1: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2: %% Trim Size: 9.75in x 6.5in
3: %% Text Area: 8in (include Runningheads) x 5in
4: %% ws-ijtaf.tex : 19-5-05
5: %% Tex file to use with ws-ijtaf.cls written in Latex2E.
6: %% The content, structure, format and layout of this style file is the
7: %% property of World Scientific Publishing Co. Pte. Ltd.
8: %% Copyright 1995, 2002 by World Scientific Publishing Co.
9: %% All rights are reserved.
10: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
11:
12: \documentclass{ws-ijtaf}
13:
14:
15: \begin{document}
16:
17: \markboth{Luca Capriotti}
18: {The exponent expansion: an effective approximation of transition ... }
19:
20:
21: %%%%%%%%%%%%%%%%%%%%% Publisher's Area please ignore %%%%%%%%%%%%%%%
22: \catchline{}{}{}{}{}
23: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
24:
25: \title{THE EXPONENT EXPANSION: AN EFFECTIVE APPROXIMATION OF TRANSITION
26: PROBABILITIES OF DIFFUSION PROCESSES AND PRICING KERNELS OF
27: FINANCIAL DERIVATIVES}
28:
29:
30: \author{LUCA CAPRIOTTI}
31:
32: \address{Global Modelling and Analytics Group, Credit Suisse \\
33: One Cabot Square, London, E14 4QJ, United Kingdom \\
34: \email{luca.capriotti@credit-suisse.com} }
35:
36: \maketitle
37:
38: \begin{history}
39: \received{(19 September 2005)}
40: \revised{(18 January 2006)}
41: \end{history}
42:
43: \begin{abstract}
44: A computational technique borrowed from the physical sciences
45: is introduced to obtain accurate closed-form approximations
46: for the transition probability of arbitrary diffusion processes.
47: Within the path integral framework the same technique allows
48: one to obtain remarkably good approximations
49: of the pricing kernels of financial derivatives.
50: Several examples are presented, and the application of these
51: results to increase the efficiency of numerical approaches
52: to derivative pricing is discussed.
53: \end{abstract}
54:
55: \keywords{computational Finance; stochastic processes; derivative pricing; path integral Monte Carlo.}
56:
57: \section{Introduction}
58:
59: Continuous-time diffusion processes are at the basis of much of the
60: modeling work performed every day in Finance and Economics.
61: Indeed, since Bachelier's pioneering work on the
62: application of probability theory to the dynamics of stock prices \cite{bachelier},
63: many economic variables subject to unpredictable fluctuations,
64: have been modeled by stochastic differential equations of the form
65: \begin{equation}
66: dY_t = \mu_y(Y_t) dt + \sigma_y(Y_t) dW_t~.
67: \label{diffusionY}
68: \end{equation}
69: Here $\mu_y(y)$ is the drift, describing a deterministic trend,
70: and $\sigma_y(y) \ge 0$ is the volatility function, describing the level of randomness
71: introduced by the Wiener process (i.e., white noise), $dW_t$. The main reason for the
72: popularity of this class of models is probably that in continuous time one can perform analytic calculations
73: using the instruments of stochastic calculus, and the powerful framework of partial differential equations.
74: In particular, for the few cases for which the process (\ref{diffusionY}) is exactly solvable,
75: one can derive closed-form solutions for the associated transition probability. The latter
76: contains all the statistical properties of the financial quantity modeled
77: by the diffusion, and can be exploited in a variety of ways, including
78: the derivation of no-arbitrage prices for financial derivatives in complete markets.
79: The milestone results derived by Black, Scholes and Merton \cite{bs}, Cox, Ingersoll and Ross \cite{cir},
80: or Vasicek \cite{vasicek}, are among the most significant examples of the amount of
81: progress in Economics that has been
82: done using integrable continuous-time diffusion processes.
83:
84: Nonetheless, an accurate description of the market observables requires in general more
85: sophisticated models than those for which an analytic solution is available. These
86: are usually tackled by means of numerical schemes ultimately relying on a
87: discretization of the diffusion, obtained by replacing the infinitesimal time $dt$ with
88: a finite time step, $\Delta t$. The approximate results obtained in this way
89: become exact only approaching the limit $\Delta t \to 0 $, and this can be done
90: with some computational effort.
91:
92: In this paper, we utilize the {\em exponent expansion} -- a technique introduced in chemical physics
93: by Makri and Miller \cite{makri} -- to derive a short-time approximation of the transition
94: probability of the diffusion process (\ref{diffusionY}).
95: The aim is to obtain an analytic approximation which is
96: as accurate as possible for a time step $\Delta t $ as large as possible.
97: On one hand, this allows one to derive approximations of financial quantities
98: that are very accurate even for sizable values of the time step. On the other, it allows a
99: reduction of the computational burden of numerical schemes
100: as the limit $\Delta t \to 0 $ can be achieved with larger time steps, i.e., with
101: less calculations.
102:
103: The possibility to use Makri and Miller's technique to derive approximations
104: of the transition probability was originally hinted by Bennati
105: {\em et al.} in Ref.~\cite{rosaclot}. Here we explore this possibility,
106: giving derivations for a generic diffusion process with state-dependent
107: drift and volatility, and we study the reliability of the exponent expansion
108: by applying it to several diffusion processes of financial interest.
109:
110: Through the exponent expansion, the transition probability is obtained as a power
111: series in $\Delta t$ which becomes asymptotically exact if an increasing number
112: of terms is included, and provides remarkably accurate results even truncating it
113: to the first few (say, $n=3$) terms. Two derivations are offered, the first by means of Kolmogorov's
114: forward equation \cite{shreve} (Sec.~\ref{expa}),
115: and the second introducing a slightly different formalism (Sec.~\ref{alternative}).
116: The latter, once the problem is formulated in terms of
117: Feynman's path integrals \cite{pireference}, allows the generalization of the exponent expansion to
118: the calculation of the pricing kernel of financial derivatives whose underlying follows the
119: considered diffusion. This allows in turn the derivation of simple approximations for
120: the price of such contingent claims (Sec.~\ref{prickern}).
121: In Sections \ref{app1} and \ref{app2}, we illustrate the exponent expansion through the application to the Vasicek, the Cox-Ingersoll-Ross, and the Constant Elasticity of Variance models, and in Section \ref{montecarlo}
122: we discuss its application to Monte Carlo calculations within the path integral framework \cite{montagna,baaquiep,matacz,dash}. Finally, we draw our conclusions and we discuss further possible developments in Section \ref{conclusion}.
123:
124:
125: \section{Exponent expansion and transition probabilities of diffusion processes}
126:
127: \label{expa}
128:
129: In this Section, we derive the exponent expansion for the
130: transition probability, $\rho_y(y,\Delta t|y_0)$, giving the likelihood
131: that the random walker following the process (\ref{diffusionY})
132: ends up in the position $y$ at time $t=\Delta t$, given that it was in $y_0$ at time $t = 0$.
133: Although the derivation can be generalized to the case where drift and
134: volatility have an explicit time dependence, here for simplicity we will
135: restrict the discussion to the time homogeneous case. In order to make the derivation easier,
136: it is convenient to transform the original process in an auxiliary one, say $X_t$,
137: with constant volatility $\sigma_x$. Following A\"it-Sahalia \cite{aitsahalia}, this
138: can be achieved in general through the following integral transformations
139: \begin{equation}
140: X_t = \gamma(Y_t) \equiv \pm \sigma_x \int^{Y_t} \frac{dz}{\sigma_y(z)}~,
141: \label{inttransf}
142: \end{equation}
143: where the choice of the sign is just a matter of
144: convenience depending on the specific problem considered.
145: The latter relation defines a one to one mapping between the $x$ and $y$
146: processes as the condition $\sigma_y(z) \ge 0$ ensures that the function
147: $x = \gamma(y)$ defined by (\ref{inttransf}) is monotonic, and therefore
148: invertible. \footnote{For a discussion of the regularity conditions on the drift and volatility
149: functions see e.g., Ref.~\cite{aitsahalia}.} A straightforward application of Ito's Lemma \cite{shreve}
150: allows one to write the diffusion process followed by $X_t$ as
151: \begin{equation}
152: dX_t = \mu_x(X_t) dt + \sigma_x dW_t~,
153: \label{processX}
154: \end{equation}
155: with
156: \begin{equation}
157: \mu_x(x) = \sigma_x \left[\mp \, \frac{\mu_y(\gamma^{-1}(x))}{\sigma_y(\gamma^{-1}(x))} - \frac{1}{2}\frac{\partial \sigma_y}{\partial y}(\gamma^{-1}(x))\right]~,
158: \label{drift}
159: \end{equation}
160: where $y = \gamma^{-1}(x)$ is the inverse of the transformation (\ref{inttransf}).
