You are on page 1of 12

ON AFFINE INTEREST RATE MODELS

PAUL LESCOT Abstract. Bernstein processes are Brownian diusions that appear in Euclidean Quantum Mechanics. The consideration of the symmetries of the associated Hamilton-Jacobi-Bellman equation allows one to obtain various relations between stochastic processes (Lescot-Zambrini, Progress in Probability, vols 58 and 59). More recently it has appeared that each onefactor ane interest rate model (in the sense of Leblanc-Scaillet) could be described using such a Bernstein process. MSC 91G30, 60H10 Keywords : interest rate models, isovectors, Bessel processes.

hal-00431615, version 4 - 26 Oct 2011

1. Introduction The relationship between Bernstein processes and Mathematical Finance was rst glimpsed in 5 of [9], where the AliliPatie family of transformations S , (see [1],2, for the particular case = 1, and [9], pp.6061, for the general case) was reinterpreted in the context of isovectors for the (backward) heat equation. Furthermore, Paties composition formula ([9], p.60) for these transformations was derived from the commutation relations of the canonical basis of the aforementioned isovector algebra. As the motivation for Paties work had been to compute certain option prices in the framework of an ane interest rate model (see [9], pp. 101104), it was natural to look for a direct parametrization of such a model by a Bernstein process. It turns out that, for any onefactor interest rate model in the sense of Leblanc and Scaillet ([4],p. 351), the associated square root process coincides, up to some (possibly innite) random time, with a Bernstein process (3, Theorem 3.4). The 1 potential 2 (that had been considered in [8], p. 220, on physical grounds) appears q here naturally (Theorem 3.4) : this is one more example of the deep links between Euclidean Quantum Mechanics and Mathematical Finance. 2. An isovector calculation We shall work within the context of [8], 3. We consider the HamiltonJacobi Bellman equation (HJ B V ) with potential C + Dq 2 , q2 by . where we have replaced (as explained in [5]) V (t, q ) =
Date : November 17th, 2010.
1

1 S 2 2 2 S S ( + ) +V =0 t 2 q 2 2 q

PAUL LESCOT

In order to determine the Lie algebra HV of pure isovectors for (HJ B V ), i.e. the algebra of [8], we have to solve the auxiliary equation (3.29) from [8],p. 215, i.e. :
.. . 2C C 2 .. 1 . C 1 ... . TN q 2 lq + + TN q ( 3 +2Dq )+ l( 3 +2Dq )+ TN ( 2 + Dq 2 ) TN = 0 , 4 2 q q q 4

that is :
.. . 2 ... 2Cl 1 ... . q 2 (2DTN TN ) + q ( l + 2Dl) + TN 3 = 0 . 4 4 q

hal-00431615, version 4 - 26 Oct 2011

Two dierent cases now appear : 1) C = 0 Then one must have l = 0, in which case the second condition holds automatically, and the system reduces itself to : 2 .. . = TN 4 ... . T N = 8D TN 1)a) D > 0 Setting = 8D, we nd TN = C1 et + C2 et , whence TN = and = therefore : 2 . TN + C4 4 C1 t C2 t e e + C3
.

As TN , l and depend only upon t, the system is equivalent to : 2Cl = 0 .. l = 2Dl 2 .. . = TN 4 ... . TN = 8D TN .

where (Cj )1j 4

2 C1 t 2 C2 t e + e + C4 4 4 denote arbitrary (real) constants. In particular = dim(HV ) = 4 .

1)b)D = 0 ... Then from TN = 0 follows TN = C1 t2 + C2 t + C3

ON AFFINE INTEREST RATE MODELS

for constants C1 , C2 , C3 . Then = and


.

2 C1 2 .. TN = 4 2

2 C1 t + C4 . 2 Therefore, here too, dim(HV ) = 4; furthermore, we get an explicit expression for the isovectors : N t = TN = C1 t2 + C2 t + C3 , = Nq 1 . q TN + l 2 1 q (2C1 t + C2 ) = 2 C2 q , = C1 tq + 2 =

hal-00431615, version 4 - 26 Oct 2011

and N S = . 1 .. = q 2 TN q l + 4 2 1 = q 2 .2C1 + C1 t + C4 4 2 C1 2 2 ( t q ) + C4 . = 2 A canonical basis for HV is thus given by (Mi )1i4 , where Mi is characterized by Cj = ij (Kroneckers symbol). Using the notation of [7], it appears that M1 = M2 = 1 N6 , 2 1 N4 , 2

M 3 = N1 , and 1 N3 , 2 therefore HV is generated by N1 , N3 , N4 and N6 . We thereby recover the result of [8], p. 220, modulo the correction of a misprint. This list ties in nicely with the symmetry properties of certain diusions related to Bessel processes (see [6] for a detailed explanation). 1)c)D < 0 Setting now = 8D, we nd M4 = TN = C1 cos (t) + C2 sin (t) , whence
.

PAUL LESCOT

TN = and, as above :

C2 C1 sin (t) cos (t) + C3 , 2 . TN + C4 , 4

= therefore : where (Cj )1j 4 =

2 C2 2 C1 cos (t) + sin (t) + C4 4 4 denote arbitrary (real) constants. In particular , dim(HV ) = 4 .

2)C = 0 Then the system becomes

..

l = 2Dl

hal-00431615, version 4 - 26 Oct 2011

The equation for l on the one hand, and the system for (, TN ) on the other hand, are independent, and, as above, the rst one has a twodimensional space of solutions and the second one a fourdimensional space of solutions, i.e. dim(HV ) = 6 .

2 .. . = TN 4 ... . T N = 8D TN .

Whence

Theorem 2.1. The isovector algebra HV associated with V has dimension 6 if and only if C = 0 ; in the opposite case, it has dimension 4.

ON AFFINE INTEREST RATE MODELS

3. Parametrization of a onefactor affine model As general references we shall use, concerning Bernstein processes, our recent survey ([5]), and, concerning ane models, H enons PhD thesis([3]) as well as Leblanc and Scaillets seminal paper ([4]). An onefactor ane interest rate model is characterized by the instantaneous rate r(t), satisfying the following stochastic dierential equation : (3.1) dr(t) = under the riskneutral probability Q ( = 0 corresponds to the socalled Vasicek model, and = 0 corresponds to the CoxIngersollRoss model ; cf. [4]). Assuming > 0, let us set =def + , 4 =def , 0. and let us also assume that The following two quantities will play an important role : C := = and D := Let us set Xt = r(t) + . Proposition 3.1. Let r0 R ; then the stochastic dierential equation (3.2) dr(t) = has a unique strong solution such that r(0) = r0 . Furthermore, in case that one has r(t) + 0 for all t 0 ; in particular, r(t) satises (3.1). Proof. Let us set Xt = r(t) + ; it is easy to see that, in terms of Xt , equation 3.2 becomes: dXt = = = dr(t) ( Xt )dt) Xt )dt . |Xt |dw(t) + ( |Xt |dw(t) + ( r0 + 0 , |r(t) + | dw(t) + ( r(t)) dt 2 . 8 2 3 ( )( ) 8 4 4 4 ( 1)( 3) 128 r(t) + dw(t) + ( r(t)) dt

hal-00431615, version 4 - 26 Oct 2011

, and We are therefore in the situation of (1), p.313, in [2], with c = , a = b = ; the result follows. In case = 0, one may also refer to [3],p.55, Proposition 12.1, with = , = . and a =

PAUL LESCOT

We shall henceforth assume all of the hypotheses of Proposition 3.1 to be satised. Corollary 3.2. One has Xt = et Y ( (e4 1) ) for = 0 , and 2 Xt = Y ( 4 t ) for = 0 where Y is a BESQ (squared Bessel process with parameter ) having initial value Y0 = r0 + . Proof. In case = 0, one applies the result of [3], p. 314. For = 0, let Zt := it appears that whence Zt is a BESQ process. As dZt = 2 |Zt |dw(t) + dt , Xt := 4 Xt ; 2
2 t

hal-00431615, version 4 - 26 Oct 2011

2 Zt , 4 the scaling property of Bessel processes yields the result. Theorem 3.3. If 2 one has, almost surely : t > 0 Xt > 0 ; on the other hand, if < 2, almost surely there is a t > 0 such that Xt = 0. Proof. We apply Corollary 1, p. 317, from [4](12.2) yielding that (3.3) t > 0 Xt > 0 . One may also use [3], p.56, from which follows that (3.3) is equivalent to 2a 1; 2 but, according to the above identications, 2 2a = 2 = . = 2 2 2 Our main result is the following : Theorem 3.4. Let us dene the process z (t) = and the stopping time as seen in Theorem 3.3, T = + a.s. for 2, and T < + a.s. for < 2. Then there exists a Bernstein process y (t) for = 2 T = inf {t > 0|Xt = 0}; Xt

ON AFFINE INTEREST RATE MODELS

and the potential V (t, q ) = such that t [0, T [ z (t) = y (t) . In particular, for 2, z itself is a Bernstein process. Proof. One has (cf. Proposition 3.1 and its proof) dXt = Xt )dt Xt dw(t) + ( z (t)2 )dt . z (t)dw(t) + ( C + Dq 2 . q2

= Taking now f (x) = x, we have

hal-00431615, version 4 - 26 Oct 2011

1 1 3 x > 0 f (x) = and f (x) = x 2 , 2 x 4 1 3 therefore, for all t ]0, T [, f (Xt ) = and f (Xt ) = 1 . The application 4 z (t) 2z (t) of It os formula now gives :

dz (t) = = = =

d(f (Xt ))
1 f (Xt )dXt + f (Xt )(dXt )2 2 1 z (t)2 )dt) 1 z (t)3 2 z (t)2 dt (z (t)dw(t) + ( 2z (t) 8 1 4z (t)2 2 )dt . dw(t) + (4 2 8z (t)

Let us now dene by (t, q ) := = 1 q 2 2 t 2 q 2 e 2 t q 1 2 q 2 . e 4

It is easy to check that solves the equation 2 for V = in other words, 2 ln( ) 1 t q 2 2 + ln(q )) = 2 ( 4 2 2 t q 2 1 = + 2 ( ) ln(q ) 16 4 8 satises (HJ B V ). Furthermore we have : S := 4 2 +V = t 2 q 2 C + Dq 2 ; q2

PAUL LESCOT

:= = = =

q 2 S q q 2 ( 1) + 2 8q 1 2 ( 2 4q 2 ) 8q

whence

1 2 4z (t)2 ) (4 8z (t) and z satises the stochastic dierential equation associated with : (t, z (t)) = B (t, z (t))dt t ]0, T [ dz (t) = dw(t) + B (as in [5], 1, equation (B )) ; the result follows.

hal-00431615, version 4 - 26 Oct 2011

Proposition 3.5. The isovector algebra HV associated with V has dimension 6 if { , 3 }, i.e. {1, 3} ; in the opposite case, it has dimension 4. and only if 4 4 Proof. It is enough to apply Theorem 2.1, observing that the condition C = 0 is { , 3 }. equivalent to 4 4 In the context of H enons already mentioned PhD thesis ([3],p.55) we have = a and the condition C = 0 is = a, = , = 2 et = 0, whence equivalent to 2 3 2 }. a { , 4 4 Let us analyze more closely the situation in which C = 0 ; the general case will be commented upon in [6]. = , i.e. = 1 . 1) 4 Then y (t) is a solution of dw(t) y (t)dt , 2 2 i.e. y (t) is an OrnsteinUhlenbeck process (it was already known that the Ornstein Uhlenbeck process was a Bernstein process for a quadratic potential). Therefore z (t) co ncides, on the random interval [0, T [, with an OrnsteinUhlenbeck process. Here t q 2 2 . (t, q ) = e 4 dy (t) = From y (t) = e 2 (y0 +
t t

e
0 2

s 2

dw(s))

= e 2 (z0 + w (

(et 1) )) 4

ON AFFINE INTEREST RATE MODELS

(w denoting another Brownian motion), it appears that y (t) follows a normal law 2 t et ) with mean e 2 z0 and variance (1 . The density t (q ) of y (t) is therefore 4 given by : t 2(q e 2 z0 )2 2 exp ( t (q ) = ). 2 (1 et ) 2 (1 et ) Whence t > 0 (t, q ) = = t (q ) (t, q ) 1 sinh ( t ) 2 q 2 q 2 et + 4qz0 e 2 (1 et ) e
(
t 2

2 t 2z0 e

hal-00431615, version 4 - 26 Oct 2011

and one may check that, as was to be expected, satises the following equation (V ) (C2 in [5]) : (3.4) 2 4 2 + V . = t 2 q 2

= 3 , i.e. = 3. 2) 4 In that case, according to Theorem 3.2, T = + whence y = z . Furthermore 32 t ( q2 ) . (t, q ) = qe 2 4 t 1 ; s(t) = e 2 z (t)

Let us dene

then an easy computation, using It os formula in the same way as above, shows that t ds(t) = e 2 s(t)2 dw(t) ; 2 in particular, s(t) is a martingale. Referring once more to Proposition 3.1 and its proof, we see that dXt = Xt dw(t) + ( 32 Xt )dt . 4

Let us now assume X0 = 0 and = 0 ; then, according to Corollary 3.2, Xt = et Y ( 2 (et 1) ) 4

where Y is a BESQ3 (squared Bessel process with parameter 3) such that Y (0) = 0. But, for each xed t > 0, Yt has the same law as tY1 , and Y1 = ||B1 ||2 is the square of the norm of a 3dimensional Brownian motion ; the law of Y1 is therefore
u 1 e 2 u1u0 du . 2

10

PAUL LESCOT

Therefore the density t (q ) of the law of z (t) is given by : 16 1 t (q ) = 2 3 (1 et ) 2


3 2 3 2

2q 2 2 2 (1 et ) 3 q e

and

16 t (q ) = (t, q ) 3 2 Here, too, one may check directly that satises equation (3.4) above. t > 0 (t, q ) =

q 2 3t 2 3 4 tanh( t 2 ) . (1 et ) 2 qe

hal-00431615, version 4 - 26 Oct 2011

ON AFFINE INTEREST RATE MODELS

11

4. Acknowledgements I am grateful to the colleagues who invited me to present preliminary versions of this work and thereby provided me with a muchneeded moral support : Professor Barbara R udiger (Koblenz, July 2007), Professor Pierre Patie (Bern, January 2009) and Professors Paul Bourgade and Ali S uleiman Ust unel (Institut Henri Poincar e, February 2009). Comments by Professor Pierre Patie led to many improvements in the formulations. I am also indebted to Mohamad Houda for a careful reading of previous versions of the paper.

hal-00431615, version 4 - 26 Oct 2011

12

PAUL LESCOT

References
[1] L. Alili and P. Patie, Sur les premiers instants de croisement du mouvement brownien et dune famille de courbes continues, C. R. Acad. Sci. Paris Sr. I Math., 340(3), pp. 225-228, 2005 [2] A. G oingJaeschke and M. Yor, A survey and some generalizations of Bessel processes, Bernoulli, 9(2), 2003, 313349 [3] S.H enon, Un mod` ele de taux avec volatilit e stochastique, PhD thesis, 2005 [4] B. Leblanc and O. Scaillet, Path dependent options on yields in the ane term structure model, Finance and Stochastics, 2(4), 1998, 349367 [5] P. Lescot, Bernstein Processes, Euclidean Quantum Mechanics and Interest Rate Models, to appear in the Proceedings of the 1st workshop Statistical Physics and Mathematics for Complex Systems (23-24 october 2008, ISMANS, Le Mans, France), 2010 ; available at http://www.univ-rouen.fr/LMRS/Persopage/Lescot/procdef.pdf. [6] P. Lescot and P. Patie, (in preparation) [7] P.Lescot and J.-C. Zambrini, Isovectors for the HamiltonJacobiBellman Equation, Formal Stochastic Dierentials and First Integrals in Euclidean Quantum Mechanics, Proceedings of the Ascona conference (2002), 187202. Birkha user (Progress in Probability, vol 58), 2004. [8] P.Lescot and J.-C. Zambrini, Probabilistic deformation of contact geometry, diusion processes and their quadratures, Seminar on Stochastic Analysis, Random Fields and applications V, 203-226. Birkha user(Progress in Probability, vol. 59), 2008. [9] P.Patie, On some rst passage times problems motivated by nancial applications, PhD Thesis, ETHZ, 2004. Laboratoire de Math ematiques Rapha el Salem, UMR 6085 CNRS, Universit e de Rouen, le du Madrillet, Avenue de lUniversit Technopo e, B.P. 12, 76801 Saint-Etienne-duRouvray (FRANCE), T el. 00 33 (0)2 32 95 52 24, Fax 00 33 (0)2 32 95 52 86, Paul.Lescot@univrouen.fr,

hal-00431615, version 4 - 26 Oct 2011

You might also like