Markov Functional interest rate
models with stochastic volatility
New College University of Oxford
A thesis submitted in partial fulfillment of the MSc in Mathematical Finance
December 9, 2009
To Rahel
Acknowledgements
I would like to thank my supervisor Dr Jochen Theis for advising me throughout the project and proof–reading of this dissertation. Furthermore I want to extend my gratitude to d–fine GmbH for giving me the opportunity to attend the MSc in Mathematical Finance programme. But above all I am indebted to my family, especially to my wife Rahel, for their great support and patience.
Abstract
With respect to modelling of the (forward) interest rate term structure under consideration of the market observed skew, stochastic volatility Libor Market Models (LMMs) have become predominant in recent years. A powerful rep- resentative of this class of models is Piterbarg’s forward rate term structure of skew LMM (FL–TSS LMM). However, by construction market models are high– dimensional which is an impediment to their efficient implementation.
The class of Markov functional models (MFMs) attempts to overcome this in- convenience by combining the strong points of market and short rate models, namely the exact replication of prices of calibration instruments and tractabil- ity. This is achieved by modelling the numeraire and terminal discount bond (and hence the entire term structure) as functions of a low–dimensional Markov process whose probability density is known.
This study deals with the incorporation of stochastic volatility into a MFM framework. For this sake an approximation of Piterbarg’s FL–TSS LMM is de- vised and used as pre–model which serves as driver of the numeraire discount bond process. As a result the term structure is expressed as functional of this pre–model. The pre–model itself is modelled as function of a two–dimensional Markov process which is chosen to be a time–changed brownian motion. This ap- proach ensures that the correlation structure of Piterbarg’s FL–TSS is imposed onto the MFM, especially the stochastic volatility component is inherited.
As part of this thesis an algorithm for the calibration of Piterbarg’s FL–TSS LMM to the swaption market and the calibration of a two–dimensional Libor MFM to the (digital) caplet market was implemented. Results of the obtained skew and volatility term structure (Piterbarg parameters) and numeraire dis- count bond functional forms are presented.
Contents
1 Introduction 1
2 A review of Libor Market and Markov Functional Models 3 2.1 TheLiborMarketModel...... 3 2.1.1 Non–log–normal forward Libor dynamics ...... 5 2.1.2 Incorporation of stochastic volatility ...... 6 2.2 MarkovFunctionalModels...... 7 2.2.1 Definition and examples of Markov Functional Models ...... 8 2.2.2 ALiborMarkovFunctionalModel ...... 10 2.2.3 Multi–dimensional Markov Functional Models ...... 12 2.3 A Libor Market Model as pre–model for a Markov Functional Model . . . . 14
3 Piterbarg’s term structure of skew forward Libor model 17 3.1 TheforwardLibordynamics...... 17 3.2 Swap rate dynamics under the FL-TSS model ...... 19 3.2.1 Derivation of the forward swap volatility level ...... 20 3.2.2 Derivation of the forward swap skew ...... 21 3.3 The effective skew and volatility formulation ...... 23 3.3.1 Theeffectiveforwardswapskew ...... 23 3.3.2 The effective forward swap volatility ...... 26 3.4 CalibrationoftheFL-TSSmodel ...... 30 3.4.1 Forward rate volatility calibration ...... 31 3.4.2 Forwardrateskewcalibration ...... 35 3.4.3 Calibrationresults ...... 39
4 A Markov functional model with stochastic volatility 41 4.1 Piterbarg’s FL–TSS Libor Market Model as pre–model ...... 42 4.2 The pre–model with two Brownian drivers ...... 43 4.3 A simplification of the pre–model process ...... 44
i 4.4 Construction of a two–dimensional Libor Markov functional model . . . . . 47 4.4.1 A two–dimensional Libor Markov functional model in the terminal measure...... 48 4.4.2 Calibrationresults ...... 50
5 Conclusion 53
A Mathematical details 55 A.1 ThedrifttermintheLiborMarketModel ...... 55 A.2 The derivative of the forward swap rate w.r.t the forward Liborrates. . . . 58
A.3 Derivation of the coefficient cmn ...... 59 A.4 Proofofcorollary3.3.2...... 60 A.5 A recursion scheme for a system of time dependent Riccati equations . . . . 63
A.5.1 An analytic solution for Di t, Ti+1 ...... 64
A.5.1.1 The case gi =0 ...... 65 6 A.5.1.2 The case gi =0 ...... 66
A.5.2 An analytic solution for Ai t, Ti+1 ...... 66
A.5.2.1 The case gi =0 ...... 66 6 A.5.2.2 The case gi =0 ...... 68 A.5.3 Summaryofthesolution...... 70 A.6 Derivationofrelation(3.28) ...... 70 A.7 2d–Markov functional integration ...... 72
B The Heston Model 75 B.1 Specificationofthemodeldynamics ...... 75 B.2 Thecharacteristicfunction ...... 77 B.3 ThesolutionoftheHestonODE ...... 78 B.3.1 Boundaryconditions ...... 78 B.3.2 AsystemofRiccatiODEs...... 79 B.4 Option pricing by transformation techniques ...... 80 B.5 CalibrationoftheHestonModel ...... 86
C Tables and figures 87
Bibliography 93
ii List of Figures
3.1 Volatility level λ10(t) and skew β10(t) of forward rate F10(t) for times T0 = 0y t < T = 10y...... 40 ≤ 10 ¯ 4.1 Proxy forward rate F10(T10) (4.11) as function of zt = 0,z2) at reset time T = 10y. This corresponds to the zero correlation case, Γ(s) 0...... 47 10 ≡ ¯ ¯ 4.2 The numeraire discount bond as functional of F10(T10, zT 10): D(T10, T11; F10(T10, zT 10)). 51
¯ ¯ C.1 The numeraire discount bond as functional of F1(T1, zT 1): D(T1, T11; F1(T1, zT 1)). 89 ¯ ¯ C.2 The numeraire discount bond as functional of F3(T3, zT 3): D(T3, T11; F3(T3, zT 3)). 89 ¯ ¯ C.3 The numeraire discount bond as functional of F5(T5, zT 5): D(T5, T11; F5(T5, zT 5)). 90 ¯ ¯ C.4 The numeraire discount bond as functional of F10(T7, zT 7): D(T7, T11; F7(T7, zT 7)). 90
iii
Chapter 1
Introduction
This study is dedicated to the incorporation of stochastic volatility into a Markov func- tional framework. The class of Markov functional models (MFM) was introduced by Hunt, Kennedy and Pelsser in [15]. A major motivation which lead to their development was the desire to have models that can exactly replicate prices of liquid calibration instruments in a similar fashion to market models while maintaining the efficiency of short rate models in calculating derivative prices [13], [18]. Latter are formulated in terms of the short rate or instantaneous forward rate which cannot be traded in the market. As a consequence the prices of derivatives in these models are quite involved functions of the underlying process which is being modelled. This fact makes it difficult to capture the most characteristic features of a derivative product with models of this kind. However, their strong point is that the short rate process is easy to follow and hence implemention is straightforward [13]. Unlike short rate models the class of market models is formulated in terms of market rates which are directly related to tradable assets. Thus they exhibit better calibration properties than short rate models. However, as these models capture the joint distribution of market rates, they are high–dimensional by construction and tedious to implement. The first formulation of a market model was provided by Brace, Gatarek and Musiela in the context of forward Libors (LMM) [4]. A forward swap market model was developed by Jamshidian in 1997 [16]. In these approaches the underlying rates are modelled as log– normal martingales under their own probability measure. However, the presence of a volatility skew in the caplet and swaption markets indicate that a pure log–normal forward dynamics is not appropriate. In this respect modified forward rate dynamics were introduced, e.g in the context of constant elasticity of variance (CEV) and displaced diffusion models in which mixtures of pure normal and log–normal dynamics are considered [5], [23]. Aiming at a proper modelling of the skew term structure, stochastic volatility extensions of Libor and Swap Market models were introduced. This was
1 achieved by modelling the forward rate variance as CEV process. Approaches of this kind are the ones by Andersen and Andreasen [2] and Piterbarg [19], [20]. Especially Piterbarg’s stochastic volatility extension accounts for a Libor specific term structure of forward rate skews and volatilities. In the study at hand this forward Libor term structure of skew model (FL–TSS) is used in the construction of a Markov functional model with stochastic volatility. The MFM framework is based on formulating the numeraire and the terminal discount bond as functionals of a low dimensional Markov process whose dynamics can be followed easily. The functional forms in turn are obtained by calibration to prices of liquid derivatives at particular dates which are relevant to the product being priced. As the discount bonds at earlier times are obtained by applying the martingale property of numeraire rebased assets the resulting model is arbitrage free by construction. Thus, MFMs combine the strong points of market and short rate models. The incorporation of stochastic volatility into a MFM is based on the concept of a pre–model which depends on a low–dimensional Markov process [14],[17]. By regarding the forward Libors and hence the discount bond processes as functions of the pre–model process, the calibration can be formulated in terms of the latter. Thus the correlation structure of the pre–model is incorporated into the MFM. In this study a pre–model is selected which is an approximation of a calibrated Piterbarg FL–TSS model and depends on a two–dimensional brownian motion. This proxy is then employed in constructing a Libor MFM which inherits the stochastic volatility structure of Piterbarg’s FL–TSS. Tractability is maintained since calibration involves the integration of the known probability distribution of the two–dimensional Markov process. The thesis is structured as follows: In chapter 2 the concepts of LMMs and MFMs are reviewed. Working under the terminal (forward) measure drift terms for a stochastic volatil- ity LMM are derived. Furthermore multi–dimensional extensions of MFMs are discussed. Chapter 3 is dedicated to the detailed study of Piterbargs FL–TSS LMM. This model is calibrated to the swaption market, and the resulting term structure of skews and volatilites is presented. The incorporation of stochastic volatility into a MFM is the topic of chap- ter 4. Here a two–dimensional Libor MFM is constructed which uses an approximation of Piterbarg’s FL–TSS as pre–model. The model is calibrated to the (digital) caplet market, and resulting numeraire discount bond funtionals are presented. The thesis concludes with chapter 5. Besides result tables and figures, the appendices contain mathematical details and a thorough presentation of the Heston Model.
2 Chapter 2
A review of Libor Market and Markov Functional Models
In this chapter we review the class of Libor Market and Markov functional models which have become prevalent in the last ten to fifteen years. In section 2.1 we discuss the LMM under consideration of non–log–normal forward rate dynamics and stochastic volatility. In particular, working under the terminal (forward) measure forward rate drift terms are derived. The Markov functional framework is introduced in section 2.2. Therein multi– dimensional extensions are discussed as well. In section 2.3 the idea of constructing MFMs in terms of a pre–model is introduced.
2.1 The Libor Market Model
As already mentioned in the introduction the Libor Market Model focusses on modelling the dynamics of forward Libor rates Fi t; Ti, Ti+1 which reset at times Ti, i = 1,...,N. In a deterministic volatility setting employing K independent Brownian drivers these are modelled as log–normal variables with respect to their martingale measure Qi+1 which is induced by taking the discount bond D t, Ti+1 as numeraire, K i+1 T i+1 dFi t; Ti, Ti+1 = λi(t)Fi t; Ti, Ti+1 σi,k(t) dWk (t)= λi(t)Fi t; Ti, Ti+1 σi(t) dWt , k=1 X i+1 for times T0 t < Ti TN where dW is a K–dimensional vector of orthogonal Brow- ≤ ≤ t i+1 nian motions under Q and λi(t) are positive continuous, real valued functions. In particular the relation dWk,dWl = δkl dt holds. The K–dimensional vector σi(t) con- tains the load factors of the orthogonal brownian motions onto forward rate Fi which K 2 satisfy k=1 σi,k(t) = 1. Thus they define a correlation matrix through the relation T 1 ρij(t) =Pσi(t)σj(t) = ρij i=1,...,K . Indeed, in this formulation the covariance of for- { }j=1,...,K
1We assume that K = N, i.e., that each forward rate has its own driving brownian motion.
3 ward Libor yields is given by K dFi(t) dFj(t) i+1 i+1 , = λi(t)λj(t) σi,k(t)σj,l(t) dWk (t),dWl (t) Fi(t) Fj(t) k,l=1 X δkl dt K | {z } = λi(t)λj(t) σi,k(t)σj,k(t) dt = λi(t)λj(t)ρij(t) dt, Xk=1 =ρij (t) from where it becomes apparent that the| brownian{z correlatio} ns ρij(t) as well as the volatil- ities λi(t) contribute to the forward rate correlation. Since the forward Libors follow a log–normal process in their own martingale measure caplet prices are given by the Black76 formula, comp. [3]. Therefore the volatilities λi(t) can be obtained by calibration to quoted caplet prices. For this a widespread approach is to assume a parametric shape for the volatilities as function of time to expiry which correctly captures their dynamics observed in the market. The parameters are then determined in the course of the calibration process, comp. [21]. With respect to the correlation function the most reasonable approach is to model it in parametric form as well. The reason for this is that it is not easy at all to extract information on the instantaneous correlation ρij(t) out of quoted derivative prices, e.g., swaption volatilities, because the latter depend on the history of λi(t)λj(t)ρij(t) on the time interval which starts at T0 and ends on swaption expiry. The parametric form proposed by Rebonato in [21] and [22] is given by
ǫ ǫ ρij(t)= ρ + (1 ρ )exp δ (Ti t) (Tj t) , (2.1) ∞ − ∞ − − − − with constants 1 ρ 1, δ,ǫ> 0, and will be adopted in what follows. − ≤ ∞ ≤ Of course, to use the LMM in practice the dynamics of all forward Libor rates have to be formulated in a single measure. In this respect a convenient choice is the terminal measure N+1 Q which is induced by taking the terminal discount bond D t, TN+1 as numeraire. As a consequence only the forward Libor FN is a martingale, and according to Girsanov’s theorem all other forward Libors will be modified by additional drift terms, comp. [13]. Indeed, when N+1 changing from the terminal measure Q to the martingale measure of forward rate Fi N+1 N+1 i+1 the process dWk (t)+ µi (t) dt also is a brownian motion under Q . Hence in the terminal measure the forward rate processes become
T N+1 N+1 dFi t; Ti, Ti+1 = λi(t)Fi t; Ti, Ti+1 σi(t) µi (t) dt + dWt , N+1 T N+1 = Fi t; Ti , Ti+1 λi(t) µi (t) dt + σi(t) dWt , (2.2) N N+1 αl λl(t) ρil(t) Fl N+1 µi (t)= e , µN (t) = 0, − 1+ αlFl(t) l=Xi+1 ,e T0 t < Ti TN , i =e 1,...,N ≤ ≤
4 N+1 where the drifts µi (t) are given by (A.10) which is derived in appendix A.1.
2.1.1 Non–log–normale forward Libor dynamics
The forward Libor process (2.2) presented above models the forward Libors Fi as a log– normal processes with respect to their martingale measure. In the terminal measure a log–normal behaviour is maintained for forward Libor FN . However, the caplet/floorlet market displays a volatility surface (in terms of terms of implied Black volatilities), i.e., the volatility varies as option expiries and moneyness changes. Specifically, the observed volatilities are monotone decreasing functions of the forward Libor level, a behaviour which is denoted as volatility skew. Its presence indicates that the forward Libors do not follow a log–normal process, for in that case the implied Black volatilities should be constant. One proposal for an alternate forward rate dynamics provided by Rubinstein [23] is a displaced diffusion which combines a log–normal and a normal process. Another approach is to model the forward Libors as constant elasticity of variance (CEV) processes which was proposed by Cox and Ross [5]. In both cases the forward Libor change can be written in terms of a volatility function ϕ Fi which imposes a rate level dependence onto the forward rate volatilities, i.e the forward Libors are modelled as
T i+1 dFi t; Ti, Ti+1 = ϕ Fi λi(t)σi(t) dWt , i = 1,...,N. For example a displaced diffusion model is established with a function of the form
ϕ Fi t; Ti, Ti = β Fi t; Ti, Ti + 1 β Fi T0; Ti, Ti , +1 +1 − +1 where the displacement parameter β is a real valued constant. Obviously the case β = 1 corresponds to a log–normal dynamics. For β = 0 a normal process is recovered. A generalization of this with time dependent paramter β(t) will be considered in the next chapter where Piterbarg’s forward Libor term structure of skew model (FL–TSS) will be presented. A CEV–model is obtained by defining
β ϕ Fi t; Ti, Ti+1 = Fi t; Ti, Ti+1 , (2.3) with 0 β 1. As with the displaced diffusion model the case β = 1 corresponds to ≤ ≤ log–normal dynamics whereas β = 0 results in a normal model. In this model the yield or dF ϕ(F ) β 1 percentage volatility, i.e., the volatility of i is given by i λ (t)= λ (t)F t; T , T − Fi Fi i i i i i+1 which for 0 <β< 1 is a monotone decreasing function of forward Libor Fi in accordance with the observed market behaviour. A model with this kind of local volatility function was first introduced by Dupire in modelling equities, comp. [7].
5 Based on these ideas a generalised forward Libor process can be formulated in the terminal measure QN+1,
T N+1 N+1 dFi t; Ti, Ti+1 = ϕ Fi λi(t)σi(t) µi (t) dt + dWt , N+1 T N+1 = ϕ Fi λi(t) µi (t) dt + σi(t) dWt , (2.4) N N+1 αl λl(t) ρil(t) ϕ Fl N+1 µi (t)= e , µN (t) = 0, − 1+ αlFl(t) l=i+1 X e T0 t < Ti TN , i = 1,...,Ne ≤ ≤ which was proposed by Andersen and Andreasen, comp. [1]. The drift terms are given by equation (A.9) which is derived in appendix A.1.
2.1.2 Incorporation of stochastic volatility
As the market observed volatility skews cannot be solely captured by the introduction of a (local) volatility function stochastic volatility extensions of the LMM were devised. In this respect one approach is to extend the forward Libor process (2.4) with a stochastic variable which accounts for the volatility level and as such modulates the local volatility function
ϕ(Fi). A convenient choice is the square root of a variance process Σt which follows a one–dimensional CEV–process, comp. [1]. Assuming that the brownian driver of the variance process is correlated with each for- N+1 N+1 ward rate driver, i.e., dV (t),dWk (t) = Γk(t) dt (k = 1,...,N), the process (2.4) can be generalised to
N+1 T N+1 dFi t; Ti, Ti+1 = ϕ Fi λi(t) Σt µi (t) dt + σi(t) dWt N+1 T T N+1 T N+1 = ϕ Fi λi(t)pΣt µi (t) dt + σi(t) Ω(t) dZt + σi(t) Γ(t) dVt , e (2.5a) p e N+1 b dΣt =Θ Σ0 Σt dt + η Σt dV , (2.5b) − t p N+1 N+1 N+1 N+1 with dV (t),dWk (t) =Γk(t) dt, and T 0 t < Ti TN , i = 1, . . . , N, k = 1,...,N, ≤ ≤ N+1 where the K–dimensional vector of brownian drivers dWt was decomposed into orthog- N+1 N+1 onal components dZt and dVt according to (A.3). The drift terms are given by 6 equation (A.8) which is derived in appendix A.1: N N+1 αl λl(t) ϕ Fl µi (t)= Σt − 1+ αlFl(t) l=i+1 p X T T e σl(t) Ω(t) Ω(t)σi(t) + σl(t)Γ(t) σi(t)Γ(t) , (2.5c) × µN+1(t) = 0,h i N b b 2 with matrix Ω(t)= 1 Γk(t) δkj k=1,...,K and vector Γ(t)= Γk(t) k=1,...,K . e − j=1,...,K { } np o b Of course, the introduction of an additional Brownian driver increases the dimensionality of the model, and the additional correlation coefficients Γk(t) enlarge the parameter space. However, for the stochastic volatility model we will work with in the following chapters, namely Piterbarg’s FL–TSS LMM, the rate and variance processes are independent. Thus Γk(t) = 0 and the drift reduces to N N+1 αl λl(t) ρil(t) ϕ Fl N+1 µi (t)= Σt , µN (t) = 0. (2.6) − 1+ αlFl(t) l=i+1 p X e e Nevertheless the process (2.5) will be referenced in section 2.3 where the idea of a pre–model in a Markov functional context is discussed. 2.2 Markov Functional Models The class of Markov functional interest rate models was originally introduced by Hunt, Kennedy and Pelsser in [15]. A major motivation which lead to their development was the desire to have models that can fit observed prices of liquid instruments in a similar fashion to the market models while maintaining the efficiency of short rate models in calculating derivative prices, comp. [13], [18]. This is achieved by specifying a low dimensional process which is Markovian in some martingale measure and formulating pure discount bond prices as functions of this process. Since efficient algorithms to compute conditional distribution functions are known for this set up, the valuation of derivatives in a Markov functional framework is much more efficient when compared to pricing using market models. Although market models are Markovian as well, they are naturally of high dimension. Moreover an essential feature of these types of models is the freedom to choose the functional form of the discount bond prices in such a way that market prices of calbration instruments are replicated. This distinguishes Markov functional from short rate models in which the functional form of discount bond prices with respect to the Markovian short rate is fixed. Therefore Markov functional models combine the strong points of market as well as short rate models, namely the fitting to observed prices of liquid instruments and tractability. 7 2.2.1 Definition and examples of Markov Functional Models In this section we want to give a formal definition of Markov functional models and also present some examples. We begin by citing the definition given by Hunt and Kennedy: Definition 2.2.1 (Hunt and Kennedy [13]). An interest rate model is said to be Markov– functional if there exists some numeraire pair (N, N) and some process x sucht that: 1. the process x is a (time–inhomogeneous) Markov process under the measure N; 2. the pure discount bond prices are of the form DtS = DtS (xt), T0 t ∂S S, ≤ ≤ ≤ for some boundary curve ∂S : [0,∂ ] [0,∂ ] and some constant ∂ ; ∗ → ∗ ∗ 3. the numeraire N, itself a price process, is of the form Nt = Nt(xt) T0 t ∂∗. ≤ ≤ Obviously the boundary curve ∂S is introduced so that the model does not need to be defined over the entire time domain 0 t S. The most common choice for the boundary ≤ ≤ curve is S, if S T ∂S = ≤ , T, if S > T for some constant T. Thus the main ingredients of a Markov functional model are the driving Markov process xt which describes the state of the economy and the functional forms of 1. the discount bond D∂ S x∂ D ∂S, S; x∂ on the boundary curve ∂S; S S ≡ S 2. the numeraire Nt(xt) for times T0 t ∂ . ≤ ≤ ∗ The reason for this is that the functional forms of bonds at earlier times t<∂S are deter- mined by the functional form of the discount bond on the boundary curve by the martingale property of numeraire rebased assets, N D(∂S, T ; x∂S ) D t, T ; xt = Nt(xt) E t , T0 t ∂S T, (2.7) N (x ) F ≤ ≤ ≤ ∂S ∂S comp. [10]. One particular choice of measure is the (terminal) forward measure N = QN+1 which is induced by taking the discount bond D t, TN+1 as numeraire. Taking TN+1 as boundary we thus have ∂S = min S, TN+1 , Nt(xt)= D t, TN+1; xt , T0 t TN+1, ≤ ≤ 8 and the price of a discount bond maturing at S T becomes ≤ N+1 QN+1 D(S, S; xS ) D t, S; xt = D t, TN+1; xt E xt D S, T ; x N+1 S QN+1 1 = D t, TN+1; xt E xt , T0 t S TN+1, (2.8) D S, T ; x ≤ ≤ ≤ N+1 S where the expectation is conditioned on x t because of the Markov property of the underlying process2. If one considers interest rate derivatives like caplets/floorlets or swaptions expiring at time Tm with strike K their payoff function Vm(Tm, K) depends on the discount bonds D(Tm, Tj; xt), with j = m + 1 for caplets/floorlets and j > m in the case of swaptions, and by (2.8) is a function of the numeraire discount bond D(Tm, TN+1; xTm ) at time Tm. Thus Vm(Tm, K) Vm Tm,K,D(Tm, TN+1; xT ) and by application of the fundamental theorem ≡ m of asset pricing the derivative value at time t is given by QN+1 Vm Tm,K,D(Tm, TN+1; xTm ) Vm t, K; xt = D t, TN+1; xt E xt . (2.9) D T , T ; x m N+1 Tm Hence, if the Markov process xt and thus the conditional probability distribution p(xTm xt) | is specified, this relation provides a means to extract the functional form of the numeraire discount bond D(Tm, TN+1; xTm ) at time Tm from market observed derivative prices since the payoff function Vm Tm,K,D(Tm, TN+1; xTm ) is known. However, in order to proceed along these lines one has to assume that the discount bonds are monotone functions of the underlying Markov process. It has to be emphasized that due to the functional dependence the specified underlying Markov process xt determines the probability distribution of discount bonds. As the prices of multi–temporal interest derivatives depend on the joint probability distribution of forward rates (and thus on the joint distribution of discount bonds) at those times relevant to the product at hand3, the driving Markov process encodes all information on the correlation structure. So in designing a Markov functional model for a specific product class the process xt has to be chosen in such a way as to capture the characteristic product features while retaining low dimensionality. Referring to the process dimension, the underlying Markov process should not have more than two brownian drivers. An example of a simple one dimensional underlying process is considered in the next section where a Libor Markov functional model is explored. 2 For a Markov process xt the relation E f(xt) Ft = E f(xt) xt holds. 3E.g., for a bermudan swaption the single exercise dates are the relevant times. Therefore the probability distribution of the process xt only needs to be known on these dates. 9 2.2.2 A Libor Markov Functional Model We know consider the set of forward Libor rates Fi t; Ti, Ti+1 which reset a times Ti, i = 1,...,N, and specify a Markov functional model. For this we choose the final time N+1 ∂S = TN+1 and work in the measure N = Q which is induced by taking the discount bond D t, TN+1 as numeraire. The underlying Markov process xt is chosen to be a time changed brownian motion. Assuming that σ(t) is a deterministic, positive real valued function on [T0, TN+1] we define t xt := σ(s) dWs, ZT0 N+1 where dWs is a brownian motion under Q . Clearly, due to the Markovian character of the brownian motion xt is a Markov process whose conditional probability distribution is normal, 1 1 (x x )2 p x x s t , s t. s t = 1 exp s − 2 (2.10) 2 −2 σ(u) du ≥ (2π)2 s σ(u)2 du t t R Having specified the underlyingR process the functional form of the numeraire discount bond D t, TN+1; xt remains to be determined. This will be done at discrete times Ti (i = 1,...,N ) according to a recursion scheme in which the functional from will be determined by calibration to digital caplet prices. To start with we observe that at time TN we observe that the forward rate FN t; TN , TN+1; xTN N+1 is a log–normal martingale under Q . Thus its dynamics is governed by dFN t; TN , TN+1 = σ(t)FN t; TN , TN+1 dWt, T0 t TN , ≤ ≤ N+1 where dWt is a brownian motion under Q , which integrates to TN TN 1 2 FN TN ; TN , TN+1; xT = FN T0; TN , TN+1 exp σ(s) ds + σ(s) dWs N −2 ZT0 ZT0 1 TN = F T ; T , T exp σ(s)2 ds + x . N 0 N N+1 −2 TN ZT0 With this relation at hand the functional form of forward rate FN at time TN is known. Because forward rates and discount bond prices are related by 1 D(Tm, Tm+1; xTm ) Fm Tm; Tm, Tm+1; xTm = − , 1 m N, αm(Tm, Tm+1)D(Tm, Tm+1; xTm ) ≤ ≤ the functional form of D TN , TN+1; xTN unfolds itself as monotone decreasing function of xTN , 1 D TN , TN+1; xTN = . 1 TN 2 1+ αN (TN , TN+1)FN T0; TN , TN+1 exp σ(s) ds + xT − 2 T0 N R (2.11) 10 This result serves as the basis for the recursive calculation of functional forms at earlier times Tm < TN which will be extracted from market observed digital caplet prices. The payoff of a digital caplet expiring at time Tm with strike K is given by V T , K; x = D T , T ; x 1 , m m Tm m m+1 Tm Fm(Tm;Tm,Tm+1;xTm )>K and at time T0 < Tm < TN the numeraire rebased derivative value is therefore Vm T0, K; xT N+1 Vm Tm,K,D(Tm, TN+1; xT ) 0 = EQ m x D T , T ; x D T , T ; x T0 0 N+1 T0 m N+1 Tm N+1 D Tm, Tm+1; xTm = EQ 1 x D T , T ; x Fm(Tm;xTm )>K T0 m N+1 Tm QN+1 Q N+1 1 = E E xTm 1F (T ;x )>K xT , D T , T ; x m m Tm 0 m+1 N+1 Tm+1 (2.12) where in the last line the martingale property of numeraire rebased discount bonds (2.8) was used. Thus if the functional form of the numeraire bond D Tm+1, TN+1; xTm+1 is known at time Tm+1, the expected value on the right hand side can be calculated (2.12) because the conditional probability distribution of xt is known. Indeed, since the forward rate Fm is a monotone function of xTm through its dependence on the discount bond D Tm, Tm+1; xTm , there exists a unique value xT∗m for which the forward rate matches the digital caplet strike 4 value K, Fm t; Tm, Tm+1; xT∗m = K . Thus (2.12) is equivalent to Vm T0, K; x T0 ∞ ∞ 1 = p xTm+1 xTm dxTm+1 p xTm xT0 dxT0 , D T0, TN+1; xT x D Tm+1, TN+1; xT 0 Z T∗m Z m+1 −∞ (2.13) which provides a relation between xT∗m and market derived derivative values. Indeed, be- cause D T0, TN+1; xTi and Vm T0, K; xTi can be observed in the market at time T0, the left hand side of (2.13) is known. With the known functional form D Tm+1, TN+1; xTm+1 the intergrals on the right hand side can be calculated numerically for varying values of xT∗m . That value of xT∗m for which the left and right hand side of (2.13) are equal is the 5 desired target value which satisfies the relation Fm Tm; Tm, Tm+1; xT∗m = K. Conducting this matching procedure for a series of options with differen t strike values Kj (j = 1,...M) and values Vm T0, Kj; xT0 results in a set xT∗m,j Fm Tm; Tm, Tm+1; xT∗m,j = Kj 1 D Tm, Tm+1; xT∗m,j − = x∗ D Tm, TN+1; x∗ = 1+ αN (TN , TN+1)Kj , Tm,j Tm,j D T , T ; x m N+1 ∗Tm,j (2.14) 4 As mentioned above we assume that the discount bonds are monotone functions of xt. Due to their relation this behaviour transfers to the forward Libor rate. 5 In practice the solution for x∗Tm is found by a numerical root finding method, e.g., the Brent algorithm. 11 where in the second line the relation 1 D(Tm, Tm+1; xT∗m,j ) Kj = Fm Tm; Tm, Tm+1; xT∗m,j = − αm(Tm, Tm+1)D(Tm, Tm+1; x∗Tm,j ) D(T ,T ;x ) 1 m m+1 T∗m,j D Tm,TN+1;x∗ − D Tm,TN+1;x∗ = Tm,j Tm,j , 1 m N, D(T ,T ;x ) m m+1 T∗m,j ≤ ≤ αm(Tm, Tm+1) D T ,T ;x m N+1 T∗m,j D(T ,T ;x ) was used. Because m m+1 T∗m,j corresponds to the inner integral (the bracket term) D T ,T ;x m N+1 T∗m,j of equation (2.13) its value has already been determined in the course of finding x . T∗m,j Therefore the set identity (2.14) defines the numeraire discount bond at time Tm as function of xT∗m,j . Obviously, because only a finite number of options with different strikes can be observed in the market the sets in (2.14) are discrete. Therefore continuous functional forms have to be obtained by interpolation between the single set elements. Following the above reasoning the recursion scheme starts at time TN 1. Performing − the calibration according to (2.13) the expected value of the inverse of the already known functional D TN , TN+1; xTN given by (2.11) is calculated. As a result the set (2.14) and thus the functional forms D TN 1, TN+1; xT at time TN 1 are obtained. These in turn − N − serve as input for the calibration at time TN 2 where they enter the inner integral of (2.13). − Pursueing the recursion along these lines until time T1 the functional forms of the numeraire discount bond D Ti, TN+1; xT are established for times TN , TN 1,...,T1. i − 2.2.3 Multi–dimensional Markov Functional Models In the previous section we presented the Libor Markov functional model in which a one– dimensional Markov process was used as underlying for the discount bond term structure. However it is also possible to consider higher dimensional underlying processes. As men- tioned in the previous section the driving process should be selected in such a way that the essential features of the derivative product for which the model is designed are captured. As an example for which a two dimensional driving process is required the class of spread options can be considered. The payout structure of this kind of derivatives can depend on the level of two different rate types which follow individual dynamics. Therefore a two– dimensional process is required in order to model the separate rate components. Another example is the incorporation of stochastic volatility which we focus on in this thesis. In this context a two dimensional Markov process would encompass a rate and a volatility component. As discussed above the specification of the Markov process is only one part in the specif- cation of a Markov functional model. The second is the determination of the functional 12 forms for the numeraire and the discount bond on the boundary curve. These are ob- tained by a calibration procedure for which it is essential that the functionals are monotone functions of the driving Markov process. But when higher dimensional Markov processes zt are considered the monotonicity of functionals can no longer be maintained because e.g., in a two dimensional extension of the matching equation (2.13) more than one tuple zt = (zt,1,zt,2) would be obtained as target value. For an n–dimensional Markov process zt this problem can be overcome by the introduction of a function n π : R R R, (t, zt) π(t, zt) =: xt, × −→ 7−→ which serves as projector to the one dimensional real axis, comp. [13]. In general, the process xt = π(t, zt) thereby defined will not be Markovian. However this fact poses no impediment since the function π merely serves as a means to facilitate the calibration to derivative prices observed in the market. Following this approach the numeraire discount bond at time Tm becomes a functional of the multi–dimensional Markov process, D Tm, TN+1; zTm , and in the calibration procedure the expected values are calulated with respect to the conditional distribution of zt. Therefore the only modification which needs to be applied to the calibration equation (2.12) is the change from one– to multi–dimensional integrals: Vm T0, K; xT0 = π(T0, zT0 ) D T , T ; x = π(T , z ) 0 N+1 T0 0 T0 N+1 Vm Tm,K,D(Tm, TN+1; xT ) = EQ m x D T , T ; x T0 m N+1 Tm QN+1 QN+1 1 = E E xTm 1F (T ;x )>K xT D T , T ; x m m Tm 0 m+1 N+1 Tm+1 QN+1 QN+1 1 = E E xTm 1x >x xT D T , T ; x Tm T∗m 0 m+1 N+1 Tm+1 QN+1 QN+1 1 = E E zTm 1π(Tm,z )>x zT D T , T ; z Tm T∗m 0 m+1 N+1 Tm+1 ∞ ∞ 1 = p zTm+1 zTm d zTm+1 p zTm zT0 dzT0 , x =π(T ,z ) D Tm+1, TN+1; zT Z T∗m m T∗m Z m+1 −∞ (2.15) x F 1 T T , T K . n with T∗m = m− ( m; m m+1; ) The variable zT∗m describes a curve in the –dimensional state space for which xT∗m = π(Tm, zT∗m ) at time Tm. Thinking about possible choices for the function π(t, zt) the concept of a pre–model was devised [14]. It is based on the idea that π can be defined as approximation to a model which has already been calibrated to market prices, e.g., a Libor Market Model where the drift terms have been frozen to their initial values. Since the so defined pre–model is an 13 approximation only it is not arbitrage free. However, since the no–arbitrage requirement is inherent in equation (2.15) a calibration to market quotes via the pre–model will result in an arbitrage free model. 2.3 A Libor Market Model as pre–model for a Markov Func- tional Model In this section an approximation of a forward Libor Market Model (LMM) will be considered which will then be used as pre–model for a Markov functional model. Working in the N+1 terminal measure Q the forward rate Fi t; Ti, Ti+1 is modelled according to (2.4), Fi t; Ti, Ti+1 = Fi T0; Ti, Ti+1 t t N+1 1 2 2 exp µ (s) λi(s) σi(s) ds + λi(s)σi(s) dW s , × i − 2 | | ZT0 ZT0 (2.16a) e K αl(Tl, Tl+1)λl(t)ρli(t)Fl t; Tl, Tl+1 µN+1(t)= , (2.16b) i − 1+ α (T , T )F t; T , T l=i+1 l l l+1 l l l+1 X e T0 t Ti, i = 1,...,N, 1 K N, ≤ ≤ ≤ ≤ N+1 where dW s is a K–dimensional Brownian motion under Q and σi(s) the load vector which encodes the effect of the individual orthogonal Brownian drivers on forward rate Fi. From this dynamics it is obvious that the change of forward rate Fi depends on the state of all forward rates Fl (l>i) at time t which is why the individual processes Fi(t) are not Markovian6. Therefore one usually resorts to Monte–Carlo methods in numerical evalua- tions which gets quite expensive as the number of factors increases. However, computations can be alleviated by referring to an approximation proposed by Rebonato due to which the forward rates Fl t; Tl, Tl+1 in the drift term (2.16b) are replaced by their time T0 values Fl T0; Tl, Tl+1 , a process which is also denoted as partial freezing [22]. A further simplifi- cation can be achieved by replacing the Brownian motion terms with normally distributed variables which exhibit the same mean and variance. Employing these ideas the forward rate vector can be approximated as N+1 F(t) F(t)= F(T )exp µ0 (t)+ M(t) zt , ≈ 0 · 6However, the process for the entire forward rate vector F(t) is Markovian since its change at time t only depends on the state of the forward rate vector at time t. 14 with vectors t N+1 N+1 1 2 2 µ0 (t)= µ0i (s) λi(s) σi(s) ds , T0 − 2 i=1,...,N Z h i K eαl(Tl, Tl+1)λl(t)ρli( t)Fl T 0; Tl, Tl+1 where µ N+1(t)= , 0i − 1+ α (T , T )F T ; T , T l=i+1 l l l+1 l 0 l l+1 X e zt = t,i , where t,i 0,t T0 , Z i=1,...,K Z ∼N − and the matrix 1 t 2 1 2 2 M(t)= λi(s) σik(s) ds . t T0 T i=1,...,N − Z 0 k=1,...,K Indeed, the terms Mik(t) t,k satisfy Z K K K t E Mik(t) t,k = Mik(t)E t,k =0= E λi(s)σik(s) dWk(s) , Z Z k=1 k=1 k=1 ZT0 X X t X = E λi(s)σi(s) dW s ZT0 t because λi(s)σik(s) dWk(s) are Ito integrals, ZT0 K K K and var Mik(t) t,k = E Mik(t)Mil(t) t,k t,l = Mik(t)Mil(t)E t,k t,l Z Z Z Z Z k=1 k,l=1 k,l=1 X X X K = Mik(t)Mil(t)cov t,k t,l Z Z k,l=1 X =δkl(t T0) − K t 2 | 2 {z } = λi(s) σik(s) ds T0 Xk=1 Z K t 2 2 = E λi(s) σik(s) ds T0 Xk=1 Z K t 2 = E λi(s)σik(s) dWk(s) by Ito’s isometry T0 Xk=1 Z K t 2 = λi(s) σik(s)σil(s) E dWk(s)dWl(s) k,l=1 ZT0 X =cov[ dWk(s),dWl(s)] K t | {z } = var λi(s)σik(s) dWk(s) k=1 ZT0 Xt = var λi(s)σi(s) dW s , i = 1,...,N. (2.17) ZT0 Above proxy processes can be related to the original forward rates Fi(t) by introducing monotone functions gi which act as perturbations on F¯i(t, zt). Thus Fi(t) = gi F¯i(t, zt) , 15 and one can define projector functions πi := gi F¯i by ◦ K πi : R R R, (t, zt) πi(t, zt) := gi F i(t, zt)), × −→ 7−→ K N+1 with F i(t, zt)= Fi(T0)exp µi (t)+ Mik(t)zt,k , for i = 1,...,N, Xk=1 which were introduced in the previous section. Hence the numeraire discount bond becomes a functional of the Markov process zTm through its dependence on Fm(Tm)= πm(Tm, zTm )= gm F m(Tm, zTm ) . Based on this approach the incorporation of stochastic volatility into a Markov functional framework will be devised in chapter 4. 16 Chapter 3 Piterbarg’s term structure of skew forward Libor model This chapter is dedicated to a survey of Piterbarg’s term structure of skew forward Libor model (FL–TSS) which was introduced in [19]. As with other forward Libor Market Models the main motivation is to capture the dynamics of the joint distribution of forward Libors throughout time. To facilitate this the forward Libor dynamics has to be flexible enough to capture information on the marginal distributions which is encoded in caplet and/or swaption prices. Since for these products the market implied volatilities exhibit a skew, i.e., the Black implied volatilities appear to be functions of the option strikes, Piterbarg considers a weighted sum of log–normal and normal dynamics for the forward Libors. Mathematically this is expressed by the introduction of a skew parameter. Reference to the swaption market necessitates this parameter to be time–dependent in order to reproduce swaption skews across expiries and underlying swap maturities. It is this time–dependent skew parameter which renders the model more flexible over earlier (constant) skew models, e.g., Andersen and Andreasen [2]. Furthermore, to account for the market observed variability of volatility levels, similar to the formulation of [2] these are modelled as stochastic CEV processes whose Brownian components are assumed to be uncorrelated with the stochastic drivers of the forward rate processes. Hence the FL-TSS model belongs to the class of stochastic volatility models. 3.1 The forward Libor dynamics Following the qualitative description of Piterbarg’s model we know set out to specify the dynamcis of forward Libors Fi t; Ti, Ti+1 resetting at times Ti, i = 1,...,N. Working in N+1 the measure Q induced by choosing the terminal discount bond D t, TN+1 as numeraire the forward rates are modelled as 17 K N+1 N+1 dFi(t)= βi(t)Fi(t)+ 1 βi T0 Fi T0 λi(t) Σt µ (t) dt + σi,l(t)dW (t) , − i l " l=1 # p X e (3.1a) N+1 dΣt =Θ Σ0 Σt dt + η Σt dV , (3.1b) − t p for times T0 t < Ti TN i = 1,...,N . All forward rates are driven by K independent ≤ ≤ N+1 Brownian motions dWl ( t) which are assumed to be uncorrelated with the stochastic N+1 N+1 N+1 volatility driver dVt and therefore dVt ,dWl (t) = 0. Their influence on the forward Libors Fi is mediated by the load factors σi,l(t) which contain information on K forward Libor correlation since the relation l=1 σi,l(t)σj,l(t) = ρij(t) holds. The drift N+1 terms µi (t) i = 1,...,N result from workingP in the terminal measure and are given by expression (2.6). It vanishes for the last forward rate FN t; TN , TN+1 since it is a log–normal e N+1 N+1 martingale under Q . Hence µN (t) = 0. The model parameters are hence the time–dependent forward skews βi(t) and volatility e levels λi(t) as well as the time–independent volatility of variance η and mean reversion speed Θ. Piterbarg chooses the latter to be constant which reduces the degrees of freedom to the set of tuples βi(t), λi(t) i=1,...,N associated with forward Libors Fi t; Ti, Ti+1 for times T0 t Ti TN and i = 1,...,N. These parameters characterise the distribution ≤ ≤ ≤ of each forward Libor as they are not affected by a change of measure. Reference to (3.1a) clearly reveals the role of the skews βi(t) as parameters mixing a purely log–normal with a normal forward Libor dynamics. From there it also is apparent that the parameters λi(t) determine the level of the stochastic volatility √Σt which is governed by (3.1b). Together with correlations ρij(t) among forward rates Fi and Fj these parameters determine the joint distribution for times T t T . 0 ≤ ≤ N The term structure of forward Libor skews and volatility levels have to be obtained by calibration to market prices of caplets/floorlets or swaptions. As outlined above the FL- TSS model is calibrated to the swaption market. Since forward swap rates are the natural swaption underlyings, it is therefore necessary to formulate a consistent forward swap rate dynamics and relate the resulting forward swap skews and volatility levels to their forward Libor counterparts. This is done by requiring that the process governing the forward swap rates should have the same structure as the forward Libor dynamics. The detailing of this idea will be presented in the following section. 18 3.2 Swap rate dynamics under the FL-TSS model To derive a consistent forward swap rate dynamics we first observe that the par rate Smn(t) of a forward swap starting at time Tm >t T0 and maturing at Tn > Tm can be expressed ≥ as a weighted sum of its constituent forward rates Fm,...Fn 1. : − D(t, Tm) D(t, Tn) Smn(t)= n 1 − l=−m αl(Tl, Tl+1)D(t, Tl+1) n 1 P− αl(Tl, Tl+1)D(t, Tl+1) = Fl t; Tl, Tl+1 n 1 α T , T D t, T l=m l=−m l( l l+1) ( l+1) X P =:wl(t) n 1 − | {z } = wl(t)Fl t; Tl, Tl+1 , (3.2) l=m X which follows from the forward rate definition D(t, Tl) D(t, Tl+1) Fl t; Tl, Tl+1 = − , αl(Tl, Tl+1)D(t, Tl+1) where αl(Tl, Tl+1) denotes the year fraction of the period [Tl, Tl+1] and D(t, Tj) stands for the discount factor corresponding to time Tj. Thus the forward swap rate is a function of its constituent forward rates, Smn(Fm,...,Fn 1), − and a stochastic differential equation is arrived at by application of Ito’s lemma: n 1 n 1 2 − ∂Smn(t) 1 − ∂ Smn(t) dSmn(t)= dFl(t)+ dFl(t)dFk(t) ∂Fl(t) 2 ∂Fl(t)∂Fk(t) lX=m l,kX=0 n 1 − ∂Smn(t) = βl(t)Fl(t)+ 1 βl T0 Fl T0 λl(t) Σt ∂Fl(t) − l=m X =: ϕ(Fl(t)) p T m n m n | {z σ (t) µ}( , )dt + dW( , )(t) × l · l n 1 2 h i 1 − ∂ Smn(t) + ϕ(Fl(t))ϕ(Fk(t))ρlk(t)dt 2 ∂Fl(t)∂Fk(t) l,kX=m n 1 − ∂S (t) T m n = mn ϕ(F (t))λ (t) Σ σ (t)dW( , )(t) + drift terms. (3.3) ∂F (t) l l t l l m l X= p Above expression is formulated in the swap measure Q(m,n) which is induced by us- n 1 1 ing the present value of a basis point Pmn(t)= l=−m αl(Tl, Tl+1)D(t, Tl+1) as numeraire . (m,n) The K–dimensional drift vectors µl accountP for the change from the terminal measure m n QN+1 to Q(m,n), under which dW( , )(t)isa K–dimensional Brownian motion. Additional 1 Strictly speaking the quantity Pmn(t) refers to a notional of 1 and therefore represents the present value of 10,000 basis points. 19 T K drift terms arise from the non–zero correlations ρlk(t) = σl (t)σk(t) = j=1 σl,j(t)σk,j(t) between forward rates Fl and Fk. But since covariances and therefore volatilitiesP and corre- lations remain invariant under a change of measure, one can consider above swap dynamics under a new measure Q with associated K–dimensional Brownian motion dW(t) in which the drift terms vanish. Under Q (3.3) transforms into e f n 1 −e ∂S (t) T dS (t)= mn ϕ(F (t))λ (t) Σ σ (t)dW(t). (3.4) mn ∂F (t) l l t l l m l X= p f As was already mentioned in the introduction we are looking for a forward swap rate dynamics which has the same structure as the forward rate process. This is equivalent to requiring a dynamics of the form K dSmn(t)= βmn(t)Smn(t)+ 1 βmn T0 Smn T0 λmn(t) Σt σmn,l(t)dWl(t), − l=1 =: ϕ(S (t)) p X e mne e e f | {z } (3.5) with forward swap rate skews βmn(t), volatility levels λmn(t), and Brownian loadings σmn,l(t) (l = 1,...,K). In order to ensure consistency between the two formulations we set (3.4) e e e equal to (3.5) and in addition match the slopes of both expression with respect to all forward rates Fl. This reasoning results in direct relations between swap and forward rate volatility levels and skews, respectively, which will be presented in the following subsections. 3.2.1 Derivation of the forward swap volatility level Matching of both expressions for the forward swap rates results in the requirement n 1 − ∂S (t) T T mn ϕ(F (t))λ (t) Σ σ (t)dW(t)= ϕ(S (t))λ (t) Σ σ (t)dW(t) ∂F (t) l l t l mn mn t mn l m l X= p p f e f from which the forward swap volatility level can be derived: e n 1 T − ∂Smn(t) ϕ(Fl(t)) T λmn(t)σmn(t)= λl(t) σl (t) (3.6) ∂Fl(t) ϕ(Smn(t)) lX=m e n 1 e − ∂Smn(t) ϕ(Fl(t)) T = λmn(t)= σl (t)σmn(t)λl(t), (3.7) ⇒ ∂Fl(t) ϕ(Smn(t)) lX=m T e e since σmn(t)σmn(t) = 1. Whereas the load factors of the orthogonal forward Libor drivers can be extracted from the correlation matrix 2 this does not apply to the Brownian motions e e 2 We assume that the correlations between forward Libors Fi and Fj are given in the parametric form ǫ ǫ (2.1) propsed by Rebonato [21]: ρij (t) = ρ + (1 − ρ ) exp −δ (Ti − t) − (Tj − t) ∞ ∞ 20 driving the swap rate. In above case of K > 1 Brownian swap rate drivers we therefore have to retrieve the swap rate volatility by referring to (3.6) instead of the volatility level λmn(t) itself. It is obvious that forward skew and volatility parameters as well as their swap counterparts are intertwined in above expressions, which is due to the presence of the skew e functions ϕ(Fl(t)) and ϕ(Smn(t)). This dependency is resolved when (3.6) is considered at the money, i.e., at the time T0 forward Libor and swap rate points. In this case ϕ(Fl(t)) ∂Smn(t) and ϕ(Smn(t)) become skew independent, and by further freezing (which are given ∂Fl(t) by (A.11)) at their initial values (3.6) simplifies to n 1 T − ∂Smn(t) Fl(T0) T λmn(t)σmn(t)= λl(t) σl (t), (3.8) ∂Fl(t) Smn(T0) l=m t=T0 X e which is the expression one which calibration will be based. 3.2.2 Derivation of the forward swap skew The forward swap and Libor skew can be interpreted as slopes of the respective skew functions ϕ(Smn(t)) and ϕ(Fl(t)) which are linear in the underlying rate variable. Hence we have ∂ βmn(t)Smn(t)+ 1 βmn T0 Smn T0 ∂ϕ(Smn(t)) − βmn(t)= = , ∂Smn(t) ∂ϕ( Smn(t)) e e e ∂ βl(t)Fl(t)+ 1 βl T0 Fl T0 ∂ϕ(Fl(t)) − βl(t)= = for l = m,...,n 1, ∂Fl(t) ∂F l(t) − and a relation between both quantities can be established by matching the slope of dSmn(t) in formulations (3.4) and (3.5) with respect to the forward rates Fk(t). Referring to (3.4) and assuming that the derivatives ∂Smn(t) (l = m,...,n 1) do not vary significantly as ∂Fl(t) − Fl(t) varies with time, we thus obtain n 1 2 ∂ dSmn(t) − ∂Smn(t) ∂ϕ(Fl(t)) ∂ Smn(t) T = + ϕ(Fl(t) λl(t) Σ σ (t)dW(t) ∂F (t) ∂F (t) ∂F (t) ∂F (t)∂F (t) t l k l=m l k k l X p 0 ≈ f n 1 − ∂Smn(t) | T {z } βk(t)δlkλl(t) Σt σl (t)dW(t) ≈ ∂Fl(t) l=m t=T0 X p ∂Smn(t) T f = βk(t)λk(t) Σt σk (t)dW(t), (3.9) ∂Fk(t) t=T0 p f where our assumptions took effect in the second line by freezing ∂Smn(t) at their initial values ∂Fl(t) and neglecting second derivatives of the swap rate with respect to the forward Libors. 21 Focussing on the formulation in terms of swap volatility levels and skews (3.5), a similar analysis yields ∂ dSmn(t) ∂ϕ(S (t)) ∂S (t) T = mn mn λ (t) Σ σ (t)dW(t) ∂F t ∂S t ∂F t mn t mn k( ) mn( ) k( ) p ∂Smn(t) e T f βmn(t) λmn(t) Σt σemn(t)dW(t). (3.10) ≈ ∂Fk(t) t=T0 p e e f Matching of the forward swap slopes with respect to Liborse Fm,...,Fn 1 in equations − (3.10) and (3.9) results in a system of n m equations between forward swap and Libor − skews: βmn(t)λmn(t)σmn(t)= βk(t)λk(t)σk(t) (k = m,...,n 1), − to which no uniquee solutione exists.e Therefore one has to revert to a least squares optimisa- tion to find an approximate solution. For this sake we consider the functional n 1 − 2 J βmn(t) = βmn(t)λmn(t)σmn(t) βk(t)λk(t)σk(t) , − k=m X e e e and obtain the optimal solution for the forwarde swap skew by requiring the derivative of J with respect to βmn(t) to vanish: n 1 dJe T − 0= = 2 λmn(t)σmn(t) βmn(t)λmn(t)σmn(t) βk(t)λk(t)σk(t) dβmn(t) − kX=m n 1 T e e− λ (t)σe (t)e λ (t)eσ (t) e 1 mn mn k k = βmn(t)= T βk(t). (3.11) ⇒ n m λ (t)σ (t) λ (t)σ (t) − k=m mn mn mn mn X e e e As the central result of this sectione wee hence havee establishe ed the relation between for- ward swap skews and volatilities to their forward Libor counterparts for each swap rate Smn(t) over time interval T0 t Tm. It is this time dependence of the forward Libor ≤ ≤ parameters which defines the central idea of a skew term structure in Piterbarg’s FL–TSS model. As mentioned earlier, in calibrating the FL–TSS model to the swaption market Piterbarg uses skew and volatility parameters of a Heston model as market input. By construction Hestons’s model parameters are time–independent which necessitates the in- troduction of a time-averaging method for the swap rate parameters within the FL–TSS. This requirement results in the formulation of effective skew and volatility parameters which are central in paving the way to calibration. 22 3.3 The effective skew and volatility formulation This section is dedicated to the formulation of swap rate effective skews and volatilities within Piterbarg’s FL–TSS model. In summary, these facilitate the transition from a swap rate dynamcis with time–dependent parameters (3.5), T dSmn(t)= βmn(t)Smn(t)+ 1 βmn T0 Smn T0 λmn(t) Σt σ (t)dW(t), − mn p to a formulation basede on time–independente ones, e e f T dSmn(t)= β Smn(t)+ 1 β Smn T0 λmn Σt σ (t)dW(t), mn − mn mn p where for each swap rate Smn(t) the effective skews βmn and volatilitiese λmnfare in principle given as weigthed time averages of the time–dependent quantities (3.8) and (3.11). The detailing of this central concept will be provided in the following, where at first attention will be paid to the effective skew in subsection 3.3.1 after which effective volatility is covered in 3.3.2. 3.3.1 The effective forward swap skew Piterbarg arrives at the effective swap rate skew by considering two diffusion processes where one has a time–dependent local volatility function and the other a time–independent one. The latter is defined as a weighted average of its time–dependent counterpart. Focussing on the weight function w(t) on time interval [0, T ], Piterbarg derives an explicit expression such that the average of differences between european swaption prices across an infinite range of strikes, calculated with respect to the respective processes, tends to zero as valuation time approaches T0. In detail, the following theorem holds: Theorem 3.3.1 (Piterbarg [19]). For T > 0, let f C1 [0, T ] R, R+ be a local volatility ∈ × function satisfying the usual growth requirements. Let σ(t), t [0, T ] be a function of time ∈ only. Fix x R. For any ǫ> 0 define a re–scaled local volatility function 0 ∈ 2 fǫ(t,x)= f tǫ ,x + (x x )ǫ , 0 − 0 and assume without loss of generality f(t,x ) 1, t [0, T ] 0 ≡ ∈ which implies fǫ(t,x ) 1, t [0, T ]. 0 ≡ ∈ 23 Let w(t), t [0, T ] be a weight function satisfying ∈ T w(t) dt = 1, Z0 and define an averaged local volatility function T 2 2 f ǫ(x) = fǫ(t,x) w(t) dt. (3.12) Z0 Further define two families of diffusions indexed by ǫ, dXǫ(t)= fǫ t, Xǫ(t) λ(t) Σ(t)dW (t), (3.13a) dYǫ(t)= f ǫ Yǫ(t) λ(t) pΣ(t)dW (t), (3.13b) Xǫ(0) = x0 , p (3.13c) Yǫ(0) = x0, (3.13d) for t [0, T ] with ∈ dΣt =Θ Σ0 Σt dt + η Σt dV (t), dV (t),dW (t) = 0. − p If the weights w(t) are given by the expression 2 2 v(t) λ(t) 2 2 w(t)= , with v(t) = E Σ(t) X0(t) x0 , (3.14) T 2 2 − 0 v(t) λ(t) dt h i then R ∞ + + 2 E Yǫ(T ) K E Xǫ(T ) K dK = O ǫ (3.15) − − − Z −∞ for ǫ 0.3 → By applying above theorem to the swap rate Smn(t) with local volatility function ϕ(Smn(t)) f t, Smn(t) = and replacing time zero with T0, a formulation with time–independent Smn(T0) effective skew parameters is obtain according to T dSmn(t)= βmn(t)Smn(t)+ 1 βmn T0 Smn T0 λmn(t) Σt σ (t)dW(t), − mn T = S (T )f t, S ( t) λ (t ) Σ σ (t)dW(t), p mne 0 1 mn mne t mn e e f T Smn(T0)f Smn(t) λmn(t) pΣt σ (t)dW(t), ≈ 1 e mne f T = β Smn( t)+ 1 β Spmn T0 λmn(t) Σt σ (t)dW(t), mn −e mn e f mn 3 p While above theorem provides a relation between the twoe diffusions, ite has to bef observed that (3.15) is not formulated in terms of absolute values. Hence the fact that Yǫ(t) complies with relation (3.15) in the limit ǫ → 0 does not generally ensure convergence to Xǫ(t) in probability. 24 Tm with βmn = 0 βmn(t)w(t) dt. By referring to above theorem one has to bear in mind that the derived relationR between the time–dependent and time averaged re–scaled local volatility e function holds in the limit ǫ 0. This corresponds to regarding the local volatility function → at time zero, although we are interested in a time–independent proxy for f(t,x) which would be obtained in the limit ǫ 1. In this sense the so derived effective skew parameters have → to be regarded as approximate results since their derivation is based on the consideration of f1(t,x), the re-scaled local volatility function with ǫ = 1. Their derivation is presented as corollary: Corollary 3.3.2 (Piterbarg [19]). The effective skew βmn for the equation T dSmn(t)= β Smn(t)+ 1 β Smn T0 λmn(t) Σt σ (t)dW(t), mn − mn mn p over a time horizon [T0, Tm] is given by e e f Tm βmn = βmn(t)wmn(t) dt, ZT0 where the weights w(t) are given by e 2 2 vmn(t) λmn(t) wmn(t)= , (3.16) Tm 2 2 vmn(t) λmn(t) dt 0 e t t eΘ(s T0) e Θ(s T0) 2 R 2 2 2 Θ(t T0) 2 − − − vmn(t) = Σ T0 λmne (s) ds + Σ T0 η e− − λmn(s) − ds. 2Θ ZT0 ZT0 e e (3.17) Proof. The proof is provided in section A.4 of Appendix A. In calibrating the model to the market the swap rates Smn(t) will be considered at discrete points in time Tk, T Tk < TK = t Tm, between which the volatility λmn(t) 0 ≤ ≤ will be assumed to be constant. For such a piecewise constant swap rate volatility function e 25 2 vmn(t) simplifies to: t t eΘ(s T0) e Θ(s T0) 2 2 2 2 Θ(t T0) 2 − − − v (t) = Σ T λ (s) ds + Σ T η e− − λ (s) − ds mn 0 mn 0 mn 2Θ ZT0 ZT0 K 1 2 − e 2 e = Σ T0 λmn(Tk) Tk+1 Tk − Xk=0 K 1 e − Tk+1 eΘ(s T0) e Θ(s T0) 2 Θ(t T0) 2 − − − + Σ T η e− − λ (Tk) − ds 0 mn 2Θ k=0 ZTk X K 1 e 2 − 2 = Σ T0 λmn(Tk) Tk+1 Tk − Xk=0 K 1 η2 e − Θ(t T0) 2 Θ(Tk+1 T0) Θ(Tk+1 T0) + Σ T e− − λ (Tk) e − + e− − 0 2Θ2 mn k=0 X Θ(Tek T0) Θ(Tk T0) e − e− − . (3.18) − − Above corollary veers towards a formulation in terms of time–independent, effective parameters in providing a link between time–dependent and effective swap rate skews. As a result the swap dynamics takes the form T dSmn(t)= β Smn(t)+ 1 β Smn T0 λmn(t) Σt σ (t)dW(t), (3.19) mn − mn mn p which is time–independent as far as the skews are concernede bute whichf still refers to the time–dependent swap volatility levels λmn(t). This gap will be closed in the following by deriving an effective volatility which completes our work towards a swap rate dynamics with e time–independent parameters. 3.3.2 The effective forward swap volatility We begin by observing that (3.19) represents a displaced diffusion process with displacement 1 βmn parameter γmn = − Smn T0 , comp. [23]. Indeed (3.19) can be recasted, βmn