- Email: [email protected]

Physica A 353 (2005) 403–412 www.elsevier.com/locate/physa

Time and foreign exchange markets Luca Berardia, Maurizio Servab, a

Dipartimento di Ingegneria Elettrica, Universita` degli Studi, L’Aquila, 67040 Poggio di Roio, AQ, Italy Dipartimento di Matematica and I.N.F.M., Universita` degli Studi, L’Aquila, 67010 Coppito, AQ, Italy

b

Received 29 July 2004; received in revised form 23 December 2004 Available online 14 March 2005

Abstract The deﬁnition of time is still an open question when one deals with high-frequency time series. If time is simply the calendar time, prices can be modeled as continuous random processes and values resulting from transactions or given quotes are discrete samples of this underlying dynamics. On the contrary, if one takes the business time point of view, price dynamics is a discrete random process, and time is simply the ordering according to which prices are quoted in the market. In this paper, we suggest that the business time approach is perhaps a better way of modeling price dynamics than calendar time. This conclusion comes from testing probability densities and conditional variances predicted by the two models against the experimental ones. The data set we use contains the DEM/USD exchange quotes provided to us by Olsen & Associates during a period of one year from January to December 1998. In this period, 1,620,843 quotes entries in the EFX system were recorded. r 2005 Elsevier B.V. All rights reserved. PACS: 89.65.Gh; 05.40.Fb Keywords: Forex markets; Time; Lags; High frequency

1. Introduction In the high-frequency arena there are two mainstreams about modeling the stochastic properties of quotes. The ﬁrst approach is to consider quotations as Corresponding author. Tel.: +39 0862 43 3153; fax: +39 0862 43 3180.

E-mail address: [email protected] (M. Serva). 0378-4371/$ - see front matter r 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.physa.2005.01.049

ARTICLE IN PRESS 404

L. Berardi, M. Serva / Physica A 353 (2005) 403–412

sampled values of an underlying continuous-time random process [1,2]. Sampling is itself a random operation, thus introducing a twofold uncertainty in the price determination [3,4]. In this framework, time in the model ﬂows continuously, and is called calendar time. In the second approach, quoted prices are modeled through a discrete-time stochastic process [5]; in this setting, time is just the natural total order relation among quotations, and it is isomorphic with the set of non-negative integers (time being 0, the time associated to the ﬁrst considered quotation). This is the business time approach, and randomness only enters in the determination of prices. It should be pointed out, however, that the waiting times between two quotes are also random quantities, but they are assumed to not contribute to the price determination process. Whether a calendar-time or a business-time framework should be adopted in modeling the stochastic nature of ﬁnancial quotes has been a longly debated issue by the ﬁnance research community, and it clearly depends on many factors, like, for example, (a) adherence to the physical behavior of reported prices, (b) usefulness in terms of a theory to be developed, and (c) last but not least, a matter of taste. See, for example [6–8]. In this paper, we suggest that business time is perhaps a better tool for modeling the asset dynamics than calendar time. In order to support our claim, we consider (1) returns corresponding to a given calendar time lag and any business time lag, (2) returns corresponding to the same calendar time lag but having a ﬁxed business time lag. We ﬁnd out that their statistical properties are different consistently with the business hypothesis and inconsistently with the calendar one. In practice, we estimate some variances and some probability densities whose behavior is different in the two scenarios. The data set we use contains the DEM/USD exchange quotes taken from Reuters’ EFX pages (the data set having been supplied by Olsen & Associates) during a period of 1 year from January to December 1998. In this period, 1,620,843 quotes entries in the EFX system were recorded. The data set provides a continuously updated sequence of bid and asks exchange quotation pairs from individual institutions whose names and locations are also recorded. The reason for using FX data is that this market is not subject to any working time restriction; in fact, it is open 24 h a day, seven days a week. This is in contrast to stock markets, where artiﬁcial time regulation would have made it more difﬁcult, if not impossible, to ﬁnd out the results outlined in this paper.

