1: \documentclass{appolb}
2: \usepackage{epsfig}
3:
4: \begin{document}
5:
6: \title{Asymmetric matrices in an analysis of financial correlations}
7:
8: \author{J.~Kwapie\'n$^1$, S.~Dro\.zd\.z$^{1,2}$, A.Z.~G\'orski$^1$,
9: P.~O\'swi\c ecimka$^1$
10: \address{$^1$ Institute of Nuclear Physics, Polish Academy of Sciences,
11: Krak\'ow, Poland \\
12: $^2$ Institute of Physics, University of Rzesz\'ow, Rzesz\'ow, Poland}
13: }
14:
15: \maketitle
16:
17: \begin{abstract}
18:
19: Financial markets are highly correlated systems that reveal both the
20: inter-market dependencies and the correlations among their different
21: components. Standard analyzing techniques include correlation coefficients
22: for pairs of signals and correlation matrices for rich multivariate data.
23: In the latter case one constructs a real symmetric matrix with real
24: non-negative eigenvalues describing the correlation structure of the data.
25: However, if one performs a correlation-function-like analysis of
26: multivariate data, when a stress is put on investigation of delayed
27: dependencies among different types of signals, one can calculate an
28: asymmetric correlation matrix with complex eigenspectrum. From the Random
29: Matrix Theory point of view this kind of matrices is closely related to
30: Ginibre Orthogonal Ensemble (GinOE). We present an example of practical
31: application of such matrices in correlation analyses of empirical data. By
32: introducing the time lag, we are able to identify temporal structure of
33: the inter-market correlations. Our results show that the American and
34: German stock markets evolve almost simultaneously without a significant
35: time lag so that it is hard to find imprints of information transfer
36: between these markets. There is only an extremely subtle indication that
37: the German market advances the American one by a few seconds.
38:
39: \end{abstract}
40:
41: \PACS{89.75.-k, 89.75.Da, 89.75.Fb, 89.65.Gh}
42:
43: \section{Introduction}
44:
45: A number of studies have shown that different financial markets reveal
46: hierarchical structure~\cite{mantegna99,bonanno00,giada01,plerou02,%
47: onnela03,dimatteo04,kim05,mizuno05} that can be approximated by factor and
48: group models (e.g.~\cite{roll80,noh00} for the stock market case). At the
49: level of financial data, these structures are determined principally by
50: strength of correlations in returns of different stocks, currencies or
51: other assets. The most popular methods of such an analysis are based on
52: the calculation of correlation matrices from multivariate time series of
53: returns. The correlation matrices can then be diagonalized in order to
54: obtain spectra of their eigenvalues and
55: eigenvectors~\cite{laloux99,plerou02,kwapien06} or can serve as
56: a source for the construction of minimal spanning
57: trees~\cite{mantegna99,bonanno01,mizuno05,gorski06}. In the standard
58: approach, in which the correlations between all analysed assets are taken
59: into consideration, the correlation matrix is by construction symmetric
60: due to the correlation coefficient invariance under a swap of signals.
61: This obviously leads to a real eigenspectrum of the matrix. Usually
62: properties of the empirical correlation matrix are compared with universal
63: predictions of the adequate, Wishart ensemble of random matrices and the
64: identified deviations are considered as an indication of actual
65: correlations among data.
66:
67: In principle, however, there is no restriction imposed on the symmetry
68: property of a correlation matrix: it may well be antisymmetric or even
69: completely asymmetric, depending on which signals are used in the
70: calculations. For example, if there are two separate sets of signals and
71: the correlations are calculated only across these two sets, the resulting
72: matrix can no longer be symmetric and, consequently, its eigenspectrum can
73: be complex. However, there is still a non-zero probability that some of
74: the eigenvalues and eigenvectors are real. As long as a distribution of
75: the correlation matrix elements is close to a Gaussian, the most relevant
76: random matrix ensemble, against which the results should be tested, is the
77: Ginibre Orthogonal Ensemble (GinOE)~\cite{ginibre65}. For the financial
78: data characterized by fat tails of p.d.f. this assumption can also be made
79: provided the time series under study are sufficiently long.
80:
81: At present one observes in literature a growing interest in theoretical
82: research on properties of real asymmetric and, more generally,
83: non-Hermitean random matrices. This interest is motivated by a broadening
84: spectrum of applications of such matrices which includes, among others,
85: random networks~\cite{timme04}, quantum chaos~\cite{fyodorov97}, quantum
86: chromodynamics~\cite{stephanov96,akemann04} and brain
87: research~\cite{kwapien00}. An issue which we address in this work and
88: which can serve as an example of application of the asymmetric correlation
89: matrices to empirical data can be related to a globalization of financial
90: markets. We investigate the cross-market correlations between returns of
91: stocks traded on two large but geographically distant markets: New York
92: Stock Exchange and Deutsche B\"orse. Our objective is to identify the
93: strength of the instanteous as well as the time lagged dependencies
94: between evolution of these two markets.
95:
96: \section{Methods}
97:
98: We begin with presenting a brief construction scheme of an asymmetric
99: correlation matrix and a short description of basic properties of GinOE.
100: Let us consider the two disjoint sets ${\rm X}, {\rm Y}$ each consisting
101: of $N$ assets and denote by $\{x_i^{(s)}\}_{i=1,...,T}$ and
102: $\{y_i^{(t)}\}_{i=1,...,T}$ the time series of normalized logarithmic
103: returns of assets $s \in {\rm X}$ and $t \in {\rm Y}$ ($s,t=1,...,N$).
104: For each set we construct an $N \times T$ data matrix ${\bf M}$ and
105: the correlation matrix ${\bf C}_{\rm XY}$ according to formula
106: \begin{equation}
107: {\bf C}_{\rm XY} = {1 \over T} {\bf M}_{\rm X} {\bf M}_{\rm Y}^{\rm T}.
108: \label{corrmatx}
109: \end{equation}
110: Each matrix element $-1 \le C_{s,t} \le 1$ is the Pearson
111: cross-correlation coefficient for assets $s$ and $t$ ($C_{s,t} \neq
112: C_{t,s}$). In the next step the correlation matrix can be diagonalized by
113: solving the eigenvalue problem
114: \begin{equation}
115: {\bf C}_{\rm XY} {\bf v}_k = \lambda_k {\bf v}_k, \ \ k = 1,...,N
116: \end{equation}
117: which provides us with a complete spectrum of generally complex
118: eigenvalues $\lambda_k$ and pairs of conjugated eigenvectors ${\bf v}_k$.
119: The assumption ${\rm Y} = {\rm X}$ in Eq.(\ref{corrmatx}) leads to the
120: standard definition of a symmetric correlation matrix ${\bf C}_{\rm XX}$
121: with a real eigenspectrum.
122:
123: \begin{figure}
124: \epsfxsize 8cm
125: \hspace{2.0cm}
126: \epsffile{fig1.eps}
127: \caption{Probability density function of complex eigenvalues of
128: exemplary GinOE random matrix ($N=30$) obtained by averaging the spectra
129: over 100000 individual matrix realizations. Colors range from black
130: (probability density close to zero) to red (highest probability density).}
131: \end{figure}
132:
133: Properties of the empirical correlation matrix have to be tested against a
134: null hypothesis of completely random correlations characteristic for
135: independent signals. Random Matrix Theory (RMT) offers some analytic
136: results for a corresponding ensemble of real asymmetric matrices, i.e. the
137: Ginibre Orthogonal Enseble~\cite{ginibre65} defined by the Gaussian
138: probability density
139: \begin{equation}
140: P_{\rm GinOE}(\mathcal{C}) = (2 \pi)^{-N^2/2} \exp [-{\rm
141: Tr}(\mathcal{CC}^{\rm
142: T}/2)],
143: \label{pdfginoe}
144: \end{equation}
145: where $\mathcal{C}$ stands for $N \times N$ real matrix. In the limit of
146: $N \to \infty$ the eigenvalue spectrum of a GinOE matrix is homogeneous
147: and assumes a regular elliptic shape in the complex plane~\cite{sommers88}
148: \begin{displaymath}
149: p(\lambda) = \left\{ \begin{array}{cc}
150: (\pi a b)^{-1}\ , & ({{\rm Re} z \over a})^2 + ({{\rm Im} z
151: \over b})^2 \le 1\\
152: 0\ , & ({{\rm Re} z \over a})^2 + ({{\rm Im} z \over b})^2 > 1,
153: \end{array} \right.
154: \label{ginoelimit}
155: \end{displaymath}
156: where $a=1+\gamma,\ b=1-\gamma$ and $\gamma$ parametrizes a degree of
157: matrix symmetry ($\gamma=1,\ \gamma=-1$ correspond to, respectively,
158: symmetric matrix with all eigenvalues being real and antisymmetric matrix
159: with imaginary eigenvalues, while $\gamma=0$ means full asymmetry). In
160: physical situations with finite $N$, these spectra, however, loose their
161: homegenity due to excess of real eigenvalues $\lambda_{\rm Re}$ whose
162: expected number expressed as a fraction of $N$ in the $N \to \infty$ limit
163: reads~\cite{edelman94}
164: \begin{equation}
165: \lim_{N \rightarrow \infty} {E_{\lambda_{\rm Re}} (N) \over N} = \sqrt{2
166: / (N \pi)}\ .
167: \label{ginoeexpect}
168: \end{equation}
169: A typical eigenvalue p.d.f. in the complex plane of a random matrix ($N =
170: 30$) obtained from 100000 independent matrix realizations is displayed in
171: Figure 1.
172:
173: \begin{figure}
174: \epsfxsize 6.5cm
175: \hspace{-0.5cm}
176: \epsffile{fig2a.eps}
177: \hspace{0.0cm}
178: \epsfxsize 6.95cm
179: \epsffile{fig2b.eps}
180: \caption{(a) Probability density function of empirical correlation matrix
181: (histogram) together with fitted Gaussian distribution (red solid line)
182: for $\Delta t=3$ s and for zero time lag. (b) Spectrum of complex
183: eigenvalues of correlation matrix for the same data as in (a). The largest
184: real eigenvalue is pointed by an arrow. Dashed circle denotes theoretical
185: eigenvalue spectrum for GinOE multiplied by standard deviation of
186: matrix elements.}
187: \end{figure}
188:
189: \section{Results}
190:
191: Our example of an application of the asymmetric correlation matrix is
192: based on high frequency data from NYSE and Deutsche B\"orse~\cite{data}
193: spanning the interval 1 Dec 1997 $-$ 31 Dec 1999. We analyze $N=30$ stocks
194: belonging to the Dow Jones Industrials group and the same number of stocks
195: constituting the main German DAX30 index. We calculate each element of
196: ${\bf C}$ by cross-correlating the time series pairs representing all
197: possible combinations of an American and a German stock. We neither
198: consider the correlations inside the German market nor inside the American
199: one. In order to investigate temporal dependencies between both markets we
200: introduce a time lag $\tau$ and associate it with the German stocks, i.e.
201: we look at $\{x_i^{(s)}\}_{i=1,...,T}$ and
202: $\{y_{i+\tau}^{(t)}\}_{i=1,...,T}$, where $\tau$ can assume both positive
203: and negative integer values. Thus, $\tau>0$ denotes a retardation of all
204: the signals corresponding to German stocks while $\tau < 0$ denotes the
205: opposite case.
206:
207: \begin{figure}
208: \epsfxsize 6.5cm
209: \hspace{-0.5cm}
210: \epsffile{fig3a.eps}
211: \hspace{0.0cm}
212: \epsfxsize 6.5cm
213: \epsffile{fig3b.eps}
214: \caption{Eigenvalue spectra in complex plane for empirical correlation
215: matrix calculated for different values of time lag: $\tau = 120$ s (a)
216: and $\tau=300$ s (b).}
217:
218: \end{figure}
219:
220: Since the two markets under study are separated by a few time zones, their
221: activities overlap only for a relatively short period of a trading day
222: (only the days that were common to both markets are considered). For most
223: time it was only 90 minutes a day from 9:30 to 11:00 New York time (15:30
224: to 17:00 Frankfurt time) and only after changing the trading hours in the
225: German floor starting from 20 Sep 1999 the overlap interval increased to
226: 120 minutes (from 9:30 to 11:30 in New York and 15:30 to 17:30 in
227: Frankfurt). This means that actually we can analyze the time series
228: spanning 47700 minutes total. A good time resolution should be a crucial
229: aspect of our analysis hence we consider only short time scales of
230: returns: from $\Delta t=120$ seconds down to $\Delta t=3$ seconds. Shorter
231: time scales cannot be used due to a fact that transaction times in the TAQ
232: database are stored with only 1 s resolution.
233:
234: \begin{figure}
235: \epsfxsize 6.5cm
236: \hspace{-0.5cm}
237: \epsffile{fig4a.eps}
238: \hspace{0.0cm}
239: \epsfxsize 6.5cm
240: \epsffile{fig4b.eps}
241:
242: \epsfxsize 6.5cm
243: \hspace{-0.5cm}
244: \epsffile{fig4c.eps}
245: \hspace{0.0cm}
246: \epsfxsize 6.5cm
247: \epsffile{fig4d.eps}
248: \caption{$|\lambda_1(\tau)|$ (vertical lines and full circles) and
249: $|\lambda_2(\tau)|$ (red solid line) for a few different time scales of
250: returns: $\Delta t=120$ s (a), $\Delta t=60$ s (b), $\Delta t=30$ s (c)
251: and $\Delta t=15$ s (d).}
252: \end{figure}
253:
254: First of all let us look at the correlation matrix and its eigenvalues for
255: $\tau=0$ (no time lag, synchronous evolution of both markets). Figure 2(a)
256: presents p.d.f. of the matrix elements $C_{s,t}$ for $\Delta t=3$ s
257: (histogram) together with a fitted Gaussian distribution (red solid line).
258: Except the central part of the empirical distribution, where there are
259: excessive small positive elements and lacking small negative ones, the
260: Gaussian is well approximated by the histogram (the same refers to the
261: other time scales). Thus, the correlation matrix can be
262: treated~\cite{drozdz00} as a sum of an essentially random core matrix and
263: a non-random part carrying the actual inter-market correlations. This
264: suggests that we can expect the eigenvalue spectrum consisting of an RMT
265: bulk and at least one significant non-random eigenvalue responsible for
266: the correlations. In fact, exactly this type of spectrum can be seen in
267: Figure 2(b). All except one eigenvalues are localized inside the RMT
268: prediction for a completely asymmetric matrix and the remaining largest
269: one is distant and resides on the real axis. By an analogy to a
270: symmetric matrix we are justified to associate this eigenvalue with the
271: coupling strength of the two markets (the global market factor).
272: Interestingly, there is no other factor which can influence the behaviour
273: of some smaller parts of the markets like e.g. specific economic sectors.
274:
275: \begin{figure}
276: \epsfxsize 10cm
277: \hspace{1.0cm}
278: \epsffile{fig5.eps}
279: \caption{$|\lambda_1(\tau)|$ (vertical lines and full circles) and
280: $|\lambda_2(\tau)|$ (red solid line) for the shortest time scale $\Delta
281: t=3$ s.}
282: \end{figure}
283:
284:
285: Figure 3 shows examples of the eigenspectra for two different positive
286: time lags. As we increase $\tau$ from 0 s up to 5 min, we observe a
287: gradual decrease of $|\lambda_1|$ which remains real even for $\tau>120$
288: s, but eventually looses its identity by drowning in the sea of random
289: eigenvalues for $\Delta t=5$ min. From the market perspective, after such
290: a time interval the stocks traded in Frankfurt forget about what happened
291: earlier in New York. We however still cannot say anything decisive about
292: the possible directional information flow between the markets. It requires
293: a more systematic investigation in which the largest eigenvalue
294: $\lambda_1$ (i.e. the one with the largest absolute magnitude) becomes a
295: function of variable $\tau$. Figure 4 displays $\lambda_1(\tau)$ for
296: different time scales of the returns. It can be seen that with the
297: resolutions of $\Delta t=120$ and 60 s the maximum coupling between the
298: markets occurs for synchronous signals and the non-random correlations
299: exist for $-3 \le \tau \le 3$ minutes. For $\Delta t=30$ s a weak trace of
300: asymmetry in both the maximum position and the memory length can be
301: identified, which is confirmed in the plot for $\Delta t=15$ s. Going down
302: to the shortest time scale of 3 s, this asymmetry becomes clear. Figure 5
303: documents that the stocks from both markets are maximally correlated if
304: the American market is retarded by about 3-15 seconds in respect to its
305: German counterpart. This observation is somehow counterintuitive because
306: one might expect that the American stock market, being the largest in the
307: world and representing the world's largest economy, is less dependent on
308: external influence than is the German market. We cannot give a
309: straightforward explanation of this phenomenon, though. Its source can lie
310: in memory properties of the American market as well as in some specific
311: behaviour of investors in the beginning of a trading day in New York. For
312: example, they may carefully observe the evolution of the European markets
313: which in the years 1998-99 used to finish their activity rather soon after
314: the American markets had been opened. We also cannot exclude the
315: possibility that the reason for this is a possible existence of artifacts
316: in the trade recordings in TAQ or KKMDB databases which cannot be
317: identified in data. Finally, the observed asymmetry of the curve tails in
318: Figures 4 and 5 with respect to $\tau=0$ can be explained, at least in
319: part, by different autocorrelation properties of the two markets under
320: study. This is evident in Figure 6, where the largest eigenvalue of the
321: symmetric matrix ${\bf C}_{\rm XX}$ is calculated separately for the
322: German and for the American markets. Here $\tau$ assumes only non-negative
323: values due to a symmetry of the problem; $\lambda_1(\tau)$ is a
324: multivariate counterpart of the autocorrelation function. It is clear from
325: Figure 6 that the German market has considerably longer and stronger
326: memory than its American counterpart; in fact, this memory can lead to
327: longer-lasting cross-dependencies presented in Figures 4 and 5 if the
328: German market is retarded. On the other hand, investors in Frankfurt may
329: need a longer time to collect all the information needed before they make
330: investment decisions if they take more markets and more information into
331: consideration. It is also possible that the American stock market is
332: technically more advanced and, on average, allows the investors to react
333: quicker than in Germany.
334:
335: \begin{figure}
336: \epsfxsize 10cm
337: \hspace{1.0cm}
338: \epsffile{fig6.eps}
339: \caption{$|\lambda_1(\tau)|$ for symmetric correlation matrix
340: ${\bf C}_{\rm XX}$ calculated for the American (black circles) and the
341: German (blue squares) stocks separately. Longer and stronger memory in the
342: latter case is visible.}
343: \end{figure}
344:
345: \section{Conclusions}
346:
347: We construct an asymmetric real correlation matrix from time series of
348: returns representing two separate groups of stocks: German and American ones.
349: Nonexistence of a symmetry condition allows us to concentrate solely on
350: the inter-market correlations without mixing them with the correlations
351: that are inner to only one market, and to study temporal properties of
352: such correlations. We introduce a time lag associated with German stocks
353: and investigate traces of direct information transfer from one market to
354: the other which can manifest itself in the existence of significant
355: non-synchronous couplings between the markets represented by a
356: $\tau$-shifted maximum in the largest eigenvalue of the empirical
357: correlation matrix. We identified such delayed correlations indicating
358: that the same information is shared by both markets with the American one
359: following its German counterpart only after a few seconds. This
360: observation, however, cannot be treated as a fully convincing one due to a
361: significant broadening of the $\lambda_1(\tau)$ maximum and an
362: unintuitive direction of this transfer from a smaller towards a larger
363: market. Another conclusion from our results is that the coupling between
364: the two analyzed markets is only of a one-factor type. We do not noticed
365: other, more subtle partial couplings that can involve a subset of stocks.
366:
367: Our results can be compared with the results of ref.~\cite{toth06} in
368: which an analysis of the delayed correlations between different stocks
369: traded on the American market are studied by means of the correlation
370: coefficients. It is worth mentioning that a similar analysis can also be
371: performed by applying the asymmetric correlation matrices used in our
372: work.
373:
374: \begin{thebibliography}{99}
375:
376: \bibitem{mantegna99} R.N.~Mantegna, Eur.~Phys.~J.~B {\bf 11}, 193-197
377: (1999)
378:
379: \bibitem{bonanno00} G.~Bonanno, N.~Vandewalle, R.N.~Mantegna, Phys.~Rev.~E
380: {\bf 62}, R7615-R7618 (2000)
381:
382: \bibitem{giada01} L.~Giada, M.~Marsili, Phys.~Rev.~E {\bf63}, 061101
383: (2001)
384:
385: \bibitem{plerou02} V.~Plerou, P.~Gopikrishnan, B.~Rosenow, L.A.N.~Amaral,
386: T.~Guhr, H.E.~Stanley, Phys.~Rev.~E {\bf 65}, 066126 (2002)
387:
388: \bibitem{onnela03} J.-P.~Onnela, A.~Chakraborti, K.~Kaski, J.~Kert\'esz,
389: A.~Kanto, Phys.~Rev.~E {\bf 68}, 056110 (2003)
390:
391: \bibitem{dimatteo04} T.~Di Matteo, T.~Aste, R.N.~Mantegna, Physica A {\bf
392: 339}, 181-188 (2004)
393:
394: \bibitem{kim05} D.-H.~Kim, H.~Jeong, Phys.~Rev.~E {\bf 72}, 046133 (2005)
395:
396: \bibitem{mizuno05} T.~Mizuno, H.~Takayasu, M.~Takayasu, preprint
397: physics/0508164 (2005)
398:
399: \bibitem{roll80} R.~Roll, S.A.~Ross, J.~Finance {\bf 35}, 121-130 (1980)
400:
401: \bibitem{noh00} J.-D.~Noh, Phys.~Rev.~E {\bf 61}, 5981-5982 (2000)
402:
403: \bibitem{laloux99} L.~Laloux, P.~Cizeau, J.-P.~Bouchaud, M.~Potters,
404: Phys.~Rev.~Lett. {\bf 83}, 1467-1470 (1999); V.~Plerou, P.~Gopikrishnan,
405: B.~Rosenow, L.A.N.~Amaral, H.E.~Stanley, Phys.~Rev.~Lett. {\bf 83},
406: 1471-1474 (1999)
407:
408: \bibitem{kwapien06} J.~Kwapie\'n, S.~Dro\.zd\.z, P.~O\'swi\c ecimka,
409: Physica A {\bf 359}, 589-606 (2006)
410:
411: \bibitem{bonanno01} G.~Bonanno, F.~Lillo, R.N.~Mantegna, Quant.~Finance
412: {\bf 1}, 96-104 (2001)
413:
414: \bibitem{gorski06} A.Z.~G\'orski, S.~Dro\.zd\.z, J.~Kwapie\'n,
415: P.~O\'swi\c ecimka, Acta.~Phys.~Pol.~B, this issue (2006)
416:
417: \bibitem{ginibre65} J.~Ginibre, J.~Math.~Phys.~{\bf 6}, 440-449 (1965)
418:
419: \bibitem{timme04} M.~Timme, F.~Wolf, T.~Geisel, Phys.~Rev.~Lett.~{\bf
420: 92}, 074101 (2004)
421:
422: \bibitem{fyodorov97} Y.~Fyodorov, B.~Khoruzhenko, H.-J.~Sommers,
423: Phys.~Rev.~Lett.~{\bf 79}, 557-560 (1997)
424:
425: \bibitem{stephanov96} M.~Stephanov, Phys.~Rev.~Lett.~{\bf 76}, 4472 (1996)
426:
427: \bibitem{akemann04} G.~Akemann, T.~Wettig, Phys.~Rev.~Lett.~{\bf 92},
428: 102002 (2004)
429:
430: \bibitem{kwapien00} J.~Kwapie\'n, S.~Dro\.zd\.z, A.A.~Ioannides,
431: Phys.~Rev.~E {\bf 62}, 5557-5564 (2000)
432:
433: \bibitem{sommers88} H.-J.~Sommers, A.~Crisanti, H.~Sompolinsky, Y.~Stein,
434: Phys.~Rev.~Lett.~{\bf 60}, 1895-1898 (1988)
435:
436: \bibitem{edelman94} A.~Edelman, E.~Kostlan, M.~Shub,
437: J.~Am.~Math.~Soc.~{\bf 7}, 247-267 (1994)
438:
439: \bibitem{kanzieper05} E.~Kanzieper, G.~Akemann, Phys.~Rev.~Lett.~{\bf 95},
440: 230201 (2005)
441:
442: \bibitem{data} http://www.taq.com (data from NYSE) and H. Goeppl,
443: Karlsruher Kapitalmarktdatenbank (KKMDB), Institut f\"ur
444: Entscheidungstheorie u. Unternehmensforschung, Universit\"at Karlsruhe
445: (TH) (data from Deutsche B\"orse)
446:
447: \bibitem{drozdz00} S.~Dro\.zd\.z, A.Z.~G\'orski, F.~Ruf, J.~Speth, Physica
448: A {\bf 287}, 440-449 (2000)
449:
450: \bibitem{toth06} B.~T\'oth, J.~Kert\'esz, Physica A {\bf 360}, 505-515
451: (2006)
452:
453: \end{thebibliography}
454:
455: \end{document}
456: