Pricing Default Events: Surprise, Exogeneity and Contagion

Transcription

1 1/31 Pricing Default Events: Surprise, Exogeneity and Contagion C. GOURIEROUX, A. MONFORT, J.-P. RENNE BdF-ACPR-SoFiE conference, July 4, 2014

2 2/31 Introduction When investors are averse to a given risk, a security whose payoffs are exposed to this risk are less valuable than those whose payoffs are not. A defaultable bond exposes its holder to two risks: (a) the risk that future probabilities of default change, (b) the risk that the bond issuer effectively defaults. In order to derive closed form expressions of the prices of credit derivatives, most reduced-form models of credit risk price risk (a) but not the default events themselves (risk (b)). That is, they implicitly consider that investors are not averse to the default-event surprise (or that these surprises can be diversified away).

3 3/31 Introduction A few papers mention this approximation and try to take into account the surprise, e.g.: Jarrow, Yu (2001, JoF) [ Counterparty Risk and the Pricing of Defaultable Securities ]. For 2 debtors only. A series of paper by Bai, Collin-Dufresne, Goldstein, Helwege (2013), with a very specific modeling of default dependence. In general: Default dependence is difficult to specify. Derivative prices have no closed-form expressions.

4 4/31 Introduction This paper solves this problem for credit derivatives (CDS and CDO) written on a pool of credits, which can be partitioned into J large homogenous segments. The model accommodates different forms of contagion: exposure to common factors (frailty); self-exciting defaults; contagion across sectors. Based on U.S. bond data, an application illustrates that this feature provides an explanation for the so-called credit-spread puzzle.

5 5/31 Introduction Outline of the presentation 1. Introduction. 2. The standard reduced-form approach and its limitations. 3. Modeling Framework and Derivative Pricing. 4. Applications.

6 6/31 The Standard Approach and its Limitations 2. The Standard Approach and its Limitations

7 7/31 The Standard Approach and its Limitations Notations A pool of I entitites i = 1,..., I. Default indicators d i,t : { 1 if entity i is in default at date t, d i,t = 0 otherwise. n t the number of defaults occurring at date t. t N t = n τ the cumulated number of defaults. τ=1

8 8/31 The Standard Approach and its Limitations Assumptions on the historical distribution i) Homogenous portfolio The default indicators d i,t+1 are independent, identically distributed given F t+1, d t. ii) Default dependence driven by F This conditional distribution depends on factor F t+1 only. iii) F is exogenous The conditional distribution of F t+1 given (F t, d t ) is equal to the conditional distribution of F t+1 given F t. Remark: (i) and (ii) will be relaxed in our general model.

9 9/31 The Standard Approach and its Limitations The standard pricing approach Assumption on the stochastic discount factor: m t,t+1 = m(f t+1 ). Then the price of a payoff g(n t+h ) at date t is: Π(g, h) = E t [ m t,t+1... m t+h 1,t+h g(n t+h )] where : g(f t+h ) = E[g(N t+h ) F t+h ]. Therefore : Π(g, h) = Π( g, h). = E t [ m t,t+1... m t+h 1,t+h g(f t+h )], It is equivalent to price g(n t+h ) or to price its prediction g(f t+h ).

10 10/31 The Standard Approach and its Limitations Risk premia associated with default events What is the change in pricing formula, when m t,t+1 = m(f t+1, n t+1 )? Let us consider the projected sdf: Then: m t,t+1 = E[m(F t+1, n t+1 ) F t+1 ]. Π(g, h) = Π(g, h) + Π(g g, h }{{}}{{} ). standard the price of formula surprise

11 11/31 The Standard Approach and its Limitations Relaxing the exogeneity assumption New assumption: The conditional historical distribution of F t+1 given F t, d t is equal to the conditional distribution of F t+1 given F t, n t. A more complete decomposition of the derivative price : Π(g, h) = Π(g, h) + [Π( g, h) Π( g, h)] + Π(g g, h) price = standard + causality + surprise price adjustment adjustment

12 12/31 The Standard Approach and its Limitations Moreover, we show that the standard formula for pricing a corporate zero-coupon bond: B(t, h) = ( E Q t [exp( r t... r t+h 1 )1l di,t+h=0 ] ) cannot be used in general. Default intensities = E Q t [exp( r t... r t+h 1 λ Q t+1... λq t+h )], If Ω t = (F t+1, d t), the historical intensity λ t+1 is defined by: P(d t+1 = 0 d t = 0, Ω t ) = exp( λ t+1 ). The risk-neutral intensity λ t+1 is defined by: Q(d t+1 = 0 d t = 0, Ω t ) = exp( λ Q t+1), If m t,t+1 = exp(δ 0 + δ F F t+1 + δ S n t+1 ), the risk-neutral intensity is: λ Q t+1 = λ t+1 + log{exp( λ t+1 ) + [1 exp( λ t+1 )] exp(δ S ).}

13 13/31 Modeling Framework and Derivative Pricing 3. Modeling Framework and Derivative Pricing

14 14/31 Modeling Framework and Derivative Pricing To get (quasi) closed form expressions for derivative prices, we need affine processes. The joint process (d 1t,..., d It, F t ) cannot be affine, but the aggregate process (n t, F t ) can be if the size of the homogenous pool is large.

