No-Arbitrage Approach to Pricing Credit Spread Derivatives
CHI CHIU CHU AND YUE KUEN KWOK
CHI CHIU CHU No arbitrage in the case of pricing credit spread should the actual spot spread breach some is a Ph.D. student in derivatives refers to determination of the time-depen- strike level. Alternatively, the payoff may mathematics at Hong dent drift terms in the mean reversion stochastic pro- depend on the difference between the spot Kong University of Sci- ence and Technology in cesses of the instantaneous spot rate and spot spread spreads of two reference entities. Credit spread Hong Kong, China. by fitting the current term structures of default-free options may be used either to earn income [email protected] and defaultable bond prices. The riskless rate and the from the option premium, to bet one’s view credit spread of a reference entity are taken to be cor- on credit spread change, or to target the pur- YUE KUEN KWOK related stochastic state variables in this pricing model. chase of a reference entity on a forward basis is an associate professor of at a favored price. mathematics at Hong When the spot rate and spot spread both follow the By following a reduced-form approach Kong University of Sci- ence and Technology. Hull and White model, one can derive an analytic that models the occasion of default as a point [email protected] representation for the time-dependent drift terms and process, Jarrow, Lando, and Turnbull [1997] analytic price formulas for credit spread options. Algo- construct a Markov chain model for valuation rithms for the numerical valuation of credit spread of risky debt and credit derivatives that incor- derivatives are developed, and the pricing behaviors porates the credit ratings of a firm as an indi- of credit spread options are examined. cator of the likelihood of default. The Markov model provides the evolution of an arbitrage- free term structure of the credit spread. redit spread derivatives are credit Problems arise in numerical implemen- rate-sensitive financial instruments tation of the model because bonds with high designed to hedge against or cap- credit ratings may experience no default within italize on changes in the credit the sample period, so the risk premium adjust- Cspread of a reference entity such as a risky ments become ill-defined as the estimated bond. The credit spread of a risky bond is the default probabilities are zero. Kijima and difference between the current yield of the Komoribayashi [1998] propose a technique of risky bond and a benchmark rate of compa- risk premium adjustment to overcome this dif- rable maturity. The spread represents the risk ficulty with high-rated bonds. premium the market demands for holding the In related work, Duffie and Singleton risky bond. [1999] and Schönbucher [1998] use intensity Unlike default swaps, credit spread models to analyze the term structure of derivatives do not depend upon any specific defaultable bonds and compute the price of credit event occurring. The buyer of a credit credit derivatives. Schönbucher [1999] con- spread option pays an up-front premium, and structs a two-factor tree model for fitting both in return the writer agrees to pay an amount defaultable and default-free term structures of
SPRING 2003 THE JOURNAL OF DERIVATIVES 51 bond prices, following the Hull and White framework formulas for credit spread options. Mougeot [2000] with time-dependent drift terms. He uses the Hull and extends the two-factor Longstaff-Schwartz model to a White [1990] model for the dynamics of the risk-free three-factor model for pricing options on the spread of interest rate and default intensity. Garcia, van Gindersen, the values of two reference assets. and Garcia [2001] use a two-factor tree algorithm that is We extend the Longstaff and Schwartz model by composed of a Hull and White tree for the risk-free assuming time-dependent drift functions in the mean interest rate dynamics and a Black and Karasinski [1991] reversion diffusion processes of the riskless interest rate tree for the intensity dynamics. and credit spread. The spot riskless interest rate is taken Das and Sundaram [2000] propose