1: %\documentstyle[prl,twocolumn,aps]{revtex}
2: \documentstyle[preprint,aps]{revtex}
3: \tightenlines
4: \begin{document}
5: \title{Financial Market Dynamics}
6: \author{Fredrick Michael and M.D. Johnson\cite{byline}}
7: \address{Department of Physics, University of Central Florida, Orlando,
8: FL 32816-2385}
9: \date{May 23, 2001}
10: %\date{\today}
11: \maketitle
12: \begin{abstract}
13: Distributions derived from non-extensive Tsallis statistics are
14: closely connected with dynamics described by a nonlinear
15: Fokker-Planck equation. The combination shows promise in
16: describing stochastic processes with power-law distributions and
17: superdiffusive dynamics. We investigate intra-day price changes in
18: the S\&P500 stock index within this framework by direct analysis
19: and by simulation. We find that
20: the power-law tails of the distributions, and the index's
21: anomalously diffusing dynamics, are very accurately described
22: by this approach. Our results show good
23: agreement between market data, Fokker-Planck dynamics, and simulation.
24: Thus the combination of the Tsallis non-extensive entropy and the
25: nonlinear Fokker-Planck equation unites in a very natural way the
26: power-law tails of the distributions and their superdiffusive dynamics.
27: \end{abstract}
28: \pacs{89.65.Gh,05.10.Gg,05.20.-y,05.40.Fb}
29: %\narrowtext
30:
31: \section{Introduction}
32: \label{intro}
33:
34: In anomalously diffusing systems, a mean-square displacement
35: scales with time according to a power law, $t^{\alpha}$,
36: with $\alpha>1$ (superdiffusion) or $\alpha<1$ (subdiffusion).
37: (The case $\alpha=1$ corresponds to normal diffusion.)
38: Anomalous diffusion has been observed in systems as widely varied as
39: plasma flow\cite{plasma}, surface growth\cite{surface}, and
40: financial markets\cite{econophysics}.
41: A general framework for treating superdiffusive systems is provided
42: by the nonlinear Fokker-Planck equation, which is associated
43: with an underlying Ito-Langevin
44: process\cite{plastino,tsallis-bukman,borland}.
45: This in turn has a very interesting connection to the
46: nonextensive entropy proposed by Tsallis\cite{tsallis}:
47: the nonlinear Fokker-Planck equation is solved by time-dependent
48: distributions which maximize the Tsallis entropy\cite{tsallis-bukman,zanette}.
49: This unexpected connection between thermostatistics and anomalous
50: diffusion gives an entirely new way to approach the study of
51: dynamics of anomalously diffusing systems. In this paper we
52: use this viewpoint to address the dynamics of financial markets.
53:
54: Several financial markets' indices as well as their member stocks are
55: characterized by price changes whose variances have been shown to undergo
56: anomalous (super) diffusion under time
57: evolution\cite{econophysics,dacrrogna,mantegna1,mantegna}.
58: Moreover, the probability distributions of the price changes have power law
59: tails\cite{econophysics,mandlelbrot}. An open long-term question is how best to
60: describe these distributions and their time evolution.
61: Earlier work has shown that the power-law tails can be described
62: by a distribution which maximizes the Tsallis entropy\cite{ramos,friedrich}.
63: The connection mentioned above then suggests that the market dynamics
64: might be controlled by a nonlinear Fokker-Planck equation.
65: In this description the power-law tails
66: and the anomalous diffusion arise together quite naturally.
67: Here we show that this approach seems to accurately
68: describe high-frequency
69: intra-day price changes for the S\&P500 index. It may also be suitable
70: to describe the superdiffusion and power law behaviors
71: observed in a broad range of markets and exchanges\cite{econophysics}.
72:
73: The description we will use was developed in the general context of
74: anomalously diffusing systems by Tsallis and Bukman\cite{tsallis-bukman}
75: and Zanette\cite{zanette}.
76: Here we briefly summarize their results, which
77: are based on a maximization of entropy subject to certain constraints.
78: The central ingredient is a time-dependent probability
79: distribution $P(x,t)$ of a stochastic variable $x$. The
80: point of departure from normal maximum entropy approaches is the feature that
81: the entropy used is the non-extensive Tsallis entropy,
82: \begin{equation}
83: S_{q}=-{\frac{1}{1-q}}\left( 1-\int P(x,t)^{q}\,dx\right) .\label{sq}%
84: \end{equation}
85: The Tsallis parameter $q$ characterizes the
86: non-extensivity of the entropy.
87: In the limit $q\rightarrow1$ the entropy becomes the usual
88: logarithmic expression S=-$\int P\ln P$.
89:
90: Associated with the non-extensive entropy is the use of the constraints
91: \begin{eqnarray}
92: && \int P(x,t)\,dx=1,\label{norm}\\
93: && \langle x-\bar{x}(t)\rangle_{q}\equiv\int[x-\bar{x}(t)]P(x,t)^{q}%
94: \,dx=0,\label{mean}\\
95: && \langle(x-\bar{x}(t))^{2}\rangle_{q}\equiv\int[x-\bar{x}(t)]^{2}%
96: P(x,t)^{q}\,dx=\sigma_{q}(t)^{2}.\label{constraints}%
97: \end{eqnarray}
98: The first of these is simply the normalization of the probability. However, in
99: the latter two equations the probability distribution function is raised to
100: the power $q$. Unless $q=1$ these are not the usual constraints leading to the
101: mean and variance, and $\sigma_{q}^2$ (the `$q$-variance') is not the ordinary
102: variance. Notice that the Tsallis nonextensivity
103: parameter $q$ is independent of time.
104:
105: Maximizing the Tsallis entropy subject to these constraints, for fixed $q$,
106: yields\cite{tsallis-bukman}
107: \begin{equation}
108: P(x,t)={\frac{1}{Z(t)}}\left\{ 1+\beta(t)(q-1)[x-\bar{x}(t)]^{2}\right\}
109: ^{-\frac{1}{q-1}}.\label{pdf}
110: \end{equation}
111: Here parameters $Z$ (a normalization constant) and $\beta$ are
112: Lagrange multipliers associated with the first and third constraints,
113: and are given by
114: \begin{eqnarray}
115: Z(t) &=& {\frac{B\left( {\frac{1}{2}},{\frac{1}{q-1}}-{\frac{1}{2}}\right)
116: }{\sqrt{(q-1)\beta(t)}}},\label{znorm}\\
117: \beta(t) &=& {\frac{1}{2\sigma_{q}(t)^{2} Z(t)^{q-1}}},\label{betaprime}
118: \end{eqnarray}
119: where $B(x,y)=\Gamma(x)\Gamma(y)/\Gamma(x+y)$ is Euler's Beta function.
120:
121: The ordinary variance of the distribution Eq.~(\ref{pdf}) is
122: \begin{equation}
123: %\sigma^{2}(t)=\int_{-\infty}^{\infty}(x-\bar{x}(t))^{2}P(x,t)dx=\left\{
124: \sigma^{2}(t)=\langle(x-\bar{x}(t))^{2}\rangle_1=\left\{
125: \begin{array}
126: [c]{cl}%
127: \frac{1}{(5-3q)\beta(t)} & ,\quad q<\frac{5}{3}\\
128: \infty & ,\quad q\geq\frac{5}{3}.
129: \end{array}
130: \right. \label{variance}%
131: \end{equation}
132: Hence for applications to data of finite variance, the Tsallis parameter $q$
133: must lie within the range
134: \begin{equation}
135: 1\leq q<5/3.
136: \label{qrange}
137: \end{equation}
138: The Tsallis function Eq.~(\ref{pdf}) can be viewed as a {\it least biased}
139: probability distribution compatible with observed stochastic data with a
140: certain mean $\bar{x}(t)$ and variance $\sigma^2(t)$ (but with a
141: particular choice of $q$).
142:
143: An important property of the probability distribution function Eq.~(\ref{pdf})
144: is that, with appropriate
145: time-dependent parameters, it is the solution of a time evolution equation
146: which leads naturally to anomalous diffusion\cite{tsallis-bukman,zanette}.
