Persistent Correlations in Major Indices of the World Stock Markets

  • Maciej Janowicz
  • Leszek J. Chmielewski
  • Joanna Kaleta
  • Luiza Ochnio
  • Arkadiusz Orłowski
  • Andrzej Zembrzuski
Chapter
Part of the Studies in Systems, Decision and Control book series (SSDC, volume 125)

Abstract

Time-dependent cross-correlation functions have been calculated between returns of the major indices of the world stock markets. One-, two-, and three-day shifts have been considered. Surprisingly high and persistent-in-time correlations have been found among some of the indices. Part of those correlations can attributed to the geographical factors, for instance, strong correlations between two major Japanese indices have been observed. The reason for other, somewhat exotic correlations, appear to be as much accidental as it is apparent. It seems that the observed correlations may be of practical value in the stock market speculations.

Keywords

Stock market indices Correlation functions Pearson correlation Technical analysis 

1 Introduction

As is well known, the time-evolution of the stock markets, and, in particular, the stock market indices, exhibit considerable short-time correlations. These are sufficient to disprove claims of the reliability of the random-walk approximations to that evolution or strong versions of the efficient-market hypothesis [1, 2]. On the other hand, there exist a set of usually quite simple computational and visualization techniques, called technical analysis [3, 4], which aims to obtain the approximate predictions of trends and their corrections in market the data. It is so even though the data appear as realizations of a random process. It is to be noticed that some recent publications, e.g. [5, 6, 7, 8] have lead to considerable revision of the previously ultra-critical stand of the many experts regarding technical analysis.

One of the many possible strategies of “beating the market” which are close to the spirit of technical analysis consists of identification of correlated pairs of stocks. In fact, if we find that, during a specific, long, time interval, the increase (decrease) of the value of one stocks has been followed by a corresponding change in another stock, we may attempt to guess that this relation may persist, at least for short additional time.

To quantify the above intuitive remarks, we have computed the time-dependent cross-correlation functions of returns of the major world stock market indices. Among the latter, the following indices have been included: ALL ORDINARIES (ALL-ORD), AMEX MAJOR (AMEX-MAJ), BOVESPA, B-SHARES, BUENOS, BUX, CAC40, DAX, DJIA, DJTA, DJUA, EOE, FTSE100, HANGSENG, MEXICIPS, NASDAQ, NIKKEI, RUSSELL, SASESLCT, SMI, SP500, TOPIX, and TSE300. Let us notice that in one of our previous work a similar research but involving normalized values has been reported. Here, however, we deal with correlations of daily returns.

The main body of this work is organized as follows. In Sect. 2 we describe our procedure; rhe description as it is, in fact, very simple, and involves well-known quantities. Section 3 is devoted to the presentation of results. Finally, Sect. 4 comprises some concluding remarks.

2 Correlation Functions of World Stock Market Indices

Let \(K_{n}^{(a)}\) denote the closing value at the trading day n of a stock market index (a), and let \(Z_{n}^{(a)}\) denote the corresponding relative return, i.e. \(Z_{n}^{(a)} = (K_{n}^{(a)} - K_{n-1}^{(a)})/K_{n-1}^{(a)}\). We define the cross-correlation function of returns of two indices (a) and (b) as:
$$\begin{aligned} C^{(a,b)} (n, M, t) = \frac{Cov \left( Z_{n-M}^{(a)} Z_{n}^{(b)} \right) }{\sqrt{ Var \left( Z_{n-M}^{(a)} \right) Var \left( Z_{n}^{(b)} \right) }}, \end{aligned}$$
(1)
where t denotes the time interval over which the averaging in the calculations of the variances and covariance has been performed (unfortunately, we have no ensemble-averaging in our disposal here).

Thus, we have defined the correlation function in terms of the Pearson R coefficient [9] made of two sequences of the length t. A similar definition can be given in terms of, e.g., Spearman coefficient.

It is to be noted that \(C^{(a,b)} (n, M, t)\) is a function of three time variables. In the following, we have considered only three values of M, \(M = 1, 2, 3\). The averaging time t has been varied from \(t = 60\) to \(t = N_{0}\), with \(N_{0}\) being the largest value of the trading sessions common to all indices. \(N_{0}\) has been equal to 3452 enclosing the time interval from 10th of October, 2001 to 31st of October, 2016). The data for stock market indices have been downloaded from [11]. The correlation functions have been computed using pearsonr function from the Python (2.7) module scipy.stats, version 0.14.0 [10].

3 Results and Discussion

We have started our study with n equal to t. Firstly, we have considered Pearson’s R coefficient using all available data.

In Tables 1, 2 and 31 we have provided pairs of indices with largest values of the correlation functions for \(n = t = u\), \(u = N_{0} - 4\), for \(M = 1, 2, 3\).
Table 1

List of indices with largest correlation functions for \(n = t = N_0 - 4\), \(M = 1\). In all cases p-value has been smaller than 0.05

Index (a)

Index (b)

\(C^{(a,b)}(u, 1, u)\)

NIKKEI

TOPIX

0.785

NASDAQ

SP500

0.694

NASDAQ

RUSSELL

0.690

NASDAQ

DJTA

0.651

DJIA

NASDAQ

0.474

DJIA

RUSSELL

0.411

Even though the number of cross-correlation functions with the values greater than 0.4, say, has been rather small, in particular cases those correlations have been very significant for \(M = 1\) as specified in Table 1. One might say, for instance, that during several years one can predict the behavior of the TOPIX index from the knowledge of the NIKKEI 225 index change one day before. Remarkably, it has worked only one way. What is more, for t significantly smaller than u a quite interesting structure in the n-dependence of the correlation function occurs, please see below.
Table 2

List of indices with largest correlation functions for \(n = t = N_0 - 4\), \(M = 2\). In all cases p-value has been smaller than 0.05

Index (a)

Index (b)

\(C^{(a,b)}(u, 2, u)\)

DJIA

SP500

0.388

DJIA

DJUA

0.366

DJIA

DJTA

0.335

DJIA

RUSSELL

0.325

DJIA

MEXICIPC

0.256

The above relatively high values of the correlation functions between the returns of DJIA and the returns of other American (and Mexican) indices certainly cannot be called counterintuitive. However, it is quite remarkable that the pairs listed in Table 2 do not appear in the list of highly correlated pairs for the shift \(M \ne 2\). Please see also further comments following Fig. 2.
Table 3

List of indices with largest correlation functions for \(n = t = N_0 - 4\), \(M = 3\). In all cases p-value has been smaller than 0.05

Index (a)

Index (b)

\(C^{(a,b)}(u, 3, u)\)

SMI

TSE-300

0.159

FTSE100

ALL-ORD

0.135

DJIA

MEXICIPC

0.105

EOE

SMI

0.092

SP500

TSE-300

0.091

AMEX-MAJ

TSE-300

0.090

Table 4

List of indices with largest correlation functions for \(n = t = N_0 - 4\), \(M = 1\) with the numbers of correct and incorrect predictions of the behavior one the second-column indices based on the change of the corresponding first-column ones

Index (a)

Index (b)

No. of correct preditions

No. of incorrect predictions

NIKKEI

TOPIX

2852

596

NASDAQ

SP500

2296

1089

NASDAQ

RUSSELL

2368

1080

NASDAQ

DJTA

2301

1147

DJIA

NASDAQ

2359

1089

DJIA

RUSSELL

2281

1167

For \(M = 3\) we have not observed any large values of the R coefficient, for \(M > 3\) they become even smaller. Let us also report that we have not obtained any negative R coefficient with significant absolute value for any M. The indices definitely tend to be correlated rather than anti-correlated.

In the Tables 4, 5 and 6 we have listed the same pairs of indices as in Tables 1, 2 and 3 but accompanied with the numbers of correct and incorrect predictions. Our definition of a correct prediction is trivial: if the sign of the product of returns of index (a) at time t and index (b) at time \(t+M\) is positive, we say that the prediction is correct; otherwise the prediction is incorrect.
Table 5

List of indices with largest correlation functions for \(n = t = N_0 - 4\), \(M = 1\) with the numbers of correct and incorrect predictions of the behavior one the second-column indices based on the change of the corresponding first-column ones

Index (a)

Index (b)

No. of correct preditions

No. of incorrect predictions

DJIA

SP500

2086

1362

DJIA

DJUA

1953

1495

DJIA

DJTA

2042

1406

DJIA

RUSSELL

1953

1495

DJIA

MEXICIPC

1887

1561

Table 6

List of indices with largest correlation functions for \(n = t = N_0 - 4\), \(M = 3\) with the numbers of correct and incorrect predictions of the behavior one the second-column indices based on the change of the corresponding first-column ones

Index (a)

Index (b)

No. of correct preditions

No. of incorrect predictions

SMI

TSE-300

1833

1516

FTSE100

ALL-ORD

1790

1658

DJIA

MEXICIPC

1828

1620

EOE

SMI

1777

1671

SP500

TSE-300

1885

1563

AMEX-MAJ

TSE-300

1813

1635

The content of the Tables 4, 5 and 6 largely agrees with the conclusions one can draw from three previous tables. For the first five rows of Table 4 we can realize that the number of correct predictions is at least two times larger that the incorrect ones. Quite obviously, even such a ratio could not guarantee any successful trading on any index-based financial instruments. On the other hand, even the smallest advantage like that provided by Table 6 may sometimes be sufficient to get the edge on trading competitors.

Let us notice that the difference between correct and incorrect predictions in the case of EOE-SMI pair is so small that p-value in the G-test is \({\approx }\)0.07. Thus, one cannot exclude the zeroth hypothesis that the difference between the numbers of predictions if purely accidental. In all other cases we have had p-value in the G-test smaller than 0.05.

We have also performed calculations for the values of n and t different from u. For instance, we have obtained the R coefficient from the last 480 trading sessions (approximately two years) preceding 31st of October, 2016. For the pair CAC40—EOE a record value of R equal to 0.94 has been obtained with the percent of correct predictions equal to 0.875. For \(M = 2\) the most considerable correlation has been obtained between the DAX and SMI indices; R has been equal to 0.526, and the ratio of correct to incorrect predictions has almost exactly been equal to 2:1.

We have not, however, been satisfied with the above results for it has not been clear precisely how the correlation function depends on the time n if the averaging window t is smaller than n.

In the following figures we have displayed selected cross-correlation function for given M and t as functions of n. These have been compared with the time evolution of the indices themselves. Note that it is the closing values of the index (not the returns) which have been shown in the central and lower parts of each figure.
Fig. 1

Upper part: time-dependent correlation function of returns for NIKKEI 225 and TOPIX indices as a function of discrete time n for \(M = 1\), \(t = 240\); central part: simple moving average (with window 240) of closing values of the NIKKEI 225 index in the same time interval; lower part: corresponding values of the TOPIX index

Naturally, the graphs in Fig. 1 provide additional insight into the results tabulated in Tables 1 and 4. In fact, we can see that correlation function involving NIKKEI and TOPIX has not been constant, and there have been two time intervals in which the correlation dropped quite significantly. The second one includes recent trading sessions. As can be seen from the central and lower parts of Fig. 1, there is nothing in the values of indices themselves which could correspond to the drops of correlations. The first period of low correlations may, perhaps, be attributed to the change of the weighting system with the help of which the TOPIX index has been calculated; that change was completed in January 2006. We have failed, however, to identify a possible reason of breaking down the correlation in the beginning of 2015.
Fig. 2

Upper part: time-dependent correlation function of returns for DJIA and SP500 indices as a function of discrete time n for \(M = 2\), \(t = 240\); central part: simple moving average (with window 240) of closing values of the DJIA index in the same time interval; lower part: corresponding values of the SP500 index

The history of correlations of the indices DJIA and SP500 (\(M = 2\)), which, perhaps, might be considered as the most important and most obvious of all, is even more striking. It appears that the quite considerable value of R coefficient calculated during the last 14 years was built during the relatively short time interval 2006–2010. Recently, DJIA-SP500 have become almost completely “unreliable”, so to say.
Fig. 3

Upper part: time-dependent correlation function of returns for MEXICIPC and SMI indices as a function of discrete time n for \(M = 3\), \(t = 240\); central part: simple moving average (with window 240) of closing values of the MEXICIPC index in the same time interval; lower part: corresponding values of the SMI index

Fig. 4

The same as in Fig. 1 but for \(t = 60\) and moving averages with the window length equal to 20

Fig. 5

The same as in Fig. 2 but for \(t = 60\) and moving averages with the window length equal to 20

We have provided Figs. 4, 5 and 6 largely to confirm the results displayed in the three previous figures. One can convince oneself that the shorter intervals and smaller windows of moving averages do not not change the overall picture and provide, on this occasion, a little more than noise.
Fig. 6

The same as in Fig. 3 but for \(t = 60\) and moving averages with the window length equal to 20

4 Concluding Remarks

In this work we have computed cross-correlation function of time series generated by the returns of important world stock market indices. We have identified pairs of indices such that correlations are both strong and persistent in time. Preliminary assessments of predictive power of the correlations have been performed. We believe that, with sufficient care, the knowledge of correlations between indices and individual stocks may be used in practice, possibly even to enhance the predictive power of technical indicators for trading purposes. Needless to say, extreme caution is required in any such attempts. In fact, they can break down at any time. For instance, when we used historical data up to the middle summer of 2016, we have observed the highest correlation function for \(M = 3\), \(n = t = u - 3\) in the pair MEXICIPC—SMI. During the subsequent mounts that correlation deteriorated quite spectacularly. One might think that it would be a valuable enterprise to find any indicators which could suggest that the correlations have just started to build up or are just about to get ruined, provided that such indicators exist. Finally, one may reasonably argue that the above very simple correlation analysis can and should be combined with the analysis of cointegration to have deeper insight into connection between shifted (i.e. with \(M \ne 0\)) and unshifted time series.2

Footnotes

  1. 1.

    Tables will be properly formatted in the camera-ready version of the paper.

  2. 2.

    References will be properly formatted in the camera-ready version of the paper.

References

  1. 1.
    Malkiel, B.: A Random Walk Down the Wall Street. Norton, New York (1981)Google Scholar
  2. 2.
    Fama, E.F., Blume, M.: Filter rules and stock-market trading. J. Bus. 39, 226–241 (1966)CrossRefGoogle Scholar
  3. 3.
    Murphy, J.: Technical Analysis of Financial Markets. New York Institute of Finance, New York (1999)Google Scholar
  4. 4.
    Kaufman, P.: Trading Systems and Methods. Wiley, New York (2013)Google Scholar
  5. 5.
    Brock, W., Lakonishok, J., LeBaron, B.: Simple technical trading rules and the stochastic properties of stock returns. J. Finan. 47(5), 1731–1764 (1992)CrossRefGoogle Scholar
  6. 6.
    Lo, A., MacKinley, A.: Stock market prices do not follow random walks: evidence from a simple specification test. Rev. Finan. Stud. 1, 41–66 (1988)CrossRefGoogle Scholar
  7. 7.
    Lo, A., MacKinley, A.: A Non-Random Walk down Wall Street. Princeton University Press, Princeton (1999)Google Scholar
  8. 8.
    Lo, A., Mamaysky, H., Wang, J.: Foundations of technical analysis: computational algorithms, statistical inference, and empirical implementation. J. Finan. 55(4), 1705–1765 (2000)CrossRefGoogle Scholar
  9. 9.
    Pearson, K.: Notes on regression and inheritance in the case of two parents. Proc. R. Soc. Lond. 58, 240–242 (1895)CrossRefGoogle Scholar
  10. 10.
    Scipy Community: Statistical functions in Python. http://docs.scipy.org/doc/scipy/reference/stats.html (2016)
  11. 11.
    Bossa.pl. http://bossa.pl/notowania/metastock. Accessed 15 Nov 2016

Copyright information

© Springer International Publishing AG 2018

Authors and Affiliations

  • Maciej Janowicz
    • 1
  • Leszek J. Chmielewski
    • 1
  • Joanna Kaleta
    • 1
  • Luiza Ochnio
    • 1
  • Arkadiusz Orłowski
    • 1
  • Andrzej Zembrzuski
    • 1
  1. 1.Faculty of Applied Informatics and Mathematics – WZIMWarsaw University of Life Sciences – SGGWWarsawPoland

Personalised recommendations