161: Finally, the transition probability for the $x$ and $y$ processes are simply related
162: by the Jacobian associated with (\ref{inttransf}) giving
163: \begin{equation}
164: \rho_y(y,\Delta t|y_0) = \sigma_x \frac{\rho_x(\gamma(y),\Delta t| x_0)}{\sigma_y (y)}~.
165: \label{jacobian}
166: \end{equation}
167:
168: %For $\Delta t\to 0$ the transition probability for the process is dominated
169: %by the diffusive Gaussian component and therefore reads:
170: %\begin{equation}
171: %\rho_x(x,\Delta T |x_0) = \left( \frac{1}{2\pi \sigma_x^2 \Delta t}\right)^{1/2} %\exp{\left(-\frac{(x-x_0)^2}{2\sigma_x^2\Delta t}\right)}~.
172: %\end{equation}
173: In order to find an expression for the transition probability associated with Eq.~(\ref{processX})
174: which is accurate for a time $\Delta t$ as long as possible, we make the following {\em ansatz}:
175: \begin{equation}
176: \rho_x(x,\Delta t|x_0) = \frac{1}{\sqrt{2\pi\sigma_x^2\Delta t}}
177: \exp {\left[ -\frac{(x-x_0)^2}{2\sigma_x^2 \Delta t}-W(x,x_0,\Delta t)\right]}~.
178: \label{ansatz}
179: \end{equation}
180:
181: Such transition probability must satisfy the Kolmogorov forward (or Fokker-Planck) equation \cite{shreve}:
182: \begin{equation}
183: \partial_t \rho(x,\Delta t|x_0) = \left[ -\partial_x \mu_x(x) +
184: \frac{1}{2}\sigma_x^2 \partial^2_x \right] \rho_x(x,\Delta t|x_0)~.
185: \label{kolmogorov}
186: \end{equation}
187: This implies in turn that the function $W(x,x_0,t)$ follows the
188: equation:
189: \begin{equation}
190: \partial_t W = - \mu_x \partial_x W
191: + \frac {1}{2} \sigma_x^2 \partial^2_x W
192: - \frac {1}{2} \sigma_x^2 \left( \partial_x W \right)^2
193: + \partial_x \mu_x
194: - \frac{x-x_0}{\Delta t} \,\left( \partial_x W+\frac{\mu_x}{\sigma_x^2} \right)~.
195: \label{eqW}
196: \end{equation}
197: Expanding the function $W(x,x_0,t)$ in powers of $\Delta t$,
198: \begin{equation}
199: W(x,x_0,\Delta t) = \sum_{n=0}^{\infty} W_n(x,x_0)\,\Delta t^n~,
200: \label{w}
201: \end{equation}
202: substituting it in Eq.~(\ref{eqW}), and equating
203: equal powers of $\Delta t$ leads in a straightforward way to a
204: decoupled equation for the order zero in $\Delta t$ giving
205: \begin{equation}
206: W_0(x,x_0) = - \frac{1}{\sigma_x^2}\int_{x_0}^{x} dz \,\mu_x(z)~,
207: \label{w0}
208: \end{equation}
209: and to the following set of recursive differential equations:
210: \begin{eqnarray}
211: (n+1)W_{n+1} &=& - (x-x_0)\partial_x W_{n+1} +
212: \left[ \frac{1}{2}\,\sigma_x^2 \partial^2_x -\mu_x\partial_x \right] W_n \nonumber \\
213: &-& \frac{1}{2}\sigma_x^2 \sum_{m = 0}^{m = n}\partial_x W_{m}\partial_x W_{n-m}
214: + \delta_{n,0} \partial_x \mu_x~.
215: \label{makrieq}
216: \end{eqnarray}
217: In particular, for $n=0,1,2$ Eqs.~(\ref{makrieq}) read:
218: \begin{eqnarray}
219: W_1(x,x_0) &=& -(x-x_0)\,\partial_x W_1(x,x_0)+ \left[ \frac{1}{2\sigma_x^2} \mu_x(x)^2 + \frac{1}{2}\partial_x \mu_x(x) \right]~,\\
220: 2W_2(x,x_0) &=& -(x-x_0)\,\partial_x W_2(x,x_0)+\frac{1}{2}\sigma_x^2\,\partial_x^2 W_1(x,x_0)~, \\
221: 3W_3(x,x_0) &=& -(x-x_0)\,\partial_x W_3(x,x_0) + \frac{1}{2}\sigma_x^2\partial_x^2W_2(x,x_0)~, \nonumber\\
222: &-& \frac{1}{2}\sigma_x^2(\partial_xW_1(x,x_0))^2 ~.
223: \end{eqnarray}
224: The differential equations above (\ref{makrieq}) are all first order, linear and inhomogeneous of the form
225: \begin{equation}
226: n W_n(x,x_0) = -(x-x_0) \partial_x W_n(x,x_0) + \Lambda_{n-1}(x,x_0)~,
227: \label{protomakri}
228: \end{equation}
229: where $\Lambda_{n-1}(x,x_0)$ is a function that is completely determined
230: by the first $n-1$ relations. It can be readily verified
231: by substitution and integration by parts that the solution of
232: (\ref{protomakri}) reads
233: \begin{equation}
234: W_n(x,x_0) = \int_0^1 \,d\xi \xi^{n-1} \Lambda_{n-1}(x_0+(x-x_0)\xi,x_0)~.
235: \end{equation}
236: This, for $n = 1,2,3$, after some manipulations, gives:
237: \begin{eqnarray}
238: W_1(x,x_0) &=& \frac{1}{\Delta x} \int_{x_0}^x dz V_{\rm eff}(z)~, \label{expansion1} \\
239: W_2(x,x_0) &=& \frac{ \sigma_x^2 } { 2\Delta x^2 } \left[ V_{\rm eff}(x)+V_{\rm eff}(x_0) - 2W_1(x,x_0) \right]~, \label{expansion2} \\
240: W_3(x,x_0) &=& -\frac{\sigma_x^2}{2\Delta x^4}
241: \left[ \Delta x \int_{x_0}^x dz V_{\rm eff}(z)^2 - \left(\int_{x_0}^x dz V_{\rm eff}(z)\right)^2\right]
242: \nonumber \\
243: &-& \frac{3 \sigma_x^2}{\Delta x^2} W_2(x,x_0) + \frac{\sigma_x^4}{4\Delta x^3}
244: \left[\partial_x V_{\rm eff}(x)-\partial_x V_{\rm eff}(x_0)\right]~,
245: \label{expansion3}
246: \end{eqnarray}
247: where $\Delta x = x-x_0$, and, for reasons that will be clearer in the next Section,
248: we have also introduced the `{\em effective potential}' as the following quantity with dimension
249: $time^{-1}$:
250: \begin{equation}
251: V_{\rm eff} (x) = \frac{1}{2\sigma_x^2} \mu_x(x)^2 + \frac{1}{2}\partial_x \mu_x(x)~.
252: \label{effpot}
253: \end{equation}
254: At this time, we just note that the first order correction can be rewritten as
255: \begin{equation}
256: W_1(x,x_0) = \frac{1}{\Delta t} \int_0^{\Delta t} dt \, V_{\rm eff} (x_0+ t \Delta x /\Delta t)~,
257: \end{equation}
258: leading to the interpretation of this term as a time-average of the effective
259: potential over the straight line, constant velocity ($\Delta x/\Delta t$)
260: trajectory between $x_0$ and $x$. Similarly, the leading term in $W_3(x,x_0)$
261: (i.e., the one proportional to the lowest power of the volatility) is proportional
262: to the variance of the effective potential over the same trajectory.
263: Finally, we observe that the corrections $W_n(x,x_0)$ are well defined in the limit $\Delta x \to 0$.
264: In particular, for $n=1,2,3$ it is not difficult to show that
265: \begin{eqnarray}
266: \lim_{x\to x_0} W_1(x,x_0) &=& V_{\rm eff}(x_0)~, \\
267: \lim_{x\to x_0} W_2(x,x_0) &=& \frac{\sigma_x^2}{12} \,\partial_x^2 V_{\rm eff}(x)~, \\
268: \lim_{x\to x_0} W_3(x,x_0) &=& -\frac{\sigma_x^2}{24} \, ( \partial_x V_{\rm eff}(x) )^2 +
269: \frac{\sigma_x^4}{240} \, \partial_x^4 V_{\rm eff}(x) ~.
270: \label{limit}
271: \end{eqnarray}
272:
273: The form of the trial transition probability represents the
274: main difference of the present approach to the one proposed in Ref.~\cite{aitsahalia},
275: which is otherwise very similar in spirit. In fact, the latter expands in powers of $\Delta t$ the exponential $\exp{\left[-W(x,x_0,\Delta t)\right]}$ rather then just the exponent, as we do here, instead. As it will be shown explicitly in the following, the present choice
276: gives rise to a distinct approximation scheme for $n>0$ providing generally a similar level of accuracy
277: but remarkably simpler mathematical expressions. This is because,
278: by keeping the exponential form of the ansatz, one formulates a guess
279: which is closer to the exact one. The latter, can be expected
280: to have an exponential form in order to satisfy the Chapman-Kolmogorov
281: property of Markov processes \cite{shreve}.
282: In addition, the exponential choice of the ansatz
283: automatically enforces the positive definiteness of the transition density which is not granted
284: in the approach of Ref.~\cite{aitsahalia}. In fact, as it will be shown explicitly in Sec.~\ref{app1},
285: in the limit of very small volatility ($\sigma_x \to 0$), when the effect of the noise disappears and
286: transition density converges to a Dirac's $\delta$ distribution, the expansion of Ref.~\cite{aitsahalia}
287: breaks down as the transition probability becomes negative. On the contrary, the exponent
288: expansion remains well defined and accurate also in this limit. Indeed, the first terms of the expansion in
289: $\Delta t$ can be also derived through a small volatility expansion of the transition
290: density, as it will be discussed in Sec. \ref{pisec}.
291:
292: Similarly to the expansion developed in Ref.~\cite{aitsahalia} the exponent expansion has in general a finite convergence radius which is a decreasing function of the volatility. As it will be shown
293: in the following, for the values of volatilities and $\Delta t$ relevant for financial applications the exponent expansion turns out to be very accurate even when truncated to the first few terms.
294:
295: \subsection{Alternative derivation}
296: \label{alternative}
297: The term $W_0(x,x_0)$ in the exponent expansion
298: is somewhat different from the higher order terms. In fact,
299: it is defined by Eq.~(\ref{w0}) which is decoupled from the recursive
300: system (\ref{makrieq}). Indeed, it is possible to obtain
301: the same result for the exponent expansion by expressing the
302: transition density as
303: \begin{equation}
304: \rho_x(x,\Delta t|x_0) = e^{-W_0(x,x_0)} \Phi_{\rho_x}(x,\Delta t| x_0)~,
305: \end{equation}
306: and looking for an approximate expansion of the form (\ref{ansatz}) for $\Phi_{\rho_x}(x,\Delta t| x_0)$.
307: It is easy to show by direct substitution in the forward Kolmogorov
308: equation (\ref{kolmogorov}) that $\Phi_{\rho_x}(x,\Delta t| x_0)$ is the solution of
309: \begin{equation}
310: {\cal H}_x \,\Phi_{\rho_x}(x,\Delta t| x_0) = - \partial_t \Phi_{\rho_x}(x,\Delta t| x_0)~,
311: \label{schro}
312: \end{equation}
313: where ${\cal H}_x$ is the ``Hamiltonian'' differential operator
314: \begin{equation}
315: {\cal H}_x = - \frac{\sigma_x^2}{2} \partial^2_x + V_{\rm eff}(x)~,
316: \end{equation}
317: and $V_{\rm eff}(x)$ is the effective potential of Eq.~(\ref{effpot}).
318: As a result one can equivalently derive the exponent expansion
319: by substituting in Eq.~(\ref{schro}) the following trial function
320: \begin{equation}
321: \Phi_{\rho_x}(x,\Delta t| x_0) = \frac{1}{\sqrt{2\pi\sigma_x^2\Delta t}}
322: \exp {\left[ -\frac{(x-x_0)^2}{2\sigma_x^2 \Delta t}-\sum_{n=1}^\infty W_n(x,x_0) \Delta t^n\right]}~,
323: \label{trial2}
324: \end{equation}
325: which does not contain the term $W_0(x,x_0)$.
326: This observation will be used in Section 3 to generalize the
327: exponent expansion to the pricing kernel of financial derivatives.
328:
329: \begin{figure}[pt]
330: \centerline{\psfig{file=fig1.ps,width=4.0in}}
331: \vspace*{-0.3cm}
332: \caption{Accuracy of the exponent expansion for the transition density of the Vasicek model (\protect\ref{vasicek}).
333: for $a=0.0717$, $b=0.261$ and $x_0=0.1$. Panel (a): comparison between the exponent expansion (stars)
334: and the approximation of Ref.~\protect\cite{aitsahalia} (squares) for $\sigma = 0.02237$ and $\Delta t = 0.5$.
335: At order zero the two schemes are identical. The uniform error of the Euler approximation is also
336: reported for comparison. Panel (b): maximum relative error for $\sigma = 0.02237$ as a function of $\Delta t$: $n=0$ (continuous), $n=1$ (dotted), $n=2$ (long dashed) and $n=3$ (short dashed). The inset is an enlargement of the 5-10 years region. Panel (c): comparison between the exponent expansion (stars) and the approximation of Ref.~\protect\cite{aitsahalia} (squares) in the regime of low volatility ($\sigma = 0.01$), for $\Delta t = 0.5$.}
337: \label{densvasicek}
338: \end{figure}
339:
340:
341: \subsection{Applications}
342: \label{app1}
343: The application of the exponent expansion to a generic
344: diffusion process of the form (\ref{diffusionY}) is rather
345: straightforward and reduces to the calculation of one dimensional
346: integrals. In this Section, we illustrate this procedure
347: for a few test cases, namely for the Vasicek \cite{vasicek}, the
348: Cox-Ingersoll-Ross \cite{cir}, and the Constant Elasticity of Variance \cite{cev}
349: diffusion processes. We will compare the results of the
350: exponent expansion with the exact results available
351: in literature, and with the approach of Ref.\cite{aitsahalia}.
352:
353: \subsubsection{Vasicek diffusion}
354: \label{vasicekapp}
355: We first consider the Ornstein-Uhlenbeck diffusion
356: proposed by Vasicek \cite{vasicek} as a model for the
357: short-term interest rate:
358: \begin{equation}
359: d X_t= a ( b - X_t) dt + \sigma dW_t~,
360: \label{vasicek}
361: \end{equation}
362: where $a$, $b$, and $\sigma$ are positive constants representing
363: the mean-reversion level, the velocity to mean reversion, and
364: the volatility, respectively. This model is integrable and
365: the corresponding probability density function is Gaussian:
366: \begin{equation}
367: \rho_{\rm ex}(x,\Delta t|x_0) = \frac{1}{(2\pi\bar\sigma^2)^{1/2}}
368: \exp{\left[-\frac{[(x_0-a) e^ {-a \Delta t}-(x-a)]^2}{2\bar\sigma^2}\right]}~,
369: \label{vasicekexact}
370: \end{equation}
371: with
372: \begin{equation}
373: \bar \sigma = \sigma \sqrt{\frac{1-e^{-2a\Delta t}}{2 a}}~.
374: \label{vasiceksig}
375: \end{equation}
376:
377: The exponent expansion of the Vasicek model can be easily derived
378: using Eqs.~(\ref{w0}), and (\ref{expansion1}-\ref{expansion3}) with the
379: effective potential, Eq.~(\ref{effpot}),
380: \begin{equation}
381: V_{\rm eff} (x) = \frac{a^2(b-x)^2}{2\sigma^2}-\frac{a}{2}~,
382: \end{equation}
383: and gives,
384: \begin{eqnarray}
385: \rho_x(x,\Delta t|x_0) &=& \frac{1}{\sqrt{2\pi\sigma^2\Delta t}}
386: \exp{\big [ -\frac{(x-x_0)^2}{2\sigma^2 \Delta t} - W_0(x,x_0) } \nonumber \\
387: &-& { W_1(x,x_0)\Delta t- W_2(x,x_0)\Delta t^2 - W_3(x,x_0)\Delta t^3 \big ]}~,
388: \end{eqnarray}
389: up to the third order in $\Delta t$ ($n=3$). Here
390: \begin{eqnarray}
391: W_0(x,x_0) &=& \frac{a (x-b)^2-a(x_0-b)^2}{2\sigma^2}~, \label{vas0} \\
392: W_1(x,x_0) &=& \frac{a^2}{6\sigma^2}((x-b)^2+(x_0-b)^2+(x-b)(x_0-b)) -\frac{a}{2}~, \label{vas1} \\
393: W_2(x,x_0) &=& \frac{\sigma^2}{2\Delta x^2}
394: \big[V_{\rm eff}(x)+V_{\rm eff}(x_0)-2W_1(x,x_0)\big]~, \label{vas2} \\
395: W_3(x,x_0) &=& -\frac{\sigma_x^2}{2\Delta x^3}\Big[ \frac{a^4}{20\sigma^4}[(x-b)^5-(x_0-b)^5] \nonumber \\
396: &+& \frac{a^2}{4}\Delta x -\frac{a^3}{6\sigma^2}[(x-b)^3-(x_0-b)^3] \Big ] \nonumber \\
397: &+& \frac{\sigma_x^2}{2\Delta x^2} (W_1(x,x_0))^2 - \frac{3\sigma_x^2}{\Delta x^2}W_2(x,x_0)
398: + \frac{a^2\sigma_x^2}{4\Delta x^2} ~,
399: \label{vas3}
400: \end{eqnarray}
401: with $\Delta x = x-x_0$.
402: It is interesting to note that the approximate transition probability obtained with
403: the present approach reproduces exactly the expansion of the exact transition
404: density Eq.~(\ref{vasicekexact}) at the same order.
405:
406: The fast convergence of the approximation
407: scheme is illustrated in Fig.~\ref{densvasicek}. Here the percentage error of the
408: exponent expansion with respect to the exact result (\ref{vasicekexact}) is plotted
409: for various $\Delta t$, and compared with the approach of Ref.~\cite{aitsahalia}. The parameter choice, also taken
410: from Ref.~\cite{aitsahalia} corresponds to a sensible parameterization for interest rate markets. We adopt one year
411: as unit of time, and we express the various parameters in this unit. The Euler approximation
412: \begin{equation}
413: p_{E}(x,\Delta t | x_0) = \sqrt{\frac{1}{2 \pi \sigma^2\Delta t}}
414: \exp{\left( -\frac{(x-x_0-\mu_x(x_0)\Delta t)^2} {2\sigma_x^2\Delta t}\right) }~,
415: \end{equation}
416: is also reported for comparison.
417: The inclusion of each successive order allows one to increase dramatically the accuracy of the approximation
418: so that the third order expansion has basically a negligible error even for a sizable time step of order
419: 6 months. Remarkably, for the considered example, the third order expansion allows one to estimate the
420: 10 years transition probability with a relative error of less than 10 basis points (Fig.~\ref{densvasicek}-(b)).
421: As illustrated in Fig.~\ref{densvasicek}-(a), in the present case the approach of Ref.~\cite{aitsahalia} provides a slightly poorer level of accuracy for $n\leq 2$. In addition, it generally produces more complicated mathematical expressions. Furthermore, as shown in Fig.~\ref{densvasicek}-(c), in the regime of small volatility the exponent expansion still provides accurate results while the performance of the approach of Ref.~\cite{aitsahalia} degrades.
422: In fact, as anticipated, in the limit of small volatility ($\sigma_x \lesssim 0.5 \%$)
423: the first order correction of the latter approach produces a negative transition probability signaling a break down of the scheme.
424:
425:
426:
427:
428: \subsubsection{Cox, Ingersoll and Ross diffusion}
429: \label{cirapp}
430:
431: The Vasicek model is probably too easy of a test case
432: as the associated transition probability is Gaussian. In fact, since
433: the exponent expansion has a leading term which is Gaussian, the higher powers in $\Delta t$
434: just have to renormalize its average and variance in order to reproduce the exact result.
435: It is interesting therefore to test the accuracy of the exponent expansion
436: for a diffusion process that, while still integrable, generates a non-normal
437: transition density. This is the case for the Feller's square root process \cite{feller}
438: \begin{equation}
439: d Y_t = a (b - Y_t) + \sigma_y \sqrt{Y_t} dW_t
440: \label{cirdiff}
441: \end{equation}
442: which is the basis of Cox, Ingersoll and Ross model
443: for the instantaneous interest rate \cite{cir}.
444: The exact transition probability is given by \cite{cir}
445: \begin{equation}
446: \rho_{\rm ex}(y,\Delta t|y_0) = c e^{-(u+v)}\left(\frac{v}{u}\right)^{\frac{q}{2}}
447: I_q ( 2\sqrt{uv})~,
448: \end{equation}
449: where $c = 2a/[\sigma_y^2(1-\exp{(-a\Delta t)})]$, $q = 2 a b/\sigma_y^2 -1 \ge 0$,
450: $ u = c y_0 \exp{(-a\Delta t})$, $v = c y$ and $I_q$ is the modified Bessel function
451: of the first kind of order $q$ \cite{abramovitz}.
452:
453: As explained in Sec.\ref{expa}, since the volatility is not uniform, it
454: is convenient to introduce the auxiliary process defined by Eq.~(\ref{inttransf}),
455: as $X_t = \gamma(Y_t) \equiv 2\sqrt{Y_t}/ \sigma_y$. The $X_t$ process follows
456: Eq.~(\ref{processX}), with $\sigma_x=1$ and
457: \begin{equation}
458: \mu_x(x) = \frac{\tilde{q}}{x}- \frac{a}{2} x~,
459: \end{equation}
460: and $\tilde{q}=q+1/2$.
461: In this case the effective potential reads:
462: \begin{equation}
463: V_{\rm eff} (x) = \frac{1}{2} \mu_x(x)^2 - \frac{\tilde{q}}{2x^2}-\frac{a}{4}~,
464: \end{equation}
465: and the four terms of the exponent expansion in Eq.~(\ref{w}) are:
466: \begin{eqnarray}
467: W_0(x,x_0) &=& - \tilde{q} \log{\frac{x}{x_0}} + \frac{a}{4} (x^2-x_0^2) \label {cir0}\\
468: W_1(x,x_0) &=& \frac{1}{2\Delta x}
469: \Big[\mu_x(x_t) - \mu_x(x_0)
470: - \tilde{q}^2 \left( \frac{1}{x} - \frac{1}{x_0}\right) \nonumber \\
471: &+& \frac{a^2}{12} \left(x^3-x_0^3\right)
472: - a\tilde{q}\left(x-x_0\right)\Big] \label {cir1} \\
473: W_2(x,x_0) & = & \frac{1}{2\Delta x^2}
474: \big[V_{\rm eff}(x)+V_{\rm eff}(x_0)-2W_1(x,x_0)\big] \label{cir2} \\
475: W_3(x,x_0) &=& -\frac{1}{2\Delta x^2}
476: \left[ \frac{G(x_t)-G(x_0)}{\Delta x} - (W_1(x,x_0))^2\right]
477: \nonumber \\
478: &-& \frac{3}{\Delta x^2} W_2(x,x_0) + \frac{1}{4\Delta x^3}
479: \left[\partial_x V_{\rm eff}(x)-\partial_x V_{\rm eff}(x_0)\right]~,
480: \label {cir3}
481: \end{eqnarray}
482: where $\Delta x = x-x_0$, $\partial_x V_{\rm eff}(z) = \tilde{q}(1-\tilde{q})/z^3+a^2 z/4$, and
483: \begin{equation}
484: G(z) = \frac{1}{5}\alpha^2 z^5 - \frac{1}{3}\frac{\beta^2}{z^3}+\gamma^2 z+2\alpha\beta z
485: - 2\frac{\beta\gamma}{z} + \frac{2}{3}\alpha\gamma z^3~,
486: \label{gz}
487: \end{equation}
488: and $\alpha = a^2/8$, $\beta = \tilde{q}(\tilde{q}-1)/2$, $\gamma = - a(\tilde{q}+1)/2$.
489: Finally, going back to the original process, using Eq.~(\ref{jacobian}),
490: the transition probability reads:
491: \begin{equation}
492: \rho_y (y, \Delta t| y_0) = \rho_x (2\sqrt{y}/\sigma_y, \Delta t| \,2\sqrt{y_0}/\sigma_y)/\sigma_y \sqrt{y}~.
493: \end{equation}
494:
495: The accuracy of the exponent expansion in this case is illustrated in Fig.~\ref{denscir}.
496: Similarly to the case of the Vasicek diffusion, the exponent expansion is characterized by a
497: remarkably fast convergence by including successive terms of the approximation so that
498: $n=3$ provides already a virtually exact representation of the transition density, for $\Delta t \simeq 1 \,yrs$.
