A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics *
Paul Malliavin 10 rue Saint Louis en l'Isle, 75004 Paris, France, e-mail:sli@ccr.jussieu.fr Maria Elvira Mancino Universit di Firenze, Italy, e-mail: mariaelvira.mancino@dmd.unifi.it Maria Cristina Recchioni Universit Politecnica delle Marche, Italy e-mail:m.c.recchioni@univpm.it

1. Abstract 2. Introduction 3. An econometrically implementable method for the computation of the iterated brackets 4. Fourier estimation methodology 5. Analysis of interest rate data 6. Fortran and Matlab codes

1. Abstract
We show that the geometry of the Heath-Jarrow-Morton interest rates market dynamics can be non-parametrically calibrated by the observation of a single trajectory of the market evolution. Then the hypoellipticity of the infinitesimal generator can be exactly measured. On a data set of actual interest rates we show the prevalence of the hypoelliptic effect together with a sharp change of regime. Volatilities are computed by applying the Fourier cross-volatility estimation methodology. In this website we show some numerical results concerning the paper [21] and we provide the FORTRAN and MATLAB software codes.

2. Introduction
An empirically relevant and specific feature of the term structure of interest rates dynamics lies, from one side, in the high dimensionality of the state space, that is the bond values for a continuum of maturities, and, from the other side, in the low dimensionality of the variance, which is found empirically to be influenced by no more than four independent factors (e.g. see [10],[6]). These stylized facts lead to the formulation of the Heath-Jarrow-Morton framework (HJM hereafter) [17], in which the yield curve is driven by a low dimensional Wiener process. From the point of view of the "regularity" of the model, these two properties imply that the ellipticity assumption, which is assumed for stock market models and pricing-hedging of contingent claims in this market, is ruled out when we deal with the interest rate market models. Therefore, we are lead to consider non-elliptic models, and the most regular models still available are the hypoelliptic ones. We consider a HJM model of the forward rates driven by a n-dimensional Brownian motion and we develop a methodology for exploring the geometric properties of the forward rates evolution's space, along the lines suggested in [22]. We apply the Fourier cross-volatilities estimator introduced in [19] to obtain a non parametric estimation of the forward rate volatility structure. Secondly, in order to estimate the iterated Lie brackets, we prove a mathematical result which reduces the computation of the iterated Lie brackets of the driving vector fields by means of iterated volatilities.

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (1 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

We illustrate the capability of our methodology to numerically achieve the geometric properties of the interest rates evolution space. The numerical results confirm the hypoellipticity of the interest rate market. In fact we find that the Lie brackets of the driving vector fields are very far from lying in the subspace generated by the first three eigenvectors. We compute the Lie brackets on an actual time series of the yield curve of Euro rates and we compute the distance of Lie brackets from the time varying subspace spanned by the three main eigenvectors. Then we compute the iterated second-order Lie brackets and their distances from the subspace generated by the first three eigenvectors and the first-order brackets. Our results show that the distance of iterated Lie brackets from this subspace is large, indicating hypoellipticity. The tools devised in our approach allow estimation of simple geometric features of the model. As an illustrating example, we compute the movements of the first eigenvectors, quantified by the angle between the same eigenvector at different instants of time. We show that, in the three-years post-EMU period considered, the second and the third eigenvectors have a very much fluctuating behavior; on the contrary the first eigenvector has in the over all a remarkable stability; this direction has therefore a meaning from an economic point of view. We decipher a remarkable change of this direction during around three months corresponding to the third trimester of the year 2001. A special economic event happened during this semester (11th September 2001). We emphasize that our methodology could be implemented in real time, therefore tracking in real time the direction of the first eigenvector could furnish to the trader an indicator on the market stability.

2. An econometrically implementable method for the computation of the iterated brackets

We consider a continuous trading economy with a trading interval [0,], for a fixed > 0. Suppose that a continuum of default free discount bonds are traded with different maturities, one for each trading date T [0,]. The instantaneous forward rate at time t for maturity date T > t, f(t,T), satisfies the following equation
d

f(t,T)-f(0,T) (s,T)ds = 0 +

j=1

(s,T)dW (s) j j 0

(2)

where (t,T) and j(t,T) are stochastic processes adapted to the filtration \cal Ft, which possibly includes all available information at time t generated by the term structure's evolution. We assume that the market is driven by a finite number of Brownian motions Wj (j = 1,, d). In the parametrization used by [7], let x = T-t and rt(x) = f(t,t+x) (3)

