right)} \right) = - \frac{1}{{\frac{1}{y} - 1}} \cdot \left( { - \frac{1}{{{y^2}}}} \right) = \frac{1}{{y\left( {1 - y} \right)}}. \) and recognizing that y is bounded in the interval [0, 1], we get
Rights and permissions
Copyright information
© 2010 Physica-Verlag HD
About this chapter
Cite this chapter
Hibbeln, M. (2010). Model-Based Measurement of Name 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_4
Download citation
DOI: https://doi.org/10.1007/978-3-7908-2607-4_4
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)