499: In this case, the approach of Ref.~\cite{aitsahalia} performs slightly worse of the exponent expansion
500: for $n=1$, and slightly better for $n=2$. However, also in this case the former breaks down
501: for small values of the volatility, generating unphysical transition densities.
502:
503: \begin{figure}[pt]
504: \centerline{\psfig{file=fig2.ps,width=4.0in}}
505: \vspace*{-0.3cm}
506: \caption{Accuracy of the exponent expansion for the transition density of the Cox-Ingersoll-Ross model
507: (\ref{cirdiff}), for $a=0.0721$, $b=0.219$, $\sigma = 0.06665$, and $x_0=0.06$. Bottom: comparison between the exponent expansion (stars)
508: and the approximation of Ref.~\protect\cite{aitsahalia} (squares) for $\Delta t = 0.5$.
509: The uniform error of the Euler approximation is also
510: reported for comparison. Top: probability density function for $\Delta t = 1.5$ as a function of
511: $n$: $n=0$ (dotted), $n=1$ (long dashed), $n=2$ (short dashed), $n=3$ (continuous),
512: Euler (dot-long dashed), exact (crosses). The inset is an enlargement of the region of the maximum.
513: On this scale, the estimates for $n=2$ and $n=3$ still appear coincident. }
514: \label{denscir}
515: \end{figure}
516:
517:
518: \subsubsection{Constant Elasticity of Variance diffusion}
519:
520: As a last example we consider the Constant Elasticity of Variance model:
521: \begin{equation}
522: d Y_t = a(b-Y_t) dt + \sigma_y Y_t^p dW_t
523: \end{equation}
524: Here we consider for brevity only the case $p>1$ and
525: the transformation to a process with constant (unit)
526: variance is $X_t = \gamma(Y_t) = Y_t^{1-p}/\sigma_y(p-1)$
527: and gives, according to Eq.~(\ref{drift})
528: \begin{equation}
529: \mu_x (x) = \frac{c_1}{x}+ c_2 x + c_3 x^{\frac{p}{p-1}}~,
530: \end{equation}
531: with
532: $c_1 = p/2(p-1)$, $c_2 = a(p-1)$, and $c_3 = - ab(p-1)^{p/(p-1)}\sigma_y^{1/(p-1)}$.
533: In this case the effective potential reads
534: \begin{equation}
535: V_{\rm eff} (x) = \frac{1}{2} \mu_x(x)^2 - \frac{c_1}{2 x^2} + \frac{c_2}{2} +
536: \frac{p}{2(p-1)} c_3 x^{1/(p-1)} ~,
537: \end{equation}
538: and the first three terms of the expansion are:
539: \begin{eqnarray}
540: W_0(x,x_0) &=& c_1\log{\frac{y_0}{y_t}} - \frac{a(p-1)}{2(2p-1)}
541: \Big[ (2p-1)(y_t^2-y_0^2) \nonumber \\
542: &+& 2 b(p-1)^{\frac{p}{p-1}} \sigma_y^{\frac{1}{p-1}}
543: \Big(x^{\frac{2p-1}{p-1}} - x_0^{\frac{2p-1}{p-1}}\Big ) \Big]~, \\
544: W_1(x,x_0) &=& \frac{1}{2\Delta x} \Big[ F(x)-F(x_0) \Big]~,\\
545: W_2(x,x_0) & = & \frac{1}{2\Delta x^2}
546: \big[V_{\rm eff}(x)+V_{\rm eff}(x_0)-2W_1(x,x_0)\big]~,
547: \end{eqnarray}
548: with
549: \begin{eqnarray}
550: F(z) &=& -\frac{c_1^2}{z} + \frac{c_2^2}{3} z^3 + \frac{c_3^2(p-1)}{3p-1} z^{\frac{3p-1}{p-1}} \nonumber \\
551: &+& 2 c_1c_2 z + \frac{2 c_1 c_3 (p-1)}{p}z^{\frac{p}{p-1}}
552: + \frac{2 c_2 c_3(p-1)}{3p-2} z^{\frac{3p-2}{p-1}} + \mu_x(z)~.
553: \end{eqnarray}
554:
555: Similarly to the examples considered previously, also for the Constant Elasticity of Variance
556: model we find a very fast convergence of the exponent expansion for $\Delta t\simeq 1$, and a
557: performance generally similar to the one of the approach of Ref.~\cite{aitsahalia}, for values
558: of the volatility large enough.
559:
560:
561: \section{Pricing kernels for contingent claims}
562: \label{prickern}
563:
564: \subsection{Path integral formulation}
565: \label{pisec}
566: The exponent expansion can be generalized to obtain an approximation
567: of the pricing kernels of `standard' derivatives.
568: This can be done by formulating the pricing problem within Feynman's path integral
569: framework \cite{rosaclot}. Here we indicate as `standard' any contingent claim written
570: on the underlying, $Y_t$, whose value at time $t=0$, $V_0$, can be expressed as an expectation
571: value of a functional of a certain type, namely
572: \begin{equation}
573: V_0(\Delta t, y_0) = E\Big [P(Y_{\Delta t}) F[Y_u] \Big| y_0 \Big ]~,
574: \label{st1}
575: \end{equation}
576: where
577: \begin{equation}
578: F[Y_u] = \exp{\left[ - \int_0^{\Delta t} du \, V_F[Y_u] \right]}~,
579: \label{st2}
580: \end{equation}
581: and $P(Y_{\Delta t}) $ is a payout function.
582: Above $\Delta t$ is the time to expiry, and the expectation value
583: is performed with respect to the probability measure defined
584: by the diffusion for $Y_t$ that we assume of the form (\ref{diffusionY}).
585: European Vanilla options, zero coupon bonds, caplets, and floorets clearly belong to
586: this family of contingent claims. In addition, other path-dependent derivatives,
587: like barrier or Asian options can be expressed in this form (see e.g., Refs.~\cite{baaquie,lin}).
588:
589: Similarly to the case of the transition probability,
590: it is in general convenient to introduce an auxiliary diffusion
591: with constant volatility of the form (\ref{processX})
592: by means of the integral transformations (\ref{inttransf}).
593: Then, the expectation value in (\ref{st1}) can be transformed
594: in an integral over the distribution generated by
595: such auxiliary diffusion by means of the usual
596: Jacobian transformation (\ref{jacobian}). As a result, the value of the option
597: can be in general written as:
598: \begin{equation}
599: V_0(\Delta t, x_0) = E\Big [ P(X_{\Delta t}) F[X_u] \Big | x_0 \Big ] = \int_D dx P(x) K(x,\Delta t|x_0)~.
600: \label{integra}
601: \end{equation}
602: where $D$ is the domain of the auxiliary process as defined by
603: the relative stochastic differential equation (\ref{processX}), and $K(x, \Delta t|x_0)$
604: is the pricing kernel. The latter can be expressed in terms of
605: a path integral as it follows\cite{rosaclot}
606: \begin{equation}
607: K(x,\Delta t|x_0) = e^{-W_0 (x,x_0)} \Phi(x,\Delta t| x_0)
608: \label{kernel}
609: \end{equation}
610: with
611: \begin{equation}
612: \Phi(x,\Delta t| x_0) = \int_{x(0)=x_0}^{x(\Delta t)=x} \, {\cal D}[x(u)]
613: \exp{\left[ -\int_{0}^{\Delta t} du \, L_{\rm eff}[x(u),\dot x(u)] \right]}.
614: \label{pi}
615: \end{equation}
616: where $W_0$ is given by Eq.~(\ref{w0}) and
617: the functional $L_{\rm eff}[x(u),\dot x(u)]$ is the effective
618: {\em Euclidean Lagrangian}
619: \begin{equation}
620: L_{\rm eff}[x(u),\dot x(u)] = \frac{1}{2\sigma_x^2} \dot x(u)^2 + V_{\rm eff}(x)
621: \end{equation}
622: ($\dot x(u) \equiv dx(u)/du$) with the {\rm effective potential}, $V_{\rm eff}(x)$, defined as:
623: \begin{equation}
624: V_{\rm eff} (x) = \frac{1}{2\sigma_x^2} \mu_x(x)^2 + \frac{1}{2}\partial_x \mu_x(x) + V_F(x)~.
625: \label{effpot2}
626: \end{equation}
627: Finally, the measure ${\cal D}[x(u)]$ is defined by discretizing
628: each path $x(u)$ connecting $x(0)=x_0$ and $x(T)=x$. This can be done by
629: dividing the time interval $[0,T]$ into $P$ intervals so that
630: $x_n = x(u_n)$ ($u_n = nT/P$ with $n=0,...,P$), and by integrating the
631: internal $P-1$ variables $x_n$ over the domain $D$. The path integral
632: $\int {\cal D}[x(u)]$ is then obtained as the limit for
633: $P\to \infty$ of this procedure, namely
634: \begin{equation}
635: \int {\cal D}[x(u)] \equiv \lim_{P\to\infty} (2\pi \sigma_x^2 \Delta t)^{-P/2} \prod_{n=1}^{P-1} \int_D \, d x_n ~.
