Вы находитесь на странице: 1из 18

STOCK PRICE CLUSTERING AND DISCRETENESS:

THE “COMPASS ROSE” AND COMPLEX DYNAMICS

by

COSTANTINOS E. VORLOW

WORKING PAPER IN ECONOMICS AND FINANCE


No. 04/07

SEPTEMBER 2004

School of Economics, Finance and Business


University of Durham
23-26 Old Elvet
Durham
DH1 3HY
UK
Working Paper
c Uviversity of Durham, School of Economics, Finance and Business
°

Stock price Clustering and Discreteness:


The “Compass Rose” and Complex Dynamics

CONSTANTINOS E. VORLOW
Durham Business School, University of Durham, Mill Hill Lane
Durham, DH1 3LB, UK.
Costas@Vorlow.org∗

Version (July 30th, 2004)

We investigate the “compass rose” (Crack, T.F. and Ledoit, O. (1996), Journal of Fi-
nance, 51(2), pg. 751-762) patterns revealed in phase portraits i.e., scatter diagrams of
daily stock returns against their lagged values. The patterns that appear have been at-
tributed to price discreteness and the tick size. We examine a novel manifestation of
price clustering and discreteness and show that the actual patterns observed may con-
tain information on the complexity of the sequences examined, that possibly enhance
the predictability of stock returns dynamics.

Keywords: Price Clustering and Discreteness; Microstructure; Compass Rose; Recurrence


Quantification Analysis.

1. Introduction
The “compass rose” pattern was identified in literature by Crack and Ledoit [15].
It was revealed in scatter diagrams of daily percentage stock returns against their
lagged values, also called “phase portraits”, “lag-plots” or “delay plots” among
other (see [43] for more definitions). Usually one can discern rays which emanate
from the center that resemble a compass rose (Fig. 1a). The formation of these rays
was attributed to price clustering and discreteness and especially the tick size.a The
above factors are considered an important part of the “market microstructure” lit-
erature with serious implications on risk evaluation, the optimal design of securities
and market efficiency.

∗ Corresponding author: http://www.vorlow.org


a Clustering in stock market prices has been highlighted by [37] and [38], motivated by the original
findings of [40]. [39] also investigated dependencies related to clustering and discreteness. Their
research was followed by [44] who conducted simulations on price rounding and discreteness and
showed that the hypothesis of a geometric Brownian motion for daily and weekly frequency price
dynamics could be rejected. Harris [25,24] has treated these issues more rigorously. He found that
clustering varies with price level, volatility and exchange listing. See also in [33-35] for further
research on this area.

1
2 Vorlow, C.E.

It has been suggested that the appearance of the compass rose pattern is to a
considerable degree subjective and of no use to forecasting [15]. In this paper, fol-
lowing the approach [30] and extending the findings of [6,5,53], we show how from
a simple recoding of the visual information provided from phase portraits we can
obtain results that can be used to enhance our knowledge on the complex dynamical
character of financial prices-returns. To achieve this we employ Recurrence Quan-
tification Analysis (RQA), which especially focuses on the examination of complex
dynamics and was originally introduced by [59] on Recurrence Plots [16,11]. The
suitability of RQA is warranted by its extreme flexibility and its minimal assump-
tions on stationarity and underlying distributional properties of the data under
examination. At the same time we demonstrate a new way to display the effects
of price clustering and discreteness, that can be more clear and precise than the
compass rose.
In the following section, we briefly review the literature around the compass rose.
In section 3 we discuss our methodology and results. We also introduce a new type
of diagram that reveals as well the effect of price clustering and discreteness and
combines both returns and price levels. In section 3.1 we briefly discuss the RQA
methodology and in section 3.2 produce both qualitative and quantitative evidence
on the complex character of the dynamics of the sequences analyzed. Finally in
section 4 we provide our conclusions and pointers for future research.

2. The Compass Rose in scientific literature


The discovery of the “compass rose” formation during the nineties by [15], has
triggered an intensive investigation of the patterns observed mainly in daily returns
sequences. These plots are a standard approach for depicting time dependencies
within the structure of deterministic-chaotic sequences and can provide qualitative
evidence for the existence of attractors [28,29].
As mentioned in the introduction, [15] first documented a distinctive pattern
that arises in these phase portraits: rays emanating from the center, generating a
compass rose like formation (fig. 1). This would not only appear in two dimensional
phase portraits but also in three dimensions. The pattern was due to the price
fluctuations being small integer multiples of a tick size (see also [36]), imposed by
the market practices or regulations. According to [15], given that Rt and Pt denote
returns and closing prices respectively at time t, and h the tick size, one can generate
the following approximation:
Rt+1 (Pt+1 − Pt )/Pt
= ≈ (2.1)
Rt (Pt − Pt−1 )/Pt−1
(Pt+1 − Pt )/Pt−1 Pt+1 − Pt nt+1 h nt+1
≈ = = = , (2.2)
(Pt − Pt−1 )/Pt−1 Pt − Pt−1 nt h nt
where nt ≡ (Pt − Pt−1 )/h is the integer number of ticks by which prices Pt and
Pt−1 differ. According to eq. (2.2), for reasonably close values of the Pt and Pt−1
Stock price Clustering and Discreteness: The “Compass Rose” and Complex Dynamics 3

closing prices, subsequent returns may generate pairs of coordinates for points in
the phase portrait that lead to alignments along the ray joining the origin with the
integer pair (nt , nt−1 ). Other past attempts to produce informative phase portraits
failed to reveal a compass rose like structure. This failure is mainly attributed to
inappropriately formatted diagrams, unfortunate choice of graphic resolutions and
connection of the dots of the portrait (see for example in [17,13,42,19,20,3]).
According to [15], three conditions are necessary for the realization of a compass
rose pattern:

(1) The daily price changes of the stock should be small relative to the price level;
(2) daily price changes should also be effected in discrete jumps of a small number
of ticks and
(3) the price of the stock should vary over a relatively wide range.

It was also reported that phase portraits of returns of some particular assets as
well as index and portfolio returns, may not necessarily reveal such a characteristic
pattern. The reason for the absence of a compass rose pattern has always been
related to the violation of any or all of the three conditions mentioned above. Szpiro
[46] argued that the only necessary condition was that price changes were realized
in discrete jumps and that the approximation of [15] through which we obtain eq.
(2.2), is unnecessary. The requirement of keeping h and nt small in relation to prices
was relaxed and hence eq. (2.1) could be written as:
µ ¶
Rt+1 nt+1 nt h
= 1− . (2.3)
Rt nt Pt
Using eq. (2.3) [46] showed that when share prices span values between P and
P (1 + 1/λ), where nt+1 = λδ, nt = λ² and δ/² determines the ray slope, clusters
of points appeared to be connecting whereas when the span range is wider, these
clusters would overlap. All clusters would connect when prices varied between P
and 2P . This view was criticized by Wang et al. [54] through the introduction of the
element of time. Using a micro-structure approach based on intra-daily UK stock
market data, [54] suggested that the compass rose should become more apparent
as prices were being sampled in higher frequencies. In the same work it was also
proposed that the original conditions of [15] would still be valid assuming that all
the potential levels of returns that are necessary for a complete phase portrait, have
historically occurred and been collected. Thus [54] suggest the compass rose is most
likely to occur if the following two conditions are fulfilled:

(1) the effective tick size is large compared to the standard deviation of the returns
and,
(2) the frequency of the observations increases.

Various other papers have appeared since 1996 on the compass rose theme. Chen
[12] investigated formations with data from Taiwan. This research revealed a new
type of compass rose which contained overlapping square patterns attributed to
4 Vorlow, C.E.

the existence of daily price limits. Clustering in some areas of the phase portraits
was also discovered and it was suggested that these are directly linked with the
predictability of stock returns. The “square compass rose” patterns were assumed
to provide useful information for forecasting with ARMA or GARCH models and
for trading purposes.
Lee et al. [32] examined the compass rose in futures markets data and concluded
that not all contracts showed the pattern. They also found that the frequency of the
data played a significant role in the appearance of the compass rose. Continuing,
Gleason et al. [22] compared intra-day with daily Forex returns and concluded that
the pattern was visible only when the tick/volatility ratio was above some threshold
level. They suggested that price discreteness alone was not a determinant in their
case. More recently, Wang and Wang [55] introduced a quality factor and showed
that with an adequate length of the returns sequence, one could obtain strong
compass rose patterns.
In [15] it was also suggested that price discreteness, as manifested in the compass
rose, could affect the power of various statistical tests. Fang [18] showed that as
autocorrelation estimates were biased because of price discreteness, a number of
random walk tests could also be biased and the underlying asymptotic theory could
be rendered invalid when transactions data was used. Earlier research by Kramer
and Runde [31] had also found that price discreteness can cause the BDS test [10]
to reject a correct null in 80% of the cases. More recently, Amilon [2], examining
low-priced Australian stocks, has shown that GARCH models are misspecified when
applied to returns calculated from discrete prices.
Price clustering and discreteness as manifested in the compass rose, has un-
doubtedly occupied the financial literature since [15]. However the compass rose as
a phase portrait is effectively a tool that illustrates time based interdependencies
in a sequence of observations. In nonlinear-dynamical time series analysis it is used
for revealing the dynamics of the data generating processes (DGP) in 2-dimensional
or 3-dimensional phase spaces. Nonlinear science articles are replete with such di-
agrams that provide qualitative evidence for the existence of complex dynamics in
the DGP.b It has been argued in [15] that the compass rose per se does not offer
a great deal of information. Rather it confirms the clustering and discreteness of
stock closing prices.
Koppl and Nardone [30] have shown that there is a way to provide a more
objective view of the discretized dynamics through the compass rose, contrary to
[15]. From a simple manipulation of the information of the phase portraits, they
obtain angular distributions on their points and produce evidence towards the non-
randomness and predictability of the dynamics. Influenced mainly by [46,30] and the

b Formore information and background theory refer to texts such as [1], [29], [28] and [52] among
others. The main limitation of phase portraits is that they can only provide a visualization of the
dynamics of a sequence up to 3 dimensions. For higher dimensional dynamics one should resort to
other kinds of diagrams such as Recurrence plots. Refer to [16,4,6] for more on this area.
Stock price Clustering and Discreteness: The “Compass Rose” and Complex Dynamics 5

broader bibliography on price clustering and discreteness, in the following section


we examine the case of the compass rose revealing even more information that it
was originally claimed.

3. Methodology and Results


Our data set refers to companies from the FTSE ALL SHARE index, spanning the
period of January 1st, 1970 to May 30th, 2003 (which provides a maximum of 8717
daily observations). Due to space limitations, we provide here graphical results on
a few stocks. The task of generating a compass rose is a trivial matter. In fig. 1 we
have generated the compass rose for the logarithmic returns (not percentage returns
as in [15]) of British-American Tobacco (subfigure a). In subfigure (b) we generated
the scatter diagram of the respective closing prices against their corresponding
logarithmic returns.
Figure 1(b), to our knowledge, depicts for the first time in financial literature,
manifestations of price clustering and discreteness such as the ones in the compass
rose. We can see very clearly the trajectories of points converging to the horizontal
axis. The curvature of these trajectories may be attributed to the slight curvature
observed in the compass rose rays as explained in [46]. In physics or mathematics
the plot of a function against its first derivative is called a “phase plot”. We can
thus loosely term 1(b) as a phase plot. What we observe in these kind of diagrams is
a set of correlation and anticorrelation patterns for various price-returns levels. It is
a really interesting formation that appears in all daily returns-prices sequences we
examined. In [6] it is suggested that there may be a more regular structure in both
compass roses and phase plots, which is revealed once a non-systematic component
is extracted from the returns sequences, using wavelet based methods.
Preliminary results have showed that when we reshuffle returns sequences using
a constraint randomization scheme used in the detection of nonlinear dynamics
[48], both the compass rose and phase plot patterns disappear [53]. This is often
an adequate indication for the presence of nonlinear-deterministic dynamics in the
DGPs of the returns processes [29].

3.1. The Compass Rose and Complex Dynamics


Following [30,5,53], we investigated the angles of the rays (line) formed by the con-
nection of each point of the compass rose with its center and the horizontal axis
(henceforth referred to as “arcs”). As we can see in the case of the BP stock in Fig.
3(b), the distribution of these arc values is multimodal. This is often linked to the
presence of nonlinear deterministic dynamics [28,29]. The time sequence of arc is
depicted in Fig. 3(c) and a sorted version of the values of the arc in Fig. 3(d). In all
subfigures we can discern the levels and positions of the clustering of the arc values.
This provides an alternative view of price discreteness and clustering as manifested
in the compass rose. However it can be regarded as more objective [30]. In [5,53]
evidence is provided in support of predictable non-stochastic dynamics. Here we
6 Vorlow, C.E.

Compass Rose
0.04
0.02
Returns

0.00
−0.02
−0.04

−0.04 −0.02 0.00 0.02 0.04

Lagged Returns

(a) Compass Rose

BAT closing prices against logarithmic returns


0.02
0.01
Returns

0.00
−0.01
−0.02

0 500 1000 1500

Prices

(b) Phase Plot

Fig. 1. Compass rose and phase diagram.

employ a different methodology in independent support the same hypothesis. We


apply Recurrence Quantification Analysis (RQA: [59,57,58]) on the time sequence
of arcs (Fig. 3c). Our approach is not without precedence (e.g., refer to [21,4,26,9];
Stock price Clustering and Discreteness: The “Compass Rose” and Complex Dynamics 7

Phase Plot (detail)


0.01 0.02 0.03
BP Returns

−0.01
−0.03

0 50 100 150 200

BP Closing Prices

Fig. 2. A detail of the phase plot of BP logarithmic prices against logarithmic returns.

for the applicability and intuition behind RQA refer also to [35,49]). RQA was born
as an attempt to quantify the information obtained through Recurrence Plots intro-
duced by [16] (see also [11] for general information on applications and extensions).
Recurrence plots are effectively a graphical representation of nonlinear time correla-
tion matrices. In order to generate such a plot we need first to embed [47,41,45] the
time series and reconstruct the phase space dynamics. More precisely, the following
steps are needed in order to create a recurrence plot:
(1) Generate m embedded vectors y from the original sequence x for time-delay τ
and embedding dimension dE :
y1 , y2 , y3 , ..., ym , (3.1)
where m = N − (dE − 1)τ and
yk = [xk , xk+τ , xk+2τ , ..., xk+(dE −1)τ ] = y(k), (3.2)
for k=1,2,...,m. This is called “embedding” of the sequence x.
(2) Choose a threshold (also termed radius or resolution) ε and then find all vectors
y that are closer than this distance
ky(i) − y(j)k < ε. (3.3)
Here i corresponds to the horizontal axis and j to the vertical one. After the
determination of i and j, the recurrence plot can be generated with a black
8 Vorlow, C.E.

dot on coordinates i, j denoting closeness of embedded vectors y(i) and y(j).c


Hence, the analytical definition of a recurrence plot is:

Ri,j = Θ(ε − ky(i) − y(j)k), i, j = 1, · · · , m, (3.4)

where y denotes the vectors obtained from embedding in equation (3.2), Θ(x) is
the Heaviside function and ε is the predefined threshold as in eq. (3.3), measured
usually in units of the standard deviation of the time series x.

Interpreting a recurrence plot may be a difficult task and may require some ex-
perience. Recurrence plots are symmetric around the 45 degree main diagonal, due
to construction. Usually, any line segments parallel to the main diagonal of the plots
are evidence of recurrent, predictable and possibly chaotic dynamics. The largest
of these segments is also inversely proportional to the largest Lyapunov exponent.
Usually complex, nonlinear deterministic dynamics should be characterized by rel-
atively short line segments whereas very long ones would imply pure determinism.
These segments refer to the time and point in the dynamics of the system where
the attractor revisits the same area of the phase space. Any lack of structure and
points scattered uniformly all over the plot, can be regarded as evidence of random
dynamics. The density of the dark areas differs around the main diagonal when
the dynamics exhibit trends, drifts and nonstationarity. Vertical and horizontal line
segments refer to the time the system remains on a stable state or exhibits slowly
drifting dynamics (“laminar states”). Line segments vertical to the main diagonal
can be an artifact of incorrect embedding.
There are two important points in RQA and recurrence plots in general. Firstly,
it is obvious that the parameters of time delay τ and embedding dimension dE
are crucial for the reconstruction of the phase space dynamics (according to [47]
we require an infinite amount of noise free data and of infinite accuracy). Due
to the noisy character of many financial time series it is very difficult to obtain
a correct embedding as usually dE can not be estimated accurately. However, we
follow here [27,51] which show that even without embedding (de = τ = 1), the
RQA results can reveal qualitatively the same dynamics with those obtained by the
correct “embedding”. The second point is the choice of radius ε. This is usually
determined on the basis of the standard deviation of the x sequence and when the
noise component is known, thresholds (radii) up to five times its standard deviation
may be appropriate [50,34]. A threshold of the magnitude of the lower 10% of the
entire distance range is suggested in [56] and we also employ this strategy for our
analysis that follows (see section 3.2). In [7] it is suggested that the time delay
τ is a more crucial parameter than the embedding dimension dE and that the
average length of line segments parallel to the main diagonal can be insensitive to
the choice of the latter. According to [49], the most crucial statistic to quantify
the predictability of the system is the distribution of diagonals. It is also suggested

c Such recurrence plots are referred to as “thresholded”


Stock price Clustering and Discreteness: The “Compass Rose” and Complex Dynamics 9

that the topological reconstruction of an attractor from the thresholded recurrence


plot is possible, whether the series exhibit deterministic or stochastic dynamics or
even a mixture of these. Hence, current developments in RQA and recurrence plots
in general, suggest that even with inaccurate embedding and loose determination
of the threshold level, we can obtain an almost complete image of the dynamical
information on the system examined through the time series.

Compass Rose Arc Distribution

600
0.02

500
0.01

400
Frequency
0.00

300
Rt−1

200
−0.01

100
−0.02

−0.02 −0.01 0.00 0.01 0.02 −3 −2 −1 0 1 2 3

Rt Arc ranges
(a) (b)

BP Compass Rose arcs BP Compass Rose (sorted) arcs


3

3
2

2
1

1
Arc value

Arc value
0

0
−1

−1
−2

−2
−3

−3

0 2000 4000 6000 8000 0 2000 4000 6000 8000

Time
(c) (d)

Fig. 3. Arc distribution for the BP stock (in radii).


10 Vorlow, C.E.

3.2. RQA results


In Fig. 4 we provide details of the recurrence plot on the arcs of the BP stock
compass rose as well as details from recurrence plot of other known processes. We
can clearly see in Fig. 4(a) and (b) that the BP plots are very different to that
of a Gaussian white noise (Fig. 4e) or a Brownian motion (Fig. 4d). In Fig. 4(c)
we have a recurrence plot of the Lorenz system which is known to exhibit chaotic
(nonlinear deterministic) dynamics whereas in Fig. 4(f) we have a recurrence plot of
a sine function which is purely deterministic (non-chaotic). The dynamics revealed
in Fig. 4 (a,b) are much more consistent with the ones of the Lorenz system of
equations than any other process. Actually the BP recurrence plot is replete with
structures that can be regarded as evidence of predictable and complex dynamics.
However, it becomes clear that the visual inspection of a recurrence plot can provide
a very subjective view of the dynamics. For this reason Zbilut and Webber [59] have
suggested a quantification of the visual information which includes entropy based
criteria, as to avoid any ambiguities in interpretation.
For our analysis, following the considerations outlined in the previous section,
we employed a threshold level (radius ε) range for eq. (3.3) starting from 1 standard
deviation s = 1.8 of the BP compass rose arcs to 2.3 (see table 1). We concentrate
only on a subset of the RQA measures [59,57,33] chosen for their clarity in inter-
pretation for the current exercise. More precisely, in table 1 we report the values of
the RQA on the BP compass rose arcs for the following statistics, according to the
radius ε:

• %REC (Recurrence): Percentage of recurrent points in the plot.


• %DET (Determinism): Percentage of recurrent points which form lines parallel
to the main diagonal. In the case of a deterministic system, these parallel lines
are an indication of the trajectories being close in phase space for time scales
that are equal to the length of these lines.
• MAXL (Maximum Line): The length of the biggest diagonal line in points.
The inverse (1/MAXL) of this measure is related to the maximal (positive)
Lyapunov exponent of the sequence.
• ENT (Entropy): The entropy of the distribution of lines in the plot, measured
in bits. This measure does not refer to the entropy of the sequence analyzed
but is intended as an indicator of “structureness” of the recurrence plot and
the complexity of the deterministic structure of the dynamics.
• TREND: This measures the paling of the plot towards the edges. It is estimated
as the slope of the linear regression line of %REC on the displacement from the
main diagonal (excluding the last 10% of the total range) and is expressed in
units of the percentage of local recurrence per 1000 points. It provides infor-
mation on the stationarity of the process and the presence of any trend or
drift.
• %LAM (“Laminarity”): Analogous to %DET above and measures the percent-
age of vertical lines which indicate the occurrence of laminar states i.e., periods
Stock price Clustering and Discreteness: The “Compass Rose” and Complex Dynamics 11

Recurrence Plot (detail) Recurrence Plot (detail)


600 100

500
80

400
60
300

40
200

20
100

0 0
0 100 200 300 400 500 600 0 20 40 60 80 100

(a) BP arcs (detail) (b) BP arcs (detail)

Recurrence Plot Recurrence Plot


3500 1000

900
3000
800
2500 700

600
2000
500
1500
400

1000 300

200
500
100

0 0
0 500 1000 1500 2000 2500 3000 3500 0 100 200 300 400 500 600 700 800 900 1000

(c) Lorenz (d) Brownian Motion

Recurrence Plot Recurrence Plot


700 900

800

700
650
600

500
600
400

300
550
200

100

500 0
500 550 600 650 700 0 100 200 300 400 500 600 700 800 900

(e) White Noise (f) Sine function

Fig. 4. Recurrence Plots.

of tranquility or slowly drifting dynamics [33].


• TRAP (Trapping Time): The average length of all vertical lines, indicating an
12 Vorlow, C.E.

average time the system is “trapped” into a laminar state as defined above [33].

Table 1. Recurrence Quantification Analysis of the arcs sequence from the compass
rose of BP returns. Results for a range of radii based on the standard deviation
of the process.

Radius % REC % DET MAXL ENT TREND % LAM TRAP


1.80 49.36 36.61 30.00 2.48 −0.01 21.78 6.20
1.83 49.99 37.27 30.00 2.50 −0.01 22.47 6.24
1.85 50.47 37.72 30.00 2.51 −0.01 23.20 6.26
1.88 50.86 38.09 30.00 2.52 0.00 23.72 6.29
1.90 51.78 39.00 30.00 2.54 −0.01 25.02 6.33
1.93 52.23 39.39 30.00 2.55 −0.01 25.50 6.35
1.95 52.68 39.79 30.00 2.56 0.00 26.15 6.38
1.98 53.24 40.38 30.00 2.57 0.03 26.91 6.40
2.00 53.61 40.67 30.00 2.58 0.04 27.36 6.42
2.02 54.02 41.05 30.00 2.59 0.03 27.91 6.45
2.05 55.14 42.30 30.00 2.63 −0.02 29.66 6.53
2.08 55.51 42.60 30.00 2.63 −0.02 30.26 6.55
2.10 55.94 43.04 30.00 2.64 −0.01 30.92 6.57
2.13 56.52 43.56 30.00 2.66 0.00 31.69 6.61
2.15 56.92 43.95 30.00 2.67 −0.00 32.26 6.64
2.17 57.65 44.78 31.00 2.69 −0.01 33.49 6.70
2.20 58.08 45.18 31.00 2.70 −0.01 34.17 6.74
2.23 58.67 45.79 31.00 2.71 −0.00 35.18 6.78
2.25 59.11 46.27 31.00 2.73 0.01 35.93 6.83
2.27 59.54 46.70 31.00 2.74 0.02 36.55 6.86
2.30 59.94 47.10 31.00 2.75 0.03 37.21 6.89

Note: Details
a
Standard deviation of process analyzed: 1.8
b
Type of Norm in eq. (3.4): Euclidean.

Table 1 shows the values of the above statistics for the range of radii chosen. It
is straightforward that the recurrence plot of the BP compass rose arcs will exhibit
significant structure (confirmed also by Fig.4a and 4a). Usually for independently
identically randomly distributed data (iid) one would expect a value of %DET close
to 0 and similarly low levels of %REC as well. The MAXL shows that there is a
maximum period of 30 days where the dynamics of the arc sequence are concentrated
in the same region of the phase space and provide a lower bound for the maximum
Lyapunov exponent of 0.033. This value, although positive, is fairly small and could
suggest some “containment” of the dynamics in the sense that a “large” MAXL
indicates less instability than a smaller one. Concentrating on the %LAM and TRAP
measures, we see that the former ranges between 21.78% and 37.21% which is a high
value and indicates the level of laminar states for the sequence. The former is a value
between 6 and 7 days, indicating the length on average that the sequence dynamics
were “trapped” on a stable condition or exhibited slowly drifting dynamics. The
evidence of RQA here corroborates with the qualitative information obtained from
Stock price Clustering and Discreteness: The “Compass Rose” and Complex Dynamics 13

Table 2. Recurrence Quantification Analysis of the arcs sequence from the compass
rose of BP returns. Results for a range of radii based on percentage of maximum
rescaled distance.
Radius % REC % DET MAXL ENT TREND % LAM TRAP
1% 2.787 0.082 7 0.728 -0.119 0 N/A
2% 4.316 0.138 7 0.696 -0.05 0 N/A
3% 6.085 0.281 8 0.702 -0.02 0.008 5
4% 8.156 0.592 9 0.859 -0.011 0.023 5
5% 9.608 0.937 9 0.929 0.012 0.076 5
6% 11.676 1.648 10 1.057 -0.014 0.123 5
7% 13.227 2.286 11 1.166 0.055 0.227 5.018
8% 15.469 3.37 13 1.283 -0.03 0.36 5.127
9% 17.048 4.33 14 1.355 -0.003 0.603 5.225
10% 18.868 5.56 15 1.435 -0.003 0.949 5.25
11% 20.551 6.809 17 1.509 0.027 1.315 5.303
12% 21.851 7.895 21 1.567 0.082 1.672 5.355
13% 24.674 10.436 21 1.696 -0.114 2.56 5.509
14% 25.799 11.544 21 1.738 -0.065 3.028 5.538
15% 27.352 12.996 21 1.794 -0.005 3.738 5.598
16% 29.099 14.836 21 1.858 -0.005 4.624 5.64
17% 30.562 16.316 22 1.915 0.013 5.289 5.687
18% 32.619 18.684 22 1.99 -0.039 6.528 5.727
19% 33.819 19.934 22 2.026 -0.008 7.219 5.754
20% 35.219 21.412 24 2.069 0.004 8.242 5.777

Note: Details
a
Standard deviation of process analyzed: 1.8
b
Type of Norm in eq. (3.4): Maximum norm.

the recurrence plot itself (Fig. 4). In this case we can not exclude that the sequence of
arcs of the BP compass rose exhibits significant non-random structure and possibly
nonlinear deterministic complexity.
To view the RQA results through a different resolution, we also produce in
table 2, the same quantification measures for radii ranging from 1% to 20% of the
maximum rescaled distance in the recurrence plot of the sequence. Moreover, we
experimented with a different type of norm for eq. (3.4). This produces effectively
an almost identical recurrence plot but because of the different magnitudes of ε and
type of norm, the RQA measures will differ. However, we can observe again here
that for a radius of 10% as in [56], we have a considerable %REC of 18.87% and
low entropy (ENT) which suggests strong structure and predictability of dynamics.
The levels of trapping time (TRAP) are not that different from those of table 1.
It is obvious that even with a different distance measure and levels of threshold,
we still obtain evidence of complex and predictable dynamics. On the basis of our
qualitative and quantitative results, we can also not refute the presence of nonlinear
determinism.
14 Vorlow, C.E.

4. Conclusions and future research


In the previous sections we have demonstrated how the information obtained from
the compass rose and RQA can be used to obtain a view of the complex dynami-
cal character of stock returns processes. Our RQA results suggest the presence of
complex non-stochastic dynamics and predictable structures in the data generating
processes. This is in line with [6,5,53,26]. An obvious problem with financial time
series is that there is no clear information on the noise component of the sequences
(level and type). While recurrence plots and RQA are quite robust to the presence
of noise, it would be an interesting exercise to experiment with blind signal sepa-
ration techniques, such as Independent Components Analysis or Wavelets, so as to
preprocess the time series before analysis (e.g., see [6]). This could eventually pro-
vide an alternative strategy to estimate systematic and unsystematic components
of stock prices and could be of potential use to risk modelling and forecasting. With
respect to the last point, the approach followed here suggests that RQA can form
the basis of a forecasting exercise. This is an area we currently are investigating.

Acknowledgments
We wish to thank James Ramsey, Enrico Capobianco, Abhay Abyankar, Timothy
Crack, Tassos Malliaris and the participants of the Microstructure workshop (orga-
nized by CentER, University of Tilburg (NL), April 2004) for their useful comments
and suggestions during the initial stages and the progress of this work. We also wish
to acknowledge the valuable help and support of Duncan Rand and the University
of Durham “High Performance Computing Service”. The author retains the respon-
sibility for any errors.
Stock price Clustering and Discreteness: The “Compass Rose” and Complex Dynamics 15

References
[1] Abarbanel, H.: 1995, Analysis of Observed Chaotic Data. New York: Springer-Verlag.
[2] Amilon, H.: 2003, ‘GARCH estimation and discrete stock prices: an application to
low-priced Australian stocks’. Economics Letters 81(2), 215–222.
[3] Andreou, A. S., G. Pavlides, and A. Karytinos: 2000, ‘Nonlinear Time-Series Analysis
of the Greek Exchange-Rate Market’. International Journal of Bifurcation and Chaos
10(7), 1729–1759.
[4] Antoniou, A. and C. E. Vorlow: 2000, ‘Recurrence plots and financial time series
analysis’. Neural Network World 10(1-2), 131–146.
[5] Antoniou, A. and C. E. Vorlow: 2004a, ‘Price Clustering and Discreteness: Is there
Chaos behind the Noise?’. Physica A: Statistical Mechanics and Its Applications.
Forthcoming.
[6] Antoniou, A. and C. E. Vorlow: 2004b, ‘Recurrence Quantification Analysis of
Wavelet pre-filtered Index Returns’. Physica A: Statistical Mechanics and Its Aplli-
cations. Forthcoming.
[7] Atay, F. M. and Y. Altıntaş: 1999, ‘Recovering smooth dynamics from time series
with the aid of recurrence plots’. Physical Review E 59(6), 6593–6598.
[8] Ball, C., W. Torous, and A. Tschoegl: 1985, ‘The Degree of Price Resolution: The
Case of the Gold Market’. The Journal of Futures Markets 5(1), 29–43.
[9] Belaire-Franch, J., D. Contreras, and L. Tordera-Lledó: 2002, ‘Assessing nonlinear
structures in real exchange rates using recurrence plot strategies’. Physica D 171(4),
249–264.
[10] Brock, W., W. Dechert, and J. Scheinkman: 1987, ‘A test for independence based
upon the correlation dimension’. Working paper, University of Winsconsin.
[11] Casdagli, M. C.: 1997, ‘Recurrence plots revisited’. Physica D: Nonlinear Phenomena
108(1-2), 12–44.
[12] Chen, A.: 1997, ‘The square compass rose: the evidence from Taiwan’. Journal of
Multinational Financial Management 7(2), 127–144.
[13] Chen, P.: 1993, ‘Searching for Economic Chaos: A Challenge to Econometric Practice
and Nonlinear Tests’. In: R. H. Day and P. Chen (eds.): Nonlinear dynamics and evo-
lutionary economics. Oxford; New York; Toronto and Melbourne: Oxford University
Press, pp. 217–53.
[14] Cho, D. C. and E. W. Frees: 1988, ‘Estimating the Volatility of Discrete Stock Prices’.
Journal of Finance 43(2), 451–66.
[15] Crack, T. F. and O. Ledoit: 1996, ‘Robust Structure without Predictability: The
“Compass Rose” Pattern of the Stock Market’. Journal of Finance 51(2), 751–762.
[16] Eckmann, J., S. O. Kamphorst, and D. Ruelle: 1987, ‘Recurrence Plots of Dynamical
Systems’. Europhysics Letters 4, 973–977.
[17] Enright, A. J.: 1992, ‘Searching for chaotic components in financial time-series’. Ph.d.
thesis, Pace University.
[18] Fang, Y.: 2002, ‘The compass rose and random walk tests’. Computational Statistics
& Data Analysis 39(3), 299–310.
[19] Franses, P. H.: 1998, Time series anslysis models for business and economic forecast-
ing. Cambridge University Press.
[20] Franses, P. H. and D. v. Dijk: 2000, Nonlinear time series models in empirical finance.
Cambridge University Press.
[21] Gilmore, C. G.: 1996, ‘Detecting Linear and Nonlinear Dependence in Stock Returns:
New Methods Derived from Chaos Theory’. Journal of Business Finance & Account-
ing 23(9–10), 1357–1377.
[22] Gleason, K. C., C. I. Lee, and I. Mathur: 2000, ‘An explanation for the compass rose
16 Vorlow, C.E.

pattern’. Economics Letters 68(2), 127–133.


[23] Gottlieb, G. and A. Kalay: 1985, ‘Implications of the Discreteness of Observed Stock
Prices’. Journal of Finance 40(1), 135–153.
[24] Harris, L.: 1990, ‘Estimation of Stock Price Variances and Serial Covariances from
Discrete Observations’. Journal of Financial and Quantitative Analysis 25(3), 291–
306.
[25] Harris, L.: 1991, ‘Stock Price Clustering and Discreteness’. Review of Financial Stud-
ies 4(3), 389–415.
[26] HoÃlyst, J. A., M. Zebrowska, and K. Urbanowicz: 2001, ‘Observations of deterministic
chaos in financial time series by recurrence plots, can one control chaotic economy?’.
The European Physical Journal B 20(4), 531–535.
[27] Iwanski, J. S. and E. Bradley: 1998, ‘Recurrence plots of experimental data: To embed
or not to embed?’. Chaos 8(4), 861–871.
[28] Kantz, H. and T. Schreiber: 1997, Nonlinear Time Series Analysis, No. 7 in Cam-
bridge Nonlinear Science series. UK: Cambridge University Press.
[29] Kaplan, D. T. and L. Glass: 1995, Understanding nonlinear dynamics, Textbooks in
mathematical sciences. New York: Springer-Verlag.
[30] Koppl, R. and C. Nardone: 2001, ‘The Angular Distribution of Asset Returns in Delay
Space’. Discrete Dynamics in Nature and Society 6, 101–120.
[31] Kramer, W. and R. Runde: 1997, ‘Chaos and the compass rose’. Economics Letters
54(2), 113–118.
[32] Lee, C. I., K. C. Gleason, and I. Mathur: 1999, ‘A Comprehensive Examination of the
Compass Rose Pattern in Futures Markets’. The Journal of Futures Markets 19(5),
541–564.
[33] Marwan, N., N. Wessel, U. Meyerfeldt, A. Schirdewan, and J. Kurths: 2002,
‘Recurrence-plot-based measures of complexity and their application to heart rate
variability data’. Physical Review E 66, 026702.1–026702.8.
[34] Matassini, L., H. Kantz, J. Holyst, and R. Hegger: 2002, ‘Optimizing of recurrence
plots for noise reduction’. Physical Review. E, Statistical, Nonlinear, and Soft Matter
Physics 65(1), 021102.
[35] McGuire, G., N. B. Azar, and M. Shelhamer: 1997, ‘Recurrence matrices and the
preservation of dynamical properties’. Physics Letters A 237(1–2), 43–47.
[36] McKenzie, M. D. and A. Frino: 2003, ‘The tick/volatility ratio as a determinant of
the compass rose: empirical evidence from decimalisation on the NYSE’. Accounting
& Finance 43(3), 331–331.
[37] Niederhoffer, V.: 1965, ‘Clustering of Stock Prices’. Operations Research 13(2), 258–
265.
[38] Niederhoffer, V.: 1966, ‘A New Look at Clustering of Stock Prices’. The Journal of
Business 39(2), 309–313.
[39] Niederhoffer, V. and M. Osborne: 1966, ‘Market Making and Reversal on the Stock
Exchange’. Journal of the Anerican Statistical Asssociation 61(316), 897–916.
[40] Osborne, M.: 1962, ‘Periodic Structure in the Brownian Mtion of Stock Prices’. Op-
erations Research 10(3), 345–379.
[41] Packard, N., J. Crutchfield, J. Farmer, and R. Shaw: 1980, ‘Geometry from a Time
Series’. Physical Review Letters 45, 712–715.
[42] Papaioannou, G. and A. Karytinos: 1995, ‘Nonlinear Time Series Analysis of the Stock
Exchange: The Case of an Emerging Market’. International Journal of Bifurcation
and Chaos 5(6), 1557–1585.
[43] Pemberton, M. and N. Rau: 2001, Mathematics for Economists. Manchester Univer-
sity Press.
Stock price Clustering and Discreteness: The “Compass Rose” and Complex Dynamics 17

[44] Rosenfeld, E.: 1980, ‘Stochastic Processes of Common Stock Returns: An Empirical
Examination’. Ph.D. thesis, MIT Sloan School of Management.
[45] Sauer, T., J. Yorke, and M. Casdagli: 1991, ‘Embedology’. Journal of Statistical
Physics 65(3-4), 579–616.
[46] Szpiro, G. G.: 1998, ‘Tick size, the compass rose and market nanostructure’. Journal
of Banking & Finance 22(12), 1559–1569.
[47] Takens, F.: 1981, ‘Detecting strange attractors in turbulence’. In: D. Rand and L. .
Young (eds.): Dynamical Systems and Turbulence. Springer, p. 366.
[48] Theiler, J., S. Eubank, A. Longtin, B. Galdrikian, and J. D. Farmer: 1992, ‘Testing
for nonlinearity in time series: the method of surrogate data’. Physica D: Nonlinear
Phenomena 58(1-4), 77–94.
[49] Thiel, M., M. C. Romano, and J. Kurths: 2004a, ‘How much information is contained
in a recurrence plot?’. Physics Letters A. Forthcoming.
[50] Thiel, M., M. C. Romano, J. Kurths, R. Meucci, E. Allaria, and F. T. Arecchi: 2002,
‘Influence of observational noise on the recurrence quantification analysis’. Physica
D: Nonlinear Phenomena 171(3), 138–152.
[51] Thiel, M., M. C. Romano, J. Kurths, and P. Read: 2004b, ‘Estimation of Dynamical
Invariants without Embedding by Recurrence Plots’. Chaos 14(2), 234–243.
[52] Urbach, R.: 2000, Footprints of Chaos in the Markets: Analyzing Non-Linear Time
Series in Financial Markets and Other Real Systems. London: Financial Times -
Prentice Hall Philadelphia.
[53] Vorlow, C. E.: 2004, ‘Stock Price Clustering and Discreteness: The “Compass Rose”
and Predictability’. Under Review in Economics Letters.
[54] Wang, E., R. Hudson, and K. Keasey: 2000, ‘Tick size and the compass rose: further
insights’. Economics Letters 68(2), 119–125.
[55] Wang, H. and C. Wang: 2002, ‘Visibility of the compass rose in financial asset returns:
A quantitative study’. Journal of Banking & Finance 26(6), 1099–1111.
[56] Webber Jr., C. L. and J. P. Zbilut: 1994, ‘Dynamical assessment of physiological
systems and states using recurrence plot strategies’. Journal of Applied Physiology
76, 965–973.
[57] Zbilut, J. P., A. Giuliani, and C. L. Webber Jr.: 1998, ‘Recurrence quantification
analysis and principal components in the detection of short complex signals’. Physics
Letters A 237(3), 131–135.
[58] Zbilut, J. P., A. Giuliani, and C. L. Webber Jr.: 2000, ‘Recurrence quantification anal-
ysis as an empirical test to distinguish relatively short deterministic versus random
number series’. Physics Letters A 267(2-3), 174–178.
[59] Zbilut, J. P. and C. L. Webber: 1992, ‘Embeddings and delays as derived from quan-
tification of recurrence plots’. Physics letters A 171(3-4), 1991–214.

Вам также может понравиться