See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225721613

Time is Money

Chapter  in  Lecture Notes in Physics · January 2008

DOI: 10.1007/3-540-45463-2_8 · Source: OAI

CITATION READS

1 2,646

3 authors:

Marcel Ausloos Nicolas Vandewalle

University of Leicester University of Liège
902 PUBLICATIONS   47,029 CITATIONS    384 PUBLICATIONS   7,326 CITATIONS   

SEE PROFILE SEE PROFILE

Kristinka Ivanova
Pennsylvania State University
25 PUBLICATIONS   433 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

VIP Gran View project

Intrinsic Entropy of Exchange- View project

All content following this page was uploaded by Nicolas Vandewalle on 17 December 2013.

The user has requested enhancement of the downloaded file.

 M PRA
Munich Personal RePEc Archive

Time is money

Marcel Ausloos and N. Vandewalle and K. Ivanova

University of Liege

2000

Online at http://mpra.ub.uni-muenchen.de/28703/
MPRA Paper No. 28703, posted 9. February 2011 09:35 UTC
Time is Money

Marcel Ausloos1
, Nicolas Vandewalle1 , and Kristinka Ivanova2,3

1
SUPRAS and GRASP, Institut de Physique B5, Université de Liège, B-4000
Liège, Belgium
2
Department of Meteorology, Pennsylvania State University, University Park, PA
16802, USA
3
permanent address : Institute of Electronics, Bulgarian Academy of Sciences,
Boul. Tzarigradsko chaussée 72, Sofia 1784, Bulgaria

Abstract. Specialized topics on financial data analysis from a numerical and phys-
ical point of view are discussed when pertaining to the analysis of coherent and
random sequences in financial fluctuations within (i) the extended detrended fluc-
tuation analysis method, (ii) multi-affine analysis technique, (iii) mobile average
intersection rules and distributions, (iv) sandpile avalanches models for crash pre-
diction, (v) the (m, k)-Zipf method and (vi) the i-variability diagram technique for
sorting out short range correlations. The most baffling result that needs further
thought from mathematicians and physicists is recalled: the crossing of two mobile
averages is an original method for measuring the ”signal” roughness exponent, but
why it is so is not understood up to now.

1 Introduction
It is fortunate to recall from the start that Louis Bachelier (1870-1946) was a
mathematician at the University of Franche-Comté in Besançon. He was the
first to develop a theory of Brownian motion, – in his Ph. D. thesis [1], in
fact for the pricing of options in speculative markets. Later on he wrote down
what is known as the Chapman–Kolmogorov equation. Alas, he motivated his
approach on what is known nowadays as the efficient market theory, basically
a Gaussian distribution for the price changes. This is known to be incorrect,
at least for economic indices [2,3].
Recently the statistical physics community has been reattracted into in-
vestigating economic and financial problems. Two modern reasons can cer-
tainly be given: (i) economic systems, like stock markets produce quite com-
plex signals due to a high number of parameters involved, and (ii) mod-
els developed so far in actual econometry do seem irrelevant for mimicking
available signals, – at least on the level expected by usual physical models
for natural phenomena. A list of recent progress is too long to be cited or
discusssed here. Several books are already of interest. One aim should be
first to review a few technical details in a global context. Even for a general
audience, with mathematical orientation, it is in fact hopefully possible to


[email protected]

M. Planat (Ed.): LNP 550, pp. 156–171, 2000.

c Springer-Verlag Berlin Heidelberg 2000


 Time is Money 157

give nonrigourous informations on how physicists pretend that an increase in

revenue can be obtained if general rules are found from non linear dynamic-
like analysis of financial time series. Some general views have been already
presented in “Money Games Physicists Play” [4]. More specialized topics are
discussed here as were already sketched in ref. [5].
There are six methods or so that we have been discussing and using in
the Liège GRASP1 , when performing investigations in the context of sorting
deterministic features from apparently stochastic noise contained in economic
and financial data. The investigations pertain to considerations on different
time correlation ranges. First it has been observed a long time ago that stock
market fluctuations were not Brownian motion-like, but some long range
correlation existed [3]. We have tested that idea on foreign exchange currency
(F XC) rates [6]. Using the detrended fluctuation analysis method (DF A), it
was shown that profit making in the F XC market can be made by bankers.
This leads to the introduction of a turbulence-like picture in order to discuss
the spareness and roughness of F XC rates. It can be shown that not all F XC
rates belong to the same so-called universality class, but nobody knows at
this time why, nor what universality classes exist.
Next, some medium range correlation can be discussed. First, a technique
due to stock analysts, known as the mobile (or moving) average technique
which allows for predicting gold or death crosses, whence suggesting buying
or selling conditions will be discussed. It can be shown to be a rather delicate
(a euphemism !) way of predicting what to do on a market. This will lead to
a very interesting, and apparently unsolved problem for mathematicians and
physicists. Moreover, the behavior of major stock market average indices will
be recalled, and it will be observed that so-called crashes have well defined
precursors. The crash of October 1987 could be seen as a phase transition [7].
The amplitude and the universality class can be discussed as well, thereby
indicating how the financial crash of October 1997 could have been (and
was) predicted. This will lead to indicate a microscopic model for such a set
of crashes, model based on sandpile avalanches on fractal lattices. This will
lead to emphasize a very interesting problem for the dynamics of numbers.
Moreover, a claim will be substantiated that the (m, k)-Zipf analysis and
low order variability diagrams can be used for demonstrating short range
correlation evidence in financial data.

2 Detrended Fluctuation Analysis Techniques

The Detrended Fluctuation Analysis technique consists in dividing a time
series or random one-variable sequence y(n) of length N into N/T nonover-
lapping boxes, each containing T points [8]. Then, the local trend in each box
1
GRASP = Group for Research in Applied Statistical Physics
http://www.supras.phys.ulg.ac.be/statphys/statphys.html
158 M. Ausloos et al.

is a priori defined. A linear trend z(n) like

z(n) = an + b, (1)

or a cubic trend like

z(n) = cn3 + dn2 + en + f , (2)

can be assumed [6,9]. In a box, the linear trend might be way-off from the
overall intuitive trend, henceforth shorter scale fluctuations might be missed
if the box size becomes quite large with respect to the apparent short time
fluctuation scale of the signal. Thus the interest of using non linear trends.
The parameters a to f are usually estimated through a linear least-square
fit of the data points in that box. The process is repeated for all boxes. The
detrended fluctuation function F (T ) is then calculated following

(k+1)T
1  2 N
2
F (T ) = |y(n) − z(n)| , k = 0, 1, 2, · · · , ( − 1). (3)
T T
n=kT +1

Usually only one type of trend is taken for the whole analysis, but mixed
situations could be envisaged. Averaging F (T ) over all N/T box sizes centered
on time T gives the fluctuations F (T ) as a function of T . If the y(n) data
are random uncorrelated variables or short range correlated variables, the
behavior is expected to be a power law

F 2 (T )1/2 ∼ T α (4)

with an exponent α =1/2 [8]. An exponent α = 1/2 in a certain range of T

values implies the existence of long-range correlations in that time interval as
e.g. in the fractional Brownian motion [10]. Correlations and anticorrelations
correspond to α > 1/2 and α < 1/2 respectively.2
If a signal has a fractal dimension D, its power spectrum is supposed to
behave like

S(f ) ∼ f −β (5)

where
(3 − β)
D=E+ (6)
2
2
Notice that the procedure to estimate α in [11] includes an a priori integration
of the tested signal, and these authors measure in fact an α
= α + 1.
 Time is Money 159

or

β = 2H + 1, (7)

in terms of the Hurst exponent H such that D = E + 1 − H [10,12-14]; e.g.

β = 2 for Brownian motion. Therefore, since α= H

β = 2α + 1. (8)

In so doing one defines pink (or black) noise depending whether H is less (or
greater) than 1/2. Black noise is related to long-memory effects, and pink
noise to anti-persistence. processes are dominant over the external influences
and perturbations [10].
Such power laws are the signature of a propagation of information across
a hierachical system over very long times. Two cases are shown in Fig. 1. For
time scales above 2 years, a crossover is however observed on studied financial
data. This crossover suggests that correlated sequences have a characteristic
duration of ca. 2 years along the whole financial evolution at least for the 16
years cases studied in ref. [6]. In order to probe the existence of correlated
and decorrelated sequences, a so-called observation box of “length” 2 year
was constructed and placed at the beginning of the data. The exponent α
for the data contained in that box was calculated at each step. The box was
then moved along the historical time axis by 20 points (4 weeks) toward the
right along the financial sequence. Iterating this procedure for the sequence,
a “local measurement” of the degree of “long-range correlations” is obtained,
i.e. a local measure of the Hurst or α exponent. The results indicate that
the α exponent value varies with the date. This is similar to what is also
observed along DN A sequences where the α exponent drops below 1/2 in
so-called non-coding regions.
The opposite has been observed for the breaking apart and disappearance
of stratus clouds (over Oklahoma) [15]. The exponent α jumps from much be-
low 1/2 to about 1/2 and drops back to a low value when the clouds scattered
all over the area. By analogy with DN A and cloud behaviors, our findings
suggest that financial markets loose the controlled structure (following some
loss of “information”) at such a time. It should be noticed that in ref. [6]
both sequences observed around 1983 and 1987 were not immediately seen
from the rough data nor the value of α, and were missed by R/S and Fourier
analysis.
Therefore, two of the main advantages of the DF A over other techniques
like Fourier transform, or R/S methods [3] are that (i) local and large scale
trends are avoided, and (ii) local correlations can be easily probed [6,9]. In
economic data like stock exchange and currency fluctuations, long or short
scale trends are a posteriori obvious and are of common evidence. The DF A
method allows one to avoid such trend effects which can be considered as the
envelope of the signal and could mask interesting details. Thus, we expect
160 M. Ausloos et al.

that DF A allows a better understanding of apparently complex financial

signals.
In so doing, correlations can be sorted out and a strategy for profit making
can be developed in terms of persistence and antipersistence signals [6].

100

10
<F>

0.1
1 10 100 1000 10000
t
1

t*
<F>

0.1

0.01
1 10 100 1000
t
Fig. 1. Linearly Detrended Fluctuation Analysis function for two typical foreign
currency exchange rates, i.e. JPY/USD and NLG/BEF between Jan. 01, 1980 and
Dec. 31, 1995. The Brownian motion behavior corrresponds to a slope 1/2 on this
log-log plot as indicated. The notation
F  is used for
F 2 (T )1/2 for conciseness
in labeling the y-axis.

3 Multiaffine Analysis Techniques

A locally varying value of α suggests a multifractal process. A multi-affine
analysis of several currency exchange rates has been performed in ref.[16–18],
and also for Gold price, and Dow Jones Industrial Average (DJIA) in ref.[17].
In order to do so the roughness (Hurst) exponent H1 and the intermittency
 Time is Money 161

exponent C1 are calculated for the correlation function c(τ ) supposed to

behave like

c1 (τ ) = |y(t) − y(t )|τ ∼ τ H1 . (9)

The technique consists in calculating the so-called ‘qth order height-height

correlation function” [19] cq (τ ) of the time-dependent signal y(t)
q
cq (τ ) = |y(t) − y(t )| τ (10)

where only non-zero terms are considered in the average .τ taken over all
couples (t, t ) such that τ = |t − t |. The roughness exponent H1 describes the
excursion of the signal. For the random walk (Brownian motion), H1 = 1/2 =
H. In the case of white noise H1 = 0 [10]. Notice also that H1 ∼ H2 = H.
The generalized Hurst exponent Hq is defined through the relation

cq (τ ) ∼ τ qHq . (11)

The C1 exponent [19–21] is a measure of the intermittency in the signal y(t)

+
dHq ++
C1 = − . (12)
dq +q=1

which can be numerically estimated by measuring Hq around q = 1.

It appears that in a (H1 , C1 ) diagram (Fig. 2) the currency exchange rates
are dispersed over a wide region around the Brownian motion case (H1 =
0.5, C1 = 0) and have a significantly non-zero thus intermittent component,
i.e. (C1 = 0) – the value of which might depend on the nature of the trading
market, thereby indicating that economic policy seems to be probed through
the analysis and its role should be taken into account in further microscopic
modeling [17].

4 Moving Average Techniques

A stock market index has often been considered as cyclic, but so-called unpre-
dictable events, like crashes are fascinating. It should be noted that they take
place at the end of a period of euphory, when some anxiety builds in. Surely it
is not obvious from the general trend nor from the apparently stochastic noise
when a crash is forthcoming. Can we find some deterministic content beside
the official trend from a basic noise characteristic, e.g. the fractal dimension
has beentaken as a fundamental question.
Roughness or Hurst exponents are commonly measured in surface science
[22] and also in time series analysis [23]. From a usual technique by analysts,
known as the mobile (or moving) average technique, an interesting way can
162 M. Ausloos et al.

0.3
Dow Jones
Gold
0.25

C 0.2 less free market

1

0.15

0.1 fBm

0.05

WN capitalist market
0
0 0.2 0.4 0.6 0.8 1
H1

Fig. 2. Roughness(H1 ), intermittency (C1 ) parameter phase diagram of a few typ-

ical financial data and mathematical, i.e.fractional Brownian motion (fBm) and
white noise (WN) signals.

be proposed for determining H and how to apply it right away to many cases
with persistent or antipersistent correlations.
Consider a time series y(n) given at discrete times n. At time n, the
mobile average ȳ is defined as
N −1
1 
ȳ(n) = y(n − i), (13)
N i=0

i.e. the average of y for the last N data points. One can easily show that
if y increases (resp. decreases) with time, ȳ < y (resp. ȳ > y). Thus, the
mobile average captures the trend of the signal over a time interval N . Such
a procedure can be used in fact on any time series like in atmospheric or
meteorological data, DN A, electronic noise, fracture, internet, traffic, and
fractional Brownian motions.
Let two different mobile averages y¯1 and y¯2 be calculated respectively over
e.g. T1 and T2 intervals such that T2 > T1 . Since T1 = T2 , the crossings of y¯1
and y¯2 coincide with drastic changes of the trend of y(n). If y(n) increases
for a long period before decreasing rapidly, y¯1 will cross y¯2 from above. This
event is called a ”death cross” in empirical finance [24,25]. On the contrary,
if y¯1 crosses y¯2 from below, the crossing point coincide with an upsurge of
the signal y(n). This event is called a “gold cross”. Financial analysts usually
try to ’“extrapolate” the evolution of y1 and y2 expecting “gold” or “death”
crosses. Most computers on trading places are equiped for performing this
kind of analysis and forecasting [25]. Even though mobile averages seem to be
”arbitrary” measures, they present some very practical interest for physicists
and raise new questions. It is paradoxical to have such a type of nalysis
 Time is Money 163

performed while on the other hand the ”efficient market hypothesis” is the
basis of most econometry theories.
It is well known [23] that the set of crossing points between a signal y(n)
and the y = 0-level is a Cantor set with a fractal dimension 1 − H. The
related physics pertain to so-called studies about first return time problems
[26]. However, we have checked that the density ρ of crossing points between
y¯1 and y¯2 curves is homogeneous along a signal and is thus not a Cantor set.
In so doing, the fractal dimension of the set of crossing points is one, i.e.
the points are homogeneously distributed in time along y¯1 and y¯2 . Due to
the homogeneous distribution of crossing points, the forecasting of ‘gold” and
“death” crosses even for self-affine signals y(n) seems unfounded.
However, it is of interest to observe how ρ behaves with respect to the
choice in the relative difference T1 − T2 . More precisely, consider the relative
difference 0 < ∆T < 1 defined as ∆T = (T2 − T1 )/T2 . It has been found
[27] that the density of crossing points ρ(∆T ) curve is fully symmetric, has
a minimum and diverges for ∆T = 0 and for ∆T = 1, with an exponent
which is the Hurst exponent. This remarkable and puzzling result does not
seem to have been mentioned previously to ref. [27] due to the fact that
some theoretical framework for the mobile average method is missing. The
behavior of ρ is analogous to the age distribution of domains after coarsening
in spin-like models [28] and to the density of electronic states on a fractal
lattice in a tight binding approximation. This method of mobile averages can
in fact serve to measure the Hurst exponent in a very fast and continuous
way.

5 Sandpile Model for Rupture and Crashes

Another investigation of the relationship between the trend and local struc-
ture of a signal, like that of stock market measures like the DJIA and the
Standard & Poor 500 (S&P500) has led us into examining regions where
huge variations were taking place. These are usually associated to rupture
phenomena‘and ‘crashes”.
It has been proposed [29] that an economic index y(t) increases as a
complex power law, i.e.
 −m .   /
tc − t tc − t
y(t) = A + B 1 + C sin ω ln +φ for t < tc
tc tc
(14)
where tc is the crash-time or rupture point, A, B, m, C, ω and φ are free
parameters. The law for y(t) diverges (converges) at t = tc if the exponent
m is positive (negative) while the period of the oscillations converges to the
rupture point at t = tc . The real part of the law is similar to that of crit-
ical points at so-called second order phase transitions [30] but generalizes
the scaleless situation to cases in which a discrete scale invariance [31] is
164 M. Ausloos et al.

presupposed when a complex exponent m + iω exists. This relationship was

already proposed in order to fit experimental measurements of sound wave
rate emissions prior to the rupture of heterogeneous composite stressed up
to failure [32]. Such log-periodic corrections have been recently reported in
biased diffusion on random lattices [33], and in our sandpile studies is found
when the underlying base is quasi-fractal [34]. Thus, an avalanche sand pile
model can be imagined for financial indices [35]
A logarithmic divergence, corresponding to the m = 0 limit, can be also
proposed, [36] i.e. the divergence of the index y for t close to tc should be
such that

 .   /
tc − t tc − t
y(t) = A + B ln 1 + C sin ω ln +φ for t < tc
tc tc
(15)
In August 1997, a series of investigations was performed in order to test
the existence of crash precursors. Daily data of the DJIA and the S&P500
was used. A strong indication of a crash event or a rupture point in between
the end of october 1997 to mid-november 1997 was numerically discovered
[37], and later predicted to occur during the week of Oct. 27, 97, and it
was observed to occur on Oct. 27, 97 [38]. This resulted from an analysis of
the similarities between two long periods: 1980-87 and 1990-97. The number
of open days per year on Wall Street is about 261 days, - the exact value
depending on the number of holidays falling on week ends. For the first period,
the analysis was performed on data ending two months before the so-called
Black Monday, i.e. October 19, 1987. For the second period, the data was
considered till August 20th, 1997. In fact, we have separated the search of
the crash day into two problems, that of the divergence itself and that of
the oscillation convergences on the other hand, i.e. (i) tdiv
c for the power (or
logarithmic) divergence and (ii) tosc
c for the oscillation convergence.
Sometimes it might be natural to be contempting, and/or displeased by,
the eye balling technique we are supposedly using [39–42]. We should totally
disagree concerning this gross misunderstanding of our technique. Our sta-
tistical analysis takes into account the approximate location of the maxima,
and in a recent paper it has been precisely shown one good way of taking
the maximum location into account. It is true that in ref.[37] the arrows
pointing at maxima and minima look rather thick, but this is for a display
purpose. In fact the statistical data analysis takes into account the num-
ber of data points in the best possible interval, as it is standard in critical
point (exponent) analysis [43,44]. In so doing the origin of the time interval
is obtained indeed. This time origin the definition of the phase) should lead
to some interesting questions in fact. It might be of interest to recall that
the closing value of the DJIA was used. This is not necessarily the intraday
maximum value in fact, nor the intraday average value. One might wonder if
the former or the latter would give a better estimate of the upper bound of
 Time is Money 165

the predicted crash day. One might search whether rather than the closing
value or the maximum the average DJIA, or average of range, or something
else over a one day interval should be better used for better predictability,
etc. It is known that there are larger fluctuations at the begining and at the
end of a day. These are left for further investigations. The stability of the
fit parameters can also be checked on these closing values with respect to
random noise and through a Monte-Carlo data in order to take into account
some sort of uncertainty in a bivariate data analysis (with error bars on the
x and y axes). However due to the apparent precision of the technique at this
time no robustness test has been performed as of now.
It should be pointed out that we do not expect any real divergence in the
stock market indices; this is total non sense of course. However a divergence is
predicted by us to occur at some upper bound of tc . This is exactly the same
as in phase transitions, where there is never any infinite divergence at the
critical temperature. The divergence of the correlation length, specific heat,
etc. is a virtual (mathematical) image of physical reality. There is no infinity
(nor zero in fact) in physics due to finite size effects, inhomogeneities, noise,
etc. Therefore to argue on the true existence of zeroes and infinity [33,40–42]
is rather meaningless. We consider that to give an upper bound is certainly
an as good predictive technique in data analysis and for modeling, as good
as to give a determinis tic finite value at tc .
Moreover a true drop certainly exists at a crash and is the signature of
the crash, and the formula of ref. [29] would seem therefore appropriate.
According to ref. [29,40], the drop goes to a finite value. Notice that there is
some sense indeed to examine the size of jumps at crashes though. Such an
attempt has been made in ref. [7].

index - (period) tdiv

c (m = 0) tdiv
c (m = 0) tosc
c
DJIA (80-87) 87.85 ± 0.02 88.46 ± 0.04 87.91 ± 0.10
DJIA (90-97) 97.92 ± 0.02 98.68 ± 0.04 97.89 ± 0.06
S&P500 (80-87) 87.89 ± 0.03 88.78 ± 0.05 87.88 ± 0.07
S&P500 (90-97) 97.90 ± 0.02 98.67 ± 0.04 97.85 ± 0.08
Table 1: Fundamental parameters found for the DJIA and S&P500 indices
during 1980-87 and 1990-97 periods. Time is expressed in years. The notations for
tc are such that e.g. 97.90 means the calendar date corresponding to the 90-th day
as if there are 100-days in 1997. Two values of tdiv
c correspond to a fit using a
logarithmic divergence (m = 0) and a fit using a power law divergence (m = 0)
respectively. The true date of the October 1987 crash in the above units gives
tc = 87.79 and for the October 1997 crash is tc = 97.81, i.e. quasi the predicted
dates.

6 (m, k)-Zipf Techniques

For testing and emphasizing short range correlations, the (m, k)-Zipf and the i-
Variability Diagram (i-VD) techniques have been used. The Zipf analysis consists
166 M. Ausloos et al.

in counting the number of words of a certain type appearing in a text, calculating

the frequency of occurence fo of each word in a given text, and sorting out the
words according to their frequency, i.e. a rank R is assigned to each word, with
R = 1 for the most frequent one, and rank RM for the word appearing the less.
Moreover call fM the frequency (occurrence) of the most often observed word.
For natural languages, one observes a power law

fo ∼ R−ζ (16)

with an exponent ζ close to one for any language. This has been applied to various
complex signals or “texts” [45–47], economy (size of sales and firms) data [48],
financial data [5], meteorological [17,49], sociological [50] or even random walk
[51] and percolation [52] after translating whatever signal into a text based on an
alphabet of k characters. The appearance of this power law is due to the presence
of a so-called hierarchical structure of long range correlations in words, sentences,
paragraphs, and so on for the given set of characters in an alphabet used for writing
a text [46]. A simple extension of the Zipf analysis is to consider m-words only, i.e.
the words strictly made of m characters without considering the white spaces.
Let for the sake of argument, only a binary alphabet with u and d characters,
and the translation of a signal into a text (Fig.3). Let the probability to find a u
in the text be p. The deviation from p = 12 , i.e. p = 12 + : where 0 ≤ : ≤ 12 is called
the bias. The bias is in fact a local measure of the trend in a stock price or index
value.
We have chosen to examine (m = 6, k = 2) cases. It may be remarked that this
is useful for attempting to observe short range (weekly) fluctuations in (weather or
financial) data for example. The aim of the study is to find the exponent ζ. By the
way, it has been conjectured [53,54] that ζ is related to the Hurst exponent H, thus
to the fractal dimension D [10,12,14] of the signal as

ζ = |2H − 1|. (17)

Therefore, for H different from 0.5, and thus ζ different from zero, the signal is
not Brownian-like, whence some predictability can be expected because non trivial
correlations exist between successive daily fluctuations.
One case can serve as an illustration herebelow. As experimental data among the
many indices and stocks available on Internet, let us choose an insurance company
Oxford Health Plan (OXHP ), treated on the NASDAQ. From Aug. 8, 1991 till
March 15, 1999, this consists in about 1900 data points [15]. The daily closing price
signal is shown in Fig. 4. The fractal dimension D, or power spectrum characterized
by β, the DF A exponent α and ζ can be examined as well. The latter from a Zipf
analysis for the OXHP closing price is given in Fig. 5, and values of exponents
summarized in Table 2. The corresponding results for the Brownian motion are
also given and serve as an estimate of the validity of the analysis.

Name of the signal ζ α β

Brownian motion 0.08 ± 0.0007 0.50 ± 0.01 1.79 ± 0.20
OXHP: Closing price 0.27 ± 0.02 0.56 ± 0.03 1.75 ± 0.25

Table 2: The ζ, α and β values for OXHP closing price for the time spans from
August 8, 1991 till March 15, 1999.
 Time is Money 167

u d u d d d u u d d u d u u d d
Fig. 3. Translation of part of a random walk sequence (“fluctuations”) into a binary
sequence made of two characters u and d.

100

90

80

70
OXHP: Closing Price

60

50

40

30

20

10

0
92 93 94 95 96 97 98 99
Time

Fig. 4. Daily closing price of Oxford Health Plan (OXHP ) stock, treated on the
NASDAQ, from mid-91 till Jan. 99, i.e. about 1900 data points.
168 M. Ausloos et al.

0.4

ζ
2

0.3
1.5
log (f /f)
M

1
0.2
10 20 30 40
0.5
N

0 0.5 1 1.5 2
log (R /R)
Fig. 5. (6,2)-Zipf analysis for the OXHP stock closing price; the estimate of the ζ
exponent is shown in the inset for R small and as a function of the number N of
data points used to calculate the best slope from the main graph.

7 Basics of i-Variability Diagram Techniques

One disadvantage of the Zipf-method is that it is not possible to distinguish
between persistent and antipersistent sequences. Only the departure from
randomness is easily observed. Another way to sort out short range correla-
tions is the i-Variability Diagram technique, used for example in heart beat
[55] and meteorological [49] studies. Recall that the first return map (ri , ri−1 )
or the τ -return map (ri , ri−τ ) of a signal are often used for revealing a pos-
sible dominant correlation between the events of the data set. This leads to
studies of strange attractors and the embedding dimension of a signal.
The return map of the first derivative of the signal, i.e. the so-called
first order variability diagram (1-V D) [55] correlates every three consecutive
points of the series,

si+1 = ri+1 − ri (18)

si = ri − ri−1

The curvature of the signal, thus relating every four consecutive events as

ui+1 = ri+1 − 2ri + ri−1 (19)

ui = ri − 2ri−1 + ri−2 .

is called the second order variability diagram (2-V D).

The links between the 4-points e.g. defining the above relationship are
seen through a phase space diagram for the curvature. A non-trivial shape of
 Time is Money 169

the point cloud and point distribution itself on such a diagram indicate an
asymmetry between the different consecutive curvatures, and can be used for
predictability.
It has been found [56] that for free √
market financial series (DJIA and Gold
price) the local trend behaves like a t. However the BGL/U SD exchange
rate variability seems to be different: the set of events leads to a line structure
with slope I = −1.21 in the curvature return map. The differences can be
conjectured to depend on economic policy grounds.
A combination of Zipf and i-V D has been recently attempted for the local
curvature correlations in financial signals [56]. This method leads to suggest
tests based on microscopic models.

Acknowledgments

KI thanks the hospitality of T. Ackerman and the Department of Meteorology

at Penn State University and NRC for a research grant during which part of
this work was completed. Thanks to the F-532 grant of the Bulgarian NFSI
as well. MA and NV thank the ARC 94-99/174 for financial support. MA
thanks the organizers of the Ecole thématique de Chapelle des Bois and in
particular Michel Planat for inviting him to present the above results.

References
1. L. Bachelier, Théorie de la spéculation, Ph. D. thesis, Acad. Paris (1900)
2. R.N. Mantegna and H.E. Stanley Nature, 376, 46-49 (1995)
3. E.E. Peters, Fractal Market Analysis : Applying Chaos Theory to Investment
and Economics, (Wiley Finance Edition, New York, 1994)
4. M. Ausloos, Europhys. News 29, 70-72 (1998)
5. N. Vandewalle and M. Ausloos, in Fractals and Beyond. Complexity in the Sci-
ences, M.M. Novak, Ed. (World Scient., Singapore, 1999) p.355-356.
6. N. Vandewalle and M. Ausloos, Physica A, 246, 454-459 (1997)
7. N. Vandewalle, Ph. Boveroux, A. Minguet and M. Ausloos, Physica A 255,
201-210 (1998)
8. H. E. Stanley, S. V. Buldyrev, A. L. Goldberger, S. Havlin, C.-K. Peng and M.
Simmons, Physica A 200, 4-24 (1996)
9. N. Vandewalle and M. Ausloos, Int. J. Comput. Anticipat. Syst. 1, 342-349
(1998)
10. B. J. West and B. Deering, The Lure of Modern Science: Fractal Thinking,
(World Scient., Singapore, 1995)
11. C.-K. Peng, J.M. Hausdorff, S. Havlin, J.E. Mietus, H.E. Stanley, and A.L.
Goldberger, Physica A 249, 491-500 (1998)
12. M. Schroeder, Fractals, Chaos and Power Laws, (W.H. Freeman and Co., New
York, 1991)
13. K. J. Falconer, The Geometry of Fractal Sets, (Cambridge Univ. Press, Cam-
bridge, 1985)
170 M. Ausloos et al.

14. P. S. Addison, Fractals and Chaos, (Inst. of Phys., Bristol, 1997)

15. K. Ivanova, unpublished
16. N. Vandewalle and M. Ausloos, Int. J. Phys. C 9, 711-720 (1998)
17. K. Ivanova and M. Ausloos, Eur. J. Phys. B 8 665-669 (1999)
18. N. Vandewalle and M. Ausloos, Eur. J. Phys. B 4, 257-261 (1998)
19. A. L. Barabási and T. Vicsek, Phys. Rev. A 44, 2730-2733 (1991)
20. A. Davis, A. Marshak, and W. Wiscombe, in Wavelets in Geophysics, E.
Foufoula-Georgiou and P. Kumar, Eds. (Academic Press, New York, 1994) pp.
249-298
21. A. Marshak, A. Davis, R. Cahalan, and W. Wiscombe Phys. Rev. E 49, 55-69
(1994)
22. A.-L. Barabási and H.E. Stanley, Fractal Concepts in Surface Growth, (Cam-
bridge Univ. Press, Cambridge, 1995)
23. J. Feder, Fractals, (Plenum, New-York, 1988)
24. E. Labie, private communication
25. A.G. Ellinger, The Art of Investment, (Bowers & Bowers, London, 1971)
26. J.-P. Bouchaud and A. Georges, Phys. Rep. 195, 127-294 (1990)
27. N. Vandewalle and M. Ausloos, Phys. Rev. E 58, 6832-6834 (1998)
28. L. Frachebourg, P.L. Krapivsky, and S. Redner, Phys. Rev. E 55, 6684-6689
(1997)
29. D. Sornette, A. Johansen, and J.-Ph. Bouchaud, J. Phys. I (France) 6, 167-175
(1996)
30. H.E. Stanley, Phase Transitions and Critical Phenomena, (Oxford Univ. Press,
Oxford, 1971)
31. D. Sornette, Physics Reports 297, 239-270 (1998)
32. J.C. Anifrani, C. Le Floc’h, D. Sornette and B. Souillard, J. Phys. I (France)
5, 631-63 (1995) 277
33. D. Stauffer and D. Sornette, Physica A 252, 271- (1998)
34. N. Vandewalle, R. D’hulst, and M. Ausloos, Phys. Rev. E 59, 631-635 (1999)
35. M. Ausloos and N. Vandewalle, unpublished
36. N. Vandewalle and M. Ausloos, Eur. J. Phys. B 4, 139-141 (1998)
37. H. Dupuis, Trends Tendances 22(38), (september 18, 1997) p.26-27
38. H. Dupuis, Trends Tendances 22(44), (october 30, 1997) p.11
39. D. Daoût, Le Vif L’Express (october 30, 1997) p.124-125
40. D. Sornette and D. Stauffer, private communicaion
41. J.-Ph. Bouchaud, P. Cizeau, L. Laloux, and M. Potters, Phys. World 12, 25-29
(1999)
42. L. Laloux, M. Potters, R. Cont, J.-P. Aguilar, and J.-P. Bouchaud, Europhys.
Lett. 45, 1-5 (1999)
43. J.R. Macdonald and M. Ausloos, Physica A 242, 150-160 (1997)
44. M. Ausloos, J. Phys. A 22, 593-609 (1989)
45. G.K. Zipf, Human Behavior and the Principle of Least Effort, (Addisson-
Wesley, Cambridge, MA, 1949)
46. W. Ebeling and A. Neiman, Physica A 215, 233-241 (1995)
47. B. Vilensky Physica A 231, 705-711 (1996)
48. M.H.R. Stanley, S.V. Buldyrev, S. Havlin, R.N. Mantegna, M.A. Salinger, and
H.E. Stanley, Ecomonics Letters, 49, 453-457 (1995)
49. K. Ivanova, M. Ausloos, A. Davis, and T. Ackerman, XXIV General Assembly
of the EGS, the Hague, Netherlands, April 19-23 (1999)
 Time is Money 171

50. M. Marsili and Y.-C. Zhang, Phys. Rev. Lett. 80, 2741-2744 (1998)
51. N. Vandewalle and M. Ausloos, unpublished
52. M. S. Watanabe Phys. Rev. E 53, 4187-4190 (1996)
53. A. Czirok, R.N. Mantegna, S. Havlin, and H.E. Stanley, Phys. Rev. E 52, 446-
452 (1995)
54. G. Troll and P.B. Graben, Phys. Rev. E 57, 1347-1355 (1998)
55. A. Babloyantz and P. Maurer, Phys. Lett. A 221, 43-55 (1996)
56. K. Ivanova and M. Ausloos, Physica A 265, 279-286 (1999).

View publication stats