2. Business time vs. calendar time 2.1. Calendar time In the calendar time framework, prices are modeled as continuous-time random processes. Clearly, market quotes are not deﬁned for every t 2 R; but only at discrete intervals, whose extensions in time are called calendar lags (usually ranging from 2 s to several minutes, sometimes hours). Nevertheless, according to the calendar time

ARTICLE IN PRESS L. Berardi, M. Serva / Physica A 353 (2005) 403–412

405

picture, prices are usually considered as discrete samples of an underlying continuous-time random process. The model of price dynamics in calendar time therefore has the following structure: Sðt þ DÞ ¼ SðtÞeRD ðtÞ ,

(1)

where SðtÞ and Sðt þ DÞ are the spot prices at times t and t þ D; D is an arbitrary calendar time lag, and RD ðtÞ is the aggregated return of prices over the time interval ½t; t þ D: Considering a framework where prices evolve over the calendar time, it is generally assumed that quotes result from a random sampling at times t0 ; . . . tn of the continuous-time underlying process SðtÞ: In a pure calendar time framework, such a random sampling is uncorrelated with the process SðtÞ itself. We observe however that this is only valid as an approximation; indeed, several studies have shown a weak correlation between the sequence of lags and that of returns, among which we cite [9]. The last assumption usually made in order to complete the model description in the calendar time setting is that the variance of RD ðtÞ is a linear function of the calendar time lag D; i.e., Var½RD ðtÞ ¼ s2 D .

(2)

If the logarithm of SðtÞ has independent increments, the above equation obviously holds and s is the constant volatility. However, it is well known that independence does not hold because of volatility clustering which is due to the correlation of the absolute values of returns [10]. As a consequence, in spite of a constant volatility, one has a time-dependent volatility. Nevertheless, the above behavior of the variance still holds true, but s2 is now the average of the squared volatility. For our purposes we only assume that the above equality holds and we do not need speciﬁc assumptions concerning volatility behavior. Let us deﬁne the process MðtÞ as follows: Mðt þ DÞ ¼ MðtÞ þ M D ðtÞ ,

(3)

where M D ðtÞ represents the number of given quotes (samples) in the interval ½t; t þ D: Clearly, MðtÞ is a non-decreasing random process assuming integer values. We also observe that MðtÞ as a function of t is piecewise constant, and its value increases by one each time a quote is given (i.e., at times t0 ; . . . tn ). Given the assumptions made so far, it follows that the process M and the process S are mutually independent. Hence, it follows that the probability density of returns corresponding to a calendar time lag D is insensitive from the condition that M D ðtÞ is also ﬁxed to a value m. In symbols, P½RD ðtÞjM D ðtÞ ¼ m ¼ P½RD ðtÞ

(4)

and, in particular, the associated variance exhibits the same insensitiveness, Var½RD ðtÞjM D ðtÞ ¼ m ¼ Var½RD ðtÞ ¼ s2 D . Therefore, we can summarize the calendar time hypothesis as follows:

(5)

ARTICLE IN PRESS L. Berardi, M. Serva / Physica A 353 (2005) 403–412

406

Hypothesis H1. The asset prices evolve over calendar time, i.e., according to the model in Eq. (1) and Eq. (2) holds. Moreover, the processes S and M are mutually independent, therefore Eqs. (4) and (5) hold. Let us anticipate that the main argument of this paper is based on the estimation of the quantities in Eqs. (4) and (5). We will show with enough evidence that, the two equalities are largely violated in a way which, on the contrary, is consistent with the business framework. 2.2. Business time In the business-time approach, price dynamics is modeled as a discrete-time random process. Indeed, the time basis is the ordered sequence of times at which prices are quoted in the markets. It is therefore a set isomorphic with the set of nonnegative integers. In such a framework, the statistic model of price dynamics in the business-time framework is the following: Sðn þ mÞ ¼ eRm ðnÞ SðnÞ ,

(6)

where SðnÞ and Sðn þ mÞ are the asset price at business times n and n þ m; while Rm ðnÞ is the aggregated return over m consecutive quotes. It is then clear that the only time dependence affecting the price process is based on the global ordering of events, while the return is independent of calendar lag. Notice that we refer to m as the business time lag as opposed to the calendar time lag D introduced in the previous section. Considering the price dynamics in a business time setting naturally leads to the following assumption: Var½Rm ðnÞ ¼ s^ 2 m

(7)

whose motivation is the same as that provided for the analogous assumption in the calendar-time hypothesis. We also deﬁne the random process as Tðn þ mÞ ¼ TðnÞ þ T m ðnÞ , where TðnÞ is the stochastic calendar time at business time n and T m ðnÞ corresponds to the calendar lag Tðn þ mÞ TðnÞ; i.e., the time elapsed from TðnÞ after the occurrence of m consecutive quotes. It can be readily seen that there is a direct connection between TðnÞ and the process MðtÞ deﬁned in the previous subsection. In fact, MðtÞ ¼ n with t 2 ½TðnÞ; Tðn þ 1ÞÞ; and, moreover, the following relation holds: M T m ðnÞ ðTðnÞÞ ¼ m for an arbitrary positive integer m. Given the assumption of statistical independence between SðnÞ and TðnÞ; for a generic D the following relation holds: P½Rm ðnÞjT m ðnÞ 2 ½D ; D þ ¼ P½Rm ðnÞ ,

(8)

where is a ﬁxed quantity. The above equation states that the probability density of returns corresponding to a business time lag m is insensitive to the condition that the

ARTICLE IN PRESS L. Berardi, M. Serva / Physica A 353 (2005) 403–412

407

calendar time lag is also ﬁxed to a value around D: In particular, we have for the variance Var½Rm ðnÞjT m ðnÞ 2 ½D ; D þ ¼ Var½Rm ðnÞ ¼ s^ 2 m ,

(9)

which is the business time analogue of Eq. (5). Given all the assumptions made so far, we are ready to formulate the hypothesis of prices dynamics in a business time setting. Hypothesis H2. Asset prices follow the model in Eq. (6) and Eq. (7) holds. Moreover, the processes S and T are independent, it thus follows that Eqs. (8) and (9) hold. Before concluding this preliminary outline of the two basic approaches used to describe price dynamics (i.e., calendar time and business time) we also give another important property of some of the quantities involved so far, which will turn useful in the remaining part of the paper. With all the positions previously made, let us ﬁrst observe that the following relation holds1: E½M D ðTðnÞÞ ¼ aD for a suitable constant a: Simply put, this property states that the expected value of the number of quotes in an interval D is proportional to D itself. Finally, considering the composition of the price process in business time and the process representing the number of quotes in a given calendar time lag D; it can be shown that Var½RM D ðtn Þ ðnÞ ¼ s^ 2 E½M D ðTðnÞÞ ¼ s^ 2 aD .

(10)

Thus, in the business time hypothesis, we also expect the variance in (10) to be proportional to D: As already anticipated, all equalities in this subsection are supported by the following statistical analysis conﬁrming the validity of the business time framework.

3. Statistical estimators In this and the next section, we carry out some experimental analysis in order to best ﬁt the description of prices dynamics choosing between the two distinct possibilities concisely modeled by hypotheses H1 and H2. In this section, in particular, we will deﬁne some statistical estimators, i.e., functions of the data contained in high-frequency time series, and relate them to their probabilistic counterparts deﬁned in the previous section. 1

This follows from the stationarity of the process M D ðTðnÞÞ: In particular, E½M D ðTðnÞÞ does not depend on TðnÞ so we drop the sub-case. Moreover, E½M kD ¼ kE½M D ; since the average number of quotes in k intervals of the same length sums up to k times the value for the single interval, from which the proportionality follows.

ARTICLE IN PRESS L. Berardi, M. Serva / Physica A 353 (2005) 403–412

408

Our data set refers to the FX ratio USD/DM over the whole year 1998 and the price S i we consider in this paper is the half-sum of bid and ask (mid-price) while ti denote the time at which the ith price is given. Some automatic ﬁltering procedure is also applied, to remove erroneous recording, which we are able to individuate since they correspond to prices macroscopically different from previous and subsequent ones. Let R ¼ fri gi¼0;1;...;L be the series of elementary returns ri deﬁned as ri ¼ log

Siþ1 Si

i ¼ 0; 1; . . . ; L

and let T ¼ fti gi¼0;1;...;L be the series of temporal lags deﬁned as ti ¼ tiþ1 ti : Now consider the series RðD; mÞ ¼ fri ðD; mÞgi¼0;1;...;LðD;mÞ ; the ri ðD; mÞ are obtained by summing m consecutive elementary returns (where m is ﬁxed) and subsequently retaining only the LðD; mÞ sums corresponding to a lag in the interval ½D ; D þ (i.e., the sum of the corresponding m elementary lags ti is in the interval ½D ; D þ ; where is also a ﬁxed quantity). The mean and variance of such a series are respectively deﬁned as mðD; mÞ ¼

LðD;mÞ X 1 ri ðD; mÞ , LðD; mÞ i¼1

vðD; mÞ ¼

LðD;mÞ X 1 ½ri ðD; mÞ mðD; mÞ2 . LðD; mÞ i¼1

We observe that vðD; mÞ represents an estimation of the quantity Var½RD ðtÞjM D ðtÞ ¼ m for the calendar time model, and, as pointed out before, we expect it to be a linear function of D; should hypothesis H1 be correct. Moreover, in this hypothesis, we expect this variance to be constant with respect to m if D is ﬁxed. Alternatively, considering the business time framework, vðD; mÞ can also be seen as an estimator of the quantity Var½Rm ðnÞjT m ðnÞ 2 ½D ; D þ deﬁned in Eq. (7); should hypothesis H2 be correct we expect, given m, that vðD; mÞ is approximately constant with respect to D: Moreover, in this hypothesis, we expect this variance to be linear in m even if D is ﬁxed. With the same set of data RðD; mÞ; we can compute the empirical pdf of returns with ﬁxed D and with ﬁxed m. This pdf is an estimator of P½RD ðtÞjM D ðtÞ ¼ m and also of P½Rm ðnÞjT m ðnÞ 2 ½D ; D þ : Consider now the series RðDÞ ¼ fri ðDÞgi¼0;1;...;LðDÞ obtained from R by summing consecutive elementary returns until the corresponding lag becomes equal to or greater than D: The number of the elements of this series is LðDÞ and the mean and variance are, respectively, deﬁned as mðDÞ ¼

LðDÞ 1 X ri ðDÞ , LðDÞ i¼1

ARTICLE IN PRESS L. Berardi, M. Serva / Physica A 353 (2005) 403–412

vðDÞ ¼

409

LðDÞ 1 X ½ri ðDÞ mðDÞ2 . LðDÞ i¼1

In the calendar time framework, vðDÞ estimates the quantity Var½RD ðtÞ; deﬁned in Eq. (2). In the business time case, instead, vðDÞ estimates the quantity Var½RM D ðtn Þ ðnÞ in Eq. (10). In both cases we expect this quantity to grow linearly with D: With the same set of data RðDÞ; we can compute the empirical pdf of returns with ﬁxed D (any m). This pdf is an estimator of P½RD ðtÞ:

4. The choice of the correct model from data analysis We now have sufﬁcient information in order to accept or discard hypotheses H1 and H2, as a result of an empirical data analysis. First, we have computed the statistical estimators vðDÞ and vðD; m ¼ 40Þ as deﬁned in the previous section and both are plotted in Fig. 1 for different values of the calendar time lag D: It can be readily seen that while vðDÞ varies linearly with D; the quantity vðD; mÞ is approximately constant. Indeed, a linear ﬁt was computed in the ﬁrst case resulting in v ¼ 6:16E 10D þ 8:26E 8 and a constant ﬁt in the second resulting in v ¼ 4:83E 7: We recall that, according to the calendar time hypothesis, the two lines should be equal and proportional to D; while in the business time case, the former should be proportional to D; while the latter should be constant. The corresponding graphs in

3e-06 2.5e-06

variance

2e-06 1.5e-06 1e-06 5e-07 0 0

500

1000

1500 2000 2500 calendar lag

3000

3500

4000

Fig. 1. We plot here the statistical estimators vðDÞ (+ symbols) and vðD; mÞ with m ¼ 40 ( symbols) for different values of the calendar time lag D: It can be readily seen that while vðDÞ varies linearly with D; the quantity vðD; mÞ is approximately constant. Therefore, if the business time lag is ﬁxed (at m ¼ 40), the variance of the returns does not scale with time lag D: This would indicate that business time lag rather than calendar time lag forms the important independent variable. A linear ﬁt was computed in the ﬁrst case resulting in v ¼ 6:16E 10D þ 8:26E 8 and a constant ﬁt in the second resulting in v ¼ 4:83E 7:

ARTICLE IN PRESS L. Berardi, M. Serva / Physica A 353 (2005) 403–412

410

2.5

x 10-3

Variance

2

1.5

1

0.5

0 0

10

20

30

40

50

m

Fig. 2. We plot here the statistical estimator vðD; mÞ with a ﬁxed D ¼ 1000 50 for different values of the business time lag m. It can be seen that vðD; mÞ grows with m (even if not linearly in all ranges considered).

Fig. 1 seem to suggest that the business time model is more likely valid, while the hypothesis of calendar time dynamics seems to be unlikely. The same kind of behavior can be found if one chooses the value of m in a range between 5 and 100. In Fig. 2 we plot the statistical estimator vðD; mÞ versus m with a ﬁxed D ¼ 1000 50: According to the calendar time hypothesis, this quantity should be constant, while, according to the business time hypothesis, should grow linearly in m. The behavior is not linear in all ranges but, anyway, vðD; mÞ grows with respect to m, which also supports the business time hypothesis. It should be noticed that the choice of other values of the ﬁxed D would not alter this picture. Second, we consider two distinct series of returns RðD; mÞ and RðDÞ (respectively, a and b) as deﬁned in the previous section. Since the minimum lag between two consecutive quotes is equal to 2 s in the given database, the two series a and b coincide for D ¼ 2 s; formally: RðD ¼ 2 s; m ¼ 1Þ ¼ RðD ¼ 2 sÞ: We have subsequently compared the estimated probability density functions (pdf) for the series a and b and the results are shown in Fig. 3. The ﬁgure is a log-linear plot of different probability densities. For D ¼ 2 s the pdf of the two cases RðD ¼ 2 sÞ and RðD ¼ 2 s; m ¼ 1Þ exactly coincide because of the data set characteristics as just explained. For D ¼ 100 s; we observe a remarkable difference between the pdf for the series RðD ¼ 100 s; m ¼ 1Þ and RðD ¼ 100 sÞ: The former, in fact, is roughly the same as RðD ¼ 2 sÞ; while the second is fatter (larger moments). This fact disagrees with Eq. (4), which is a consequence of calendar time hypothesis. In fact, according to this equation, the two pdf corresponding to RðD ¼ 100 s; m ¼ 1Þ and RðD ¼ 100 sÞ should be equal.

ARTICLE IN PRESS

probability density

L. Berardi, M. Serva / Physica A 353 (2005) 403–412

411

0.1

0.01

-0.001

-0.0005

0 return

0.0005

0.001

Fig. 3. Estimated probability density functions for RðD ¼ 2 sÞ ¼ RðD ¼ 2 s; m ¼ 1Þ; RðD ¼ 100 sÞ and RðD ¼ 100 s; m ¼ 1Þ in a log-linear plot. The ﬁrst two pdf (+ symbols) coincide because of the data set characteristics as explained in the text; the pdf for RðD ¼ 100 s; m ¼ 1Þ ( symbols) is roughly the same of the ﬁrst two while the pdf for RðD ¼ 100 sÞ (star symbols) is macroscopically different having larger moments. The signiﬁcance of the plots lies in the fact that if m ¼ 1; then a large calendar time seems to make no difference, whereas if m is allowed to vary, then the PDF becomes fat, due to return aggregation.

On the contrary, one can immediately see that this result is in accordance with Eq. (8) and, therefore, with business time hypothesis. In fact, RðD ¼ 100 s; m ¼ 1Þ and RðD ¼ 2 s; m ¼ 1Þ are roughly the same. This experimental equality simply means that given the value of m, returns are substantially insensitive to D as stated in Eq. (8). In conclusion, this experimental result provides further evidence that the correct model should be the one of the process evolving over business time (hypothesis H2).

5. Conclusions In this paper, we suggest that the business time approach is perhaps a better way of modeling price dynamics than calendar time. In order to derive some insight from data we neglect possible autocorrelation between returns and possible autocorrelation between lags assuming implicitly that they would only give a second-order correction to our ﬁndings. With this simpliﬁcation, our results altogether seem to provide enough evidence for the rejection of hypothesis H1 (calendar time model) and the acceptance of hypothesis H2 (business time model). Nevertheless, it should be noticed that hypothesis H1 assumes that the sampling process is independent of the price evolution. Therefore, our results do not rule out the continuous time model, but rather they show that the continuous time model would require correlations between processes M and S in order to ﬁt the data.

ARTICLE IN PRESS 412

L. Berardi, M. Serva / Physica A 353 (2005) 403–412

The deep reason of the behavior we point out in this paper is that, when an asset (at least a forex asset) is not traded, the prices evolution is slow while the evolution is fast when the asset is heavily traded. A faster evolution corresponds to a larger volatility in calendar time [11,12]; therefore, one could even maintain the calendar point of view, but in this case it should accept a seasonal modulation of volatility. The fact that the evolution of a price is slow when there are few transactions is very well known to practitioners, but it is still not accepted in its extremal consequence that prices are frozen when assets are not traded at all. This is because this behavior is in contrast to the stock market experience where opening prices are different from previous night closing prices. Nevertheless, the difference between the two markets is not astonishing if one thinks that the stock market is artiﬁcially time regulated, while the forex exchange market is an over the counter (OTC) market not subject to any time restriction.

Acknowledgements We would like to thank Filippo Petroni for a number of discussions on the subject and for his advice about the manipulation of the high-frequency data sets. References [1] R.C. Merton, Continuous-Time Finance, Blackwell Publishers, 1992. [2] D. Nelson, ARCH models as diffusion approximations, J. Econometrics 45 (1990) 7–38. [3] F. Mainardi, M. Raberto, R. Gorenﬂo, E. Scalas, Fractional calculus and continuous-time ﬁnance II: the waiting-time distribution, Physica A 287 (2000) 468–481. [4] E. Scalas, R. Gorenﬂo, F. Mainardi, Fractional calculus and continuous-time ﬁnance, Physica A 284 (2000) 376–384. [5] S. Taylor, Modeling Financial Time Series, Wiley, New York, 1986. [6] R. Baviera, M. Pasquini, M. Serva, D. Vergni, A. Vulpiani, Correlations and multiafﬁnity in high frequency ﬁnancial data sets, Physica A 300 (2001) 551–557. [7] M. Pasquini, M. Serva, Indeterminacy in foreign exchange markets, Physica A 277 (2000) 228–238. [8] R. Baviera, M. Pasquini, M. Serva, D. Vergni, A. Vulpiani, Forecast in foreign exchange markets, Eur. Phys. J. B 20 (2001) 473–479. [9] M. Raberto, E. Scalas, F. Mainardi, Waiting-times and returns in high-frequency ﬁnancial data: An empirical study, Physica A 314 (2002) 751–757. [10] M. Pasquini, M. Serva, Multiscaling and clustering of volatility, Physica A 269 (1999) 140–147. [11] M. Dacorogna, R. Gensay, U. Maller, R. Olsen, O. Pictet, An Introduction to High-Frequency Finance, Academic Press, 2001. [12] F. Lillo, J.D. Farmer, R. Mantegna, Muster curve for price-impact function, Nature 421 (2003) 129–130.

Copyright © 2021 COEK.INFO. All rights reserved.