that is x represents the time to maturity in contrast with T which represents the maturity time and rt(x) is the instantaneous forward rate curve. The interest rate curve takes its value in the infinite dimensional space C of continuous functions on [0,[. Nevertheless a remarkable experimental fact is that the rank n of its volatility matrix is very low n 4 (see also [6]). This result suggests that elliptic models are ruled out and hypoelliptic models are the most regular models still available (see also [22] on this point). Now consider the family of SDE, depending upon the parameter x, describing the HJM model for the forward rate curve. As shown in [13], the no-arbitrage restriction forces the drift coefficient of the forward curve evolution to be just a function of the volatility structure, for the risk neutral probability measure. Therefore the forward rate curve evolution has the following expression in the Musiela parametrization:
x drt(x) j(t,x) j(t,y) rt(x)+ 0 dy = x j=1 d

j(t,x) dWj(t). dt + j=1

(4)

In the literature there is a vast possible choice for the functional forms of the term's structure volatility. For our estimation we will restrict our attention to the Markovian models, which means that we suppose the volatility function is of the form (t,x) = ((rt))(x), (5)

where () is a smooth function. This class of volatility structure is quite general and it is considered also in [14]. Therefore we suppose the following evolution for the instantaneous forward rate:

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (2 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

x drt(x) (Aj(rt))(x) (Aj(rt))(y) rt(x)+ 0 dy = x j=1

(Aj(rt))(x) dWj(t) dt + j=1

(6)

where Aj are "driving vector fields" defined on C, the infinite dimensional space of continuous functions on [0,[. An appropriate notion of smoothness of vector fields is a necessary hypothesis in order to prove existence and uniqueness of solutions for (6). In fact the operator r [(rt)/ ( x)] is an unbounded operator, therefore Banach space type differential calculus cannot be used, and a theory of differential calculus in Frchet spaces is needed. We refer to [11],[12] for a precise treatment of this topic. We do not address this problem because our interest will be to work on a reasonable finite dimensional approximation, that is we will work on RM, where M is a fixed (large) number of maturities, but having an nfinite dimensional" point of view in mind. The main result of this section is the following: The brackets of the driving vector fields of the HJM diffusion can be numerically computed from a single time series of market data, under the only assumption on the model of the differentiability of the infinitesimal generator's coefficients. In order to prove this result we reduce the computation of the iterated Lie brackets of the driving vector fields by means of iterated volatilities. Fix M different maturities (M can be extremely large and in particular it should be d << M), then define the time dependent MM covariance matrix (t) : = [drt() drt()]

(7)

where , vary between all maturities, denotes the It contraction divided by dt and the instantaneous forward rate satisfies (6). Using the Fourier methodology for the estimation of (cross-)volatility proposed in [19] we compute non parametrically the historical variancecovariance matrix for the interest rate market evolution. We get a time dependent MM matrix (t). Denote by k,t its eigenvalues and [()\tilde]k,t the corresponding normalized eigenfunctions. We fix m < M the number of eigenvalues not very close to zero. Denote by k,t = {k,t}[()\tilde]k,t the canonical eigenvectors, we have that the econometric reconstruction of (t) is:
m

(t)

k=1

k,tk,t .

(8)

The vector-valued function t k,t is the econometric reconstruction of the vector-valued function t Ak(rt) under the assumption that the following model holds true:
d

drt(x) =

j=1

(Aj(rt))(x)dWj(t) +(A0(rt))(x)dt.

(9)

Theorem 1 (computation of first order brackets) Let us consider the econometric reconstruction of the function k,t: t Ak(rt) obtained in (8). For any { 1, , M} it holds : 1 l,t
M =1

[Al,Ak] (rt) =

l,t

(d

k,td rt

1 k,t

M =1

)-

k,t

(d

l,t

d rt ).

(10)

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (3 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

Further results can be found in [25].

4. Fourier estimation methodology

We briefly recall the methodology which has been proposed in [19] and developed in [2], [3], [20]- [22]. Assume that the market data processes p (t) = (p1(t), , pn(t)) are Brownian semi-martingales satisfying the following It SDE
d

dpj(t) =

i=1

ji(t) dWi + bj(t) dt

j = 1,, n

(11)

where W = (W1, , Wd) are independent Brownian motions, and ** and b* are adapted random processes. We recall now the Fourier method for computing multivariate volatilities. From the representation (11) we define the volatility matrix, which in our hypothesis depends upon time:
d

j,k(t) =

i=1

ji(t)ik(t).

(12)

The Fourier method reconstructs *,*(t) on a fixed time window (which we can reduce always to [0,2] by change of origin and rescaling) using the Fourier coefficients of dp*(t). First we compute the Fourier coefficients of dpj for j = 1, , n defined by cos(kt)dpj(t), bk(dpj) ak(dpj) 1 ]0,2[ = = Then we consider the Fourier coefficients of the cross-volatilities a0(i,j) = i,j(t) dt, ak(i,j) ]0,2[ = 2 1 1 cos(kt) i,j(t) dt, ]0,2[ 1 sin(kt)dpj(t). ]0,2[

(13)

(14)

bk(i,j) 1 sin(kt)i,j(t) dt. ]0,2[ =

(15)

In [19] it is proved the following result relating the Fourier coefficients of the variation of the process p* to the Fourier coefficients of the volatility matrix. Theorem 2 Fix an integer n0 > 0, the Fourier coefficients of the functions i,j(t), which are the entries of the volatility matrix, are given by the following limits in probability lim
N

a0(i,j) =

N N+1-n0 s = n 2 0

(as(dpi)as(dpj)+bs(dpi)bs(dpj))

(16)

ak(i,j) =

lim

N N+1-n0 s = n 0

(as(dpi)as+k(dpj)+as(dpj)as+k(dpi))

(17)

bk(i,j) =

lim

N N+1-n0 s = n 0

(as(dpi)bs+k(dpj)+as(dpj)bs+k(dpi)).

(18)

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (4 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

Finally using the Fourier-Fjer inversion formula it is possible to reconstruct i,j(t) from its Fourier coefficients: i,j(t) = where for any t (0,2)
N

lim (i,j)N(t)

(19)

(i,j)N(t): =

k=0

k (1N

)(ak(i,j) cos(kt)+ bk(i,j)sin(kt)).

(20)

We stress the point that the computation involved in formula (10) can be realized exploiting the fact that the Fourier methodology allows us to reconstruct cross volatilities as a function of time (20), therefore it makes possible to iterate the procedure of computing the cross-volatilities.

5. Analysis of interest rate data

We implement the described methodology on a data set of actual interest rates. To this purpose, we use a time series of Euro swap rates and Euribor rates, ranging from January 1999 (start of EMU) to December 2001, for a total of 777 days. We have Euribor rates at 3, 6 and 9 months, and swap rates yearly from 1 to 10 years. The swap rates are bootstrapped to get the yield curve, which we define as: log P(t, T) T-t where P(t,T) is the price, observed at t, of one Euro payed at T. The numerical difference between y(t,T) and f(t,T) is negligible. In our data the time to maturity x = T-t is fixed and we observe 13 maturities ranging from three months to 10 years. We start by computing the 1313 variance-covariance matrix (t) as a function of the time. We approximate the volatility matrix (t) R1313, t [0,777/259], with the matrix N(t) obtained using the procedure described in Theorem 2, formula (19), (20). To this purpose we adopt the Fourier methodology using N1 coefficients for price and N coefficients for the volatility. Note that we have H = 777 observations and the time unit is one year in the computations. The data are uniformly distributed in the time interval [0,777/259] with step-size dt = 1/259. Moreover we choose N1 = H/2 element to reconstruct the Fourier coefficients of the process and N = H/4 Fourier coefficients of the volatility matrix and we smooth the Fjer kernel in (20) replacing (1-k/N) with sin2( k)/(k)2, k = 1,2,N and = 2/259 and we consider a loss of resolution equal to = 2 corresponding to the operation of taking the volatility. Hence the prices are given with a precision 1/N1, the covariances with a precision / N1, the first order brackets with a precision 2/N1 and the second order brackets with a precision 3/N1. Let us denote by trace(N(t)) the trace of the matrix N(t), by i(t) R, i = 1,2,,13, 1(t) 2(t) 13(t) the eigenvalues of the matrix N(t) at time t and by Ai* R13, i = 1,2, ,13 the corresponding eigenvectors with Ai*(t)TAj*(t) = 0, i j and ||A*i(t)|| = 1, i,j = 1,2,,13, where the superscript T denotes the transpose operator. We define the eigenvectors Ai(t) = {i(t)}A*i(t), i = 1,2,, 13 that are orthogonal vectors with ||Ai(t)|| = {i(t)}, i = 1,2,,13.

y(t;T) = -

(21)

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (5 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

Figure 1: Eigenvalues versus time: different scales As shown in Figures 1, 2, 3, the first three eigenvalues describe almost completely the spectrum of the volatility matrix. In particular Figure 1 shows the plot of the first six eigenvalues as a function of the time variable t (the unit time of the figures is the day).