636: \end{equation}
637:
638: It is well known form the physical sciences that the path integral
639: $\Phi(x,\Delta t| r_0)$ satisfies the partial differential equation
640: Eq.~(\ref{schro}) \cite{pireference}. Note that this is consistent with the fact that,
641: by definition, for $V_F(x) \equiv 0$ the pricing kernel coincides
642: with the transition density of the underlying diffusion process for $X_t$.
643: In particular, as observed in Sec.~\ref{alternative}, one can use Eq.~(\ref{schro})
644: to derive the exponent expansion for $\Phi(x,\Delta x| r_0)$ using the trial form
645: (\ref{trial2}). As a result, the same expressions Eqs.(\ref{expansion1}-\ref{expansion3})
646: derived for the transition density hold true also
647: for the pricing kernel, provided that the effective potential (\ref{effpot2})
648: replaces the one in Eq.~(\ref{effpot}).
649:
650: \subsubsection{Correspondence with Quantum Mechanics}
651: \label{qm}
652:
653: It is interesting to note that the path integral formulation
654: of the pricing kernel (\ref{pi}) is mathematically
655: equivalent to the quantum mechanical description of the
656: thermodynamic properties of an ideal gas of particles moving
657: in the potential $\hbar V_{\rm eff}(x)$ ($\hbar$ is the reduced Planck's constant giving
658: the correct energy dimensions). The complete correspondence is
659: obtained by identifying $\sigma_x^2 \to \hbar/m$ and $\Delta t \to \hbar/k_B T$
660: where $m$ is the mass of the particle, $T$ is the temperature, and $k_B$ is the Boltzmann constant.
661: With this prescription, $\Phi(x,\Delta t| x_0)$ becomes the so-called
662: single particle density matrix, and the results of Makri and Miller \cite{makri} can
663: be readily recovered. In addition, it is straightforward to show using Eqs.~(\ref{limit})
664: that the exponent expansion of its diagonal elements, $\Phi(x_0,\Delta t| x_0)$, are consistent
665: with the so called Wigner expansion for the high-temperature limit.
666: Finally, performing the analytic continuation known as Wick
667: rotation $\Delta t \to i \hbar t$ allows one to obtain the single-particle quantum
668: propagator.
669: In this case (\ref{schro}) becomes the celebrated Schr\"odinger
670: equation.
671:
672: This correspondence provides an alternative justification of the choice of the exponential ansatz
673: in Eq.~(\ref{ansatz}). Indeed, this is the form in which can be expressed in general
674: the quantum mechanical propagator or the single particle density matrix \cite{pireference}.
675: Furthermore, it has been shown for the
676: quantum problem \cite{makri2} that the exponent expansion up to third order in $\Delta t$ and
677: first order in $\hbar/m$ can be derived starting from the short time semiclassical
678: propagator obtained \cite{schulman} through a saddle point analysis of the
679: limit $\hbar/m\to 0$. Indeed, it can be shown that the second order correction
680: $W_2(x,x_0)$, Eq.~(\ref{expansion2}), is the so-called van-Vleck determinant of the
681: saddle point expansion. This explains why the accuracy
682: of the present scheme is preserved in the corresponding regime of low volatility, as it was
683: anticipated in Sec.\ref{expa}, and illustrated in Sec.\ref{app1}.
684:
685: \begin{figure}[pt]
686: \vspace*{-1.5cm}
687: \centerline{\psfig{file=fig3.ps,width=4.0in}}
688: \vspace*{-1cm}
689: \caption{Accuracy of the exponent expansion for the pricing kernel Eq.~(\ref{kernel})
690: of the Vasicek model (\ref{vasicek}), for $a=0.0717$, $b=0.261$, and $x_0=0.1$.
691: Symbols as in Fig.~\ref{denscir}. The inset shows the maximum absolute error as a
692: function of the order of the expansion $n$. }
693:
694:
695: \label{prop}
696: \end{figure}
697:
698:
699: \subsubsection{An example: Zero Coupon Bond}
700: \label{app2}
701: In this Section we illustrate the prescriptions outlined above
702: by applying the exponent expansion to the calculation of a
703: zero coupon bond within the Vasicek and Cox-Ingersoll-Ross models.
704: The zero coupon bond is a financial instrument that provides at time $t=\Delta t$
705: a payout of one unit of a certain notional. Its value at time $t=0$ can
706: be expressed therefore as
707: \begin{equation}
708: P(0,\Delta t) = E\Big[ \exp{ - \int_0^{\Delta t} du \, X_u } \Big | r_0 \Big]
709: \end{equation}
710: which is of the standard form given by Eqs.~(\ref{st1}) and (\ref{st2}), with
711: $V_F[X_u] = X_u$, and $P[X_{\Delta t}] \equiv 1$.
712:
713: As a result, the exponent expansion for the kernel Eq.~(\ref{kernel})
714: can be easily derived giving for the Vasicek model
715: \begin{eqnarray}
716: W_1(x,x_0) &=& W_1^0(x,x_0) + \frac{x+x_0}{2}\\
717: W_2(x,x_0) &=& W_2^0(x,x_0) \\
718: W_3(x,x_0) &=& W_3^0(x,x_0) - \frac{\sigma_x^2 + 2 a^2 (x-b) }{24}
719: \end{eqnarray}
720: where $W_i^0(x,x_0)$ are the expressions obtained for the transition probability
721: Eqs.~(\ref{vas0})-(\ref{vas3}) of Sec.~\ref{vasicekapp}. For the Cox-Ingersoll-Ross model
722: instead we get:
723: \begin{eqnarray}
724: W_1(x,x_0) &=& W_1^0(x,x_0) + \frac{\sigma^2}{12}(x^2+x_0^2+xx_0) \\
725: W_2(x,x_0) &=& W_2^0(x,x_0) - \frac{\sigma^2}{24}
726: \end{eqnarray}
727: with $W_i^0(x,x_0)$ given by Eqs.~(\ref{cir0})-(\ref{cir2}), and
728: $W_3^0(x,x_0)$ related to the previous quantities as in Eq.~(\ref{cir3})
729: with
730: \begin{equation}
731: G(z)=G^0(z) + \frac{\sigma_y^2}{2}\left[ (\alpha +\frac{\sigma_y^2}{8})\frac{z^5}{5}
732: +\frac{\gamma}{3}z^3+ \beta z \right]~,
733: \end{equation}
734: $G^0(z)$ as in Eq.~(\ref{gz}), and $\partial_z V_{\rm eff}(z) = \partial_z V^0_{\rm eff}(z) + \sigma_y^2 z /2$.
735: The exponent expansion for the pricing kernel can be compared with the
736: exact results that can be shown to read for the Vasicek model
737: (\ref{vasicek})
738: \begin{eqnarray}
739: K_{\rm ex}(x,\Delta t|,x_0) &=& \frac{\exp{\left[(x-x_0)/a-\Delta t(b-\sigma^2/2a^2 )\right]}}{(2\pi\bar\sigma^2)^{1/2}}
740: \nonumber \\
741: &&\exp{\left[-\frac{\left[\left (x_0-b+\sigma^2/a^2 \right) e^ {-a \Delta t}-(y-b+\sigma^2/a^2 )\right]^2}
742: {2\bar\sigma^2}\right]}~,
743: \end{eqnarray}
744: with $\bar\sigma$ given by Eq.~(\ref{vasiceksig}), and
745: \begin{eqnarray}
746: &&K_{\rm ex}(x,\Delta t|,x_0) = \frac{2}{x}
747: \exp{\left[\frac{-\frac{a}{4}(x^2-x_0^2)+(2ab-\frac{\sigma^2}{2})\log{\frac{x}{x_0}}}{\sigma^2}\right]}
748: \frac{\gamma \sqrt{xx_0}\,e^{a^2b\Delta t/\sigma^2}}{2\sigma^2\sinh\left[\gamma\Delta t /2\right]}
749: \nonumber \\ &&
750: \exp{\left[-\frac{\gamma}{4\sigma^2}(x^2+x_0^2)\coth{\left[\gamma\Delta t/2\right]}\right]}
751: I_q\left(\frac{xx_0\gamma}{2\sigma^2\sinh{\left[\gamma\Delta t/2\right]}}\right)
752: \end{eqnarray}
753: with $\gamma =\sqrt{a^2+2\sigma^2}$, for the Cox-Ingersoll-Ross one.
754: As illustrated in Fig.\ref{prop}, similarly to the case of the
755: transition probability, the exponent expansion provides a remarkably good, and
756: fast converging approximation of the exact pricing kernel for financially sensible
757: parameterizations, and for a sizable value of the time step $\Delta t$.
758:
759:
760: \begin{figure}[pt]
761: \vspace{-1.6cm}
762: \centerline{\psfig{file=fig4.ps,width=4.0in}}
763: \vspace*{-1cm}
764: \caption{Zero Coupon Bond for the Cox-Ingersoll-Ross model. Parameters and symbols as in Fig.\protect\ref{denscir}.}
765: \label{zerof}
766: \end{figure}
767:
768:
769: Finally, the zero coupon bond can be obtained by numerical integration
770: of the pricing kernel according to Eq.~(\ref{integra}). The corresponding
771: results are shown in Fig.~\ref{zerof} confirming once more the quality of
772: the approximation. In a similar fashion, one can obtain systematic approximations
773: for caplets, floorets, and other simple interest rate derivatives whose value depends
774: only on the value of the instantaneous short rate at time $t = \Delta t$. It is also possible
775: to generalize this approach to path dependent contingent claims, like Asian options.
776:
777:
778: \section{Extending the time-step: path integral Monte Carlo methods}
779: \label{montecarlo}
780:
781: For an extended time interval $T$, the calculation of the transition density
782: or, more in general, of the pricing kernel (\ref{kernel}) can be
783: performed by discretizing the path integral (\ref{pi}), and taking the limit of large
784: number of time slices ($P\to \infty$) according to the standard Trotter product formula:
785: \begin{eqnarray}
786: & &K(x,\Delta t|x_0) \simeq (2\pi \sigma_x^2 \Delta t)^{-P/2} e^{-W_0 (x,x_0)}
787: \prod_{n=1}^{P-1} \int_D \, d x_n \nonumber \\
788: & \times &\exp{\left[-\frac{1}{2\sigma_x^2\Delta t}\sum_{k=1}^{k=P}(x_k-x_{k-1})^2 -
789: \frac{\Delta t}{2} \sum_{k=1}^{k=P} \left( V_{\rm eff}(x_k)+V_{\rm eff}(x_{k-1})\right) \right]}
790: \label{finiteT}
791: \end{eqnarray}
792: with $\Delta t = T/P$, $x_P=x$, and the effective potential given by Eq.~(\ref{effpot2}).
793: It is worth remarking that, for the case of the transition density,
794: by interpreting the latter equation as the Chapman-Kolmogorov property of Markov processes \cite{shreve} one
795: obtains the following approximation of the short-time propagator
796: \begin{equation}
797: K_{\rm Trotter}(x,\Delta t|x_0) = \frac{e^{-W_0 (x,x_0)}}{\sqrt{2\pi\sigma_x^2\Delta t}}
798: \exp {\left[ -\frac{(x-x_0)^2}{2\sigma_x^2 \Delta t}- \frac{\Delta t}{2}
799: \Big( V_{\rm eff}(x)+V_{\rm eff}(x_{0})\Big) \right]}~.
800: \end{equation}
801: However, in contrast to the $n=1$ exponent expansion, the latter expression is
802: not strictly correct up to order $\Delta t$, and only in the limit $P\to \infty$
803: the difference becomes negligible.
804:
805: In general, to obtain an accurate result for the pricing kernel
806: for an extended time period $T$ one has to increase the number of time slices,
807: or Trotter number $P$, until convergence is achieved. By replacing the Trotter formula
808: with the improved short-time kernel obtained through the exponent expansion (\ref{ansatz})
809: one achieves a faster convergence with the Trotter number, thus significantly
810: reducing the computational burden. In this case the finite-time expression of the
811: pricing kernel reads
812: \begin{eqnarray}
813: & &K(x,T|x_0) \simeq (2\pi \sigma_x^2 \Delta t)^{-P/2}
814: \prod_{n=1}^{P-1} \int_D \, d x_n \nonumber \\
815: & \times &\exp{\left[-\frac{1}{2\sigma_x^2\Delta t}\sum_{k=1}^{P} (x_k-x_{k-1})^2 -
816: \sum_{k=1}^{P} W(x_k,x_{k-1},\Delta t) \right]}
817: \label{finiteT2}
818: \end{eqnarray}
819: with $W(x,x_0,\Delta t)$ given by Eq.~(\ref{w}).
820:
821: Equation (\ref{finiteT2}) allows one to obtain the transition density
822: or the pricing kernel for a standard derivative through the calculation of a multidimensional
823: integral over the variables $x_1, \dots, x_{P-1}$. The latter integration is ideally
824: suited for Monte Carlo methods either in real, or in Fourier space \cite{quasifourier}, the specific choice
825: depending on the particular problem at hand.
826: In addition, importance-sampling schemes, e.g., by means of the Metropolis algorithm \cite{metropolis}, can be easily applied in order
827: to reduce the computation time. However, in order not to introduce a systematic bias in the result
828: a particular attention has to be paid in order to sample accurately the configuration space.
829:
830: The most straightforward way to perform a Monte Carlo quadrature of Eq.~(\ref{finiteT2}),
831: is to realize that a simple Markov chain ${\bf x} = (x_1,\dots,x_{P-1})$
832: \begin{equation}
833: x_n = x_{n-1} + \sigma_x \sqrt{\Delta t} \, Z_n~,
834: \label{markov}
835: \end{equation}
836: with $Z_n$, sampled from a standard normal distribution, generates an ensemble of walkers distributed according to
837: \begin{equation}
838: p(x_1,\dots,x_{P-1}|x_0) = (2\pi \sigma_x^2 \Delta t)^{-(P-1)/2} \exp{\left[ -\frac{1}{2\sigma_x^2\Delta t}\sum_{k=1}^{P-1}(x_k-x_{k-1})^2\right] }~.
839: \end{equation}
840: As a result, the pricing kernel (\ref{finiteT2}) can be obtained as the average over
841: the random paths generated according to Eq.~(\ref{markov}) of the following estimator:
842: \begin{equation}
843: {\cal O}({\bf x}, x_P =x ) = (2 \pi \sigma_x^2 \Delta t)^{-1/2} \exp{ \left[-\frac{1}{2\sigma_x^2\Delta t} (x-x_{P-1})^2 - \sum_{k=1}^{P} W(x_k,x_{k-1}) \right]}~.
844: \end{equation}
845: A remarkable property of the path integral approach is that $K(x,\Delta t|x_0)$
846: for any final point $x$ can be evaluated with a single Monte Carlo
847: simulation by appropriately reweighting the accumulated estimator. In fact the distribution
848: of walkers $p(x_1,\dots,x_{P-1}|x_0)$ is independent of the final point $x_P$ so that
849: $K(x^\prime,\Delta t|x_0)$ can be calculated by averaging ${\cal O}({\bf x}, x_P = x^\prime)$. In addition,
850: the latter quantity can be efficiently obtained by means of the following reweighting procedure
851: \begin{equation}
852: {\cal O}({\bf x}, x_P = x^\prime) = {\cal O} ({\bf x}, x_P =x ) \frac{{\cal W}(x^\prime,{\bf x})}{ {\cal W}(x,{\bf x})}
853: \end{equation}
854: with
855: \begin{equation}
856: {\cal W}(x,{\bf x}) = \exp{\left[ -\frac{1}{2\sigma_x^2\Delta t} (x-x_{P-1})^2 -W(x,x_{P-1}) \right]}~.
857: \end{equation}
858:
859:
860: Expectation values of the form (\ref{integra}) on a time horizon $T$ can be calculated
861: by integrating over the final variable giving:
862: \begin{eqnarray}
863: V_0(T , x_0) &=& \int_D dx P(x) K(x,T|x_0) \simeq
864: (2\pi \sigma_x^2 \Delta t)^{-P/2} \prod_{n=1}^{P} \int_D \, d x_n P(x_P) \nonumber \\ & & \exp{\left[-\frac{1}{2\sigma_x^2\Delta t}\sum_{k=1}^{P} (x_k-x_{k-1})^2 -
865: \sum_{k=1}^{P} W(x_k,x_{k-1}) \right]}~.
866: \end{eqnarray}
867: This can be simulated by extending the Markov chain (\ref{markov}) up to step $x = x_P$, and
868: accumulating the estimator
869: \begin{equation}
870: {\cal P}({\bf x}, x_P = x^\prime) = P(x) \exp{\left[- \sum_{k=1}^{P} W(x_k,x_{k-1}) \right]}~.
871: \end{equation}
872: Remarkably, within the path integral approach, the sensitivities of such
873: expectation values with respect to the model parameters (the so-called Greeks)
874: can be computed in the same Monte Carlo simulation, thus
875: avoiding any numerical differentiation.
876: Indeed, indicating with $\theta$ a generic parameter,
877: under quite general conditions \cite{glassermann}, one has
878: \begin{equation}
879: \partial_\theta V_0(T,x_0,\theta)= \int_D dx \left[ K_\theta(x,T|x_0) \partial_\theta P(x,\theta) + P(x,\theta)\partial_\theta {K_\theta(x,T|x_0)}\right]~.
880: \end{equation}
881: As a result the sensitivity
882: $\partial_\theta V_0(T,x_0,\theta)$ can be calculated by means of the estimator:
883: \begin{equation}
884: {\cal G}({\bf x}, x_P = x^\prime) = \exp{\left[- \sum_{k=1}^{P} W(x_k,x_{k-1})\right]}
885: \left (\partial_\theta P + P \partial_\theta \log{K_\theta} \right)~.
886: \end{equation}
887: Higher order sensitivities can be obtained in a similar fashion.
888:
889: The convergence with the Trotter number $P$ of the path integral Monte Carlo estimates
890: is illustrated in Fig.~\ref{pimc} for the calculation of the first five moments
891: of the $T=40\,yrs$ transition probability of the Cox-Ingersoll-Ross model (\ref{cirdiff}).
892: The finite $P$ estimates converge very rapidly with $1/P$. In particular, for the case considered,
893: $P=20$ already provides estimates in agreement with the exact result
894: within statistical uncertainties. In general, as also
895: shown in the figure, a convenient indicator of the convergence is the zeroth-moment or normalization of
896: the distribution. The calculation of this quantity allows in general to assess the level
897: of convergence without performing a complete scaling with $P$, thus saving computational time.
898:
899: \begin{figure}[pt]
900: %\vspace{-1.6cm}
901: \centerline{\psfig{file=fig5.ps,width=4.0in}}
902: %\vspace*{-1cm}
903: \caption{Convergence with the Trotter number $P$
904: of the normalization, and of the first 4 moments
905: of the 40 years transition density for the CIR model.
906: Parameters as in Fig.\protect\ref{denscir}. In the first top panels the
907: relative error with respect to the exact result is reported. Lines are quadratic fits.}
908: \label{pimc}
909: \end{figure}
910:
911:
912: \section{Conclusions}
913: \label{conclusion}
914:
915: The exponent expansion is an approximation of the quantum mechanical
916: propagator known in physical chemistry \cite{makri,makri2}.
917: We have generalized this approach to produce closed-form approximation
918: of the transition probability of arbitrary non linear processes,
919: and we have shown that it produces very accurate results for integrable diffusions
920: of financial interest, like the Vasicek and the Cox-Ingersoll-Ross models. In contrast to
921: previously developed approximation schemes that share a similar rationale \cite{aitsahalia},
922: the exponent expansion always generates positive definite transition probabilities, and
923: remains stable also in the limit of low volatility.
924:
925: By introducing a path integral
926: framework we have generalized the exponent expansion to the calculation of
927: pricing kernels of financial derivatives, and we have shown how to obtain
928: accurate approximations for the price of simple contingent
929: claims. We have also shown how the exponent expansion can be naturally used
930: to increase the efficiency of path integral Monte Carlo simulations. The latter
931: allow one to extend the calculations to arbitrarily large time steps,
932: and to efficiently calculate the Greeks of contingent claims,
933: avoiding any numerical differentiation. A systematic study of the efficiency
934: of this approach for the pricing of exotic derivatives will be the object of
935: future investigations.
936:
937: The exponent expansion can be generalized to multifactor
938: diffusion processes, and to time dependent drift and volatility
939: functions. This would allow one to increase the efficiency of numerical
940: simulations of a larger family of diffusion processes relevant
941: for Financial applications, including local volatility or Libor Market Models.
942: Work is currently in progress along this line of research.
943:
944: {\bf Acknowledgments} It is a pleasure to thank Gabriele Cipriani for sparking my interest
945: in the path integral approach to option pricing, and for continuous stimulating discussions.
946: Very useful comments from the anonymous referee are also gratefully acknowledged.
947:
948: \begin{thebibliography}{00}
949:
950: \bibitem{bachelier} L. Bachelier, {\em Theorie de la Sp\'eculation}, Annales de l'Ecole Normale Sup\'erieure (1900).
951: \bibitem{bs} F. Black and M. Scholes, {\em Journal of Political Economy} {\bf 81}, 637 (1973);
952: R. C Merton, {\em Bell Journal of Economics} {\bf 4}, 141 (1973).
953: \bibitem{cir} J.C. Cox, J.E. Ingersoll, and S.A. Ross, {\it Econometrica} {\bf 53}, 385 (1985).
954: \bibitem{vasicek} O. Vasicek, {\it Journal of Financial Economics} {\bf 5}, 177 (1977).
955: \bibitem{makri} N. Makri and W.H. Miller, {\it J. Chem. Phys.} {\bf 90}, 904 (1989).
956: \bibitem{rosaclot} M. Bennati, M. Rosa-Clot, and S. Taddei, {\it Int. J. of Theor. and App. Finance} {\bf 2}, 381 (1999); M. Rosa-Clot, and S. Taddei, {\it Int. J. of Theor. and App. Finance} {\bf 5}, 123 (2002).
957: \bibitem{shreve} S. E. Shreve, {\em Stochastic Calculus for Finance} (Springer-Verlag, New York, 2004).
958: \bibitem{pireference} R. P. Feynman, {\em Rev. Mod. Phys.} {\bf 20}, 367 (1948). See also: R. P. Feynman and
959: A. R. Hibbs, {\em Quantum Mechanics} (MCGraw-Hill, New York, 1965).
960: \bibitem{montagna} G. Montagna, O. Nicrosini, and N. Moreni, {\it Physica A} {\bf 310}, 450 (2002); G. Bormetti, G. Montagna, N. Moreni, and N. Nicrosini, e.print arXiv:cond-mat/0407231 (2004).
961: \bibitem{baaquiep} B.E. Baaquie, C. Corian\`o, and M. Srikant, {\it Physica A} {\bf 334}, 531 (2004).
962: \bibitem{matacz} A. Matacz, e.print arXiv:cond-mat/0005319 (2000).
963: \bibitem{dash} J. W. Dash, {\em Quantitative Finance and Risk Menagement: a Physicist's approach} (World Scientific, Singapore, 2004).
964: \bibitem{aitsahalia} Y. A\"{\i}t Sahalia, {\it Journal of Finance} {\bf 54}, 1361 (1999).
965: \bibitem{cev} J. C. Cox and S. A. Ross, {\em Journal of Financial Economics} {\bf 3}, 145-166 (1976).
966: \bibitem{feller} W. Feller, {\it Annals of Mathematics} {\bf 55}, 468 (1952).
967: %\bibitem{jasmidian} F. Jamshidian, {\em Journal of Finance} {\bf 44}, 205 (1989).
968: \bibitem{abramovitz} M. Abramovitz, I. A. Stegun, {\em Handbook of Mathematical Functions} (National Bureau of Standars, Applied Mathematics Series, 1965).
969: \bibitem{makri2} N. Makri and W.H. Miller, {\it Chem. Phys. Lett} {\bf 151}, 1 (1989).
970: \bibitem{baaquie} B. E. Baaquie, {\em Quantum Finance} (Cambridge University Press, 2004).
971: \bibitem{lin} V. Linetsky, {\it Computational Economics} {\bf 11}, 129 (1998).
972: \bibitem{schulman} L.S. Schulman, {\em Techniques and applications of path integration} (Wiley, New York, 1981).
973: \bibitem{quasifourier} R.D. Coalson, {\it J. Chem. Phys.} {\bf 85}, 926 (1986).
974: \bibitem{metropolis} N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, H. Teller, and E. Teller,
975: {\it J. Chem. Phys.} {\bf 21}, 1087 (1953).
976: \bibitem{glassermann} P. Glasserman, {\em Monte Carlo Methods in Financial Engineering} (Springer-Verlag,
977: New York, 2004).
978: \end{thebibliography}
979:
980: \end{document}
981:
982:
983:
984:
985: