You are on page 1of 18

A Refinement to Aı̈t-Sahalia’s (2002) “Maximum

Likelihood Estimation of Discretely Sampled Diffusions:
A Closed-Form Approximation Approach”
Gurdip Bakshi and Nengjiu Ju∗

First Draft: June 11, 2002
Final Version: December 15, 2003
Journal of Business (forthcoming)

Abstract

This paper provides a closed-form density approximation when the underlying
state variable is a one-dimensional diffusion. Building on Aı̈t-Sahalia (2002), we show
that our refinement is applicable under a wide class of drift and diffusion functions.
In addition, it facilitates the maximum likelihood estimation of discretely sampled
diffusion models of short interest-rate or stock volatility with unknown conditional
densities. Our interest-rate examples demonstrate that the analytical approximation
is sufficiently accurate.


Bakshi and Ju are both at Department of Finance, Smith School of Business, University of
Maryland, College Park, MD 20742. Bakshi can be reached at Tel: (301) 405-2261, email: gbak-
shi@rhsmith.umd.edu, Website: www.rhsmith.umd.edu/finance/gbakshi/; and Ju at Tel: (301) 405-2934,
email: nju@rhsmith.umd.edu, and Website: www.rhsmith.umd.edu/finance/nju/. The detailed comments
of the anonymous referee have substantially improved the paper. We thank Doron Avramov, Peter Carr,
David Chapman, Steve Heston, Soeren Hvidkiaer, Dilip Madan, Albert Madansky (the Editor), and Hui
Ou-Yang for their useful suggestions. All computer codes are available from the authors.

x|x0 . θ) and σ(Xt . θ) to which the method can be applied (in a manner to be made precise shortly). x|x0 . Aı̈t-Sahalia (2002) has proposed a general method to derive the density function of a one-dimensional diffusion process. θ) are re- spectively the coefficients of drift and diffusion. applying Ito’s lemma to Yt . it is shown that our closed-form approximation approach for determining the transition density expands the class of µ(X. n}. Specifically. he derives the 1 .I Introduction and the Approach in Aı̈t-Sahalia (2002) In a recent contribution. The intent of this article is to modify and refine the methodology in Aı̈t-Sahalia (2002). The functions µ(Xt . θ). To highlight the general difficulty in determining the transition density of Xt and to un- derstand the innovations in Aı̈t-Sahalia (2002). θ) = . Let this conditional density be denoted by pX (∆. we follow his notation and consider a one-dimensional diffusion process for some state variable Xt : dXt = µ(Xt . the density function pX (∆. · · · . (3) σ(u. (1) where Wt is a standard Brownian motion. θ) is pre-specified from (1). Building on Aı̈t-Sahalia (2002). Second. The maximum likelihood estimator of θ is obtained by maximizing the function. θ). θ) and σ(Xt . we offer one primary contribution. To describe his closed-form approx- imation approach. θ) where σ(u. θ) has remained intractable for a sufficiently wide class of µ(Xt . θ) dt + σ(Xt . θ) and σ(X. (2) i=1 Up until the work of Aı̈t-Sahalia (2002). he constructs a transformed process Y defined below: Z X du Y ≡ γ(X. Xi∆ |X(i−1)∆ . let us first outline the basic steps underlying his approach. First. θ) . n X n o ℓn (θ) = log pX (∆. Aı̈t-Sahalia (2002) is interested in characterizing the conditional density function of Xt+∆ given Xt . and θ represents the parameter vector in an open bounded set Θ ⊂ RK . θ) dWt . Under the assumption that Xt is observed at dates {t = i ∆ | i = 0.

J X (J) (j) pZ (∆. θ) and γ −1 (y. θ) ≡ − (γ (y. θ). ηZ (∆. as in Aı̈t-Sahalia (1999). θ) are known analytically. θ). (9) j! −∞ Fourth. Suppose. can be expanded around a standard normal density using Hermite polynomials (up to the Jth term). θ). (5) σ(γ −1 (y. θ) ≡ φ(z) ηZ (∆. θ) = (1 − ξ)−1 X 1−ξ and 1 γ −1 (y. its density. z|y0 . z|y0 . j ≥ 0. θ) is derived by solving the integral in (3). θ) dt + dWt . It is assumed that γ(X. θ). the expectation for the Hermites (j = 1. z 2 /2 dj −z 2 /2 Hj (z) ≡ e [e ]. θ) dz. θ) = (y(1 − ξ)) 1−ξ . Accordingly. y0 is approximately the mean of Yt+∆ . 2. θ). and Hj (z) are the Hermite polynomials. he rescales and standardizes Y via. and ∆ its variance. (4) where µ(γ −1 (y.unit diffusion process for Yt : dYt = µY (Yt . y0 . Aı̈t-Sahalia (2002) shows that Z is close to a standard normal random variable. Z = ∆−1/2 (Y − y0 ) . z|y0 . (7) j=0 2 /2 e−z where φ(z) ≡ √ 2π . σ(X. θ) represents its inverse function. Proceeding to the next step. y0. (6) where y0 is the initial value of Yt+∆ . and γ −1 (y. θ) ≡ Hj (z) pZ (∆. θ) 2 ∂x Notice that γ(X. the expansion coefficients. (8) dz j (j) As is standard. pZ (∆. y0. θ) = X ξ with 0 < ξ < 1 then γ(X. θ). · · ·) are approximated using Taylor 2 . θ) Hj (z). θ) 1 ∂σ −1 µY (y. are recovered through the orthog- onal conditions of the Hermites: Z (j) 1 ∞ ηZ (∆. For small ∆.

