riskcourse lmm

Upload: suedesuede

Post on 12-Oct-2015

42 views

Category:

Documents


0 download

DESCRIPTION

Risk training course on LMM

TRANSCRIPT

  • LIBOR market model andits uses

    Andrew Lesniewski, Ph.D.Managing Director

    [email protected]

    Ellington Management Group53 Forest Avenue

    Old Greenwich, CT 06870

    Libor market model p. 1

  • Overview

    Dynamics of Libor forward rates and martingale measuresVolatility structures of Libor forwardsAsymptotic solution to the LMMThe vanilla systemValuation of caps and floorsAsymptotic valuation of swaptionsParametrization of instantaneous volatilities and correlations

    Stable least square optimization

    Libor market model p. 2

  • Overview

    Discretization of the LMM and effective methods for generatingMC pathsAccurate and fast asymptotic calculation of the drift termsValuation of non-callable securities

    Valuation of securities with embedded Bermudan optionsLMM and securities with uncertain time horizons: prepaymentrisk, credit risk

    LMM and the SABR model of stochastic volatilityRisk management with LMM: delta, gamma, vega, etc.

    Libor market model p. 3

  • Dynamics of LMM

    We consider a sequence of approximately equally spaced dates

    0 T0 < T1 < . . . < TN ,

    which are called the standard tenors. A standard Libor forward rate

    Lj , j = 0, 1, . . . , N 1,

    is associated with a forward rate agreement which starts on Tj andends on Tj+1. Usually, we assume N = 120 and the Lj s are 3 monthLibor forward rates. These dates are the actual start and end dates ofthe contracts rather than the fixing dates. For simplicity, we disregardthis distinction. Proper implementation takes it into account.

    Libor market model p. 4

  • Dynamics of LMM

    We model Lj as a continuous time stochastic process Lj (t),0 t Tj1 (killed at t = Tj1!). The dynamics of the forward processis driven by an N -dimensional, correlated Wiener processW0 (t) , . . . ,WN1 (t). The probability measure associated with thisWiener process is denoted by P. We let jk denote the instantaneouscorrelation between Wj (t) and Wk (t), i.e.

    EP [dWj (t) dWk (t)] = jkdt .

    To motivate the form of the stochastic differential equations describingthe dynamics of the Lj s, consider the world in which there is novolatility of interest rates, i.e. for all js Lj (t) = Lj0, or

    dLj (t) = 0.

    Libor market model p. 5

  • Dynamics of LMM

    The fact that the rates are stochastic leads us to replace this with asystem of stochastic differential equations:

    dLj (t) = j (L (t) , t) dt+ Cj (L (t) , t) dWj (t) .

    The first term on the right hand side is called the drift term, and thesecond term is called the diffusion term. The no arbitrage requirementforces a relationship between the drift and the diffusion terms. Theform of the drift term depends on the choice of numeraire.

    Define : [0, TN ] Z by

    (t) = m, if t [Tm1, Tm) .

    Libor market model p. 6

  • Dynamics of LMM

    We consider two types of numeraire:Forward. For k = 1, . . . , N , choose as numeraire the zerocoupon bond

    Pk (t) = Pstub (t)

    (t)ik

    1

    1 + iLi (t),

    where Pstub (t) is the stub adjustment for the period [t, T(t)].Spot. Choose as numeraire the rolling bank account

    P0 (t) =P(t)1 (t)

    1i(t)1 Pi (Ti1).

    Let Qk, k = 0, 1, . . . , N denote the appropriate equivalent martingalemeasure.

    Libor market model p. 7

  • Dynamics of LMM

    Under the forward measure Qk, k = 1, . . . , N , the dynamics of theforward rate Lk (t) is a martingale. Indeed,

    Lk (t) =1

    k

    Pk1 (t) Pk (t)Pk (t)

    ,

    i.e. it is the difference of two assets. Consequently, the dynamics ofLk (t) should be driftless:

    dLk (t) = Ck(Lk (t) , t

    )dWk (t) ,

    where Ck(Lk (t) , t

    )is an instantaneous volatility function.

    Libor market model p. 8

  • Dynamics of LMM

    This and the change of numeraire techniques allow us to determinethe dynamics for j 6= k. Changing numeraire from Pj (t) to Pk (t), wefind that:

    dLj (t) = Cj(Lj (t) , t

    )

    j+1ik jiiCi

    (Li (t) , t

    )1 + iLi (t)

    dt+ dWj (t) , if j < k,

    k+1ij

    jiiCi(Li (t) , t

    )1 + iLi (t)

    dt+ dWj (t) , if j > k .

    Libor market model p. 9

  • Dynamics of LMM

    Finally, under the spot measure (k = 0), the LMM dynamics reads:

    dLj (t) = Cj(Lj (t) , t

    ) (t)ij

    jiiCi(Li (t) , t

    )1 + iLi (t)

    dt+ dWj (t)

    .

    These equations have to be supplied with initial values for the Liborforwards:

    Lj (0) = Lj0,

    where Lj0 is the current value of the forward which is implied by thecurrent yield curve.

    Libor market model p. 10

  • Factor reduction

    In a market spanning 30 years, there are 120 quarterly Libor forwards(i.e. 120 stochastic factors). This poses severe problems with themodels implementation:

    The curse of dimensionality kicks in, leading to unacceptablyslow performance.The parameters of the model are severely underdetermined andthe calibration of the model becomes unstable.

