1: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2: % Articolo %
3: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
4: % %
5: % Articolo sulla caratterizzazione temporale delle serie FX %
6: % %
7: % %
8: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
9:
10: \documentclass[11pt]{article}
11:
12: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
13: % INCLUDE %
14: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
15:
16:
17: \usepackage{graphicx}
18: \usepackage{amssymb}
19: \usepackage{amsmath}
20:
21: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
22: % PAGE LAYOUT %
23: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
24:
25: %% \hoffset -20mm
26: \voffset -10mm
27: %% \textwidth=160mm
28: \textheight=230mm
29: %% \headheight = 10mm
30: \footskip = 50pt
31: %% \headsep = 0mm
32:
33: \linespread{1.3}
34:
35:
36: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
37: % TITLE %
38: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
39:
40: \begin{document}
41: \title{ Time and foreign exchange markets }
42: \author{Luca Berardi$^\star$ and Maurizio Serva$^\dagger$}
43: \date{}
44: \maketitle
45:
46: \begin{small}
47: \begin{tabular}{cc}
48: ${}^\star \,$ Dipartimento di Ingegneria Elettrica &
49: ${}^\dagger \,$ Dipartimento di Matematica and I.N.F.M. \\
50: Universit\`a degli Studi, L'Aquila &
51: Universit\`a degli Studi, L'Aquila \\
52: 67040 Poggio di Roio (AQ) - Italy &
53: 67010 Coppito (AQ) - Italy
54: \end{tabular}
55: \end{small}
56: \bigskip
57:
58:
59:
60: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
61: % BODY %
62: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
63:
64:
65:
66:
67: %%%%----------------------------------------------%%%%%%%%%%%%%%%%
68:
69:
70: \begin{abstract}
71:
72: The definition of time is still an open question when one deals with
73: high frequency time series. If time is simply the calendar time,
74: prices can be modeled as continuous random processes and values
75: resulting from transactions or given quotes are discrete samples of this
76: underlying dynamics. On the contrary, if one takes the business time
77: point of view, price dynamics is a discrete random process, and time
78: is simply the ordering according which prices are quoted in the
79: market. In this paper we suggest that the business time approach is
80: perhaps a better way of modeling price dynamics than calendar time.
81: This conclusion comes out from testing probability densities and conditional
82: variances predicted by the two models against the experimental ones.
83: The data set we use contains the DEM/USD
84: exchange quotes provided to us by Olsen \& Associates during a period of one
85: year from January to December 1998. In this period 1,620,843 quotes
86: entries in the EFX system were recorded.
87:
88: \end{abstract}
89:
90:
91:
92: \bigskip
93: \bigskip
94: \bigskip
95: \bigskip
96: PACS numbers: 89.65.Gh; 05.40.Fb
97:
98: \bigskip
99: Keywords: Forex markets, time, lags, high-frequency.
100:
101: \bigskip
102: \bigskip
103: \bigskip
104: \bigskip
105: \bigskip
106: \bigskip
107: \bigskip
108: Corresponding author:
109:
110: Maurizio Serva, email: serva@univaq.it,
111: tel. +390862433153, fax. +390862433180
112:
113:
114: \newpage
115:
116: \section{Introduction}
117:
118: In the high-frequency arena there are two main-streams
119: about modeling the stochastic properties of quotes.
120: The first approach is to consider quotations as sampled values
121: of an underlying continuous-time random process \cite{Merton90},
122: \cite{Nelson90}. Sampling is itself a random operation, thus
123: introducing a twofold uncertainty in the price determination
124: \cite{Mainardi00}, \cite{Scalas00}. In this framework, time in the
125: model flows continuously, and is called {\bf calendar time}.
126:
127: \noindent In the second approach, quoted prices are modeled through a
128: discrete-time stochastic process \cite{Taylor86}; in this setting,
129: time is just the natural total order relation among quotations, and it
130: is iso-morphic with the set of non-negative integers (being time $0$,
131: the time associated to the first considered quotation). This is the
132: {\bf business time} approach, and randomness only enters in the
133: determination of prices. It should be pointed out, however, that the
134: waiting times between two quotes are also random quantities, but they
135: are assumed to not contribute to the price determination process.
136:
137: Whether a calendar-time or a business-time framework should be adopted
138: in modeling the stochastic nature of financial quotes, has been a
139: longly debated issue by the finance research community, and it clearly
140: depends on many factors, like, for example: a) adherence to the
141: physical behavior of reported prices, b) usefulness in terms of a
142: theory to be developed, and c) last but not least, a matter of
143: taste. See, for example \cite{Baviera99}, \cite{Pasquini00},
144: \cite{Baviera01}.
145:
146: In this paper, we suggest that business time is perhaps a better
147: tool for modeling the asset dynamics than calendar-time.
148: In order to support our claim, we consider:
149: 1) returns corresponding to a given calendar time lag and any business time lag,
150: 2) returns corresponding to the same calendar time lag but having a fixed
151: business time lag.
152: We find out that their statistical
153: properties are different consistently with the business hypothesis
154: and inconsistently with the calendar one.
155: In practice, we estimate some variances and some
156: probability densities
157: whose behavior is different in the two scenarios.
158:
159: The dataset we use contains the DEM/USD exchange quotes
160: taken from Reuters' EFX pages (the dataset having been supplied by
161: Olsen \& Associates) during a period of one year from January to
162: December 1998. In this period 1,620,843 quotes entries in the EFX
163: system were recorded. The dataset provides a continuously updated
164: sequence of bid and ask exchange quotation pairs from individual
165: institutions whose names and locations are also recorded.
166: The reason for using FX data is that this market is not subject to
167: any working time restriction; in fact, it is open 24 hours a day, seven days a week.
168: This is in contrast to stock markets, where artificial time regulation
169: would have made more difficult, if not impossible, to find out
170: the results outlined in this paper.
171:
172: \section{Business time vs Calendar Time}
173:
174: \subsection{Calendar Time}
175:
176: In the calendar time framework, prices are modeled as continuous-time
177: random processes. Clearly, market quotes are not defined for every $t
178: \in \mathbb{R}$, but only at discrete intervals, whose extensions in
179: time are called \emph{calendar lags} (usually ranging from 2sec. to
180: several minutes, sometimes hours). Nevertheless, according to the
181: calendar time picture, prices are usually considered as discrete
182: samples of an underlying continuous-time random process.
183:
184: The model of price dynamics in calendar time therefore has the
185: following structure:
186: %
187: \begin{equation}
188: S(t+\Delta) = S(t) e^{R_\Delta(t)}
189: \label{eq:CT-modelB}
190: \end{equation}
191: %
192: where $S(t)$ and $S(t+\Delta)$ are the spot prices
193: at times $t$ and $t+\Delta$,
194: $\Delta$ is an arbitrary calendar time lag and
195: $R_\Delta(t)$ is the {\it aggregated} return of
196: prices over the time interval $[t,t+\Delta]$.
197:
198: Considering a framework where prices evolve over the calendar time, it
199: is generally assumed that quotes result from a random sampling at
200: times $t_0, \ldots t_n$ of the continuous-time underlying process
201: $S(t)$. In a pure calendar time framework such a random sampling is
202: uncorrelated with the process $S(t)$ itself. We observe however that
203: this is only valid as an approximation; indeed, several studies have
204: shown a weak correlation between the sequence of lags and that of
205: returns, among which we cite \cite{Raberto02}.
206:
207: The last assumption usually made in order to complete the model
208: description in the calendar time setting is that the variance of
209: $R_\Delta(t)$ is a linear function of the
210: calendar time lag $\Delta$, i.e.:
211: %
212: \begin{equation}
213: Var[R_\Delta(t)] = \sigma^2 \Delta
214: \label{eq:CT-variance}
215: \end{equation}
216: %
217: If the logarithm of $S(t)$ has independent increments the above
218: equation obviously holds and $\sigma$ is the constant volatility.
219: However, it is well known that independence does not hold
220: because of volatility clustering which is due to the correlation
221: of the absolute values of returns\cite{Pasquini99}.
222: As a consequence, in spite of a constant volatility, one has a time
223: dependent volatility.
224: Nevertheless, the above behavior of the variance
225: still holds true but $\sigma^2$
226: is now the average of the squared volatility.
227: For our purposes we only assume that the above equality holds
228: and we do not need of specific assumptions concerning volatility behavior.
229:
230: Let us define the process $M(t)$ as follows:
231: %
232: \begin{equation}
233: M(t+\Delta) = M(t) + M_\Delta(t)
234: \end{equation}
235: %
236: where $M_\Delta(t)$ represents the number of given quotes (samples) in the
237: interval $[t,t+\Delta]$. Clearly, $M(t)$ is a non-decreasing random
238: process assuming integer values. We also observe that $M(t)$ as a
239: function of $t$ is piecewise constant, and its value increases by one
240: each time a quote is given (i.e. at times $t_0, \ldots t_n$).
241:
242: Given the assumptions made so far, it follows that the process $M$ and
243: the process $S$ are mutually independent. Hence, it follows that
244: the probability density of returns corresponding
245: to a calendar time lag $\Delta$ is insensitive from the condition
246: that $M_\Delta(t)$ is also fixed to a value $m$.
247: In symbols:
248: %
249: \begin{equation}
250: P[R_\Delta(t) | M_\Delta(t) = m] = P[R_\Delta(t)]
251: \label{eq:CT-cond_density}
252: \end{equation}
253: %
254: and, in particular, the associated variance exhibits the same
255: insensitiveness:
256: %
257: \begin{equation}
258: Var[R_\Delta(t) | M_\Delta(t) = m] = Var[R_\Delta(t)] =
259: \sigma^2 \Delta
260: \label{eq:CT-cond_variance}
261: \end{equation}
262: %
263:
264: \noindent Therefore, we can summarize the calendar time
265: hypothesis as follows:
266:
267: \paragraph{Hypothesis H1:}
268: The asset prices evolve over calendar time, i.e. according to the
269: model in Eq.~(\ref{eq:CT-modelB}) and Eq.~(\ref{eq:CT-variance}) holds.
270: Moreover the processes $S$ and $M$ are mutually independent, therefore
271: Eq.~(\ref{eq:CT-cond_density}) and Eq.~(\ref{eq:CT-cond_variance}) hold.
272:
273: \bigskip
274:
275: Let us anticipate that the main argument of the paper
276: is based on the estimation of the quantities in
277: Eq.~(\ref{eq:CT-cond_density}) and Eq.~(\ref{eq:CT-cond_variance}).
278: We will show with enough evidence that the two equalities are
279: largely violated in a way which, on the contrary, is consistent
280: with the business framework.
281:
282: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
283: \subsection{Business Time}
284:
285:
286: In the business-time approach, price dynamics is modeled as a
287: discrete-time random process. Indeed, the time basis is the ordered
288: sequence of times at which prices are quoted in the markets. It is
289: therefore a set isomorphic with the set of non-negative integers. In
290: such a framework the statistic model of price dynamics in the
291: business-time framework is the following:
292: %
293: \begin{equation}
294: S(n+m) = e^{R_m(n)}S(n)
295: \label{eq:BT-model}
296: \end{equation}
297: %
298: where $S(n)$ and $S(n+m)$ are
299: the asset price at business times $n$ and $n+m$ while
300: $R_m(n)$ is the \emph{aggregated} return over $m$ consecutive quotes.
301: It is then clear that the only time-dependence affecting the price process is
302: based on the global ordering of events while the
303: return is independent from calendar lag.
304: Notice that we refer to $m$ as
305: the business time lag as opposed to the calendar time lag
306: $\Delta$ introduced in the previous section.
307:
308: Considering the price dynamics in a business time setting naturally
309: leads to the following assumption:
310: %
311: \begin{equation}
312: Var[R_m(n)] = \hat{\sigma}^2 m
313: \label{eq:BT-variance}
314: \end{equation}
315: whose motivation is the same of that provided for the analogous
316: assumption in the calendar time hypothesis.
317: %
318: We also define the random process:
319: $$
320: T(n+m) = T(n) + T_m(n)
321: $$ where $T(n)$ is the stochastic calendar time at
322: business time $n$ and $T_m(n)$ corresponds
323: to the calendar lag $T(n+m) - T(n)$, i.e. the time
324: elapsed from $T(n)$ after the occurrence of $m$ consecutive quotes.
325: It can be readily seen that there is a direct
326: connection between $T(n)$ and the process $M(t)$ defined in the
327: previous subsection. In fact, $M(t)=n$ with $t \in [T(n),T(n+1))$,
328: and, moreover, the following relation holds:
329: $$
330: M_{T_m(n)}(T(n)) = m
331: $$
332: for an arbitrary positive integer $m$.
333:
334: \bigskip
335: Given the assumption of statistical independence between $S(n)$ and
336: $T(n)$, for a generic $\Delta$ the following relation holds:
337: %
338: \begin{equation}
339: P[R_m(n) | T_m(n) \in [\Delta-\epsilon, \Delta+\epsilon] ] = P[R_m(n)]
340: \label{eq:BT-pdf}
341: \end{equation}
342: %
343: where $\epsilon$ is a fixed quantity.
344: The above equation states that the probability density of returns corresponding to a
345: business time lag $m$ is insensitive to the condition that
346: the calendar time lag is also fixed to a value around $\Delta$.
347: In particular we have for the variance:
348: \begin{equation}
349: Var[R_m(n) | T_m(n) \in [\Delta-\epsilon, \Delta+\epsilon] ] = Var[R_m(n)]
350: = \hat{\sigma}^2 m
351: \label{eq:BT-cond_variance}
352: \end{equation}
353: %
354: which is the business time analogue of
355: Eq.~(\ref{eq:CT-cond_variance}).
356:
357: Given all the assumptions made so far, we are ready to formulate the
358: hypothesis of prices dynamics in a business time setting.
359:
360: \bigskip
361: \paragraph{Hypothesis H2:}
362: Asset prices follow the model in Eq.~(\ref{eq:BT-model}) and
363: Eq.~(\ref{eq:BT-variance}) holds. Moreover, the processes $S$ and $T$
364: are independent, it follows that Eq.~(\ref{eq:BT-pdf}) and
365: Eq.~(\ref{eq:BT-cond_variance}) hold.
366:
367: \bigskip
368:
369: Before concluding this preliminary outline of the two basic approaches
370: used to describe price dynamics (i.e. calendar time \& business time)
371: we also give another important property of some of the quantities
372: involved so far, which will turn useful in the remaining part of the
373: paper. \\ With all the positions previously made, let us first observe
374: that the following relation holds:~\footnote{This follows from the
375: stationarity of the process $M_\Delta(T(n))$. In particular:
376: $E[M_\Delta(T(n))]$ does not depend on $T(n)$ so we drop the
377: sub case. Moreover, $E[M_{k \Delta}] = k E[M_\Delta]$, since the
378: average number of quotes in $k$ intervals of the same length sums up
379: to $k$ times the value for the single interval, from which the
380: proportionality follows.}
381: $$
382: E[M_\Delta(T(n))] = \alpha \Delta
383: $$
384: for a suitable constant $\alpha$. Simply put, this property states
385: that the expected value of the number of quotes in an interval
386: $\Delta$ is proportional to $\Delta$ itself.
387:
388: Finally, considering the composition of the price process in business
389: time and the process representing the number of quotes in a
390: given calendar time lag $\Delta$, it can be shown that:
391: %
392: \begin{equation}
393: Var[R_{M_\Delta(t_n)}(n)] = \hat{\sigma}^2 E[M_\Delta(T(n))] =
394: \hat{\sigma}^2 \alpha \Delta
395: \label{eq:BT-CT-crossover}
396: \end{equation}
397: %
398: Thus, in the business time hypothesis, we also expect the variance in
399: (\ref{eq:BT-CT-crossover}) to be proportional to $\Delta$.
400:
401: As already anticipated, all equalities in this subsection
402: are supported by the following statistical analysis confirming
403: the validity the business time framework.
404:
405:
406: \section{Statistical Estimators}
407:
408:
409: In this and next section we carry out some experimental
410: analysis in order to best fit the description of prices dynamics
411: choosing between the two distinct possibilities concisely modeled
412: by hypotheses H1 and H2.
413:
414: In this section, in particular, we will define some statistical
415: estimators, i.e. functions of the data contained in high frequency
416: time series, and relate them to their probabilistic counterparts
417: defined in the previous section.
418:
419: Our dataset refers to the FX ratio USD/DM over the whole year 1998
420: and the price $S_i$ we consider in this
421: paper is the half sum of bid and ask (mid-price) while $t_i$ denote
422: the time at which the $i$-th price is given. Some automatic filtering
423: procedure is also applied, to remove erroneous recording, which we are
424: able to individuate since they correspond to prices macroscopically
425: different from previous and subsequent ones.
426:
427: \bigskip
428: Let $\mathcal{R} = \left\{ r_i \right\}_{i=0,1,\ldots,L}$ be the
429: series of elementary returns $r_i$ defined as:
430: %
431: $$
432: r_i = \log \frac{S_{i+1}}{S_i} \qquad i=0,1,\ldots,L
433: $$
434: and let $\mathcal{T} =
435: \left\{ \tau_i \right\}_{i=0,1,\ldots,L}$
436: be the series of temporal lags defined as: $\tau_i=t_{i+1}-t_i$.
437:
438:
439: Now consider the series $\mathcal{R}(\Delta,m) = \left\{ r_i(\Delta,m)
440: \right\}_{i=0,1,\ldots,L(\Delta,m)}$; the $r_i(\Delta,m)$ are obtained
441: by summing $m$ consecutive elementary returns (where $m$ is fixed) and
442: subsequently retaining only the $L(\Delta,m)$
443: sums corresponding to a lag in the
444: interval $[\Delta-\epsilon, \Delta+\epsilon]$ (i.e. the sum of the
445: corresponding $m$ elementary lags $\tau_i$ is in the interval
446: $[\Delta-\epsilon,\Delta+\epsilon]$, where $\epsilon$ is also a fixed
447: quantity).
448:
449: The mean and variance of such a series are respectively defined as:
450: %
451: \begin{eqnarray*}
452: \mu(\Delta,m) &=& \frac{1}{L(\Delta,m)} \sum_{i=1}^{L(\Delta,m)}
453: r_i(\Delta,m) \\
454: v(\Delta,m) &=& \frac{1}{L(\Delta,m)} \sum_{i=1}^{L(\Delta,m)}
455: [r_i(\Delta,m) - \mu(\Delta,m)]^2
456: \end{eqnarray*}
457: %
458: We observe that $v(\Delta,m)$ represents an estimation of the quantity
459: $Var[R_\Delta(t) | M_\Delta(t)=m]$ for the calendar time model; and,
460: as pointed out before, we expect it to be a linear function of
461: $\Delta$, should hypothesis H1 be correct.
462: Moreover, in this hypothesis, we expect this variance to be constant
463: with respect to $m$ if $\Delta$ is fixed.
464:
465: Alternatively, considering the business time framework, $v(\Delta,m)$
466: can also be seen an estimator of the quantity $Var[R_m(n) | T_m(n) \in
467: [\Delta-\epsilon, \Delta+\epsilon] ]$ defined in
468: Eq.~(\ref{eq:BT-variance}); should hypothesis H2 be correct we expect,
469: given $m$, that $v(\Delta,m)$ is approximately constant with respect to
470: $\Delta$.
471: Moreover, in this hypothesis, we expect this variance to be linear
472: in $m$ even if $\Delta$ is fixed.
473:
474: With the same set of data $\mathcal{R}(\Delta,m)$
475: we can can compute the empirical pdf
476: of returns with fixed $\Delta$ and with fixed $m$.
477: This pdf is an estimator of $ P[R_\Delta(t) | M_\Delta(t) = m] $
478: and also of
479: $P[R_m(n) | T_m(n) \in [\Delta-\epsilon, \Delta+\epsilon] ]$.
480:
481: \bigskip
482: Consider now the series $\mathcal{R}(\Delta) = \left\{ r_i(\Delta)
483: \right\}_{i=0,1,\ldots,L(\Delta)}$ obtained from $\mathcal{R}$ by
484: summing consecutive elementary returns until the corresponding lag
485: becomes equal or greater than $\Delta$.
486: The number of the elements of this series is $L(\Delta)$ and
487: the mean and variance are respectively defined as:
488: %
489: \begin{eqnarray*}
490: \mu(\Delta) &=& \frac{1}{L(\Delta)}
491: \sum_{i=1}^{L(\Delta)} r_i(\Delta) \\
492: v(\Delta) &=& \frac{1}{L(\Delta)}
493: \sum_{i=1}^{L(\Delta)} [r_i(\Delta) -
494: \mu(\Delta)]^2
495: \end{eqnarray*}
496: %
497: In the calendar time framework, $v(\Delta)$ estimates the quantity
498: $Var[R_\Delta(t)]$, defined in Eq.~(\ref{eq:CT-variance}). In the
499: business time case, instead, $v(\Delta)$ estimates the quantity
500: $Var[R_{M_\Delta(t_n)}(n)]$ in Eq.~(\ref{eq:BT-CT-crossover}). In both
501: cases we expect this quantity to grow linearly with $\Delta$.
502:
503: With the same set of data $\mathcal{R}(\Delta)$
504: we can can compute the empirical pdf
505: of returns with fixed $\Delta$ (any $m$).
506: This pdf is an estimator of $ P[R_\Delta(t)] $.
507:
508: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
509:
510: \section{The choice of the correct model from data analysis}
511: \label{sec:test_results}
512:
513:
514: \begin{figure}
515: \begin{center}
516: \includegraphics[width=0.7\textwidth,angle=-90]{fig1.ps}
517: \caption{ \footnotesize We plot here the statistical estimators
518: $v(\Delta)$ (+ symbols) and $v(\Delta,m)$ with $m=40$ ($\times$ symbols)
519: for different values of the calendar time lag $\Delta$. It can be
520: readily seen that while $v(\Delta)$ varies linearly with $\Delta$, the
521: quantity $v(\Delta,m)$ is approximately constant.
522: Therefore, if the business time lag is fixed (at $m=40$),
523: the variance of the returns does not scale with time lag $\Delta$.
524: This would indicate that business time lag rather than
525: calendar time lag forms the important independent variable.
526: A linear fit was
527: computed in the first case resulting in $v = 6.16E-10 \Delta + 8.26E-8
528: \quad$ and a constant fit in the second resulting in $v = 4.83E-7$.}
529: \label{fig:variance_Delta}
530: \end{center}
531: \end{figure}
532:
533:
534: We have now sufficient information in order to accept or discard
535: hypothesis H1 and H2, as a result of an empirical data analysis.
536:
537: First, we have computed the statistical estimators $v(\Delta)$
538: and $v(\Delta,m=40)$
539: as defined in the previous section and both plotted
540: in Fig.~\ref{fig:variance_Delta} for different values of the
541: calendar time lag $\Delta$. It can be readily seen that while
542: $v(\Delta)$ varies linearly with $\Delta$, the quantity
543: $v(\Delta,m)$ is approximately constant.
544: Indeed, a linear fit was computed in the first case
545: resulting in $v = 6.16E-10 \Delta + 8.26E-8 $
546: and a constant fit in the second resulting in
547: $v = 4.83E-7$.
548:
549:
550: \begin{figure}
551: \begin{center}
552: \includegraphics[width=0.7\textwidth,angle=0]{fig2.ps}
553: \caption{ \footnotesize We plot here the statistical estimator
554: $v(\Delta,m)$ with a fixed $\Delta=1000 \pm 50$
555: for different values of the business time lag $m$.
556: I can be seen that $v(\Delta,m)$ grows with $m$
557: (even if not linearly in all range considered).}
558: \label{fig:variance_m}
559: \end{center}
560: \end{figure}
561:
562:
563: We recall that, according to the calendar time hypothesis the two
564: lines should be equal and proportional to $\Delta$, while in the
565: business time case the former should be proportional to $\Delta$,
566: while the latter should be constant. The
567: corresponding graphs in fig.~\ref{fig:variance_Delta}
568: seem to suggest that the business time model
569: is more likely valid, while the hypothesis of
570: calendar time dynamics seems to be unlikely.
571: The same kind of behavior can be found if one
572: chooses the value of $m$ in a range between 5 and 100.
573:
574: In fig.~\ref{fig:variance_m} we also plot the statistical
575: estimator $v(\Delta,m)$ versus $m$
576: with a fixed $\Delta=1000 \pm 50$.
577: According to the calendar time hypothesis
578: this quantity should be constant
579: while, according to the business time hypothesis ,
580: should grow linearly in $m$.
581: The behavior is not linear in all range
582: but, anyway, $v(\Delta,m)$ grows
583: with respect to $m$, which also supports
584: the business time hypothesis.
585: It should be noticed that the choice of other values
586: of the fixed $\Delta$ would not alter this picture.
587:
588:
589:
590: \begin{figure}
591: \begin{flushleft}
592: \includegraphics[width=0.7\textwidth,angle=-90]{fig3.ps}
593: \caption{ \footnotesize Estimated probability density functions for
594: $\mathcal{R}(\Delta=2sec) = \mathcal{R}(\Delta=2sec,m=1)$,
595: $\mathcal{R}(\Delta=100sec)$ and $\mathcal{R}(\Delta=100sec,m=1)$ in a
596: log-linear plot. The first two pdf (+ symbols) coincide because of the
597: data set characteristics as explained in the text; the pdf for
598: $\mathcal{R}(\Delta=100sec,m=1)$ ($\times$ symbols) is roughly the same
599: of the first two while the pdf for $\mathcal{R}(\Delta=100sec)$
600: (star symbols) is macroscopically different having larger moments.
601: The significance of the plots lies in the fact that
602: if $m=1$, then a large calendar time seems to make no difference,
603: whereas if $m$ is allowed to vary, then the PDF becomes fat,
604: due to return aggregation. }
605: \label{fig:densities}
606: \end{flushleft}
607: \end{figure}
608:
609:
610: \bigskip
611: Second, we consider two distinct series of returns
612: $\mathcal{R}(\Delta,m)$ and $\mathcal{R}(\Delta)$
613: (respectively $a$ and $b$) as defined in the previous section.
614:
615: Since the minimum lag between two consecutive quotes is equal to $2$
616: seconds in the given database, the two series $a$ and $b$ coincide
617: for $\Delta=2$ sec; formally: $\mathcal{R}(\Delta=2 sec, m=1) =
618: \mathcal{R}(\Delta=2 sec)$.
619:
620: We have subsequently compared the estimated probability density
621: functions (pdf) for the series $a$ and $b$ and the results are shown
622: in fig.~\ref{fig:densities}.
623:
624:
625: The figure is a log-linear plot of different probability densities,
626: For $\Delta=2 sec$ the pdf the two cases $\mathcal{R}(\Delta=2sec)$
627: and $\mathcal{R}(\Delta=2sec,m=1)$
628: exactly coincide because of the data set
629: characteristics as just explained. For $\Delta=100 sec$ we observe a
630: remarkable difference between the pdf for the series
631: $\mathcal{R}(\Delta=100sec,m=1)$ and $\mathcal{R}(\Delta=100sec)$.
632: The former, in fact, is roughly the same as
633: $\mathcal{R}(\Delta=2sec)$, while the second is fatter (larger
634: moments).
635:
636: This fact disagrees with Eq.~(\ref{eq:CT-cond_density})
637: which is a consequence of calendar time hypothesis.
638: In fact, according to this equation the two pdf
639: corresponding to $\mathcal{R}(\Delta=100sec,m=1)$
640: and $\mathcal{R}(\Delta=100sec)$ should be equal.
641:
642: On the contrary, one can immediately see that this result is
643: in accordance with Eq.~(\ref{eq:BT-pdf}) and, therefore,
644: with business time hypothesis.
645: In fact, $\mathcal{R}(\Delta=100sec,m=1)$ and
646: $\mathcal{R}(\Delta=2sec,m=1)$ are roughly the same.
647: This experimental equality simply means that given the value of $m$
648: returns are substantially insensitive to $\Delta$
649: as stated in Eq.~(\ref{eq:BT-pdf}).
650:
651: In conclusion, this experimental result
652: provides further evidence that the correct model should be the one of
653: the process evolving over business time (hypothesis H2).
654:
655:
656:
657:
658:
659:
660: \section{Conclusions}
661:
662: In this paper we suggest that the business time approach is perhaps a
663: better way of modeling price dynamics than calendar time.
664: In order to derive some insight from data we neglect
665: possible autocorrelation between returns and possible autocorrelation
666: between lags assuming implicitly that they would only give a second
667: order correction to our findings. With this simplification our
668: results altogether seem to provide enough evidence for the rejection
669: of hypothesis H1 (calendar time model) and the acceptance of
670: hypothesis H2 (business time model).
671: Nevertheless, it should be noticed that hypothesis H1
672: assumes that the sampling process is independent of the price evolution.
673: Therefore, our results do not rule out the continuous time model, but rather
674: they show that the the continuous time model would require
675: correlations between processes $M$ and $S$ to in order to fit the data.
676:
677: The deep reason of the behavior we point out in this paper is that
678: when an asset (at least a forex asset) is not traded,
679: the prices evolution is
680: slow while the evolution is fast when the asset is heavily traded.
681: A faster evolution corresponds to a larger volatility in calendar time
682: \cite{Dacorogna01,Lillo02}, therefore, one could even maintain
683: the calendar point of view, but in this case
684: it should accept a seasonal modulation of volatility.
685: The fact that the evolution of a price is slow
686: when there are few transactions
687: is very well known to practitioners but it is
688: still not accepted in its extremal consequence that prices are frozen
689: when assets are not traded at all.
690: This is because this behavior is
691: in contrast to the stock market experience where opening prices are
692: different from previous night closing prices. Nevertheless the
693: difference between the two markets is not astonishing if one thinks
694: that the stock market is artificially time regulated, while the forex
695: exchange market is an over the counter (OTC) market not subject to any
696: time restriction.
697:
698: \bigskip
699:
700: \bigskip
701:
702: {\bf Acknowledgements}
703:
704: We would like to thank Filippo Petroni for a number
705: of discussion on the subject and for his advice
706: about the manipulation of the high frequency data sets.
707:
708: \bigskip
709:
710:
711: \begin{thebibliography}{99}
712:
713:
714: \bibitem{Baviera99}
715: R. Baviera, M. Pasquini, M. Serva, D. Vergni and A. Vulpiani,
716: {\it Correlations and multyaffinity in high frequency
717: financial data sets},
718: Physica A {\bf 300}, 551-557 (2001).
719:
720: \bibitem{Baviera01}
721: R. Baviera, M. Pasquini, M. Serva, D. Vergni and A. Vulpiani,
722: {\it Forecast in foreign exchange markets},
723: Eur. Phys. J. B {\bf 20} 473-479 (2001).
724:
725: \bibitem{Dacorogna01}
726: M. Dacorogna, R. Gensay, U. Maller, R. Olsen and O. Pictet,
727: {\it An Introduction to High-Frequency Finance},
728: Academic Press. 2001.
729:
730: \bibitem{Lillo02}
731: F. Lillo, J.D. Farmer and R. Mantegna,
732: {\it Muster curve for price-impact function},
733: Nature {\bf 421}, 129-130 (2003)
734:
735: \bibitem{Mainardi00}
736: F. Mainardi, M. Raberto, R. Gorenflo and E. Scalas,
737: {\it Fractional calculus and continuous-time finance II: the
738: waiting-time distribution},
739: Physica A, {\bf 287}, 468-481 (2000).
740:
741: \bibitem{Merton90}
742: R.C. Merton, {\it Continuous-Time Finance},
743: Blackwell Publishers, 1992.
744:
745: \bibitem{Nelson90}
746: D. Nelson,
747: {\it ARCH Models as Diffusion Approximations},
748: Journal of Econometrics, {\bf 45}, 7-38 (1990).
749:
750: \bibitem{Pasquini99}
751: M. Pasquini and M. Serva,
752: {\it Multiscaling and clustering of volatility},
753: Physica A, {\bf 269}, 140-147 (1999).
754:
755: \bibitem{Pasquini00}
756: M. Pasquini and M. Serva,
757: {\it Indeterminacy in foreign exchange markets},
758: Physica A {\bf 277}, 228-238 (2000).
759:
760: \bibitem{Raberto02}
761: M. Raberto, E. Scalas, and F. Mainardi,
762: {\it Waiting-times and returns in high-frequency financial data: an
763: empirical study},
764: Physica A, {\bf 314}, 751--757 (2002).
765:
766: \bibitem{Scalas00}
767: E. Scalas, R. Gorenflo and F. Mainardi,
768: {\it Fractional calculus and continuous-time finance},
769: Physica A {\bf 284}, 376-384 (2000).
770:
771: \bibitem{Taylor86}
772: S. Taylor,
773: {\it Modeling financial time series},
774: John Wiley \& Sons, New York, (1986).
775:
776:
777: \end{thebibliography}
778:
779:
780:
781:
782:
783: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
784:
785:
786: %% \bibliography{../bibliography/quantitative-finance}
787: %% \bibliographystyle{plain}
788:
789:
790: \end{document}
791:
792:
793:
794: