1: \documentclass[preprint,aps]{revtex4}
2: \usepackage{graphicx}
3: \usepackage{dcolumn}
4: \usepackage{amsmath, amssymb}
5:
6: \let\epsilon\varepsilon
7: \newcommand{\myhat}[1]{\bar{#1}}
8: \newcommand{\hatepsilon}{\myhat{\epsilon}}
9: \renewcommand{\hatepsilon}{{\boldsymbol \epsilon}}
10: \newcommand{\ie}{{\it i.e.}}
11: \newcommand{\comment}[1]{}
12:
13: \begin{document}
14:
15: \title{Nonstationary Increments, Scaling Distributions, and Variable Diffusion Processes in Financial Markets}
16:
17: \author{Kevin E. Bassler$^{1,2}$, Joseph L. McCauley$^{1,3}$, and Gemunu H. Gunaratne$^{1,4}$\footnote{Corresponding Author - Electronic Address: gemunu@uh.edu} }
18:
19: \affiliation{$^1$ Department of Physics,
20: University of Houston,
21: Houston, TX 77204}
22: \affiliation{$^2$ Texas Center for Superconductivity at the University of Houston,
23: Houston, TX 77204}
24: \affiliation{$^3$ Senior Fellow, Department of Economics,
25: J.E.Cairnes Graduate School of Business and Public Policy,
26: NUI Galway, Ireland}
27: \affiliation{$^4$ The Institute of Fundamental Studies,
28: Kandy, Sri Lanka}
29: \maketitle
30:
31: \nobreak
32:
33: {\bf Arguably the most important problem in quantitative finance is
34: to understand the nature of stochastic processes that underlie
35: market dynamics. One aspect of the solution to this problem involves
36: determining characteristics of the distribution of
37: fluctuations in returns. Empirical studies conducted
38: over the last decade have reported that they are
39: non-Gaussian, scale in time, and have power-law
40: (or fat) tails~\cite{mand,mccAgun,manAsta,friApei,borl}. However,
41: because they use sliding interval methods of analysis,
42: these studies implicitly assume that the underlying process has
43: stationary increments. We explicitly show that this assumption is not valid
44: for the Euro-Dollar exchange rate between 1999-2004.
45: In addition, we find that fluctuations in returns of the exchange rate
46: are uncorrelated and scale as power-laws for certain time intervals during each day.
47: This behavior is consistent with a diffusive process with a diffusion
48: coefficient that depends both on the time and the price
49: change. Within scaling regions, we find that sliding interval methods
50: can generate fat-tailed distributions as an artifact, and that the type
51: of scaling reported in many previous studies does not exist.}
52:
53: Our analysis is conducted on one-minute intra-day prices of
54: the Euro-Dollar exchange rate (obtained from Olsen
55: and Associates, Z\"urich) which is traded 24-hours a day.
56: Let $P(t)$ represent the exchange rate at time $t$ and define
57: the return of the exchange rate as
58: ${\bar x}(\tau; t) \equiv \log \left [P(\tau + t)/P(t) \right]$.
59: Here $t$ represents a time during the day and $\tau$ a time increment
60: that is initiated at $t$. The analysis presented below is predicated on
61: the assumption, for which we provide evidence,
62: that the stochastic dynamics of ${\bar x}(\tau; t)$
63: is the same between trading days. Then, we find that the average movement
64: taken over the approximately 1500 trading days during 1999-2004,
65: $\langle {\bar x} (\tau; t) \rangle$ nearly vanishes for each value of
66: $t$. A value of $\tau = 10\ min$ is used so that the
67: autocorrelations in the signal $P(t)$ have decayed sufficiently.
68: The rest of our analysis is conducted on fluctuations
69: $x(\tau;t) = {\bar x}(\tau;t) - \langle {\bar x} (\tau; t) \rangle$
70: about the mean.