a discrete-time to follow the Hull-White model, while the spot credit Heath-Jarrow-Morton [1992] model for valuation of spread can be taken to follow either the Hull-White model credit derivatives. They derive risk-neutral drifts for the [1990] or the Black-Karasinski model [1991]. When both processes of the risk-free forward rate and forward spread. state variables follow the Hull-White model, we can Combining a recovery of market value condition, they derive analytic expressions for the time-dependent drift obtain a recursive representation of drifts that are path- functions that fit the current term structures of default- dependent. free and defaultable bond prices. Also, we can obtain ana- This formulation has the advantage that it can easily lytic pricing formulas for credit spread options. handle the American exercise feature in the pricing of Unfortunately, the time-dependent drift function the credit derivatives, and the corresponding algorithm also in the spot spread process is not analytically tractable when includes information about default probabilities inferred the logarithm of the spot spread is assumed to follow a from market data. Its major disadvantage is that its com- mean-reverting diffusion (this is the assumption in the putational complexity grows exponentially with the Black-Karasinski model). By extending the Hull-White number of time steps, since the corresponding binomial trinomial tree fitting methodologies [1994], we extend tree is non-recombining. the fitting algorithms to allow numerical valuation of A structural approach derives the term structure of credit spread options. the credit spread of a credit entity by examining the credit quality of the issuer. In structural models, the credit spread I. CONTINUOUS MODELS is not one of the state variables in the model, but rather is a derived quantity. Default occurs when the issuer firm We begin with the pricing of credit spread deriva- value falls below some threshold value. The structural tives under the Hull-White [1990] model. This is an exten- approach has been used by Das [1995] to price credit risk sion of the one-factor Hull-White model to the two-factor derivatives. version, where both the instantaneous spot riskless interest Longstaff and Schwartz [1995] develop a valuation rate and the credit spread follow a mean-reverting model model for pricing credit spread derivatives by assuming that with time-dependent drift terms. Like the one-factor ver- the riskless interest rate and the spread follow correlated sion, the two-factor model exhibits nice analytical mean reversion diffusion processes. From empirical analysis tractability and ease of numerical implementation. of observed credit spreads, they argue that the logarithm of the credit spread follows a mean reversion process. They Pricing Framework Under assume the parameters in the characterization of the pro- Two-Factor Hull-White Model cesses to be constant, and derive closed-form price formulas for the credit spread options. A disadvantage of the process We specify the processes of the instantaneous spot is that their model does not incorporate information about riskless interest rate r and the spot spread s to follow the the current term structures of bond prices. extended Vasicek [1977] model, where the drift terms are Chacko and Das [2002] have since then derived assumed to be time-dependent functions. Under a risk- price formulas for credit spread options based on the neutral valuation framework, we assume that the stochastic uncorrelated Cox-Ingersoll-Ross [1985] processes for the processes followed by r and s are given by spot rate and spot spread, and again the parameters in the stochastic processes are assumed to be constant. Tahani dr =[θ(t) − ar]dt + σrdWr, (1-A) [2000] models the process of the logarithm of credit spread using the GARCH model and obtains closed-form price
52 NO-ARBITRAGE APPROACH TO PRICING CREDIT SPREAD DERIVATIVES SPRING 2003 then ds =[φ(t) − bs]dt + σsdWs, (1-B)