    We need a small number d of independent Brownian motions Za (t),a = 1, . . . , d,

    E [dZa (t) dZb (t)] = ab dt ,

    driving the process. Typically, d = 1, 2, 3, or 4.

    Libor market model p. 11

  • Factor reduction

    We express the true Brownian motions in terms of the Z s:

    dWj (t) =

    1ad

    Uaj dZa (t) ,

    where U is an N d matrix so that UU . We rewrite the dynamicsof the model in terms of the independent Brownian motions:

    dLj (t) = j (L (t) , t) dt+

    1ad

    Bja(Lj (t) , t

    )dZa (t) ,

    whereBja

    (Lj (t) , t

    )= Uaj C

    j(Lj (t) , t

    ).

    Libor market model p. 12

  • Factor reduction

    This system gives the factor reduced LMM dynamics. We interpretd UU as the correlation matrix corresponding to the factor reducedLMM dynamics. Clearly, its rank is equal to d, and

    d =d

    1ad

    EaEa,

    where Ea denotes the a-th column of the matrix U .

    Libor market model p. 13

  • Low noise expansion

    There is no known closed form solution to the LMM dynamics. We canconstruct an approximate solution by means of a low noise expansion:We introduce a smallness parameter in front of the diffusioncoefficients. This parameter is set to 1 at the end of the computation.Since the drift coefficients are quadratic in the diffusion coefficients, wemultiply them by 2. Expanding in powers of then yields:

    Lj (t) = Lj0 +

    1ad

    t0

    Bja (L0, s) dZa (s) +

    t0

    j (L0, s) ds

    +

    1a,bd1kN

    0ust

    Bka (L0, u)Bjb (L0, s)

    LkdZa (u) dZb (s)

    + . . . .

    Libor market model p. 14

  • Low noise expansion

    The low noise solution implies the following asymptotic expansion forthe expected value of E

    [Lj (t)

    ]:

    E[Lj (t)

    ]= Lj0 +

    t0

    j (L0, s) ds

    +

    1a,bd1kN

    ab

    t0

    Bka (L0, s)Bjb (L0, s)

    Lkds+ . . . .

    This formula shows that the naive expected value, namely todaysforward, has to be adjusted by a convexity correction which dependson the market volatility.

    Libor market model p. 15

  • Low noise expansion

    The covariance matrix Cov[Lj (t) , Lk (t)

    ]is asymptotically given by:

    Cov[Lj (t) , Lk (t)

    ]=

    1a,bd

    ab

    t0

    Bja (L0, s)Bkb (L0, s) ds

    +

    1a,b,a,bd1i,iN

    aabb

    t0

    t0

    Bia (L0, s)Bia (L0, s

    )

    Bjb (L0, s)

    LiBkb

    (L0, s)

    Lids ds + . . . .

    The first term in the formula above is what one would naively expect,while the second term is a convexity correction to this quantity.

    Libor market model p. 16

  • Optimization and Tikhonov regularization

    In the following, we will face various multidimensional optimizationproblems, all of which are ill posed (i.e. the number of unknownparameters is smaller than the number of data points). Their solutionrequires minimizing a loss function L0 (x):

    L0 (x) =i

    ri (x)2 ,

    with nonlinear residuals ri (x). Such problems are typically poorlyconditioned and numerically finicky. Instead, we will consider amodified loss function:

    L (x) = L0 (x) + Sx2,

    where S is a suitable smoothing operator.Libor market model p. 17

  • Vanilla system

    Vanilla system consists of:Forward curve stripper and swap pricing model. For ourpurposes, the forward curve is represented as a sequence of NLIBOR forwards Lj0, j = 0, . . . , N . These forwards are used tocalibrate the model, and as the initial value to evolve the LIBORforwards.

    Cap volatility stripper and cap / floor pricing model. The output ofthe cap vol stripper is a sequence of N caplet volatilities. Theyare used to calibrate the volatility structure of the model.Swaption pricing model. Swaption volatilities are also used tocalibrate the volatility structure of the model.

    Libor market model p. 18

  • Stripping cap volatility

    Caps and floors are baskets of calls and puts on Libor forwards struckat the same rate. Because of the complexity of their description (a 10year cap involves 39 caplets!), the market quotes them in terms of apremium or a single flat volatility. This flat volatility has the propertythat, when inserted in the pricing formula, it reproduces the optionpremium. In reality, caplet volatility exhibits a very pronounced termstructure. Typical of it is the persistence of a volatility hump whichusually peaks somewhere between 6 months and 2 years. Theprocess of constructing implied caplet volatility from market quotes issometimes called stripping.

    Libor market model p. 19

  • Stripping cap volatility

    We use a two step stripping algorithm.We fit the caplet volatilities to the hump function:

    H (t) = (+ t) exp (t) + .

    Once , , , and have been calibrated, we parameterize thecap volatility curve by means of cubic B-splines and optimize tonail down the fit to market prices.

    The outcome of stripping is a sequence j , j = 0, . . . , N 1 of N at themoney caplet volatilities. Specifically, j is the implied volatility of thecaplet expiring at Tj .

    Libor market model p. 20

  • Approximate valuation of swaptions

    A key ingredient of efficient calibration of the LMM is rapid andaccurate swaption valuation. This is hard, as:

    A swap rate is a non-linear function of the underlying Liborforward rates.

    The stochastic differential equation for the swap rate cannot besolved in closed form, and pricing swaptions requires MonteCarlo simulations.This poses a serious issue, as such simulations are costly.

    We describe an approximation for rapid swaption valuation, based onthe low noise expansion, which we use to calibrate the model.

    Libor market model p. 21

  • Approximate valuation of swaptions

    Since the purpose of the approximate formulas is to provide anefficient calibration tool, we focus on benchmark swaptions only.Consider a benchmark forward starting swap. The start date of theswap is denoted by Tm, and its end date is denoted by Tn. The levelfunction of the swap is defined by:

    Amn (t) =

    mjn1

    jPj (t) ,

    where j are the day count fractions for fixed rate payments. Typically,the payment frequency on the fixed leg is not the same as that on thefloating leg. This fact causes a bit of a notational nuisance but needs tobe taken properly into account for accurate pricing.

    Libor market model p. 22

  • Approximate valuation of swaptions

    The forward swap rate is given by:

    Smn (t) =Pm (t) Pn (t)

    Amn (t)

    =1

    Amn (t)

    mjn1

    jLj (t)Pj (t) .

    Incidentally, the formula above shows that the swap rate is amartingale under the measure associated with the level function.

    The swap rate process in LMM can be written as:

    dS (t) = (L, t) dt+

    mjn1

    j (L, t) dWj (t) .

    Libor market model p. 23

  • Approximate valuation of swaptions

    Here

    =

    mjn1

    S

    Ljj +

    1

    2

    mj,kn1

    jk2S

    LjLkCjCk ,

    andj =

    S

    LjCj .

    Explicitly, j is given by

    j (L, t) = Cj(Lj , t

    ) (Rj (L, t) + j (L, t)

    ),

    where the leading term is

    Rj (L, t) =Pj (t)

    A (t),

    Libor market model p. 24

  • Approximate valuation of swaptions

    and the correction term is given by:

    j =j

    1 + jLj

    S

    jln1

    lRl

    jln1

    lLlRl

    .

    To use this dynamics effectively, we approximate it by quantities withtractable analytic forms. The simplest approximation consists inreplacing the values of the stochastic forwards Lj (t) by their initialvalues Lj0. This amounts to freezing the curve at its current shape.

    Libor market model p. 25

  • Approximate valuation of swaptions

    Within this approximation, the coefficients are deterministic:

    j (L, t) j (L0, t) ,(L, t) (L0, t) .

    The expected value of the swap rate is given by

    E [S (t)] = S0 + t

    0

    (L0, s) ds,

    and its variance is

    Var [S (t)] =

    mj,jn1

    jj

    t0

    j (L0, s) j (L0, s) ds.

    Libor market model p. 26

  • Approximate valuation of swaptions

    The normal volatility mn of a swaption with expiry Tm is

    mn =

    1

    TmVar [Smn] .

    Consequently, its frozen curve approximation 0,mn is given by

    20,mn =1

    Tm

    mj,jn1

    jj

    t0

    j (L0, s) j (L0, s) ds .

    This formula is easy to implement in code, and leads to remarkablyaccurate results. One could, of course, go beyond this approximation,at the expense of producing a complicating and increasingly unwieldyanalytic expressions.

    Libor market model p. 27

  • Structure of instantaneous correlations

    An important input into the model calibration is determining theinstantaneous correlation matrix = {jk}0j,kN1. Thedimensionality of is N (N + 1) /2, posing an issue of finding a stablecalibration procedure. Possible strategies include:

    Semi-definite programming. Try to imply the correlations fromthe cap / floor and swaption markets. This approach leads tonon-intuitive results and is prone to overfitting.Historical data. This approach, in conjunction with the principalcomponent analysis leads to stable correlation structure. It maynot reflect the current market correlations.

    Parameterized correlations. May be the best choice.

    Libor market model p. 28

  • Structure of instantaneous correlations

    A realistic and flexible parametric form of instantaneous correlations isgiven by

    jk = + (1 ) exp( |Tj Tk|

    1 + min (Tj , Tk)

    ).

    Caution: this parametrization produces a matrix that is onlyapproximately positive definite. It has some clear advantages:

    It is intuitive: measures the overall level of correlations, isthe decay rate of correlations, and describes the short enddecorrelation.

    It is easy to calibrate: only three parameters are involved.Perturbing the parameters , , leads to meaningful riskmeasures.

    Libor market model p. 29

  • Parametrization of instantaneous volatilities

    So far we have been working with a general instantaneous volatilityCj(Lj (t) , t

    )for the forward Lj (t). In the implementation, we assume

    Cj(Lj (t) , t

    )to be one of the following standard models:

    Cj(Lj (t) , t

    )=

    j (t) (normal model),j (t)Lj (t)j (CEV model),j (t)Lj (t) (lognormal model),j (t)Lj (t) + j (shifted lognormal model),

    where the functions j (t) are deterministic, and where 0 j 1,j 0. These functions have to be suitably regularized at Lj 0 toprevent negative rates.