71:
72: A stochastic process has stationary increments if the
73: distribution of $x(\tau; t)$ is independent of $t$;
74: otherwise, increments are nonstationary.
75: Figure 1(a) shows the behavior of the standard deviation
76: $\sigma(\tau;t) \equiv \sqrt {\langle x (\tau; t)^2 \rangle}$
77: of the Euro-Dollar rate as a function
78: of the time of day.
79: If the stochastic increments are stationary, the curve would be flat.
80: Clearly, it is not. Instead $\sigma (\tau; t)$ exhibits complicated
81: nonstationary behavior while changing by more than a factor of
82: 3 during the day.
83:
84: Our assumption of daily repetition of the stochastic process
85: is validated by conducting a corresponding analysis
86: of fluctuations throughout a trading week~\cite{galAcal}. Figure 1(b) shows the
87: standard deviation of returns averaged over the 300 weeks studied. The approximate
88: daily periodicity of $\sigma(\tau; t)$ is evident,
89: thereby justifying our approach. Similar observations were made on
90: price increaments for Euro-Dollar rate in Ref.~\cite{galAcal}.
91:
92: The standard deviation scales as power-laws with time during several intervals
93: within the day. Power-law fits to the data in some of these intervals are
94: shown by colored lines in Fig.~1(a). We focus our analysis on the
95: time interval {\bf I} which begins at 9:00 AM
96: New York time and lasts approximately $3$ hours. The data shown in red in
97: Fig.~2(a) shows that the standard deviation within this interval
98: scales like $t^{-\eta}$ where $t$ is measured from the beginning of the
99: interval and the index $\eta = 0.13 \pm 0.04$.
100: This scaling extends for more than $1.5$ decades in time.
101: Note that the value of $\eta$ is different for the other time intervals
102: during which the standard deviation scales in time.
103: Similar variation in scaling exponents during the
104: day has been reported previously \cite{carAcas}.
105:
106: The scaling index within {\bf I} does not change
107: significantly during the six years studied. This is demonstrated by
108: independently analyzing three two-year periods 1999-2000, 2001-2002,
109: and 2003-2004. Figure 2(b) shows that the scaling index
110: remains nearly unchanged between these two-year periods.
111:
112: We have also analyzed the behavior of other moments
113: $\langle |x(\tau; t) |^{\beta}\rangle^{1/\beta}$ of the returns.
114: Figure 2(a) shows that each of the moments $\beta=0.5,\ 1.0,\ 2.0$, and $3.0$
115: also scales as a power-law in time, and furthermore that the scaling index
116: for each of them is consistent with the value of $\eta = 0.15$.
117: This nearly uniform scaling of the different moments suggests that the return distribution
118: itself scales in time. Denote the distribution of $x(\tau; t)$ by
119: $W(x, \tau; t)$, where the final argument reiterates that
120: the distribution can depend on the starting time of the interval. In
121: particular, when the increments are nonstationary $W(x,\tau; t)$ depends
122: on $t$. Our scaling anzatz is
123: \begin{equation}
124: W(x, \tau; 0) = \frac{1}{\tau^{H}} {\cal F} (u)
125: \label{scaling}
126: \end {equation}
127: where $H$ is the {\it scaling index}, $u=x/\tau^{H}$
128: the {\it scaling variable} and ${\cal F}$ the {\it scaling function}.
129: Note that the scaling anzatz is for a time interval
130: {\it starting} from the beginning of {\bf I}.
131:
132: In addition to scaling, the stochastic dynamics appears to have
133: no memory. This can be demonstrated by
134: evaluating the auto-correlation function
135: \begin{equation}
136: A_{\tau}(t_1,t_2) = \frac{\langle x(\tau;t_1) x(\tau;t_2) \rangle} {\sigma(\tau;t_1) \sigma(\tau;t_2)}.