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (6 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

Figure 2: Percentage of variance explained by the first three eigenvectors

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (7 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

Figure 3: Ratios ri(t), i = 1 (dotted line), i = 2 (dashed-dot line), i = 3 (solid line) As previously mentioned, Figure 1 shows that only the first three eigenvalues are significant. It is a well known fact that less than four factors are enough to account for the volatility of the whole yield curve. This is shown in Figure 2, where we plot the time-varying percentage of the variance explained by the fist three factors, that is we show the ratios 1(t)/trace(N(t)), (1(t)+2(t))/trace(N(t)), (1(t)+2(t)+3(t))/trace(N(t)). Moreover Figure 3 shows the ratios ri(t) = i(t)/(1(t)+2(t)+3(t)), i = 1,2,3. Now we consider the econometric reconstruction of Aj(t), j = 1,2,3 (see formulae (19), (20)). Let ti = 777 i/100, i = 1,2,,100, Figure 4 shows the cosine between Aj(ti) and Aj(ti+1), j = 1,2,3, i = 1,2,,100.

Figure 4: Variation of the first three eigenvectors Moreover we observe the surfaces generated by the cosines between the eigenvectors A1(t), A1(t ). Let us denote by c1(t,t ) the cosines c1(t,t ) = A1 (t)TA1(t )/(||A1(t)|| ||A1(t )||). Figure 5 shows the surface c1(t,t ) when t = t10*i, i = 1,2,,10 t = t10*k, k = 1,2,,10. Due to the extreme rapid variation of the angle of the two last vectors A2, A3 it is not relevant to draw the analog of the plot 5. t = t10*i, i = 1,2,,10 t = t10*k, k = 1,2, ,10. Figure 5 shows a remarkable change of the direction of the first eigenvector during around three months corresponding to the the end of the third trimester and the beginning of the last trimester of the year 2001. We conjecture that a special economic event happened during this trimester probably the crisis of the September 11th. We emphasize that our methodology could be implemented in real time: therefore tracking in real time the direction of the first eigenvector could furnish to the trader an indication on the market stability. Now we consider the first order brackets: B1,2(t) = [A1(t),A2(t)], (22)

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (8 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

B1,3(t) = [A1(t),A3(t)],

(23)

B2,3(t) = [A2(t),A3(t)].

(24)

Figure 5: c1(t,t ) surface, t [0,777], t [0,777] and a contour plot We can compute these brackets (see Corollary 2.4) thank to the Fourier method proposed since, roughly speaking, they are the covariance between the eigenvectors and the yield curve. We focus on the time varying subspace G1,2,3(t) spanned by the first three eigenvectors A1(t), A2(t), A3(t). We compute the projection of the vectors Bi,j(t), i = 1,2, j = i+1,,3 on the subspace G1,2,3(t): Aj(t)TB1,2 (t) j(t) Aj(t), (25)

v1,2(t) =

j=1

v1,3(t) =

Aj(t)TB1,3 (t) j(t) Aj(t), (26)

j=1

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (9 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

v2,3(t) =

Aj(t)TB2,3 (t) j(t) Aj(t), (27)

j=1

Figure 6: Lengths of the brackets Bi,j(t), i = 1,2, j = i+1,3 versus time and then we compute the cosine between each bracket and its projection on the subspace G1,2,3 as function of time and we denote this cosine by ci,
j(t)

= Bi,j(t)Tv1,2(t)/(||Bi,j(t)||||vi,j(t)||), i = 1,2, j = i+1,3. Figure 6 shows the lengths of the first order brackets and Figure 7 shows the cosines ci,j(t),

i = 1,2, j = i+1,3 as functions of t.

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (10 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

Figure 7: Cosines between the brackets Bi,j(t), i = 1,2, j = i+1,3 and their projections on G1,2,3(t) versus time

6. FORTRAN and MATLAB codes

The archive Fortran_code.zip contains the Fortran code (main.f), the file containing the interest rates data (interest_data.txt) and the output files generated by running the code. The code computes the eigenvalues, the eigenvectors of the volatility matrix and the first order brackets for several time values. The Fortran code requires the use of the IMSL routine for computing the eigenvalues and the eigenvectors and it provides several output files (see readme_fortran.txt) The archive Matlab_code.zip contains the Matlab code (main.m), the file containing the interest rates data (interest_data.txt) and a readme file. The Matlab code computes the volatility matrix for several time values and it plots the first six eigenvalues versus time.

References
[1] Airault, H. and Malliavin, P. (2004) Backward regularity for some infinite dimensional hypoelliptic semi-group. Stochastic analysis and related topics in Kyoto. In honour of K. It. Lectures given at the conference in Kyoto, Japan (2002). Math. Soc. of Japan. Advanced Studies in Pure Math. 41: 1-11. [2] Barucci, E. and Ren, R. (2001) On measuring volatility of diffusion processes with high frequency data. Economics Letters, 74: 371-378. [3] Barucci, E. and Ren, R. (2002) On measuring volatility and the GARCH forecasting performance. Journal of International Financial Markets, Institutions and Money,12: 183-200. [4]

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (11 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

Baudoin, F. and Teichmann, J. (2005) Hypoellipticity in infinite dimensions and an application in interest rate theory. Ann. Appl. Prob., 15. [5] Bjrk, T. and Svensson, L. (2001) On the existence of finite dimensional realizations for nonlinear forward rate models. Mathematical Finance,11: 205-243. [6] Bouchaud, J.P., Cont, R., El-Karoui, N., Potters, M. and Sagna, N. (1999). Phenomenology of the interest rate curve. Appl. Math. Finance, 6, no.3, 209-232. [7] Brace, A. and Musiela, M. (1994). A Multi-factor Gauss Markov implementation of Heath, Jarrow and Morton. Math. Finance, 4: 259283. [8] Carmona, R. and Tehranchi, M. (2004) A characterization of hedging portfolios for interest rate contingent claims. Ann. Appl. Probab., 14, no.3: 1267-1294. [9] Cont, R. (2005). Modeling term structure dynamics: an infinite dimensional approach. Int. J. Theor. Appl. Finance, 8. [10] Duffie, D. and Kan, R. (1995) Multifactor models of the term structure. in Howison, Kelly & Wilmott (Eds.) Math. Models in Finance, London: Chapman & Hall. [11] Filipovic, D. and Teichmann, J. (2003a) Existence of invariant manifolds for stochastic equations in infinite dimension. J. Funct. Anal., 2: 398-432. [12] Filipovic, D. and Teichmann, J. (2003b) Regularity of finite dimensional realizations for evolution equations. J. Funct. Anal., 2: 433-446. [13] Heath, D., Jarrow, M. and Morton A. (1992) Bond pricing and the term structure of interest rates: a new methodology for contingent claims valuation. Econometrica, 60, no.1: 77-105. [14] Jeffrey, A., Kristensen, D., Linton, O., Nguyen, T. and Phillips, P.C.B. (2004) Nonparametric Estimation of a Multifactor Heath-JarrowMorton Model: an Integrated Approach. Journal of Financial Econometrics, 2: 251-289. [15] Malliavin, P. (1978a) Stochastic Calculus of Variations and Hypoelliptic operators. Proc. Internat. Symposium on Stochastic Differential Equations, Res. Inst. Math. Sci., Kyoto Univ., Kyoto, 1976, Wiley, New York, 1978, 195-263. [16] Malliavin, P. (1978b) Ck-hypoellipticity with degenaracy. II, Stochastic analysis. Proc. Internat. Conf., Northwestern Univ., Evanston, Ill., 1978 , Academic Press, New York, 1978, 327-340. [17] Malliavin, P. (1997) Stochastic analysis. A series of comprehensive studies in mathematics, vol.313. Springer-Verlag, 1997. [18] Malliavin, P. (2006) It atlas, its application to mathematical finance and to exponentiation of infinite dimensional Lie algebras. Proceedings of The Abel Symposium 2006 - Stochastic Analysis and Applications - A Symp. in Honour of Kiyoshi It. Springer-Verlag. [19] Malliavin, P. and Mancino, M.E. (2002). Fourier series method for measurement of multivariate volatilities. Finance and Stochastics, 4: 49-61. [20] Malliavin, P. and Mancino, M.E. (2005). A Fourier transform method for nonparametric estimation of volatility. Preprint Dept. Math. for
file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (12 di 13)09/05/2007 12.26.55

A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics

Decisions, University of Firenze. [21] Malliavin, P., Mancino, M.E. and Recchioni, M.C. (2007). A non parametric calibration of the HJM geometry: an application of It calculus to financial statistics. Japanese Journal of Mathematics, 1: 55-77. [22] Malliavin, P. and Thalmaier, A. (2005). Stochastic Calculus of Variations in Mathematical Finance, Springer Finance, 2005. [23] Musiela, M. (1993). Stochastic PDEs and term structure models. Journes Internationales de Finance, IGR-AFFI, La Baule, 1993.

file:///C|/Documents and Settings/ik5mic/Desktop/Sito_Mav/Sito_Mav/home.html (13 di 13)09/05/2007 12.26.55