147: Consider the nonlinear Fokker-Planck equation
148: \begin{equation}
149: {\frac{\partial P(x,t)^{\mu}}{\partial t}}=-{\frac{\partial}{\partial x}%
150: }\left[ F(x)P(x,t)^{\mu}\right] +\frac{D}{2}{\frac{\partial^{2}P(x,t)^{\nu}%
151: }{\partial x^{2}}},\label{fokker}%
152: \end{equation}
153: where $F(x)$ is a linear drift force, $F(x)=a-bx$. One can
154: show\cite{tsallis-bukman,zanette} that a
155: function of the form Eq.~(\ref{pdf}) solves this
156: as long as
157: \begin{equation}
158: q=1+\mu-\nu
159: \label{q}
160: \end{equation}
161: and the time dependences of the parameters are given by
162: \begin{eqnarray}
163: -{\frac{\mu}{\mu+\nu}}{\frac{dZ^{\mu+\nu}(t)}{dt}}&+&2\nu D\beta(t_0)Z(t_0)^{2\mu
164: }-bZ^{\mu+\nu}=0,\label{z}\\
165: {\frac{\beta(t)}{\beta(t_0)}}&=&\left( {\frac{Z(t_0)}{Z(t)}}\right)^{2\mu},\label{beta}\\
166: {\frac{d\bar{x}}{dt}}&=&a-b\bar x.\label{xbar}%
167: \end{eqnarray}
168: By comparing
169: Eqs.~(\ref{znorm}) and (\ref{beta}) one easily shows that the nonlinear
170: Fokker-Planck equation preserves the norm ($\int P(x,t) dx$)
171: only if $\mu=1$. Hereafter we specialize to this case.
172: Then Eqs.~(\ref{q},\ref{z},\ref{beta}) give
173: \widetext
174: \begin{equation}
175: \beta(t)^{-{3-q\over2}} = \beta(t_0)^{-{3-q\over2}}e^{-b(3-q)(t-t_0)}
176: -2Db^{-1}(2-q)
177: \left[ \beta(t_0)Z^2(t_0) \right]^{q-1\over2}
178: \left( e^{-b(3-q)(t-t_0)}-1 \right). \label{betat}
179: \end{equation}
180: $Z(t)$ is then obtained from Eq.~(\ref{znorm}).
181: \narrowtext
182:
183: \section{Application to a Market Index}
184:
185: It is apparent from the results sketched in the previous section that the
186: Tsallis probability distribution function $P(x,t)$ has properties which
187: make it a good candidate for describing the anomalous diffusion of
188: financial market indices, member stocks, and currency exchanges.
189: First, in the large
190: $x$ limit $P(x,t)$ becomes a power law distribution,
191: \begin{equation}
192: P(x,t)\sim x^{-{\frac{2}{q-1}}}.\label{powerlaw}%
193: \end{equation}
194: This is in keeping with the observed power law tails in market distributions
195: and, indeed, certain market distributions have been well fit by
196: a distribution of the Tsallis form\cite{ramos,friedrich}. In fact this fit to
197: market data has a longer history --- the Tsallis function
198: is an extension to continuous values of $q$ of
199: the student-$t$ distribution, which has been known for
200: some time to provide a good fit to certain market data\cite{studentt,econophysics}.
201:
202: What has not been widely recognized in the finance-related applications
203: of statistics is that the Tsallis distribution solves
204: the nonlinear Fokker-Planck equation. Consequently
205: it may provide a framework for understanding the dynamics of certain market
206: data, including anomalous diffusion. It is this possibility that we will
207: now discuss.
208:
209: For application to market indices the Tsallis parameter $q$ must
210: lie within the range Eq.~(\ref{qrange}),
211: ensuring that the regular variance remains finite and evolves for
212: moderately long times as
213: \begin{equation}
214: \sigma^{2}(t)\sim1/\beta(t)\sim t^{{\frac{2}{3-q}}}\label{sigmat}
215: \end{equation}
216: [using Eq.~(\ref{variance},\ref{betat}) in the case $b\ll 1$].