137: \nonumber
138: \end{equation}
139: We find that for $\tau=10$, $A_{\tau}(t_1, t_2) = 1$ if $t_1=t_2$, and of the
140: order of $10^{-3}$ when $|t_1-t_2|\ge 10$. This observation eliminates fractional
141: Brownian motion~\cite{manAvan} as a description for the underlying
142: stochastic dynamics, and strongly indicates that
143: ${\partial W (x, \tau; 0)}/{\partial \tau}$ depends only on
144: $x(\tau; 0)$ and $\tau$. If, in addition, $W(x, \tau;0)$ has finite
145: variance (see Fig. 4),
146: it has been analytically established that the
147: evolution of $W(x,\tau; 0)$ is given by a diffusion equation \cite{chan,gunAmcc}
148: \begin {equation}
149: \frac {\partial W(x, \tau; 0)}{\partial \tau} = \frac{1}{2} \frac{\partial^2}{\partial x^2} \left (D(x, \tau) W(x, \tau; 0)\right),
150: \label{FPeqn}
151: \end{equation}
152: where $D(x, \tau)$ is the diffusion coefficient.
153: There is no drift term in Eq.~(2) because
154: $x(\tau; t)$ has zero mean for all $t$.
155: Note that the stochastic dynamics is completely determined by the diffusion coefficient,
156: which, as shown below, depends on $H$. Hence, $H$ can be considered to
157: be the {\it dynamical} scaling index.
158:
159: %With Eq.~(1), if $x=0$ at $\tau=0$,
160: %then the variance $\langle x(\tau; 0)^2\rangle \sim \tau^{2H}$ \cite{basAgun}.
161: %\footnote{Finally, we point out that, generally, the second argument
162: %in the diffusion coefficient is the time $(t+\tau)$ since the stochastic process
163: %was initiated.}
164:
165: Because we have found scaling, consider solutions of the form (1) to Eq.~(2).
166: When $H = 1/2$, the diffusion coefficient has been
167: shown to be a function of $u$; i.e., $D(x, \tau) = {\cal D}(u)$ \cite{gunAmcc}.
168: If, in addition, ${\cal D}(u)$ is symmetric in $u$, it is related to the
169: scaling function by
170: ${\cal F}(u) = D(u)^{-1} \exp \left (-\int^{u} dy \ y/D(y) \right )$ \cite{gunAmcc,aleAbas}.
171: When $H \ne 1/2$, we can
172: ``rescale" time intervals by $\tilde \tau = \tau^{2H}$~\cite{galAcal,basAgun}. In $\tilde \tau$,
173: the stochastic process has a scaling index $1/2$ and a diffusion
174: coefficient of the form ${\cal D}(x/\sqrt{\tilde \tau})$.
175: Converting back to $\tau$,
176: $D(x, \tau) = 2 H \tau^{2H-1} {\cal D}(u)$ \cite{basAgun}.
177:
178: Statistical analyses of financial markets have often been conducted using
179: sliding interval methods \cite{mccAgun,manAsta,friApei,borl,borl2,galAcal,ghaAbre},
180: which implicitly assume that increments are stationary even if they are not. For example,
181: they compute the distribution $W_S(x, \tau) = \langle W (x, \tau; t)\rangle_{t}$,
182: where $\langle . \rangle_t$ indicates an average over $t$.
183: Many of these studies have reported that $W_S(x, \tau)$ scales as
184: \begin{equation}
185: W_S (x, \tau) = \frac{1}{\tau^{H_S}} {\cal F}_S (v),
186: \label{siscaling}
187: \end{equation}
188: where $v = x / {\tau^{H_S}}$ and $H_S \approx 1/2$.
189: It has also been reported that the scaling function ${\cal F}_S$ has power-law (or fat)
190: tails~\cite{friApei,borl}. However, it is important to understand that $W_S(x, \tau)$ is
191: a solution of Eq.~(2) only when the stochastic process has stationary
192: increments, in which case $H = H_S = 1/2$.
193: In general, $H_S$ and $W_S (x, \tau)$ are different from
194: $H$ and $W (x, \tau; 0)$. Next, we give an explicit example where this is the case,
195: and, in addition, $W_S(x, \tau)$ appears to have fat-tails even though
196: $W(x, \tau; 0)$ does not.
197:
198: Consider a diffusive process initiated at $x=0$ that has a variable
199: diffusion coefficient $2 H \tau^{2H-1} (1 + |u|)$. Its distribution
200: has a scaling index $H$ and a scaling function
201: ${\cal F}(u) = \frac{1}{2}\exp (-|u|)$~\cite{gunAmcc,aleAbas}.
202: (See the discussion following Eq.~(2).)
203: Numerical integration of the stochastic process for $H=0.35$
204: confirms this claim, see Fig.~3(a).
205: In contrast, $W_S(x, \tau)$ calculated from {\it the same data}
206: appears to scale with an index $H_S = \frac{1}{2}$.
207: Unlike ${\cal F}$ which is bi-exponential, the apparent scaling function ${\cal F}_S$
208: (shown in Fig.~3(b)) has fat-tails. However, a careful analysis
209: reveals that distributions $W_S(x,\tau)$ do not scale in the tail region,
210: and hence that ${\cal F}_S$ is not well-defined.
211: Differences analogous to those between $H$ and $H_S$ have been
212: noted for L\'evy processes~\cite{fogAboh} and for the R/S analysis of Tsallis
213: distributions~\cite{borl2}.
214:
215: The behavior of $\sigma(\tau; t)$ (Fig.~2(a)) can
216: be calculated for variable diffusion processes. Assuming that
217: $\tau$ is small, Ito calculus gives
218: $\delta x^2 \equiv x(\tau; t)^2 = D(x, t) \tau$.
219: Averaging over returns at $t$ gives
220: \begin{equation}
221: \langle \delta x^2 \rangle = \left[\int dx W(x, t; 0) D(x, t) \right] \tau.
222: \label{mean_abs1}
223: \end{equation}
224: In a variable diffusion process,
225: $W(x,t; 0) = t^{-H} {\cal F}(u)$ and $D(x; t) = 2 H t^{2H -1} {\cal D}(u)$;
226: consequently
227: \begin{equation}
228: \sqrt {\langle \delta x^2 \rangle} \sim t^{H - 1/2},
229: \label {mean_abs2}
230: \end{equation}
231: independent of the exact form of ${\cal D}(u)$.
232: Results for the Euro-Dollar rate within the interval {\bf I} (Fig.~2(a))
233: which showed that $\eta \approx 0.15$ are therefore consistent with a
234: scaling index $H = \frac{1}{2} - \eta \approx 0.35$. Note that, unlike for
235: L\'evy processes and fractional Brownian motion, $H < 1/2$, and is substantially less than $H_S$
236: reported in previous analyses of the Euro-Dollar exchange rate
237: (between 0.5 and 0.6)~\cite{galAcal,ghaAbre,mulAdac}.
238: A general calculation for the moments of a variable diffusion process gives
239: \begin{equation}
240: \langle |\delta x|^{\beta} \rangle^{1/\beta} \sim t^{H -1/2},
241: \label {mean_abs3}
242: \end{equation}
243: for all $\beta$, consistent with results shown in Fig.~2(a).
244:
245: In order to estimate $H_S$ for an arbitrary variable diffusion process, we note first that
246: $\langle x(t+\tau;0)^2\rangle = \langle x(t;0)^2 \rangle + \langle x(\tau;t)^2 \rangle$
247: for any diffusive process without memory (see Ref.\cite{gunAmcc}).
248: Then, using the scaling anzatz (1), setting $c = \int du \ u {\cal F}(u)$,
249: and taking the sliding interval average
250: \begin{equation}
251: \langle x(\tau; t)^2 \rangle_t = \langle c (t+\tau)^{2H} - c t^{2H}\rangle_t \approx 2Hc \langle t^{2H-1}\rangle_t \tau,
252: \label{expn}
253: \end{equation}
254: where the last approximation is valid when $\tau \ll t$, a condition
255: that is true for most intervals of
256: length $\tau$ in a sliding interval calculation. Hence
257: $\langle x(\tau; t)^2 \rangle_{t} \sim \tau$.
258: Consequently, $H_S = 1/2$ regardless of the value of $H$!
259:
260: Finally, we introduce a method to extract the empirical scaling function ${\cal F}$ from
261: the Euro-Dollar time series. Unfortunately, the available data
262: are insufficient to determine ${\cal F}(u)$ accurately using the usual method
263: of collapsing $W(x, \tau; 0)$ for multiple values of $\tau$. However, since we have
264: determined $H (\approx 0.35)$ independently, we can use Eq.~(\ref{scaling}) for
265: multiple values of $\tau$ in the interval {\bf I} (i.e., $\tau$
266: between approximately $10$ and $160$ minutes) to determine ${\cal F}$.
267: The result is shown in Fig.~4(a). Note that the distribution has an approximate bi-exponential
268: form. Since exponential distributions have finite variance, all assumptions
269: needed for the derivation of Eq.~(\ref{FPeqn}) are justified.
270: However, it is asymmetric and decays more slowly on the negative side.
271: By contrast, the empirical sliding interval scaling function ${\cal F}_S(v)$
272: for the same time interval is shown in Fig.~4(b).
273: For this case, the scaling collapse is achieved for $H_S=1/2$. $F_S(v)$ appears to have fat tails,
274: consistent with previous reports~\cite{mulAdac,borl}.
275: However, in light of the example discussed earlier and the fact that
276: $H \ne 1/2$, it is unlikely that ${\cal F}_S$ is well-defined
277: for this financial market data within the interval {\bf I}.
278:
279: Variable diffusion processes exhibit another signature (stylized
280: fact) of market fluctuations. Although their
281: autocorrelation vanishes, a large fluctuation will typically produce
282: a large value of $|x|$, and hence a return with
283: a large diffusion coefficient. Consequently, a large
284: fluctuation is likely to be followed by additional large fluctuations
285: whose signs are uncorrelated to the first~\cite{gunAmcc}. As a result, the
286: autocorrelation function for the signal $|x(\tau;t)|$ (or for the signal $x(\tau;t)^2$)
287: will decay slowly in $t$. Such behavior,
288: referred to as the ``clustering of volatility" is seen in the
289: Euro-Dollar exchange rate and has been reported in
290: empirical studies of other financial markets~\cite{conApot,heyAyan,heyAleo}.
291:
292: The analysis given here applies to stochastic dynamics of a single
293: scaling interval. However, the daily fluctuations in the
294: Euro-Dollar rate are a combination of scaling intervals with distinct
295: scaling indices, and possibly regions with no scaling.
296: We have not yet determined how to extend our analysis
297: beyond a single scaling region. Bacuase of this, it is not clear how
298: to interpret the distributions over intervals longer than
299: a scaling region, including inter-day data.
300:
301: We have shown that stochastic fluctuations in the Euro-Dollar
302: rate have uncorrelated nonstationary increments during the course of a trading day,
303: and that there are intervals during which their absolute moments scale like
304: a power-law in time. The stochastic dynamics during these scaling intervals
305: can be described by a diffusion process with variable diffusion coefficient.
306: We have also shown that sliding interval analysis of variable diffusion processes can give an
307: incorrect scaling exponent and in addition can give scaling functions with
308: fat-tails even when the underlying dynamics do not have them. Indeed, this
309: appears to be the case within the interval {\bf I}.
310:
311: %stochastic processes are time dependent and hence nonstationary~\cite{basAgun}.
312: %Since $H$ and $H_S$ can be different,
313: %it is necessary to clarify which index is related to various characteristics
314: %of the stochastic processes underlying financial markets.
315: %For example, the fact that the dynamical scaling index $H$ can
316: %be changed by rescaling $\tau$ implies that its value has no bearing
317: %on whether or not the process has memory~\cite{basAgun,couAdav}.
318: %Whether $H_S$ is well-defined and if the presence of memory is
319: %related whether or not $H_S=1/2$ needs to be studied.
320:
321: The authors would like to thank A. A. Alejandro-Quinones for discussions.
322: They also acknowledge support from the Institute for Space Science
323: Operations (KEB, GHG) and the NSF through grants DMR-0406323 (KEB),
324: DMR-0427938 (KEB) and DMS-0607345 (GHG).
325:
326: \begin{thebibliography}{99}
327:
328: \bibitem{mand} B. B. Mandelbrot, The Variation of Certain Speculative Prices,
329: {\em J. Bus.} {\bf 36}, 394 (1963).
330:
331: \bibitem{mccAgun} J. L. McCauley and G. H. Gunaratne,
332: An Empirical Model of Volatility Returns and Options Pricing,
333: {\em Physica A} {\bf 329}, 170 (2003).
334:
335: \bibitem{manAsta} R. N. Mantegna and H. E. Stanley,
336: Scaling Behavior in the Dynamics of an Economic Index, {\em Nature} {\bf 376}, 46 (1995);
337: Turbulence in Financial Markets, {\em Nature} {\bf 383}, 587 (1996).
338:
339: \bibitem{friApei} R. Friedrich, J. Peinke, and Ch. Renner,
340: How to Quantify Deterministic and Random Influences on the Statistics
341: of the Foreign Exchange Market,
342: {\em Phys. Rev. Lett.} {\bf 84}, 5224 (2000).
343:
344: \bibitem{borl} L. Borland, A Theory of Non-Gaussian Option Pricing,
345: {\em Quan. Finance} {\bf 2}, 415 (2002).
346:
347: \bibitem{galAcal} S. Galluccio, G. Caldarelli, M. Marsili, and Y. C. Zhang,
348: Scaling in Currency Exchange,
349: {\em Physica A} {\bf 245}, 423 (1997).
350:
351: \bibitem{carAcas} A. Carbone, G. Castelli, and H. E. Stanley,
352: Time-dependent Hurst Exponents in Financial Time Series,
353: {\em Physica A} {\bf 344}, 267 (2004).
354:
355: \bibitem{manAvan} B. Mandlebrot and J. W. van Ness,
356: Fractional Brownian Motion, Fractional Noise and Applications,
357: {\em SIAM Rev.} {\bf 10}, 422 (1968).
358:
359: \bibitem{gunAmcc} G. H. Gunaratne, J. L. McCauley, M. Nicole, and A. T\"or\"ok.
360: Variable Step Random Walks and Self-Similar Distributions,
361: {\em J. Stat. Phys.} {\bf 121}, 887 (2005).
362:
363: \bibitem{chan} S. Chandrasekhar,
364: Stochastic Problems in Physics and Astronomy,
365: {\em Rev. Mod. Phys.}, {\bf 15}, 1 (1943).
366:
367: %\bibitem{gunAmcc2} G. H. Gunaratne and J. L. McCauley,
368: % Theory of Fluctuations in Stock Prices and Valuation of Their Options,
369: % {\em Proceedings of the SPIE} {\bf 5848}, 131 (2005).
370:
371: \bibitem{aleAbas} A. A. Alejandro-Quinones, K. E. Bassler, M. Field, J. L. McCauley, M. Nicol,
372: I. Timofeyev, A. T\"or\"ok, and G. H. Gunaratne,
373: A Theory of Fluctuations in Stock Prices,
374: {\em Physica A} {\bf 363}, 383 (2006).
375:
376: \bibitem{basAgun} K. E. Bassler, G. H. Gunaratne, and J. L. McCauley,
377: Markov Processes, Hurst Exponents, and Nonlinear Diffusion Equations with
378: Applications to Finance, Physica A, {\bf 369}, 343 (2006).
379:
380: \bibitem{borl2} L. Borland,
381: Microscopic Dynamics of the Nonlinear Fokker-Planck Equation: A Phenomenological Model,
382: {\em Phys. Rev. E} {\bf 57}, 6634 (1998).
383:
384: \bibitem{ghaAbre} S. Ghashghaie, W. Breymann, J. Peinke, P. Talkner, and Y. Dodge,
385: Turbulent Cascades in Foreign Exchange Markets,
386: {\em Nature} {\bf 381}, 767 (1996).
387:
388: \bibitem{fogAboh} H. C. Fogedby, T. Bohr, and H. J. Jensen,
389: Fluctuations in a L\'evy Flight Gas,
390: {\em J. Stat. Phys.} {\bf 66}, 583 (1992).
391:
392: \bibitem{mulAdac} U. A. M\"uller, M. M. Dacorogna, R. B. Olsen, O. V. Pictet, M. Schwarz, and C. Morgenegg,
393: Statistical Study of Foreign Exchange Rates, Empirical Evidence of a Price Change Scaling Law,
394: and Inter-day Analysis,
395: {\em J. Bank. Fin.} {\bf 14}, 1189 (1990).
396:
397: \bibitem{conApot} R. Cont, M. Potters, and J.-P. Bouchard,
398: Scaling in Stock Market Data: Stable Laws and Beyond,
399: in "Scale Invariance and Beyond", eds. B. Dubrulle, F. Graner, and D. Sornette,
400: Springer, Berlin, 1997.
401:
402: \bibitem{heyAyan} C. C. Heyde and Y. Yang,
403: On Defining Long Range Dependence,
404: {\em J. Appl. Prob.} {\bf 34}, 939 (1997).
405:
406: \bibitem{heyAleo} C. C. Heyde and N. N. Leonenko,
407: Student Processes,
408: {\em Adv. Appl. Prob.} {\bf 37}, 342 (2005).
409:
410: \bibitem{couAdav} M. Couillard and M. Davison, A Comment on Measuring the Hurst
411: exponent of Financial Time Series,
412: {\em Physica A} {\bf 348}, 404 (2005).
413:
414: \end{thebibliography}
415:
416: \begin{figure}
417: \includegraphics[width=2.5in]{Fig1a.eps}\\
418: \includegraphics[width=2.5in]{Fig1b.eps}
419: \caption {(a) The standard deviation
420: $\sigma(\tau;t) \equiv \sqrt {\langle x (\tau; t)^2 \rangle}$
421: of the daily Euro-Dollar exchange as a function of the time of day (in GMT).
422: Here $\tau = 10\ min$ to ensure that autocorrelations in $P(t)$ have decayed sufficiently.
423: Our statistical analysis assumes that $x (\tau; t)$ follows the same stochastic
424: process each trading day. The average indicated by
425: the brackets $\langle . \rangle$ is taken over the
426: approximately 1500 trading days between 1999-2004, and the standard error
427: at each point is typically 3\%. Note that, if the stochastic dynamics had stationary
428: increments, $\sigma(\tau;t)$ would be constant. Instead, it varies by more
429: than a factor of 3 during the day, thus
430: showing explicitly that the exchange rate has nonstationary
431: increments. Notice also that $\sigma(\tau;t)$ scales in time during several
432: intervals, four of which are highlighted by colored lines that are power-law fits.
433: Our analysis focuses on the interval {\bf I} shown by the horizontal solid line.
434: (b) The weekly behavior of $\sigma(\tau;t)$ for the same data. Observe that it exhibits an
435: approximate daily periodicity, thereby justifying our assumption of the daily repeatability
436: of the stochastic process underlying the Euro-Dollar exchange rate.}
437: \label{figure1}
438: \end{figure}
439:
440:
441: \begin{figure}
442: \includegraphics[width=2.5in]{Fig2a.eps}\\
443: \includegraphics[width=2.5in]{Fig2b.eps}
444: \caption {(a) A log-log plot of $\langle x(\tau; t)^{\beta} \rangle^{1/\beta}$
445: for $\beta = 0.5, 1.0, 2.0$, and $3.0$, demonstrating power law decay $t^{-\eta}$ for each index.
446: Here $t$ is measured in local New York time stating at 9:00AM.
447: The data for $\beta = 0.5, 1.0, 2.0$, and $3.0$, shown in blue, green, red, and
448: black, respectively, have scaling indices (given by the slopes of the solid lines)
449: $\eta = 0.15\pm0.02$, $0.14\pm0.02$, $0.13\pm0.04$ and $0.13\pm0.08$.
450: All of these values are consistent with $\eta \approx 0.15$, and hence
451: a dynamical scaling index of $H=\frac{1}{2}-\eta \approx 0.35$. The error estimates on the exponents
452: are the standard errors from the nonlinear fit including the standard
453: deviations for each time point, but neglecting any correlations between them.
454: (b) The behavior of the standard deviation $\sigma(\tau;t)$
455: in the interval {\bf I} during each of the periods 1999-2000 (blue), 2001-2002 (red), and
456: 2003-2004 (green). The scaling index from nonlinear fits for the three data sets are
457: $0.13 \pm 0.06$, $0.14 \pm 0.04$ and $0.14 \pm 0.07$.
458: The near equality of these indices shows that the scaling index is nearly
459: invariant over time. }
460: \label{Figure2}
461: \end{figure}
462:
463:
464: \begin{figure}
465: \includegraphics[width=2.5in]{Fig3a.eps}\\ ~ \\
466: \includegraphics[width=2.5in]{Fig3b.eps}
467: \caption {(a) The scaling function of the return distribution ${\cal F}$ calculated
468: from a collapse of data for $\tau=10$ (blue), $100$ (green), and $1000$ (red) units.
469: The results are from a set of 5,000,000 independent stochastic processes with
470: variable diffusion. The scaling index used was $H=0.35$,
471: and the diffusion coefficient was $2H t^{2H-1} (1+|u|)$.
472: Note that ${\cal F}$ is bi-exponential, as discussed in the text.
473: (b) The sliding interval scaling function ${\cal F}_S$ calculated from the same
474: runs. Shown are results for sliding intervals with $\tau=10$ (blue), $100$ (green) and
475: $1000$ (red) units from runs of length $10,000$ units. Unlike ${\cal F}$, it appears to have fat tails.
476: The scaling index used here for which the scaling collapse is achieved is $H_S = 1/2$
477: even though the dynamical scaling index
478: is $H = 0.35$. Note, however, although the central part of
479: the distribution scales well, the tails do not.}
480: \label{Figure3}
481: \end{figure}
482:
483:
484: \begin{figure}
485: \includegraphics[width=2.5in]{Fig4a.eps}\\ ~ \\
486: \includegraphics[width=2.5in]{Fig4b.eps}
487: \caption {(a) The empirical scaling function ${\cal F}$ for interval {\bf I} calculated assuming
488: the scaling anzatz Eq.~(1) with $H=0.35$ and values of $\tau$ between
489: 10 and 160 minutes. Note that ${\cal F}$ is slightly asymmetric and approximately
490: bi-exponential. Since exponential distributions have finite variance,
491: all assumptions needed for the derivation of Eq.~(\ref{FPeqn}) are
492: justified. (b) The empirical sliding interval scaling
493: function ${\cal F}_S$ for interval {\bf I}
494: calculated by scaling collapse of data using the anzatz Eq.~(3) for
495: $\tau$ of 10 (blue), 20 (green) and 40 (red) minutes.
496: Note that ${\cal F}_S$ has fat-tails.}
497: \label{Figure4}
498: \end{figure}
499:
500: \end{document}