K) approximation. Z K   1 ∞ 1 X ∆k k y − y0 Hj (z) pZ (∆. Finally. The (J. of Yt+∆ becomes (up to the ∆K -th term. θ) . (12) ∆1/2 y0 k=0 k! where c0 (y|y0.K) (J. pX (∆. θ) + . consult (11)-(12) in Aı̈t-Sahalia (1999) and (4. One appealing alternative is to collect all terms with different orders of the Hermite polynomials j but keep the expansion coefficients to the same order K in ∆. θ) = 1. θ)dw ck (y|y0. The regularity conditions and the convergence properties of the approximations are available in Aı̈t-Sahalia (2002). θ) − γ(x0 .K) Under mild regularity conditions. θ .K) −1/2 pX (∆. θ) ≡ − µ2Y (y.11) in Aı̈t-Sahalia (2002)):   Z X K (K) y − y0 y ∆k pY (∆. θ) ∂y is the infinitesimal generator of the diffusion Yt . the approximate density of Xt+∆ conditional on Xt = x0 is obtained from the density of Z in (7) and by the Jacobian formula: 1   (J. θ) ∆1/2 Z (J. x|x0 . y|y0. (13) y0 ∂w 2 and ! 1 ∂µY (y. θ) = j (y − y0 ) (w − y0 ) λ(w) cj−1(w) + dw. θ). pY ≡ pY .K) resulting density approximation of Z is termed as pZ . and for j ≥ 1. θ) λY (y. 2002) argues that there are different ways of collecting terms in (11). In other words.theorem up to the Kth term. θ) = p ∆. z|y0 . Aı̈t-Sahalia (1999. θ) converges to the true density function as J → ∞ and K → ∞. θ) cj (y|y0. let J → ∞ in (7) and leave K the same for all j’s in (10). Z " # −j y j−1 ∂ 2 cj−1 (w|y0. (10) j! −∞ j! k=0 k! ∆1/2 1 ∂2 ∂ where A(θ) ≡ 2 ∂y 2 + µY (y. The resulting density (K) (∞. ∆ (γ(x. x|x0 . (14) 2 ∂y 3 . θ) dz ≈ A (θ) · Hj |y→y0 . (11) σ(x. θ))|γ(x0 .10)-(4. θ) = ∆−1/2 φ exp µY (w.

σ(−∞. CASE 1 Consider the volatility function (for the short interest-rate) studied in Aı̈t-Sahalia (1996): q σ(X. (15) σ(x. β1 > 0. of the type: dSt = µS (S. θ) Y For the later accuracy and comparisons tests we refer to the approximation (15) as “Aı̈t- Sahalia enhanced” and to (11) as “Aı̈t-Sahalia basic. θ) St dt + (β0 + β1 St−β2 ) St dWt . and β3 > 0. This process nests the constant elasticity of variance model. θ) and σ(X. β2 . there are potential difficulties with implementing his approach. θ) = β0 + β1 X + β2 X β3 . β0 +β1 u+β2 u 3 CASE 2 Now consider a stochastic volatility model for the stock price. for any µ(X. (K) 1 (K) pX (∆. in theory. θ) = β0 + β1 e−β2 Xt . θ) is known in exact closed-form (see equation (5)). (16) for some constants β0 .Again. β1 . The behavior of return volatility outlined in (17) is theoretically plausible with volatility bounded between β0 and ∞. the approximate density of Xt+∆ is obtained by invoking the Jacobian formula. the density approximation assumes that the inverse of γ(X. θ) − γ(x0 . even when the integration in (3) is analytical. Letting Xt ≡ log(St ).” II A Class of Closed-form Density Approximations While Aı̈t-Sahalia (2002) has shown that his method is accurate when applied to commonly adopted interest-rate models and works. the stock return volatility is: σ(Xt . Second. and β3 . the method requires that the integration in (3) be done analytically which is not always feasible in a general setting. the integration RX du √ β cannot be performed analytically. For this volatility specification. θ). St . (17) with β0 > 0. First. θ) . Specifically. θ). γ(x. θ) = ∞ 4 . θ)|γ(x0 . The following volatility functions illustrate the practical limitations of the method and shows that it precludes an important collection of continuous-time models. θ) = p (∆. x|x0 .

(18) where Y and Σ are respectively the true mean and variance of Y . We now propose a methodological refinement that circumvents the aforementioned difficul- ties. As formalized in the following three subsections. θ) is not invertible in closed-form. In model implementations the true 5 . RX   du X 1 and σ(∞. For any given µ(Xt . and makes the method applicable to a wider class of one-dimensional diffusions. While γ(X.θ) in our standardization. when computing the moments of Z (or equivalently Y ) to obtain the expansion coefficients. Equation (21) follows from noting that 2 Σ = E[(Y − Y )2 ] = ∆ E[(Z − Z)2 ] = ∆ (E[Z 2 ] − Z ). the function γ(X. this allows us to obtain the approximate moments of Z by applying the Taylor theorem to a function of X. A Standardization of Yt As articulated above. the resulting closed-form approximation RX du is less demanding because it avoids using the drift of σ(u. θ) = β0 +β1 e−β2 u = β0 + β0 β2 log β0 + β1 e−β2 X . we use the true mean and variance of σ(u. (19) where Z is the mean of Z.θ) . θ) and σ(Xt . (20) ∆1/2   1 2 −2 −1/2 1/2 2 ρ ≡ Σ ∆ = E[Z ] − Z . and Y − y0 Z ≡ . (21) E[·] denotes expectation operator throughout. For future use. We show that this modification enhances the accuracy of the method. rewrite (18) as: ! 1/2 Y − y0 Y − y0 Zb = Σ −1/2 ∆ − ∆1/2 ∆1/2   ≡ ρ Z −Z . as opposed to Aı̈t-Sahalia (2002) who constructs RX du Z in (6). First. we regard Z as a function of X (see equation (23) below). There are two aspects to our refinement. θ) = β0 . θ). we keep the transformation from X to Y in (3) but standardize Y as: Zb = Σ−1/2 (Y − Y ). Second.

(ii) the Hermite polynomials. Third. C Recovering the Expansion Coefficients ηj (∆. θ) and σ(X. In this sense. although the densities (11) and (22) are observationally similar. x0 . x0. Suppressing this dependence of Y and Σ on ∆ for brevity of presentation. the (J) pX (∆. At the outset note that the density approximation (22) requires three inputs for implemen- tation: (i) the value of ρ to be determined from (21). the density of Zb is approximated by appealing to (7) with (j) expansion coefficients ηj (∆. θ)) determined via (9)-(10). x|x0 . θ) in terms of the initial x0 . b we then arrive at the closed-form density of X: J   (J) ρ X pX (∆. x0 . (23) ∆ x0 σ(u. θ) to approximate the required moments. x0 .mean Y and variance Σ of Y are approximated by Taylor expanding in ∆. θ) ≡ φ(ρ(Z − Z)) ηj (∆. we fix J=6 and determine the desired coefficients in terms of the 6 .θ) . θ). and (iii) the expansion coefficients ηj (∆. θ) in terms of the moments of 1 RX du ∆1/2 x0 σ(u. x0 . we now proceed to characterizing the density of Xt . x|x0 . only a numerical value of Z is needed in (22). θ) so approximated enables the maximum likelihood estimation for a wider class of one-dimensional diffusion models. θ) H j ρ(Z − Z) . θ) in terms of the moments of Z. θ) (short-hand corresponding to ηZb (∆. (22) σ(x. x0 . and thereby circumventing the use of µY (y. By a straightforward application of the Jacobian formula to the density of Z. Hj (·). we depart from Aı̈t-Sahalia (2002) by obtaining ηj (∆. θ) Several comments are in order regarding the density approximation presented in (22). x0 . as shown below: Z 1 X du Z = 1/2 . B Characterizing the Density of Xt Guided by Aı̈t-Sahalia (2002). This is done by directly using the primitive µ(X. Second. θ) ∆1/2 j=0 where we may write Z as a function of X. θ) This subsection focuses on recovering the expansion coefficients ηj (∆. First. evaluated at ρ(Z − Z). even though we treat Z as an integral of x in (23).

we rely on a result on the conditional expectation of a function of X. the moments of Zb are expressed in those of Z. (26) 2 2 1 1 1 η3 = E[H3 (zb)] = E[−zb3 + 3zb] = − E[zb3 ]. x) ∂W (t. (32)  2 3 5  E[Zb 5 ] = ρ5 E[Z 5 ] − 5ZE[Z 4 ] + 10Z E[Z 3 ] − 10Z E[Z 2 ] + 4Z . R the conditional expectation W (t. where i = 1. x) = G(w) pw (t. 4. θ) = . the remaining task is to compute E[Z i ] = E[ ∆1/2 x0 σ(u. 6. (35) 2 ∂x2 ∂x ∂t 2 subject to the boundary condition W (0. 3. x) ∂W (t. x) σ (x. (24) η1 = E[H1 (zb)] = −E[zb] = 0. (33)  2 3 4 6  E[Zb 6 ] = ρ6 E[Z 6 ] − 6ZE[Z 5 ] + 15Z E[Z 4 ] − 20Z E[Z 3 ] + 15Z E[Z 2 ] − 5Z . (29) 120 120 120 12 1 1 1 1 1 η6 = E[H6 (zb)] = E[zb6 − 15zb4 + 45zb2 − 15] = E[zb6 ] − E[zb4 ] + . θ) for compactness and using (9)): η0 = 1.moments of Zb (suppressing the arguments of ηj (·. From (19). (31)  2 4  E[Zb 4 ] = ρ4 E[Z 4 ] − 4ZE[Z 3 ] + 6Z E[Z 2 ] − 3Z . θ) dWt . θ) dt + σ(Xt . (25) 1 1 η2 = E[H2 (zb)] = E[zb2 − 1] = 0. ·. (30) 720 720 720 48 24 In the next step. (27) 6 6 6 1 1 1 1 η4 = E[H4 (zb)] = E[zb4 − 6zb2 + 3] = E[zb4 ] − . Let B(θ) ≡ 21 σ 2 (x. 2. w|x) dw satisfies the partial differential equation 1 2 ∂ 2 W (t. θ) ∂∂2 x +µ(x. 5. θ) + µ(x.θ) ]. For this. θ) ∂x ∂ rep- 7 . (34)  RX i 1 du Thus. x) = G(x). Lemma 1 For any generic one-dimensional diffusion process Xt governed by dXt = µ(Xt . (28) 24 24 24 8 1 1 1 1 η5 = E[H5 (zb)] = E[−zb5 + 10zb3 − 15zb] = − E[zb5 ] + E[zb3 ]. it is easy to see that the higher-moments of Zb and Z are linked as:  3  E[Zb 3 ] = ρ3 E[Z 3 ] − 3ZE[Z 2 ] + 2Z .

resent the infinitesimal generator of the diffusion Xt . θ). µY (y) ≤ −K y β for some K > 0 and β ∗ > 1 which is analogous to the statement that near x̄: µ(x) 1 ∂σ(x) ∗ σ(x) − 2 ∂x ≤ R β ∗ X du −K σ(u) . the desired condition µY (y) ≤ −Ky β for Y gets translated into µZ ∗ (z ∗ ) ≤ −K ′ z ∗β for Z ∗ provided R β ′ µ(x) x du K ′ > 0 and β ′ > 1. In terms of x the technical condition takes the form: σ(x) − 21 ∂σ(x) ∂x ≤ −K ′ x∗ σ(u) . if we want to compute the moments of ∗ ′ Z ∗ . the formal solution to equation (35) is given by: ∞ X tj j W (t. for −1 ∗ (Z )) Z ∗ = Γ(x) = Z + Γ(x0 ). Because the same moments are computed. (37) 4 24 E[Z 2 ] ≈ 1 + (ϕ20 + ϕ1 )∆ + (6ϕ20 ϕ1 + 4ϕ21 + 7ϕ0 ϕ2 + 2ϕ3 )∆2 /6 + (16ϕ31 + 16ϕ30 ϕ2 +32ϕ1 ϕ3 + 28ϕ20 (ϕ21 + ϕ3 ) + 8ϕ0 (11ϕ1 ϕ2 + 2ϕ4 ) + 3(7ϕ22 + ϕ5 ))∆3 /48. x) = eBt G(x) = B · G(x). we can show that the first six moments can be derived as (using Mathematica. (40) E[Z 5 ] ≈ 15ϕ0 ∆1/2 + 5(8ϕ30 + 30ϕ0 ϕ1 + 11ϕ2 )∆3/2 /4 + (8ϕ50 + 200ϕ30ϕ1 + 320ϕ20ϕ2 +350ϕ1ϕ2 + 20ϕ0 (21ϕ21 + 10ϕ3 ) + 43ϕ4 )∆5/2 /8. Hence. where Γ(x) = x∗ du/σ(u) and x is some appropriately chosen lower limit. with the aid of Lemma 1. we focus on how to map one such condition: near ȳ = +∞. and with K=3 for the expectation of the Hermites): ∆3/2 ∆5/2 E[Z 1 ] ≈ ϕ0 ∆1/2 + (2ϕ0 ϕ1 + ϕ2 ) + (4ϕ20 ϕ2 + 6ϕ1 ϕ2 + 4ϕ0 (ϕ21 + ϕ3 ) + ϕ4 ) . (41) 1 To avoid unnecessary duplication. Write Z = x0 σ(u) = Rx ∗ Γ(x) − Γ(x0 ). θ)) under which the series j=0 j! B j ·G(x) converges coincide with those RX du in Proposition 4 of Aı̈t-Sahalia (2002). the sufficient conditions P∞ tj on (µ(Xt . (36) j=0 j! for any function G(x). σ(Xt . (38) E[Z 3 ] ≈ 3ϕ0 ∆1/2 + (4ϕ30 + 18ϕ0 ϕ1 + 7ϕ2 )∆3/2 /4 + (6ϕ30 ϕ1 + 13ϕ20 ϕ2 + 15ϕ1 ϕ2 + ϕ0 (16ϕ21 + 9ϕ3 ) + 2ϕ4 )∆5/2 /4.θ) can be 1 mapped back into those on µ(Xt . Ito’s Lemma implies dZt∗ = µ(Γ 1 ∂σ σ(Γ−1 (Z ∗ )) − 2 ∂x (Γ −1 (Z ∗ )) dt + dWt . θ) and σ(Xt . (39) E[Z 4 ] ≈ 3 + 6(ϕ20 + ϕ1 )∆ + (ϕ40 + 12ϕ20 ϕ1 + 7ϕ21 + 11ϕ0 ϕ2 + 3ϕ3 )∆2 + (80ϕ40 ϕ1 + 240ϕ31 + 280ϕ30 ϕ2 + 241ϕ22 + 368ϕ1 ϕ3 + 20ϕ20 (27ϕ21 + 17ϕ3 ) + 4ϕ0 (290ϕ1 ϕ2 + 43ϕ4 ) + 31ϕ5 )∆3 /40. Realize that ∗ the above dynamics for Z coincides with that for Y . we restate this condition as a R x du definite integral so that numerical integration can be employed to check its validity. Since the desired infinite integral may not be available. Then. 8 . Restrictions on the drift of Y ≡ σ(u. Finally. Therefore. θ).

while the derived expressions appear cumbersome.E[Z 6 ] ≈ 15 + 45(ϕ20 + ϕ1 )∆ + 15(2ϕ40 + 18ϕ20 ϕ1 + 10ϕ21 + 15ϕ0 ϕ2 + 4ϕ3 )∆2 /2 + (16ϕ60 + 720ϕ40 ϕ1 + 1440ϕ31 + 1760ϕ30ϕ2 + 1291ϕ22 + 1968ϕ1 ϕ3 +60ϕ20 (59ϕ21 + 31ϕ3 ) + 24ϕ0 (275ϕ1ϕ2 + 37ϕ4 ) + 157ϕ5)∆3 /16. In this multivariate extension. they are straightforward to program. (46) ϕ4 = µ0 σ14 /σ0 − σ12 (µ1 σ1 + 3µ0 σ2 ) + σ0 (σ12 µ2 − σ13 σ2 /2 + 2µ0 σ22 + σ1 (2µ1 σ2 + µ0 σ3 )) − σ02 (2µ2 σ2 + 2σ1 (σ22 − µ3 ) + 3µ1 σ3 + 7σ12 σ3 /2 + µ0 σ4 ) + σ03 (−5σ2 σ3 /2 + µ4 − 3σ1 σ4 ) − σ04 σ5 /2. and into (31)-(34) to get the moments b Allowing for possible modifications in µ and σ . 1. θ) and their partial derivatives. (48) and µi ≡ ∂ i µ(x0 . (45) ϕ3 = −µ0 σ13 /σ0 + σ1 (µ1 σ1 + 2µ0 σ2 ) − σ0 (2µ1σ2 + σ12 σ2 /2 + µ0 σ3 ) − σ02 (σ22 /2 − µ3 + 3σ1 σ3 /2) − σ03 σ4 /2. for i = 0. θ) and σ(X. (43) ϕ1 = µ1 − µ0 σ1 /σ0 − σ0 σ2 /2. (47) ϕ5 = −µ0 σ15 /σ0 + σ13 (µ1 σ1 + 4µ0 σ2 ) − σ0 σ1 (4µ1 σ1 σ2 + σ13 σ2 /2 + 4µ0 σ22 + 2µ0 σ1 σ3 ) − σ02 (σ12 (11σ22 /2 − 5µ3 ) + 15σ13 σ3 /2 − 5µ0 σ2 σ3 + µ1 (−4σ22 + 3σ1 σ3 ) + µ0 σ1 σ4 ) − σ03 (2σ23 + 5µ2 σ3 + 37σ1 σ2 σ3 /2 − 5σ1 µ4 + 4µ1σ4 + 25σ12 σ4 /2 + µ0 σ5 ) − σ04 (5σ32 /2 + 11σ2 σ4 /2 − µ5 + 5σ1 σ5 ) − σ05 σ6 /2. x0 .θ) to be determined numerically at the last stage of the procedure is related to the multivariate expansion for the “irreducible” case explored in Aı̈t-Sahalia (2003). each expansion coefficient is obtained in closed-form in terms of µ(X. 6. i i for any µ(X. this method affords convenience and can be applied without having to construct µY (y. θ)/∂xi0 and σi ≡ ∂ i σ(x0 . One. As such. 5. For implementation. 2. Two. 1 RX du The method of this paper that allows ∆1/2 x0 σ(u. θ) and σ(X. θ) deserves several remarks. ηj (∆. x0 . 4. the same computer code can be used of Z. the expansion coefficients 9 . (44) ϕ2 = −µ1 σ1 + µ0 (σ12 /σ0 − σ2 ) − σ0 (−2µ2 + σ1 σ2 + σ0 σ3 )/2. θ). (42) where ϕ0 = µ0 /σ0 − σ1 /2. 3. θ) are obtained by substituting (43)-(48) into (37)-(42). θ) in (5). Our construction for determining the expansion coefficients ηj (∆. θ)/∂xi0 .

the Va- sicek process. III Application to Models of Interest-Rate and Stochas- tic Volatility As emphasized earlier. and certain members in the constant elasticity of variance class. θ) with respect to X. To show the promise of the proposed refinement. Examples in this class includes. It merely re- quires the successive derivatives of µ(X. (54) 10 .satisfy a cascade of differential equations. The multivariate procedure thus relies on a double Taylor series expansion in (x − x0 ) and ∆. (51) µ1 = µ′ (x0 ) = α1 + 2 α2 x0 − α3 /x20 . β3 ) can be estimated using maximum likelihood. α1 . For convenience. they are provided below (for initial value x0 ): µ0 = µ(x0 ) = α0 + α1 x0 + α2 x20 + α3 /x0 . β1 . α2 . β2 . where the drift and diffusion are: µ(X. (49) q σ(X. θ) = α0 + α1 X + α2 X 2 + α3 /X. θ) and σ(X. the expansion coefficients are Taylor expanded in (x − x0 ) which reduces the differential equations to a system of linear equations. α3 . we now present specific examples where the true density function is not analytically known but can nonetheless be approximated using our approach. To obtain analytical solutions. β0 . (50) The density function can easily be approximated based on (22). (52) µ2 = µ′′ (x0 ) = 2 α2 + 2 α3 /x30 . θ) = β0 + β1 X + β2 X β3 . the inverse square-root process of Ahn and Gao (1999). The dependence of the density function on x is through the Hermite polynomials. Example 1: Suppose the short interest-rate is modeled as in Aı̈t-Sahalia (1996). when (3) and (5) can be done analytically. (53) µ3 = µ(3) (x0 ) = −6 α3 /x40 . and the parameter vector θ ≡ (α0 . the CIR square-root process. Both our approach and Aı̈t-Sahalia (2002) however obtain Hermite coefficients ηj by a Taylor expansion in ∆. among others. the Aı̈t-Sahalia (2002) method is elegant and accurate.

(57) σ1 = σ ′ (x0 ) = (β1 + β2 β3 xβ0 3 −1 )/(2 σ0 ). (64) 2 σ(X. θ) = β0 + β1 e−β2 X . (62) σ6 = σ (6) (x0 ) = (β2 β3 (β3 − 1)(β3 − 2)(β3 − 3)(β3 − 4)(β3 − 5) x0β3−6 − 12 σ1 σ5 − 30 σ2 σ4 − 20 σ32)/(2 σ0 ). the derivatives of σ(X. is driven by a generalized constant elasticity of variance process (as outlined in Case 2). (67) σ2 = σ ′′ (x0 ) = β1 β22 e−β2 x0 . (68) 11 . (55) µ5 = µ(5) (x0 ) = −120 α3 /x60 . (63) Substituting the expressions for the derivatives of µ(X. 1 2 µ(X. θ) − σ (X. (65) Consistent with our procedure. θ). q σ0 = σ(x0 ) = β0 + β1 x0 + β2 xβ0 3 . θ) and µ(X. θ) in (57)-(63) into (43)-(48). Example 2: As another parametric example outside of the Bessel class. θ) = µS (eX . θ) with respect to X are given by σ0 = σ(x0 ) = β0 + β1 e−β2 x0 . (60) σ4 = σ (4) (x0 ) = (β2 β3 (β3 − 1)(β3 − 2)(β3 − 3) xβ0 3 −4 − 8 σ1 σ3 − 6 σ22 )/(2 σ0 ). St . The drift and diffusion of Xt ≡ log(St ) are respectively. and the derivatives of σ(X. (59) σ3 = σ (3) (x0 ) = (β2 β3 (β3 − 1)(β3 − 2) x0β3 −3 − 6 σ1 σ2 )/(2 σ0). θ) in (51)-(56). (61) σ5 = σ (5) (x0 ) = (β2 β3 (β3 − 1)(β3 − 2)(β3 − 3)(β3 − 4) x0β3 −5 − 10 σ1 σ4 − 20 σ2 σ3 )/(2 σ0 ). (66) σ1 = σ ′ (x0 ) = −β1 β2 e−β2 x0 . (56) and. µ4 = µ(4) (x0 ) = 24 α3 /x50 . (58) σ2 = σ ′′ (x0 ) = (β2 β3 (β3 − 1) x0β3 −2 − 2 σ12 )/(2 σ0 ). suppose the stock price. we get the required moments for Z.

x0 . To explore these points we first compare the approximate density using Aı̈t-Sahalia basic (i. (76) (4) µ4 = µ(4) (x0 ) = µS (ex0 . (71) σ6 = σ (6) (x0 ) = β1 β26 e−β2 x0 . we arrive at the conditional density of the log stock price by appealing to the series approximation in (22). (78) for any stock price drift function µS (S. based on (15)). (70) σ5 = σ (5) (x0 ) = −β1 β25 e−β2 x0 . θ). θ) − 3 σ1 σ2 − σ0 σ3 . (74) µ2 = µ′′ (x0 ) = µ′′S (ex0 . µ0 = µ(x0 ) = µS (ex0 . (77) (5) µ5 = µ(5) (x0 ) = µS (ex0 . Armed with these expressions. θ) One remaining central issue concerns the accuracy of the refinement and whether using our Ẑ = Σ−1/2 (Y − Y ) reduces the magnitude of the expansion coefficients ηj (∆. In sum. θ). x0. θ) − σ02 /2.. θ) − σ12 − σ0 σ2 . θ) − σ0 σ1 . with that derived in (22). We consider three candidate interest-rate models for which the conditional density function 12 . (73) µ1 = µ′ (x0 ) = µ′S (ex0 . (72) and. (75) (3) µ3 = µ(3) (x0 ) = µS (ex0 . θ) − 10 σ2 σ3 − 5σ1 σ4 − σ0 σ5 . based on (11)) and Aı̈t-Sahalia enhanced (i.e. the refinement offers flexibility and renders the basic methodology applicable to an expanded class of one-dimensional diffusion processes. IV Accuracy of the Method and Comparison of Ex- pansion Coefficients ηj (∆.e.. σ3 = σ (3) (x0 ) = −β1 β23 e−β2 x0 . (69) σ4 = σ (4) (x0 ) = β1 β24 e−β2 x0 . θ) − 3 σ22 − 4σ1 σ3 − σ0 σ4 .

2002) that approximating the density of the transformed variable Y around a standard normal distribution using the Hermites is indeed good.   √ dXt =  κ(µ − Xt ) dt + σ Xt dWt CIR. t t t t As recommended by Aı̈t-Sahalia (1999). it highlights the finding in Aı̈t-Sahalia (1999. the tests indicate that the proposed method is likely to be very accurate for many conditional densities. 0. we set J=6 and K=3. this criterion gauges the worst possible ap- proximation error.31 (×10−8 ) using Aı̈t-Sahalia enhanced. The fact that the limit J → ∞ is taken in the enhanced formula (15) explains why this method is far more accurate. x|x ) 0 − papprox (∆.25 (×10−5 ). x|x0 ) |. Second. in Table 2. our refinement reduces maximum absolute errors several folds with respect to Aı̈t-Sahalia basic but not with respect to the enhanced formulae (15). To illustrate this point we note that for the CIR model and x0 =10%. in what follows. MAXE using the approximation in Aı̈t- Sahalia basic is 5. Overall this exercise demonstrates a potential tradeoff between the methodology in Aı̈t-Sahalia enhanced and the refinement in (22): the former is more precise but the latter provides additional applicability by not requiring the closed-form integration of 1/σ(u. we report and compare the magnitude of the Hermite coefficients η1 through η6 in (25)- (30) with the counterparts in Aı̈t-Sahalia basic.is known in closed-form:   κ(µ − Xt ) dt + σ dWt  Vasicek. and compute MAXE ≡ max | pexact (∆. Given an initial value. (79)   q  κ(µ − X ) X dt + σ X 3 dW Ahn-Gao. a superior approximation method produces a lower value of maximum absolute error.14 (×10−5 ) using (22) and 0. For benchmarking this error. This improvement pattern in ηj is observed regardless of the level of x0 (by construction our η1 = 0 and η2 = 0). 13 . our yardstick for testing accuracy relies on compar- ing the maximum absolute error of the approximate density relative to its exact counterpart. That is. we also present the maximum exact conditional density denoted as max(p|x0 ). First. To understand the impact of adopting different standardization schemes for Z. In essence. With ∆ = 1/12. Observe that using the true mean and variance in the calculation of Ẑ = Σ−1/2 (Y − Y ) reduces the coefficients η3 through η6 in comparison to Z = ∆−1/2 (Y − y0 ) in Aı̈t-Sahalia basic. θ). Several points can be made based on the accuracy results shown in Table 1.

First. θ) and σ(Xt . the refinement can be employed to empirically compare alternative models of short interest-rate (or stochastic volatility models describing equity returns) for which the exact conditional density is generally unavailable. 14 . Among various applications. However. θ).V Concluding Remarks In theory. Our density characterization is based on two building blocks. θ). we show how to obtain the approximate moments of RX du/σ(u. θ) by applying the Taylor theorem to a generic function of X. θ). θ) and σ(Xt . We show how to determine the density function of Xt in a less restrictive environment. θ) dt + σ(Xt . Next. the density approximation developed in Aı̈t-Sahalia (2002) holds for any gen- eral diffusion process dXt = µ(Xt . The numerical implementation suggests that this refinement produces approximate densities that are sufficiently accurate but nonetheless less precise than the enhanced formulae when J → ∞ in the original density approximation (11). the method demands the closed-form RX inversion of the function du/σ(u. θ). A key feature of this refinement is that it renders the basic methodology applicable in a setting where the integral transformation in (3) and/or the inverse function (5) cannot be performed analytically. θ) dWt . Even when this integration is possible. we have developed a methodological refinement to Aı̈t-Sahalia (2002). the method can only be applied in the narrower diffusion class for which 1/σ(u. In this article. the likelihood function of the observations cannot be explicitly computed. Our examples establish that the refinement generates closed-form density approxima- tions for a wider class of µ(Xt . In either scenario. This approxima- tion method not only rests on sound theoretical foundations but has promising applications in the field of financial economics. for any given µ(Xt . RX we adopt a standardization that uses the true mean and variance of du/σ(u. in practice. θ) can be analytically integrated.

D. Transition densities for interest rate and other nonlinear diffusions. Y. A parametric nonlinear model of term structure dynamics. 1996. 1999. 2002. Econometrica 53: 385-408.. Princeton University. Y. J. B. 2003. Aı̈t-Sahalia. Econometrica 70: 223-262. Ingersoll. 1999.. An equilibrium characterization of the term structure. Review of Financial Studies 12: 721-762. 1977. A theory of the term structure of interest rates. Aı̈t-Sahalia. Aı̈t-Sahalia. and Ross. Journal of Finance 54: 1361-1395. References Ahn. Cox. Work- ing Paper. Y. Testing continuous-time models of the spot interest rate. 15 . Journal of Finan- cial Economics 5: 177-188. Review of Financial Studies 9: 385-426. Vasicek.. 1985. Maximum likelihood estimation of discretely sampled diffusion: A closed-form approximation approach. and Gao. Closed-form likelihood expansions for multivariate diffusions. O. Y. Aı̈t-Sahalia. J. S.

this calculation is reported for (i) the density approximation in Aı̈t-Sahalia basic.25 8.70 This article [×10−5 ] 0.31 0.17 5.33 0.31 6.77 21.71 5. The results for dXt = κ(µ − Xt ) Xt dt + σ Xt3 dWt (Ahn and Gao (1999)) are similar and not presented to save on space.11 1.51 11.24 10.04 0. and (2) dXt = κ(µ −Xt ) dt+ σ Xt dWt (CIR).258. (ii) the density ap- proximation presented in equation (22).75 3.12 0.24 3. Setting J=6 and K=3.04 1. The final expressions for the approximate densities for each interest-rate model are rather lengthy and omitted (but available from the authors).31 6.41 Aı̈t-Sahalia enhanced [×10−7 ] 1.86 3.45 0. for the Vasicek model.e.7 8. and σ = 0.31 6. we take κ = 3.17 13.81 3.0717. we set κ = 0.34 5.0828. denoted MAXE.24 2.10 1.75 −8 Aı̈t-Sahalia enhanced [×10 ] 89.59 2. The results are presented separately for two interest-rate √ processes: (1) dXt = κ(µ −Xt ) dt+ σ dWt (Vasicek).06521.14 5. µ = 0.32 1.49 10.11 5. x|x0 ) over x.71 7.28 1.12) in Aı̈t- Sahalia (2002)).14 1.82 13. and (iii) the density approximation based on Aı̈t- Sahalia enhanced (i. The structural parameters used in the density calculations for the Vasicek and the CIR model are adopted from Aı̈t-Sahalia (1999). For each initial value of the short interest- rate (denoted x0 ).18 2.49 1.65 4. x|x0 ) − papprox (∆.11)-(4.31 6. we q set κ = 0. and σ = 1.27 124.59 7. Specifically. x0 0.83 3.31 MAXE: Aı̈t-Sahalia basic [×10−4 ] 9.10 5.51 3. the entries corresponding to max(p|x0 ) represent the maximum exact conditional density.26 16 .07 This article [×10−6 ] 5. For the Ahn-Gao model.02213. µ = 0.21 8..17 5.16 0.65 0.75 6. For the CIR model.18 Vasicek max(p|x0 ) (×101 ) 6. TABLE 1: Maximum Absolute Errors for Interest Rate Models All reported calculations are based on ∆ = 1/12.06 0.145.15 0.31 6. Initial Interest-Rate.14 0. We also re- port the maximum absolute errors.55 6.00 2.31 6. and σ = 0.31 6. equations (11)-(12) in Aı̈t-Sahalia (1999) or (4.49 17. which is the maximum value of |pexact (∆.16 28.08 0.09 0.31 6.99 CIR max(p|x0 ) (×101 ) 15. x|x0 )| over x. µ = 0.11 1. That is.27 8.36 2.22 0.31 0.4387.60 6.23 0.0732.02 0.0 10.04 MAXE: Aı̈t-Sahalia basic [×10−5 ] 58.1920.10 0. max(p|x0 ) = max of pexact (∆.12 4.

81 -15. Aı̈t-Sahalia [×10−7 ] -16.18 Vasicek η1 .02 2. Aı̈t-Sahalia [×10−5 ] -6.27 0.64 -3.92 −7 η4 .39 CIR η1 .65 -8.50 -6.02 -1.30 0.06 10.04 0. Aı̈t-Sahalia [×10−3 ] 8.09 -4.19 2.42 -32.22 -6. The structural parameters for the interest rate models are the same as in Table 1.17 1.86 6.73 -11.06 -0.17 -4.38 -3.24 η3 . This article [×10 ] -205. TABLE 2: Comparison of ηj (∆.90 0.10 -2.27 -53.63 -1.46 -0.47 -3.69 2. This article [×10−9 ] 23.13 -2.03 -2.19 2.05 η6 .27 2.64 -2.96 -1.57 4.36 11.54 8.66 1.94 3.54 -12.33 8. As seen from equations (25)-(26).26 -1. This article [×10−8 ] -110.01 -1.09 η5 . This article [×10−5 ] -42.69 η6 .28 -3.18 η6 .89 -2.44 -1.06 0.58 0. Aı̈t-Sahalia [×10−1 ] -1.20 -1.16 3.36 0.16 2.92 -10. This article [×10−8 ] 1.55 -3.62 -8.94 1. This article [×10−7 ] -0.14 0.12 0.08 -3.34 -9.35 -2.42 η3 .61 -5.22 4.44 -6.02 0. x0 0.01 -2.64 -2.93 -0.29 -13.61 2.72 -1.59 7.39 0.14 -1.60 -4. Initial Interest-Rate.20 η5 . For each initial value of the short interest-rate (denoted x0 ).23 -1.75 1.84 -10.95 1.28 -0.23 -3. η1 = 0 and η2 = 0 in our approach and therefore suppressed in the tabulations.40 η6 . Aı̈t-Sahalia [×10−6 ] -1.57 -0.66 η5 .26 3.51 1.26 32.74 -1.66 -1.97 -2.67 η3 .11 -4.28 2.524 1.89 9.61 1.22 -5.61 0. x0 .94 -4.28 -3.73 2.47 -24.10 -1.77 -6.91 -9.50 11.90 η3 .37 4.44 -1.40 -2.51 2. θ) Coefficients for Interest Rate Models: Aı̈t-Sahalia Basic versus the Refinement in (22) All calculations are based on ∆ = 1/12. Aı̈t-Sahalia [×10−4 ] 9.13 -10.54 3.65 η4 . Aı̈t-Sahalia [×10−5 ] -4.07 -0.91 1.26 -1.85 0.22 -1.57 η4 . Aı̈t-Sahalia [×10−8 ] 11. we report the magnitude of ηj (∆.96 0. and (ii) the density approximation presented in equation (22).08 -2.04 17.44 -8.61 η2 .08 0.16 -3.27 0.08 13.19 -10.95 -3. Aı̈t-Sahalia [×10−2 ] -20.46 -5.18 39. Aı̈t-Sahalia [×10−4 ] 7.76 17 .39 -1. This article [×10−7 ] -2.19 η2 .71 2.46 1. Aı̈t-Sahalia [×10−3 ] 4.25 -5.98 12.10 0.04 0.86 -20.31 -5.33 15.82 -2.42 1.95 2.93 0.17 5.03 -9.09 -13.62 54.64 1.32 3.62 -6.58 38.33 -1.77 0.84 0.66 15.89 -8.68 η4 .61 1.08 0.29 -2. Aı̈t-Sahalia [×10−7 ] 2.16 0. This article [×10−8 ] 1.51 6. θ) used in (i) the density approximation in Aı̈t-Sahalia basic.21 -2.18 η5 .74 9.51 -3. x0 .42 -1.08 -1.59 -10.51 -2.09 5.93 93.35 6.18 6.