15 15/31 Modeling Framework and Derivative Pricing Assumptions (a) A Poisson regression model for the default count: n t+1 F t+1, n t P(β F F t+1 + β n n t + γ); (b) The conditional Laplace transform of F t+1 given F t is exponential affine in (F t, n t ): E t [exp(v F t+1 )] = exp[a F (v) F t + a n (v) n t + b(v)]; (c) The s.d.f. is exponential affine in both F t+1 and n t+1 : m t,t+1 = exp(δ 0 + δ F F t+1 + δ S n t+1 ).

16 16/31 Pricing Default Events: Surprise, Exogeneity and Contagion Modeling Framework and Derivative Pricing In that setup, (F t, n t ) is jointly affine. Then the price at date t of any exponential payoff exp(un t+h ) can be derived by recursion. Since the pool is homogenous, we know also how to price: individual default d 1 (single name CDS), joint defaults d 1 d Indeed: Π(d 1... d K, h) =» 1 d K Π(exp(N log v), h). I(I 1)... (I K + 1) dv K v=1 The price of a non-exponential payoff deduced by Fourier transform [Duffie, Pan, Singleton (2000)]: CDO pricing, tranches.

17 17/31 Modeling Framework and Derivative Pricing Extension: Heterogeneous pools The pool can be partitioned into J homogenous pools, with different risk characteristics. For corporations, the segment can be defined by the industrial sector, by the size, by the domestic country, but the rating cannot be used since it is time-varying. n j,t, j = 1,..., J denote the numbers of defaults in each segment, conditionally independent : n j,t+1 P[β j F t+1 + C j n t + γ j ], j = 1,..., J. Additional contagion channel: across sectors.

18 18/31 Illustrations 4. Illustrations

19 19/31 Illustrations App.1: Contagion and network An illustration with six homogenous segments of size 100. Two types of factors: F B,t F N,t a sequence of i.i.d. Bernoulli variables processes keeping memory of past default counts in each segment F N,j,t = ρf N,j,t 1 + n j,t 1, j = 1,..., 6. The distribution of the count variables with a circular structure of the network: n 1,t+1 P(0.4F N,6,t + F B,t ), n j,t+1 P(0.4F N,j 1,t ), j = 2,..., 6.

20 20/31 Illustrations App.1: Contagion and network The next figure gives the evolutions of factors and default counts. A high value of factor F B may immediately generate defaults in segment 1. These defaults propagate to the other segments by contagion.

21 21/31 Illustrations App.1: Contagion and network

22 22/31 Illustrations App.1: Contagion and network The next figure displays the term structures of: the CDS premium, the CDS without pricing the surprise, the actuarial value (physical probability) for two dates and segments.

23 23/31 Pricing Default Events: Surprise, Exogeneity and Contagion Illustrations App.1: Contagion and network Date 1 Sect.1 Sect.2 Sect.6 Sect.3 Sect.5 Sect.4 More than 5 defaults between 1 and 5 defaults No default Date 45 Sect.1 Sect.2 Sect.6 Sect.3 Sect.5 Sect.4 Panel A - Sector 1, date 1 Event: Default in the next h periods 1.2% CDS premium 1.0% w/ o pricing of the surprise 0.8% Physical probability 0.6% 0.4% 0.2% h Event: Default at period t+ h 0.11% 0.10% 0.09% 0.08% 0.07% 0.06% h Panel C - Sector 1, date 45 Event: Default in the next h periods 4% CDS premium w/ o pricing of the surprise 3% Physical probability 2% 1% h Event: Default at period t+ h 0.35% 0.30% 0.25% 0.20% 0.15% 0.10% h Panel B - Sector 4, date 1 Event: Default in the next h periods 0.6% 0.5% 0.4% 0.3% 0.2% 0.1% h 0.0% Event: Default at period t+ h 0.08% 0.06% 0.04% 0.02% h 0% Panel D - Sector 4, date 45 Event: Default in the next h periods 3% 2% 1% h Event: Default at period t+ h 0.35% 0.30% 0.25% 0.20% 0.15% 0.10% h

24 24/31 Illustrations App.2: Credit Spread Puzzle Credit-spread puzzle: observation of a wide gap between (a) Credit Default Swap (CDS) spreads, that can be seen as default-loss expectations under the risk-neutral measure, and (b) expected default losses (under P). See e.g. D Amato, Remolona (2003), Hull, Predescu, White (2005). Standard credit-risk models, that do not price default-event surprises, deal with the credit-risk puzzle by incorporating credit-risk premia. But these premia are too small for short maturities. We show that pricing default-event surprises may solve the credit-puzzle for all maturities, including the shortest ones.

25 25/31 Illustrations App.2: Credit Spread Puzzle We calibrate our model on U.S. banking-sector bond data covering the last two decades. Specifically, we consider riskfree (Treasury) bonds and bonds issued by U.S. banks ( ), rated BBB. Our results suggest that neglecting the pricing of default events is likely to result in an overestimation of model-implied physical probabilities of defaults for short-term horizons.

26 26/31 Illustrations App.2: Credit Spread Puzzle δ F,1 δ F,2 δ F,3 δ F,4 δ S δ 0 M M µ 1 ν 1 ρ 1 µ 2 ν 2 ρ 2 M M M1 (resp. M2) is the model pricing the default-event surprise, i.e. with δ S 0 (resp. δ S = 0). F 1,t and F 2,t follow independent ARG processes [(µ 1, ρ 1, ν 1 ) and (µ 2, ρ 2, ν 2 ), respectively]. The sdf is given by m t,t+1 = exp(δ 0 + δ F F t+1 + δ S n t+1 ) where F t = [F 1,t, F 1,t 1, F 2,t, F 2,t 1 ]. The conditional distribution of n t given F t, n t 1 is Poisson P(F 2,t ). Calibration is carried out to reproduce a set of unconditional moments derived from observed data (fitted moments on next slide).

27 27/31 Illustrations App.2: Credit Spread Puzzle Panel A - Unconditional moments (means / standard deviations S: Sample, M1: model pricing the surprise, M2: model not pricing the surprise. Treasuries (riskfree) yields Spreads (Banks vs. Treas.) Correlations 1 mth 1y 3y 5y 1y 3y 5y 1y 3y 5y ω S 2.7/ / / / / / / M1 2.7/ / / / / / / M2 2.6/ / / / / / / Panel B - Time-series fit (MSE divided by series variances, in %) Treasuries (riskfree) yields Spreads (Banks vs. Treas.) 1 mth 1y 3y 5y 1y 3y 5y M M M1 and M2 are estimated by weighted-moment methods (weights provided in row ω). Model M1 is better than M2 at reproducing sample moments, especially at the short-end of the term structure of spreads. Panel B reports the ratios of mean squared pricing errors (MSE) to the sample variances of corresponding yields/spreads. Pricing errors obtained with M1 are far lower than those associated with M2.

28 28/31 Illustrations App.2: Credit Spread Puzzle

29 29/31 Illustrations App.2: Credit Spread Puzzle

30 30/31 Conclusion 6. Conclusion

31 31/31 Conclusion Standard approaches of credit-risk pricing neglect default-event surprises. This paper proposes a tractable way to price these surprises. In our framework, quasi-closed-form expressions for derivative prices still exist if the sizes of the homogenous segments are sufficiently large. The specification accommodates different forms of contagion. An empirical analysis suggests that models pricing default-event surprises can generate sizable credit-risk premia at the short end of the yield curve and, hence, can solve the credit-risk puzzle.

32 32/31 Conclusion Appendix

33 33/31 Conclusion If F t is exogenous under P and δ S 0, F t is no longer exogenous under Q. The intensity λ i,t is a pre-intensity if: ( h ) P(τ i > t + h τ i > t, Ω t ) = E exp( λ i,t+k ) d i,t = 0, Ω t with τ i = inf{t : d i,t = 1}. k=1 If F t is exogenous (under P), then λ i,t+1 is a pre-intensity. If δ S 0, F t is not exogenous under Q (even if it is exogenous under P) then λ Q i,t+1 is not a pre-intensity, and the standard formula for B(t, h) is not valid.

34 34/31 Conclusion In fact, the pricing formula becomes: B(t, h) = Q E t [exp( r t... r t+h 1 λ Q t+1,t+h... λ Q t+h,t+h )], where λ Q t+1,t+h = log Q(d t+1 = 0 d t = 0, F t+h ) is doubly indexed, with the interpretation of a forward intensity.

35 35/31 Conclusion Homogenous model Factor: F t = (F 1,t, F 1,t 1, F 2,t ), where (F 1,t ) and (F 2,t ) are independent Autoregressive Gamma (ARG) processes. A lagged value of F 1 is introduced to get more flexible specifications of the s.d.f. and of the term structure of the yields. Parameter β is set in order to get : n t F t P(F 2,t ) The short term rate is: r t = K 0 + K F F t, where the coefficients K 0, K F depend on the parameters characterizing the ARG dynamics, on the β, and on the parameter in the s.d.f. to ensure the AAO.

36 36/31 Conclusion The next figure provides the evolutions of: the factors F 1, F 2, the short-term rate, the defaultable bond rate for the maturity h = 20, the spread between the latter and its riskfree counterpart (same maturity).

37 37/31 Conclusion 7% 6% 5% 4% 3% 2% 1% 0% % 6% 5% 4% 3% 2% 1% 0% Risk- free rate (h= 20) Defaultable- bond rate (h= 20) Spread

38 38/31 Conclusion The next figure compares: the (forward) CDS price, this price without pricing the surprise, the cumulated probability of default. (Forward) CDS prices to avoid the discounting effects that are implicit in the standard CDS pricing formula. Note that the forward CDS prices are not exactly equal to the risk-neutral probability of default. About half of the total credit-risk premia are accounted for by the credit-event risk premia. This proportion weakly depends on the time-to-maturity.

39 39/31 Conclusion 14% 12% (forward) CDS price Without pricing of the surprise Cumulated proba. of default 10% 8% 6% 4% 2% maturity