217: Thus the Fokker-Planck equation can describe superdiffusive processes
218: (with $1<q<5/3$).
219:
220: Here we investigate one test case, the S\&P500 stock market index,
221: using fairly high-frequency 1-minute-interval data collected from
222: July 2000 to January 2001.
223: This data consists of
224: a set of prices (index values) $p(\tau)$ at discrete times $\tau$
225: ($\tau=1,2,\dots,50000$ trading minutes) obtained from the Terra-Lycos
226: QCharts server\cite{terra}. (The time period chosen has no special
227: significance.)
228: The quantity amenable to a stochastic analysis of the Ito-Langevin
229: type is the price change during various time
230: intervals\cite{gardiner,friedrich,borland,stariolo}.
231: To agree with the notation used in the previous section we define
232: $x$ to be a price change during a time interval $t$.
233: For each fixed time interval $t$ we generate a
234: sequence of non-overlapping price changes $\{x_1,x_2,\dots\}$,
235: with $x_j=p(jt+1)-p((j-1)t+1)$.
236: We view the data thus generated for each fixed time interval $t$ as a sample
237: selected from a population with some distribution $P_{market}(x,t)$.
238: Thus in this application the presumably stochastic variable $x$
239: is a price change, and the `time' variable $t$ is really the
240: corresponding time interval.
241:
242: Anomalous diffusion occurs in random systems with correlations in time.
243: Because the price change during a longer interval is a sum of price
244: changes during shorter intervals, one has
245: \begin{equation}
246: P(x,t_1+t_2) = \int dx_1 P(x-x_1,t_2|x_1,t_1) P(x_1,t_1).
247: \end{equation}
248: where $P(\cdot|\cdot)$ is a conditional probability.
249: In the absence of correlation, price changes in different intervals
250: are independent, so that
251: $P(x-x_1,t_2|x_1,t_1)=P(x-x_1,t_2)$. In this case it is easy
252: to see that the variance grows linearly with time.
253: Anomalous diffusion thus means that price changes during successive
254: time intervals are not independent --- naturally enough, since
255: traders respond to earlier changes. The observed superdiffusion of
256: financial markets thus indicates
257: correlation, and consequently a nontrivial time dependence.
258: Autocorrelation analyses of the S\&P500 market have found
259: strong correlations for times on the order of a few minutes,
260: with weak persistent correlations at longer
261: times\cite{mantegna1,mantegna,econophysics}.
262: We find the same for our data set.
263: Consequently in this work we will concentrate on the intra-day market
264: dynamics of intervals less than one hour (1 min $\le t \le$ 60 min).
265:
266: If the dynamics are of Fokker-Planck form, then there
267: must be a consistency between the distribution $P_{market}(x,t)$
268: at a certain time $t$ and the way the distributions evolve in time.
269: One would have to find that the data at different
270: times can be fit by distributions of Tsallis form, with
271: a Tsallis parameter $q$ independent of time. If so, then
272: the variance will evolve in time according to
273: Eq.~(\ref{variance},\ref{betat}) with the same $q$ and appropriate
274: values for $D,b$.
275:
276: We investigate the degree to which the S\&P500 price changes at different time
277: intervals $t$ can be described by a Tsallis distribution
278: evolving according to a Fokker-Planck equation.
279: The approach is to bin the data and perform a direct nonlinear
280: $\chi^2$ fit of Eq.~(\ref{pdf}).
281: The `initial' distribution corresponds to the shortest value of $t$,
282: here $t_0=1$ minute. The $t_0=1\min$ price changes and a fit of Tsallis form
283: are shown in Fig.~\ref{fit}. The fit parameters are
284: $q=1.64\pm0.02$ and $\beta(t_0)=4.90\pm0.11$,
285: and consequently [from Eq.~(\ref{znorm})]
286: $Z(t_0)=1.09\pm0.02$.