σ σ (t, T )=− r [1 − e−a(T−t)]. where θ(t) and φ(t) are time-dependent functions, and a P a (6) and b are positive constant parameters that give the rates σ of mean reversion of r and s, respectively. Further, r and By following a similar procedure, it is possible to σ φ s are the constant volatility parameters for the spot rate determine the other time-dependent function (t) in the r and spot spread s, respectively, and dWr and dWs are stan- mean reversion process of s by fitting the current term dard Brownian motions under the risk-neutral measure structure of defaultable bond prices. This is achieved by ρ P. Let denote the correlation coefficient between dWr changing the probability measure to the forward-adjusted ρ T and dWs; that is, dWrdWs = dt. measure Q with delivery at future time T. Let D(t, T) Determination of the Time-Dependent Drift Terms. denote the price at time t of a defaultable bond maturing Hull and White [1990] demonstrate how to determine the at time T. By the Girsanov theorem, we have: time-dependent function θ(t) in the drift term of the default-free spot rate by fitting the current term structure T of default-free bond prices. Their procedure can be sum- D(t, T)=EP exp − [r(u)+s(u)] du marized as follows. Let P(t, T) denote the price at time t t of a discount bond maturing at time T, which is given by: T T = P(t, T)EQ exp − s(u) du (7) t T P P(t, T )=E [exp(− r(u) du)], (2) t Accordingly, the process followed by the spot spread s under the forward measure QT is given by where EP denotes the expectation under the risk-neutral measure P. By taking the risk-neutral expectation in Equa- T tion (2), where r is defined in Equation (1-A), one obtains: ds =[φ(t) − bs + ρσsσP (t, T)]dt + σsdW , (8) r(t) −a(T−t) P(t, T )=exp − [1 − e ] − T T a where dW is a standard Brownian motion under Q . Under the process defined in Equation (8), the quantity T 1 − e−a(T−u) T σ2 ∫T θ(u) du + r [1 − e−a(T −u)]2 du ts(u) du follows the normal distribution with mean: 2 t a t 2a T (3) T s(t) EQ s(u) du = 1 − e−b(T−t) + b Differentiating Equation (3) with respect to T twice, t we obtain: T 1 − e−b(T−u) [φ(u)+ρσsσP (u, T )] du t b σ2 ∂2 ∂ θ(T)= r [1 − e−2a(T−t)] − ln P(t, T) − a ln P(t, T). (9-A) 2a ∂T2 ∂T and variance (4) This equation gives the dependence of the drift term T T σ2 θ(t) using the information of the current term structure of var s(u) du = s [1 − e−b(T −u)]2 du. b2 the default-free bond price P(t, T) over the period [t, T]. t t Suppose we write the process followed by P(t, T) as: (9-B)
dP (t, T ) Since = rdt + σ (t, T)dW , P(t, T ) P r (5)
SPRING 2003 THE JOURNAL OF DERIVATIVES 53 T T dξ =[φ(t)+ρσsσP (t, T)]s(t)(−b(T − t))dt EQ exp − s(u) du = T t + σss(t)(−b(T − t))dW . (14) T T QT 1 ξ exp −E s(u) du + var s(u) du Note that T is conditionally normally distributed t 2 t 2 with mean µξ and variance σξ, where:
(10-A) ρσ σ 1 − e−(a+b)(T −t) 1 − e−b(T−t) µ = s(t)e−b(T−t) + s r − we obtain ξ a a + b b
T −b(T−u) D(t, T) s(t) T 1 − e−b(T −u) + φ(u)e du =exp − [1 − e−b(T−t)] − [φ(u)+ρσ σ (u, T )] du t P(t, T ) b s P b t σ2 ∂ D(t, T ) = s [1 − e−b(T−t)][1 − e−2bte−b(T−t)] − ln T σ2 2 + s [1 − e−b(T −u)]2 du . 2b ∂T P(t, T ) 2 t 2b ρσ σ 1 − e−(a+b)(T−t) 1 − e−b(T−t) e−at[e−a(T−t) − e−b(T −t)] (10-B) + s r − − a a + b b b Again, by differentiating Equation (10-B) with e−(a+b)te−b(T −t)[1 − e−a(T−t)] − , respect to T twice and observing Equation (6), we obtain: b (15-A) σ2 ∂ D(t, T ) ∂2 D(t, T ) φ(T)= s [1 − e−2b(T−t)] − b ln − ln 2b ∂T P(t, T) ∂T2 P(t, T) 2 2 σs −2b(T−t) 1 − e−a(T −t) 1 − e−b(T−t) σξ = [1 − e ]. (15-B) + ρσ σ + e−a(T−t) . 2b s r a b (11) Note that µξ depends not only on the time to expi- Credit Spread Put Option. Let psp(r, s, t) denote the ration T – t but also on the current time t. Once the mean ξ price of the credit spread put option, whose terminal and variance of T are available, the price formula for the payoff is given by credit spread put is given by:
psp(r, s, T )=max(s − K, 0). (12) psp(r, s, t)=P(t, T) σ 2 2 µ − K ξ −(K−µξ ) /2σ ξ √ e ξ +(µξ − K)N Here K is the strike spread at which the option holder has 2π σξ the right to put the risky bond to the option writer. It (16) is called a credit spread put option although the terminal payoff resembles a call payoff. When the spot credit spread is higher than the strike spread, this signifies that the price Combined Hull-White of the underlying risky bond drops below the strike price and Black-Karasinski Model of the bond corresponding to the strike spread. Since the terminal payoff depends only on s, we may use the default Longstaff and Schwartz [1995)] propose that the free bond price as the numeraire and obtain: interest rate and the logarithm of the credit spread follow mean-reverting Vasicek processes. We extend their model p (r, s, t)=P(t, T)EQT [max(s − K, 0)]. by assuming time-dependent functions in the drift terms; sp (13) that is, we choose the Hull-White model for the interest rate process and the Black-Karasinski model for the credit Suppose we define ξ(t) = s(t)exp(–b[T – t]). Then spread process. Assume that y = ln s, and the stochastic ξ follows the stochastic process: processes followed by r and y are governed by:
dr =[θ(t) − ar]dt + σrdWr, (17-A)
54 NO-ARBITRAGE APPROACH TO PRICING CREDIT SPREAD DERIVATIVES SPRING 2003 dy =[ψ(t) − cy]dt + σ dW , σ2 y y (17-B) σ2 = y [1 − e−2c(T −t)]. ζ 2c (20-B) σ where c is the rate of mean reversion of y, y is the con- ^ρ stant volatility parameter for y, and dWrdWy = dt. Unfor- Heath-Jarrow-Morton Model tunately, the Black-Karasinski model lacks the level of analytic tractability exhibited by the Hull-White model. The two-factor extended Vasicek model is a special Under the Black-Karasinski assumption of the credit spread case of the Heath-Jarrow-Morton (HJM) model corre- process, we find that Equation (7) has to be modified as: sponding to a specific choice of the volatility structures. First, we present the general HJM formulation of the T pricing model with the two state variables r and s. D(t, T) T = EQ exp − exp(y(u)) du . Let f(t, T ) denote the riskless instantaneous forward P(t, T) t rate. Following the HJM framework, the dynamic of f(t, T ) (18) is assumed to be
∫T ∫T Unlike ts(u) du, however, the quantity texp[y(u)]du (t, T )=α (t, T)dt + σ (t, T)dW , is no longer normally distributed. Hence, there is not a simple df f f 1 (21) relation between [D(t, T )]/[P(t, T )] and the time-depen- ψ α σ dent drift function (t), like that shown in Equation (11). where f is the drift, and f is the volatility. Also, dW1 is If θ(t) and ψ(t) are known functions, we may derive a standard Brownian motion under the risk-neutral mea- α the price formula for the credit spread put option. In sure P. Under the no-arbitrage condition, the drift f σ terms of y, the terminal payoff of the credit spread put is and volatility f must observe: given by max(ey – K, 0). The put price formula takes the usual Black-Scholes form: T αf (t, T)=σf (t, T) σf (t, u) du. (22) t µ +σ2/2 psp(s, r, t)=P(t, T ) e ζ ζ N(d) − KN(d − σζ) From the dynamics of the risky bond prices, we (19-A) can define the risky instantaneous forward rate fd(t, T ) in where the same manner as the risk-free instantaneous forward rate. Next, the instantaneous forward spread h(t, T) is 2 defined to be f (t, T ) – f(t, T ); accordingly, the spot spread − lnK + µζ + σ d d = ζ . s(t) = f (t, t) – f(t, t). Assume that h(t, T ) follows the σ (19-B) d ζ stochastic process:
2 Similar to µξ and σξ as given in Equations (15-A) 2 dh(t, T)=αh(t, T)dt + σh1 (t, T)dW1 + σh2 (t, T )dW2, and (15-B), the mean µξ and variance σξ are given by: (23)
−c(T−t) ρσ yσr µζ = y(t)e + α σ σ a where h is the drift, h1 and h2 are volatilities, and dW1 and dW are independent standard Brownian motions 1 − e−(a+c)(T −t) 1 − e−c(T−t) 2 − + under the risk-neutral measure P. Again, the no-arbi- a + c c α σ σ trage condition dictates that h, h1, and h2, satisfy (see T Schönbucher [1998]): ψ(u)e−c(T −u) du, (20-A) t
SPRING 2003 THE JOURNAL OF DERIVATIVES 55 2 T T II. CONSTRUCTION OF NUMERICAL α (t, T )= σ (t, T) σ (t, u) du + σ (t, T) σ (t, u) du h hi hi h1 f + i=1 t t ALGORITHMS T σ (t, T ) σ (t, u) du. f h1 (24) Now we can extend the Hull and White [1994] t methodology to fitting the current term structures of default-free and defaultable bond prices. By recalling that s(t) = h(t, t), one can show that the process followed by s is governed by: Two-Factor Algorithms