Vasicek Single Factor Model Vasicek Single Factor Model Alexandra Kochendörfer 7. Februar 2011 1 / 33 Vasicek Single Factor Model Problem Setting I Consider portfolio with N different credits of equal size 1. I Each obligor has an individual default probability. I In case of default of the n’th obligor we lose the whole n’th position in portfolio. I What can we say about the loss distribution? 2 / 33 Vasicek Single Factor Model Contents Default Correlation Definition Why is default correlation important Independent/perfectly dependent defaults Modelling Default Correlation Data sources Default triggered by firm’s value Vasicek Single Factor Model Loss distribution in finite portfolio Large Homogeneous Portfolio Approximation Conclusion 3 / 33 Vasicek Single Factor Model Default Correlation Definition Definition Default correlation is the phenomenon that the likelihood of one obligor defaulting on its debt is affected by whether or not another obligor has defaulted on its debts. I Positive correlation: one firm is the creditor of another I Negative correlation: the firms are competitors Drivers of Default Correlation I I State of the general economy Industry-specific factors I I I Oil industry: 22 companies defaulted over 1982-1986. Thrifts: 19 defaults over 1990-1992. Casinos/hotel chains: 10 defaults over 1990-1992. 4 / 33 Vasicek Single Factor Model Default Correlation Definition U.S. Corporate Default Rates Since 1920 5 / 33 Vasicek Single Factor Model Default Correlation Why is default correlation important Why is default correlation important? Consider, for two default events A and B I default probabilities pA and pB I joint default probability pAB I conditional default probability pA|B I correlation ρAB between default events These quantities are connected by pA|B = pAB pB 6 / 33 Vasicek Single Factor Model Default Correlation Why is default correlation important Cov (A, B) pAB − pA pB p ρAB = p =p Var (A) Var (B) pA (1 − pA )pB (1 − pB ) The default probabilities are usually very small pA = pB = p 1 p pAB = pA pB + ρAB pA (1 − pA )pB (1 − pB ) ≈ p 2 + ρAB · p ≈ ρAB · p ρAB p pA|B = pA + pA (1 − pA )pB (1 − pB ) ≈ ρAB pB The joint default probability and conditional default probability are dominated by the correlation coefficient. 7 / 33 Vasicek Single Factor Model Default Correlation Independent/perfectly dependent defaults Independent Defaults Consider N independent default events D1 , . . . , DN with pD1 = · · · = pDN = p ⇒ Number of defaults ∼ B(p, N).For N = 100, p = 0.05 p (%) 99.9(%) VaR 1 5 2 7 3 9 4 11 5 13 6 14 7 16 8 17 9 19 10 20 8 / 33 Vasicek Single Factor Model Default Correlation Independent/perfectly dependent defaults Perfectly dependent defaults Consinder default correlation ρij = 1 for all pairs ij. pD D − p 2 pD1 D2 − pD1 pD2 = 1 2 1= p p(1 − p) pD1 (1 − pD1 )pD2 (1 − pD2 ) ⇒ pD1 D2 = pD1 = pD2 = p i.e. D1 ∩ D2 = D1 = D2 a.s. ⇒ pD1 D2 D3 = pD2 D3 = p ⇒ D1 ∩ D2 ∩ D3 = D1 a.s. . . . ⇒ D1 ∩ · · · ∩ DN = D1 a.s. ⇒ pD1 ...DN = p All loans in the portfolio defaults with probability p, none with probability 1 − p. 9 / 33 Vasicek Single Factor Model Default Correlation Independent/perfectly dependent defaults Perfectly dependent defaults 10 / 33 Vasicek Single Factor Model Modelling Default Correlation Data sources Data Sources I Actual Rating and Default Events. + Objective and direct. – Joint defaults are rare events, sparse data sets. I Credit Spread. + Incorporate information on markets, observable. – No theoretical link between credit spread correlation and default correlation. I Equity correlation. + Data easily available, good quality. – Connection to credit risk not obvious, needs a lot of assumptions. 11 / 33 Vasicek Single Factor Model Modelling Default Correlation Default triggered by firm’s value Default triggered by Firm’s Value The firm value (An,t )0≤t≤1 of each obligor n ∈ {1, . . . , N} is modelled as in Black-Scholes model, hence at terminal time t = 1 with An,1 = An we have σn2 + σn Bn An = An,0 exp µn − 2 with some standard normal variable Bn . The r.v. (B1 , . . . , BN ) are jointly normally distributed with covariance matrix Σ = (ρij )ij , where ρij denotes the asset correlation between assets i and j. 12 / 33 Vasicek Single Factor Model Modelling Default Correlation Default triggered by firm’s value The obligor n defaults if the asset value falls below a perspecified barrier Cn (debts) Dn = 11{An <Cn } The default probability of the n’s debtor is pDn = P(Dn = 1) = P(An < Cn ) = P(Bn < cn ) = Φ(cn ) with default barrier cn = n − µn log ACn,0 σn We can assume the individual default probabilities pDn as given and compute cn = Φ−1 (pDn ) and vice versa. 13 / 33 Vasicek Single Factor Model Modelling Default Correlation Default triggered by firm’s value The joint distribution of Bi determines the dependency structure of default variables uniquely P(D1 = 1, . . . , DN = 1) = P(B1 < c1 , . . . , BN < cN ) = ΦN (Φ−1 (pD1 ), . . . , Φ−1 (pDn ); Σ) In case with two assets with correlation ρ1,2 = ρ2,1 , the default correlation can be computed via P(D1 = 1, D2 = 1) − pD1 pD2 ρ= p pD1 (1 − pD1 )pD2 (1 − pD2 ) = Φ2 (Φ−1 (pD1 ), Φ−1 (pD2 ); ρ1,2 ) − pD1 pD2 p pD1 (1 − pD1 )pD2 (1 − pD2 ) 14 / 33 Vasicek Single Factor Model Modelling Default Correlation Default triggered by firm’s value We need I N(N − 1)/2 asset correlations of Σ I N individual default probabilities I Additional assumptions on the structure of Bi to reduce the number of parameters. 15 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Vasicek Single Factor Model Assume, that the logarithmic return Bn can be written as p √ Bn = ρ · Y + 1 − ρ · n with some constant ρ ∈ [0, 1] and N + 1 independent standard normally distributed r.v. Y , 1 , . . . , N . Interpretation I Y is a common systematic risk factor affecting all firms (state of economy) I n are idiosyncratic factors independent across firms (management, innovations) I Corr (Bi , Bj ) = ρ controls the proportions between systematic and idiosyncratic factors, empirically around 10%. 16 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Conditional on the realisation of the systematic factor Y I the logarithmic returns Bn are independent ( for a constant y √ √ variables ρ · y + 1 − ρ · n are independent) I default variables Dn = 11{Bn <cn } are independent as function of Bn The only effect of Y is to move Bn closer or further away from barrier cn . 17 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Loss distribution in finite portfolio Theorem For ρ ∈ (0, 1) and same default probabilities p = pD1 = · · · = pDN the conditional default probability is given by −1 √ Φ (p) − ρ · y √ p(y ) := P[Bn < c | Y = y ] = Φ . 1−ρ P The number of defaults L = N i=1 Di has the following distribution m Z ∞ X N P(L ≤ m) = · p(y )k · (1 − p(y ))N−k · φ(y )dy k −∞ k=0 18 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Loss distribution in finite portfolio Proof The probability of k defaults is Z ∞ P(L = k) = E(P({L = k} | Y )) = P(L = k | Y = y )φ(y )dy , −∞ where φ is density of Y . The defaults Dn are independent conditional on Y , hence N P(L = k | Y = y ) = · p(y )k · (1 − p(y ))N−k k Thus, for m ∈ {1, . . . , N} we have m Z ∞ X N P(L ≤ m) = p(y )k · (1 − p(y ))N−k · φ(y )dy · k −∞ k=0 19 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Loss distribution in finite portfolio Loss Distibutions for different ρ 20 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Loss distribution in finite portfolio VaR Levels for different ρ with N = 100 and p = 5% ρ(%) 0 1 10 30 50 99.9(%)VaR 13 14 27 55 80 99.(%)VaR 11 12 19 35 53 Independent defaults p (%) 99.9(%) VaR 1 5 2 7 3 9 4 11 5 13 6 14 7 16 8 17 9 19 10 20 21 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation Large Homogeneous Portfolio Approximation Definition (Large Homogeneous Portfolio LHP) I pD1 = · · · = pDN = p I portfolio is weighted with ω1 , . . . , ωN , such that N X (N) lim (ωn )2 = 0 (N) N→∞ (N) (N) n=1 ωn PN = 1, n=1 The portfolio is not dominated by few loans much larger then the rest. 22 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation Definition (Loss Rate) The portfolio loss rate is defined by (N) L = N X (N) ωn Dn ∈ [0, 1] n=1 Lemma Following holds for the LHP (N) E(L | Y ) = p(Y ) = Φ Var (L(N) | Y ) = √ Φ−1 (p) − ρ · Y √ 1−ρ N X (N) (ωn )2 · p(Y ) · (1 − p(Y )) n=1 23 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation Proof Linearity of conditional expectation yields (N) E(L | Y) = = N X n=1 N X (N) ωn E(Dn | Y ) (N) ωn P(Dn | Y ) = p(Y ) n=1 N X (N) ωn = p(Y ) n=1 Dn are independent conditional on Y , thus Var (L(N) | Y ) = = N X (N) (ωn )2 Var (Dn | Y ) n=1 N X (N) (ωn )2 · p(Y ) · (1 − p(Y )) n=1 24 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation Theorem The portfolio loss rate in LHP converges in probability for N → ∞. −1 √ Φ (p) − ρ · Y P √ L(N) → p(Y ) = Φ 1−ρ Proof For the large portfolio the variation of loss rate given Y tends to 0 Var (L(N) | Y ) = ≤ N X (N) (ωn )2 · p(Y ) · (1 − p(Y )) n=1 N X 1 4 (N) (ωn )2 −−−−→ 0 n=1 N→∞ 25 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation This provides convergence in L2 : E((L(N) − p(Y ))2 ) = E((L(N) − E(L(N) | Y ))2 ) = E(E((L(N) − E(L(N) | Y ))2 | Y )) = E(Var (L(N) | Y )) −−−−→ 0 N→∞ Convergence in L2 implies convergence in probability i.e. for all > 0: lim P L(N) − p(Y ) > = 0 N→∞ The law of L(N) converges weakly to the law of p(Y ), i.e. P(L(N) ≤ x) −−−−→ P(p(Y ) ≤ x) N→∞ for all x, where the distribution function of p(Y ) is continuous. 26 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation Theorem (Approximative Distribution of Loss Rate in LHP) √ P(p(Y ) ≤ x) = Φ 1 − ρ · Φ−1 (x) − Φ−1 (p) √ ρ , x ∈ [0, 1] Proof −1 √ Φ (p) − ρ · Y √ P(p(Y ) ≤ x) = P Φ ≤x 1−ρ √ 1 − ρ · Φ−1 (x) − Φ−1 (p) =P Y ≤ √ ρ √ 1 − ρ · Φ−1 (x) − Φ−1 (p) =Φ √ ρ 27 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation Approximative density of loss rate with p = 2%, ρ = 10% 28 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation Properties of Loss Rate Distribution (N) E(p(Y )) = lim E(L N→∞ ) = lim N→∞ N X (N) ωn p = p n=1 Because of convergence we can easily compute α-Quantiles of loss rate distribution for large N √ 1 − ρ · Φ−1 (α) − Φ−1 (p) (N) P(L ≤ α) ≈ Φ √ ρ 29 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation I When ρ → 1 P(L(∞) ≤ α) = 1 − p = P(L(∞) = 0) for all α ∈ (0, 1) P(L(∞) = 1) = p All loans default with prob. p, none with 1 − p. I When ρ → 0 P(L(∞) ≤ α) = 0 for α < p P(L(∞) ≤ α) = 1 for α ≥ p ⇒ P(L(∞) = p) = 1 With the Law of Large Numbers the loss in Binomial model tends almost surly to N 1 X Di → p N i=1 30 / 33 Vasicek Single Factor Model Vasicek Single Factor Model Large Homogeneous Portfolio Approximation Simulated Loss Distibution 31 / 33 Vasicek Single Factor Model Conclusion Conclusion The Vasicek Single Factor Model provides a closed form Loss Rate Distribution (N) lim P(L N→∞ √ ≤ x) = Φ 1 − ρ · Φ−1 (x) − Φ−1 (p) √ ρ for a Large Homogeneous Portfolio, which depends only on two parameters p and ρ and gives a good fit to market data. 32 / 33 Vasicek Single Factor Model Conclusion Bibliography Vasicek : The Distribution of Loan Portfolio Value, Risk (2002). Martin, Reitz, Wehn : Kredit und Kreditrisikomkodelle, Vieweg, (2006). Schönbucher : Faktor Models: Portfolio Credit Risks When Defaults are Correlated, Journal of Risk Finance (2001). Elizalde : Credit Risk Models IV: Understanding and pricing CDOs, discussion paper (2005). 33 / 33