    Libor market model p. 30

  • Parametrization of instantaneous volatilities

    For the purpose of calibration we require that the deterministic volatilitycomponents j (t) are piecewise constant. That leads to the followingparametrization of the instantaneous volatility:

    j (t) t [T0, T1) [T1, T2) [T2, T3) . . . [TN1, TN )0 (t) 0 0 0 . . . 0

    1 (t) 1,0 0 0 . . . 0

    2 (t) 2,0 2,1 0 . . . 0

    3 (t) 3,0 3,1 3,2 . . . 0

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    N1 (t) N1,0 N1,1 N1,2 . . . 0

    Libor market model p. 31

  • Parametrization of instantaneous volatilities

    The table above contains 7140 parameters (assuming N = 120), andthe problem is vastly overparametrized !A natural remedy to the overparameterization problem is to assumethat the instantaneous volatility is stationary, i.e.,

    j,i = ji,0

    ji ,

    for all i < j. This assumption appears natural and intuitive, as it impliesthat the structure of cap volatility will look in the future exactly the sameway as it does currently. Consequently, the forward volatility problemplaguing the traditional terms structure models would disappear.

    Libor market model p. 32

  • Parametrization of instantaneous volatilities

    With the stationary volatility assumption we have the followingparametrization of the instantaneous volatility structure:

    j (t) t [T0, T1) [T1, T2) [T2, T3) . . . [TN1, TN )0 (t) 0 0 0 . . . 0

    1 (t) 1 0 0 . . . 0

    2 (t) 2 1 0 . . . 0

    3 (t) 3 2 1 . . . 0

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    N1 (t) N1 N2 N3 . . . 0

    Libor market model p. 33

  • Parametrization of instantaneous volatilities

    This assumption is unsuitable for accurate calibration of the model.The financial reason is the phenomenon of mean reversion of longterm rates. One cannot take it into account by adding anOrnstein-Uhlenbeck style drift term to the LMM dynamics as this wouldviolate the arbitrage freeness of the model. One can, however, achievea similar effect by modifying the instantaneous volatility function.We introduce a volatility kernel function K (, ) to account for thedeviation from the purely stationary model. A convenient form of thevolatility kernel is

    K (, ) = exp () .For each maturity Tj we choose a parameter j , and set

    Kj,i = K (Tj Ti, j) .Libor market model p. 34

  • Parametrization of instantaneous volatilities

    We are thus led to assume the following structure of the instantaneousvolatility:

    j (t) t [T0, T1) [T1, T2) . . . [TN1, TN )0 (t) 0 0 . . . 0

    1 (t) 1K1,0 0 . . . 0

    2 (t) 2K2,0 1K2,1 . . . 0

    3 (t) 3K3,0 2K3,1 . . . 0

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    .

    N1 (t) N1KN1,0 N2KN1,1 . . . 0

    Libor market model p. 35

  • Parametrization of instantaneous volatilities

    The lower triangular matrix above, LMMs internal representation ofvolatility, is referred to as the LMM volatility surface.The structure above may still overparameterize the instantaneousvolatility, depending on the number of benchmark options used tocalibrate the model. To further reduce the number of parameters weexpress the j s and j s as linear interpolations of a smaller numberof auxiliary parameters.

    Libor market model p. 36

  • Least square optimization

    In order to calibrate the model we seek instantaneous volatilityparameters i so that to fit the at the money caplet and swaptionvolatilities. These can be expressed in terms of the instantaneousvolatilities are as follows. The at the money volatility of the capletexpiring at Tm is given by:

    m (1, . . . , m, m)2=

    1

    Tm

    0im1

    2mi

    Ti+1Ti

    K (Tm t, m)2 dt

    1Tm

    0im1

    2miK2m,i (Ti+1 Ti) .

    Libor market model p. 37

  • Least square optimization

    The at the money volatility of the swaption expiring at Tm into a swapmaturing at Tn is approximately equal to

    m,n (, m, . . . , n1)2 =

    1

    Tm

    mj,kn1

    jkjm,n

    km,n

    m1i=0

    jikiKj,iKk,i (Ti+1 Ti) .

    Here, j are simply rescaled versions of the corresponding functionswhich were defined and calculated asymptotically.

    Libor market model p. 38

  • Least square optimization

    We calibrate the model by means of the Levenberg-Marquardtoptimization. The loss function is given by:

    L (, ) = 12

    m

    wm(m (, ) m

    )2+

    1

    2

    m,n

    wm,n(m,n (, ) m,n

    )2

    +1

    2j

    ()2j ,

    where m and m,n are the market observed caplet and swaptionvolatilities. The coefficients wm and wm,n are weights which allow oneselect the accuracy of calibration of each of the instruments. The lastterm is a Tikhonov regularizer.

    Libor market model p. 39

  • Approximation schemes for LMM

    For the purpose of asset pricing we solve numerically the SDEsdefining the LMM dynamics. Such a solution is a collection of MonteCarlo paths, each of which represents a future forward rate scenario.

    We choose a sequence of event dates t0, t1, . . . , tm, and denote byLjn Lj (tn) the approximate solution. We also use

    jn = j (Ln, tn) ,

    Cjan = Cja (Ln, tn) ,

    to denote the values of the drift and diffusion terms at time tn.

    Libor market model p. 40

  • Approximation schemes for LMM

    The simplest numerical scheme is Eulers scheme. It consists inreplacing the differentials by finite differences. For n = 1, . . . ,m, we let

    tn = tn tn1,dWna =Wa (tn)Wa (tn1) ,

    denote the time and Wiener process increments,respectively. Note thatdWna N

    (0,tn). Then, Eulers solution scheme reads:

    Ljn+1 = Ljn +

    jntn +

    1ad

    Uaj Cjn dWna .

    Eulers scheme converges at the rate of maxtn, as m.

    Libor market model p. 41

  • Approximation schemes for LMM

    A somewhat more accurate method is Milsteins scheme. It is arefinement of Eulers scheme which converges at the rate of max tn,as m. We set

    jabn = Uaj U

    bj C

    j(Ljn, tn

    ) CjLj

    (Ljn, tn

    ).

    Then, Milsteins scheme for the LMM reads:

    Ljn+1 = Ljn +

    jn 12

    1ad

    jaan

    tn

    +

    1ad

    Uaj Cjn dWna +

    1

    2

    1a,bd

    jabn dWna dWnb .

    Libor market model p. 42

  • Approximation schemes for LMM

    Having discretized the dynamics, we have to select a method forsimulating the Wiener process. An efficient method relies on thespectral decomposition of the covariance matrix of W (t) sampled att0, t1, . . . , tm. The covariance matrix is explicitly given by:

    Cij = E [W (ti)W (tj)]= min (ti, tj) .

    Consider the eigenvalue problem for C:

    CEj = jEj , j = 0, . . . , m,

    with orthonormal eigenvectors Ej s and eigenvalues

    0 . . . m 0.Libor market model p. 43

  • Approximation schemes for LMM

    We let Ej (tn) denote the n-th component of Ej , and represent thediscretized Wiener process as

    W (tn) =

    0jm

    j Ej (tn) j ,

    where j are i.i.d. random variables with j N (0, 1). These numbersare best calculated by applying the inverse cumulative normal functionto a sequence of Sobol numbers. In practice, we may want to use onlya certain portion of the spectral representation by truncating it at somep < m. This eliminates the high frequencies from W (tn), and lowersthe sampling variance. The price for this is lower accuracy.

    Libor market model p. 44

  • Computing the drift terms

    Evaluating the drift terms along each Monte Carlo path is timeconsuming and accounts for over 50% of total simulation time.Typically they are small compare to the initial values of the Liborforwards, and it is desirable to develop an efficient methodology foraccurate approximate evaluation of the drift terms.The first and simplest approach consist in freezing the values ofLj (t) at the initial value Lj0 Lj (0). We precompute the values

    j0 j (L0, 0) ,

    and use them for the drift terms throughout the simulation. Thisapproximation, the frozen curve approximation, is rather crude, anddoes not perform very well for long dated instruments.

    Libor market model p. 45

  • Computing the drift terms

    The order 3/4 approximation, uses the next order term in the low noiseexpansion:

    j1/2 (t) j0 (L0, 0) + a,j (L0, 0)Za (t) + j (L0, 0) t.

    Under the forward measure Qk, the coefficients a,j are given by:

    a,j =

    Cj j+1ik

    jiiCi

    1+iLi

    [Uaj

    Cj

    Lj + Uai

    (Ci

    Li iCi

    1+iLi

    )], if j < k,

    0, if j = k,Cj

    k+1ij

    jiiCi

    1+iLi

    [Uaj

    Cj

    Lj + Uai

    (Ci

    Li iCi

    1+iLi

    )], if j > k,

    Libor market model p. 46

  • Computing the drift terms

    and the coefficients j are given by:

    j =

    j+1ik

    jii1+iLi

    [j0C

    i Cj

    Lj +i0C

    j(Ci

    Li iCi

    1+iLi

    )], if j < k,

    0, if j = k,k+1ij

    jii1+iLi

    [j0C

    i Cj

    Lj +i0C

    j(Ci

    Li iCi

    1+iLi

    )], if j > k.

    Under the spot measure,

    a,j = Cj

    (t)ij

    jiiCi

    1 + iLi

    [Uaj

    Cj

    Lj+ Uai

    (Ci

    Li iC

    i

    1 + iLi

    )],

    Libor market model p. 47

  • Computing the drift terms

    and

    j =

    (t)ij

    jii1 + iLi

    [j0C

    i Cj

    Lj+i0C

    j

    (Ci

    Li iC

    i

    1 + iLi

    )].

    The order 3/4 approximation leads to excellent accuracy.

    On might easily refine this approach by computing terms of higherorder in stochastic Taylors expansion. This leads, however, to morecomplex and computationally expensive formulas, and the benefit ofusing an asymptotic expansion disappears. The order 1/2approximation appears to offer the best performance versus accuracyprofile.

    Libor market model p. 48

  • Valuation of non-callable securities

    Valuations within the LMM are based on the arbitrage free pricing law.For our purposes, the equivalent martingale measure Q is either one ofthe forward measures or the spot measure. The conditional expectedvalue is calculated by means of Monte Carlo simulations. A convenientchoice of the numeraire for all valuations is the spot numeraire. Thesample paths generation method described above produces lowvariance estimates of the expected value. In addition, one might use ageneric variance reduction method (such as antithetic variables) inorder to further reduce the sampling variance. For most instruments, arelatively small number of Monte Carlo paths leads to accurate andstable valuations. As few as 1000 paths are sufficient to producereliable prices.

    Libor market model p. 49

  • Valuation of Bermudan options

    Valuation of securities with embedded American (or Bermudan)options is more difficult, as the traditional Monte Carlo approach isinefficient. The problem is that while optimal exercise of a Bermudanoption requires solving a backward induction problem, Monte Carlopaths evolve forward in time. Recently, a number of efficientapproximate Monte Carlo algorithms for pricing Bermudan andAmerican options have been proposed. The approach we outline isinspired by the Longstaff - Schwarz approach. It is based on asequence of nested chaos expansions of the continuation values ateach exercise date. As few as 5000 Monte Carlo paths lead to reliablepricing.

    Libor market model p. 50

  • Valuation of Bermudan options

    We fix a measure Qk and consider a Bermudan swaption which can beexercised on dates t1, . . . , tM . If exercised on tj , it allows the holder toenter into a swap starting on (or about) tj . We let p (t) = p (L (t) , t)denote the payoff function. For a stopping time , we letp (t) = p (t ). Then its value at time t0 (today) equals

    V (t0) = supT (t1,...,tM )

    E [p |Ft0 ] ,

    where T (t1, . . . , tM ) denotes the set of stopping times taking values inthe set {t1, . . . , tM}. The stopping time at which the maximum isattained is called the optimal stopping time.

    Libor market model p. 51

  • Valuation of Bermudan options

    In order to simplify the notation, we assume one factor only, d = 1(nothing gets lost). We introduce the notation:

    Xj =W (tj)

    tj, j = 1, . . . ,M,

    so that Xj N (0, 1). From the properties of a Wiener process,

    Xj =

    tj1tj

    Xj1 +

    1 tj1

    tjj , with j N (0, 1) .

    We denote by gj (Xj) = p (L (tj) , tj) the payoff function at time tj , i.e.the present value (at time tj) of the corresponding swap.

    Libor market model p. 52

  • Valuation of Bermudan options

    Let Vj (Xj) denote the value of the option at time tj . Then,

    VM (XM ) = gM (XM ) .

    For j =M 1, . . . , 1, the value of the option is the greater of theimmediate payoff and the continuation value of the option:

    Vj (Xj) = max (gj (Xj) , D (tj , tj+1)E [Vj+1 (Xj+1) |Xj ]) ,

    where D (tj , tj+1) = Pk (tj+1) /Pk (tj). Todays value of the option is

    V0 (X0) = D (t0, t1)E [V1 (X1) |X0] .

    Libor market model p. 53

  • Valuation of Bermudan options

    The continuation value of the option at time tj is

    Cj (Xj) = D (tj , tj+1)E [Vj+1 (Xj+1) |Xj ] ,

    i.e.Vj (Xj) = max (gj (Xj) , Cj (Xj)) .

    We also have a recursion for the continuation values:

    CM (XM ) = 0,

    Cj (Xj) = D (tj , tj+1)E [max (gj+1 (Xj+1) , Cj+1 (Xj+1)) |Xj ] ,

    for j =M 1, . . . , 1. Todays option price is

    V0 (X0) = C0 (X0) .

    Libor market model p. 54

  • Valuation of Bermudan options

    We now would like to write

    Vj (Xj) =

    0k

  • Valuation of Bermudan options

    The expansions with respect to the Hermite polynomials are truncatedat some finite order , i.e.:

    Vj (Xj)

    0k

    ajk Hk (Xj) ,

    E [Vj (Xj) |Xj1]

    0k

    (tj1 t0tj t0

    )k/2ajk Hk (Xj1) .

    Relatively low values of , 5, lead to good numerical results.

    Libor market model p. 56

  • Valuation of Bermudan options

    The Fourier coefficients ak are replaced by their Monte Carloestimates. We shall choose the coefficients so as to minimize thesquare error:

    1

    2

    1iN

    Vj (X(i)j )

    0k

    ajkHk

    (X

    (i)j

    )2

    .

    This leads to the following estimate for the Fourier coefficients:

    aj G1j vj ,

    where Gj is a matrix whose components are:

    Libor market model p. 57

  • Valuation of Bermudan options

    (Gj)kl =1

    N

    1iN

    Hk

    (X

    (i)j

    )Hl

    (X

    (i)j

    ),

    and where vj is a vector with the components:

    (vj)k =1

    N

    1iN

    Vj

    (X

    (i)j

    )Hk

    (X

    (i)j

    ).

    From the performance point of view, is worthwhile to accelerate theconvergence of this series. Payoff functions of options are, typically,not smooth. This causes slowdowns of the rate of convergence at thepoints where the payoff has kinks (as in the usual calls or puts) ordiscontinuities (as in digital options).

    Libor market model p. 58

  • Valuation of Bermudan options

    Having estimated the conditional expected values (and thus calculatedthe consecutive continuation values), we construct the optimalstopping time as follows.

    Along a Monte Carlo path, find the earliest of the dates tj , wherethe immediate exercise outweighs the continuation value.Assign this date to the optimal stopping time along the path.Calculate the average over all Monte Carlo paths.

    Libor market model p. 59

  • Securities with uncertain horizons

    It is important to extent LMM so that instruments with risks other thaninterest rate risk can be valued. Certain classes of such instruments,such as mortgage backed securities (MBS) or credit sensitiveinstruments are characterized by the uncertainty of their time horizon.It is caused by an event such as prepayment in case of MBS, ordefault in case of credit sensitive instruments. Unlike the callableLIBOR exotics, where the termination decision is made optimally andconsistently with the rates process, these risks are largely exogenousto the rates process. Below we formulate some basic assumptionsabout modeling these instruments.

    Libor market model p. 60

  • Securities with uncertain horizons

    In order to make the framework consistent with LMM we make thefollowing assumptions.

    Assumption 1. The underlying uncertainty is modeled by theLIBOR market model.Assumption 2. Additional uncertainty (prepayment, default, ...)is modeled by a random time T . T is not required to be astopping time with respect to Ft. However, we assume that

    P ({T > t} |F) = P ({T > t} |Ft) .

    The second assumption means that the probability of event occurringdepends only on the information up to time t

    Libor market model p. 61

  • Securities with uncertain horizons

    Assumption 3. There exist a process (t) 0, the intensityprocess, so that

    P ({T > t} |Ft) = exp( t

    0

    (s) ds

    ).

    In the context of prepayment, (t) is the (continuous time versionof) SMM. In the context of credit, it is the hazard rate.Assumption 4. The event will occur almost surely,

    P ({T

  • Securities with uncertain horizons

    Valuation of securities with uncertain time horizons is done by meansof Monte Carlo simulations. The stochastic process

    S (t) = exp

    ( t

    0

    (s) ds

    )

    is called the survival probability, while

    F (t) = 1 S (t)

    is the event probability.

    Libor market model p. 63

  • Securities with uncertain horizons

    A cashflow of a security with event risk is given by:j

    (cjP (tj)S (tj) + rjP (tj)

    (Fj (t) F (tj1)

    )),

    where cj are the known cash amounts, and rj are the recovery rates incase an event occurs. The intensity process defining the survivalprobability is modeled outside of LMM. The price of the security iscalculated by taking the appropriate expected value of the cashflowsabove.

    Libor market model p. 64

  • LMM and the SABR model

    The SABR model describes the dynamics of a single forward withstochastic volatility. Its dynamics is given by:

    dL (t) = C (L (t)) dW (t) ,

    d (t) = (t) dZ (t) ,

    whereE [dW (t) dZ (t)] = r dt.

    One usually chooses C (L) to be of the CEV type

    C (L) = L .

    SABR captures volatility smile on vanilla options. In order to model thesmile of more complex instruments, we need to suitably extend LMM.

    Libor market model p. 65

  • LMM and the SABR model

    We consider an extension of the LIBOR market model with stochasticvolatility parameters denoted by 1t , . . . , Nt :

    dLj (t) = j(L (t) , (t) , t)dt+ Cj(Lj (t) , j (t) , t)dW j (t) ,

    dj (t) = j(L (t) , (t) , t)dt+Dj(Lj (t) , j (t) , t)dZj (t) ,

    with

    E[dW i (t) dW j (t)] = ijdt,

    E[dW i (t) dZj (t)] = rijdt,

    E[dZi (t) dZj (t)] = piijdt.

    Choose the drifts so that this dynamics is arbitrage free!

    Libor market model p. 66

  • LMM and the SABR model

    Under the forward measure Qk, the arbitrage free dynamics is:

    dLjt = Cj(Ljt ,

    jt , t)

    j+1ik jiiCi(Lit, it, t)1 + iLit dt+ dW jt , if j < k,dW jt , if j = k,k+1ij

    jiiCi(Lit,

    it, t)

    1 + iLitdt+ dW jt , if j > k ,

    djt = Dj(Ljt ,

    jt , t)

    j+1ik rjiiCi(Lit, it, t)1 + iLit dt+ dZjt , if j < k,dZjt , if j = k,k+1ij

    rjiiCi(Lit,

    it, t)

    1 + iLitdt+ dZjt , if j > k .

    Libor market model p. 67

  • LMM and the SABR model

    Under the spot measure, the dynamics reads:

    dLj (t) = Cj(Lj (t) , j (t) , t

    ) (t)ij

    jiiCi(Li (t) , i (t) , t

    )1 + iLi (t)

    dt

    + Cj(Lj (t) , j (t) , t

    )dWj (t) ,

    dj (t) = Dj(j (t) , t

    ) (t)ij

    jiiCi(Li (t) , i (t) , t

    )1 + iLi (t)

    dt

    +Dj(j (t) , t

    )dZj (t) .

    Note that the model involves a large number of parameters. Judiciouschoices have to be made in order to produce a stable framework.

    Libor market model p. 68

  • Risk management: delta

    This is the most important risk factor (and the easiest to hedge).Choose a hedging portfolio consisting of vanilla instrumentssuch as swaps, Eurodollar futures, forward rate agreements, etc:

    hedge = {B1, . . . , Bn} .

    Let C0 denote the current forward curve (the base scenario).Choose a number of new micro scenarios

    C1, . . . , Cp

    by perturbing a segment of C0. For example, C1 could result fromC0 by shifting the first 3 month segment down by 1 bp.

    Libor market model p. 69

  • Risk management: delta

    The vector of portfolios sensitivities to these scenarios is

    i = (Ci)(C0) , i = 1, . . . , p,

    where by (Ci) we denote the value of the portfolio given theforward curve Ci.The matrix B of sensitivities of the hedging instruments to thesescenarios is

    iBj = Bj (Ci)Bj (C0) .To avoid accidental colinearities between its rows or columns,one should always use more scenario than hedging instruments.

    Libor market model p. 70

  • Risk management: delta

    The vector of hedge ratios is calculated by minimizing

    L () = 12B 2 + 1

    2Q2.

    Here, is an appropriately chosen small smoothness parameter,and Q is the smoothing operator. Explicitly,

    =((B)

    tB + QtQ

    )1(B)

    t.

    One can think of the component j as the sensitivity of the portfolio tothe hedging instrument Bj . This method of calculating portfoliosensitivities is called the ridge regression.

    Libor market model p. 71

  • Risk management: gamma

    The gamma of a portfolio is sometimes calculated as its globalconvexity characteristic. This is a rather crude measure, as portfoliostypically exhibit complex convexity behaviors. A better way is toconstruct the portfolio gamma as the change in its delta underspecified macro scenarios:

    0,1, . . . ,r,

    with 0 base scenario (no change in rates). For example:+50 All rates up 50 basis points.+25 All rates up 25 basis points.25 All rates down 25 basis points.50 All rates down 50 basis points.

    Libor market model p. 72

  • Risk management: gamma

    For each of the macro scenarios, we calculate the deltas

    1, . . . ,r.

    The quantities:

    1 = 1 0,.

    .

    .

    r = r 0,

    are the portfolio gammas under the corresponding scenarios. Forintermediate market moves, the portfolio gamma can be calculated bylinearly interpolating the macro scenarios.

    Libor market model p. 73

  • Risk management: vega

    In order to quantify the vega risk we have to first design appropriatevolatility scenarios.

    Perturb vol inputs. Shift selected swaption and/or capvolatilities.

    Perturb internal parameters. As explained before, LMM buildsits internal volatility surface S. We construct volatility microscenarios by accessing S and shifting selected non-overlappingsegments. Let us call these scenarios

    S0,S1, . . . ,Sq,

    with S0 = S being the base scenario.

    Libor market model p. 74

  • Risk management: vega

    We choose a hedging portfolio hedge which may consist of liquidinstruments such as swaptions, caps and floors, Eurodollaroptions, or other instruments (Bermudan options?).The rest is a verbatim repeat of the delta story. We calculate thesensitivities of the portfolio to the volatility scenarios. Wecalculate the sensitivities of the hedging portfolio to the volatilityscenarios. Finally, we use ridge regression to find the hedgeratios.

    This method of managing the vega risk works remarkably well andallows one, in particular, to separate the exposure to swaptions fromthe exposure to caps / floors.

    Libor market model p. 75

    head {Overview}head {Overview}head {Dynamics of LMM}head {Dynamics of LMM}head {Dynamics of LMM}head {Dynamics of LMM}head {Dynamics of LMM}head {Dynamics of LMM}head {Dynamics of LMM}head {Factor reduction}head {Factor reduction}head {Factor reduction}head {Low noise expansion}head {Low noise expansion}head {Low noise expansion}head {Optimization and Tikhonov regularization}head {Vanilla system}head {Stripping cap volatility}head {Stripping cap volatility}head {Approximate valuation of swaptions}head {Approximate valuation of swaptions}head {Approximate valuation of swaptions}head {Approximate valuation of swaptions}head {Approximate valuation of swaptions}head {Approximate valuation of swaptions}head {Approximate valuation of swaptions}head {Structure of instantaneous correlations}head {Structure of instantaneous correlations}head {Parametrization of instantaneous volatilities}head {Parametrization of instantaneous volatilities}head {Parametrization of instantaneous volatilities}head {Parametrization of instantaneous volatilities}head {Parametrization of instantaneous volatilities}head {Parametrization of instantaneous volatilities}head {Parametrization of instantaneous volatilities}head {Least square optimization}head {Least square optimization}head {Least square optimization}head {Approximation schemes for LMM}head {Approximation schemes for LMM}head {Approximation schemes for LMM}head {Approximation schemes for LMM}head {Approximation schemes for LMM}head {Computing the drift terms}head {Computing the drift terms}head {Computing the drift terms}head {Computing the drift terms}head {Valuation of non-callable securities}head {Valuation of Bermudan options}head {Valuation of Bermudan options}head {Valuation of Bermudan options}head {Valuation of Bermudan options}head {Valuation of Bermudan options}head {Valuation of Bermudan options}head {Valuation of Bermudan options}head {Valuation of Bermudan options}head {Valuation of Bermudan options}head {Valuation of Bermudan options}head {Securities with uncertain horizons}head {Securities with uncertain horizons}head {Securities with uncertain horizons}head {Securities with uncertain horizons}head {Securities with uncertain horizons}head {LMM and the SABR model}head {LMM and the SABR model}head {LMM and the SABR model}head {LMM and the SABR model}head {Risk management: delta}head {Risk management: delta}head {Risk management: delta}head {Risk management: gamma}head {Risk management: gamma}head {Risk management: vega}head {Risk management: vega}