287: We then fit the data at different time intervals $t$, with
288: the Tsallis parameter $q=1.64$ fixed and $\beta(t)$ determined by
289: the fit. It is important to this analysis that the data at all
290: times continues to be well fit by the Tsallis distribution, and indeed
291: this is what we find. (See Fig.~\ref{fit}.)
292:
293: The resulting inverse variance $\beta(t)$ extracted from the fits
294: is shown in Fig.~\ref{betafig}. Time evolution controlled by the
295: Fokker-Planck equation predicts the time dependence for $\beta(t)$
296: given in Eq.~(\ref{betat}). Accordingly we fit this form to the
297: extracted values, finding $D=0.217 \pm 0.003$ and $b=0.047\pm0.004$.
298: The Fokker-Planck form for $\beta(t)$ is also shown in Fig.~(\ref{betafig}).
299:
300: It is clear from Fig.~\ref{betafig} that the agreement is quite good, indicating
301: time evolution of nonlinear Fokker-Planck form.
302:
303: As a check on our analysis we also generated simulated data
304: using a Tsallis distribution Eq.~(\ref{pdf}) evolving in time as described in
305: Eqs.~(\ref{z},\ref{betat}). We then performed the same analysis that
306: we performed on the S\&P500 data. Briefly, the approach is as follows.
307: We use the parameters $q,D,b$ and initial values $\beta(t_0),Z(t_0)$
308: obtained from
309: the S\&P500 data. Parameters $\beta(t),Z(t)$ at
310: later times are obtained from Eq.~(\ref{betat},\ref{znorm}).
311: At several times $t$
312: we generate a set of random numbers $x$ with the Tsallis distribution
313: Eq.~(\ref{pdf}). This can be done by a transformation from a
314: uniformly distributed set of random numbers\cite{press}, as follows.
315: Consider a uniform distribution
316: $P_{\mathrm uniform}(y)$ which is unity for $0\le y\le 1$ and zero otherwise.
317: By choosing an appropriate function $y(x)$ we can obtain the desired
318: distribution from the uniform distribution in the form
319: \begin{equation}
320: P(x,t)=P_{\mathrm uniform}(y)\left| \frac{dy}{dx}\right|.\label{xform}
321: \end{equation}
322: This can be solved for $y$ as a hypergeometric function:
323: \begin{equation}
324: y=\int_{0}^{x}P(x^{\prime},t)dx^{\prime}=
325: x\,\cdot\, {}_{2}F_{1}\left[\frac{1}{2},\frac{1}{q-1},\frac{3}
326: {2},-(q-1)\beta(t) x^{2}\right].
327: \end{equation}
328: For each $y$ generated uniformly, we numerically invert this to find $x$.
329: The resulting $x$'s are distributed according to the Tsallis form,
330: Eq.~(\ref{pdf}).
331: Thus the simulated price-change data is guaranteed to have a
332: Tsallis distribution evolving in time according to the Fokker-Planck
333: equation. We then analyzed this simulated data exactly as we did the
334: financial data as a check. Fig.~\ref{betafig} shows that simulated data
335: closely tracks the actual market data.
336:
337: \section{Conclusions}
338:
339: The nonlinear Fokker-Planck equation for a probability distribution
340: $P(x,t)$ has a time-dependent solution
341: equivalent to the distribution obtained by maximization
342: of the Tsallis nonextensive entropy. Consequently
343: if a stochastic process has a Tsallis distribution it is natural to
344: attempt to work backwards, and ask whether the underlying
345: dynamics are Fokker-Planck.
346: We have investigated this possibility in one market, analyzing the
347: high-frequency intra-day dynamics of the S\&P500 index
348: within this framework.
349: We found that the distribution
350: $P(x,t)$ is well fit by the Tsallis form at all times, and evolves in time
351: according to the Fokker-Planck equation. Thus the combination
352: of the Tsallis distribution and the Fokker-Planck equation
353: unites in a very natural way the power-law tails of the distributions and their
354: superdiffusive dynamics.
355:
356: It would be of interest to apply this analysis to longer time windows,
357: inter-day (long term) dynamics, and also to other markets.
358:
359: We acknowledge support from the NSF through grant DMR99-72683.
360:
361: \bigskip
362:
363: \begin{thebibliography}{99}
364:
365: \bibitem[*]{byline} Email address: mjohnson@ucf.edu
366:
367: \bibitem{plasma} J.G. Berryman, J. Math. Phys. {\bf 18}, 2108 (1977).
368:
369: \bibitem{surface} H. Spohn, J. Phys. I {\bf 3}, 69 (1993).
370:
371: \bibitem{econophysics}R.N. Mantegna and H.E. Stanley, {\it An Introduction to
372: Econophysics; Correlations and Complexity in Finance}, Cambridge University
373: Press, Cambridge (2000).
374:
375: \bibitem{plastino} A.R. Plastino, A. Plastino, Physica A {\bf 222}, 347 (1995).
376:
377: \bibitem{tsallis-bukman} C. Tsallis and D.J. Bukman, Phys. Rev. E {\bf 54},
378: 2197 (1996).
379:
380: \bibitem{borland} L. Borland, Phys. Rev. E {\bf 57}, 6634 (1998).
381:
382: \bibitem{tsallis} C. Tsallis, J. Stat. Phys. {\bf 52}, 479 (1989).
383:
384: \bibitem{zanette} D.H. Zanette, Los Alamos archive cond-mat/9905064 (1999).
385:
386: \bibitem{dacrrogna} M.M. Dacrrogna {\it et al.}, J. Intl. Money and Finance,
387: {\bf 12}, 413 (1993).
388:
389: \bibitem{mantegna1} R.N. Mantegna and H.E. Stanley, Nature, {\bf 383},
390: 587 (1996).
391:
392: \bibitem{mantegna} R.N. Mantegna, Physica A {\bf 179}, 232 (1991).
393:
394: \bibitem{mandlelbrot} B.B. Mandlelbrot, {\it Fractals and
395: Scaling in Finance}, Springer-Verlag (1997).
396:
397: \bibitem{ramos} F.M. Ramos, C. Rodrigues Neto, and C.R. Rosa,
398: Los Alamos archive cond-mat/9907348 (1999).
399:
400: \bibitem{friedrich} S. Siegert, R. Friedrich, J. Peinke, Los Alamos
401: archive cond-mat/9803250;
402: R. Friedrich, J. Peinke, and C. Renner, physics/9901034 (2000).
403:
404: \bibitem{studentt} R.C. Blattberg and N.J. Gonedes, J. Business {\bf 47},
405: 244 (1974).
406:
407: \bibitem{terra} www.quote.com and www.qcharts.com
408:
409: \bibitem{gardiner} C.W. Gardiner, {\it Handbook of Stochastic Methods},
410: 2nd Ed., Springer, 1997.
411:
412: \bibitem{stariolo} D. Stariolo, Phys. Lett. A {\bf 185}, 262 (1994).
413:
414: \bibitem{press} W.H. Press, S.A. Teukolsky, W.T. Vetterling,and B.P.Flannery,
415: {\it Numerical Recipes (Fortran)}, 2nd Ed., (1994).
416:
417: \end{thebibliography}
418:
419: \begin{figure}
420: \caption{Distribution of S\&P500 index price changes at intervals
421: $t=$1, 10, and 60 minutes, and a fit of Tsallis form [Eq.~(\ref{pdf})].
422: (Note: horizontal scales in the three plots are different.)}
423: \label{fit}
424: \end{figure}
425:
426: \begin{figure}
427: \caption{Parameter $\beta(t)$ (which is proportional to the inverse
428: of the variance) as a function of time interval $t$. Filled
429: circles: S\&P500 index data. Solid curve: Fit of the Fokker-Planck $\beta(t)$
430: from Eq.~(\ref{betat}).
431: Open diamonds: Simulated data using parameters from the
432: fit. Along most of the curve the simulated data are
433: indistinguishable from the actual market data.}
434: \label{betafig}
435: \end{figure}
436:
437: \end{document}
438: