Spurious long-range dependence in emerging financial markets

The problem with estimating the long range dependence. Motivation for long memory. Long memory VS structural breaks. Markov switching model. Generating the same statistical properties as the long memory model during processing with structural breaks.

Рубрика Финансы, деньги и налоги
Вид статья
Язык английский
Дата добавления 26.01.2017
Размер файла 130,8 K

Отправить свою хорошую работу в базу знаний просто. Используйте форму, расположенную ниже

Студенты, аспиранты, молодые ученые, использующие базу знаний в своей учебе и работе, будут вам очень благодарны.


Размещено на http://www.allbest.ru/



State University - Higher School of Economics

SPURIOUS LONG-RANGE DEPENDENCE IN EMERGING FINANCIAL MARKETS We gratefully acknowledge financial support through State University - Higher School of Economics scientific fund, grant №08-04-0043.

S. Bryzgalova, S. Gelman


  • 1. Introduction
  • 2. Motivation for long memory
  • 3. Long memory VS structural breaks
  • 4.1 ARFIMA-model
  • 4.2 Markov switching (MS) model
  • 4.3 Tests for predictive power
  • Granger - Newbold test
  • Diabold - Mariano test
  • 5. Data
  • 6. Estimation
  • Concluding remarks
  • References

1. Introduction

Long memory in financial markets has been a widely discussed issue for the last 40 years. Originally found by Mandelbrot and Wallis (1968), significant autocorrelation in the yields of various financial instruments at low frequency, is one of its manifestations. Further research confirmed that this is not the only paradoxal evidence from the memory of the process - the autocorrelation function itself decays much slower than the one corresponding to standard ARMA - GARCH processes. Long memory presence has been recently extensively used to test the market efficiency and stage of development, for it allows for significant predictability [Rogers, 1997; Eom et al., 2008] and therefore, serves as an indicator of market inefficiency, often applied to emerging markets [Dacorogna et al., 2005; Cajueiro, Tabak, 2006; Eom et al., 2008].

The problem with estimating the long range dependence is that a process with structural breaks generates exactly the same statistical properties as the long memory model, and there is still no test on their differentiation (see Section 3). However, this problem is often neglected in the empirical studies.

We suggest using predictive power of the model to give evidence in favor of the true underlying process. The logic is straightforward: since the long range dependence is supposed to yield superior predictive abilities, we should test them against those of a regime-switching process. For this purpose we consider emerging markets of BRIC countries. In particular, we estimate ARFIMA and Markov Switching model, build a series of one-step ahead forecasts and them compare them with Granger - Newbold and Diabold - Mariano tests. We show that MS process demonstrates superior predictive powers compared to those of ARFIMA for all the markets. Hence, it is quite likely that traditional long memory estimations are spurious, being due to structural breaks.

The paper is organized as follows. Section 2 discusses the properties of long memory processes and their application in finance, section 3 focuses on the dichotomy between long-range dependence and structural breaks. Section 4 describes the used methodology, section 5 provides the data description. Finally, section 6 presents the results of the estimation, and section 7 concludes, summarizing findings and outlining directions for future research.

financial market markov switching model

2. Motivation for long memory

Significant autocorrelation at low frequencies and a slow decay of the autocorrelation functions are two basic manifestations of the long memory process. Theoretically these unusual patterns could be explained by the heterogeneity of the agents present on the market. In the recent decade there have been proposed a number of models, directly incorporating it into the market structure [Brock, Hommes, 1998; LeBaron, 2000; Abel, 2008]. Finally, there have been developed Heterogeneous Market Hypothesis [Dacorogna et al., 2001] which provides a general framework for these models and postulates major characteristics of the corresponding market structure.

According to HMH, the market is populated by heterogeneous agents of different time horizons and dealing frequency, each having the short-term memory with a standard exponential decay (in different time scales). It is a known mathematical fact that the superposition of different exponential decays at various scales results in the behaviour close to a hyperbole. Hence, such a market structure could perfectly explain the hyperbolic decay of autocorrelation widely observed in real financial markets.

Long memory processes allow for significant predictability [Rogers, 1997; Eom et al., 2008] and therefore, its presence can be used as an indicator of market inefficiency. Introducing the long memory components directly into the asset pricing theory is rather difficult, but there has emerged a considerable area of research on derivative pricing, namely options, which accounts for possible long memory in stock returns [Jamdee, Los, 2007]. It has been demonstrated that the modified approach results in much better explanation of the market realizations than the standard Black and Scholes formulation.

3. Long memory VS structural breaks

Structural breaks have long been an integral feature of many financial time series. The pioneering work belongs to Chow (1960), who suggested an F-test for a linear regression model, provided the known date of the break. The limitation of exogeneity was overcome in further studies for endogenous structural breaks: Brown, Durbin and Evans (1975) proposed a procedure based on the cumulative sum of the residuals, known as CUSUM-test. Later studies accounted for possible heteroscedasticity and multiple structural breaks [Andrews, 1993; Andrews, Ploberger, 1994; Bai, Perron, 1998].

Several studies were conducted to check tests validity in the presence of the long memory. In particular, Kramer and Sibbertsen (2000) proved that in the presence of long memory, the null distribution of CUSUM statistics tends to infinity, thus, detecting a spurious structural break. Alas, not only a long memory process can successfully imitate the properties of structural breaks, but the vice versa is also possible. In particular, there have been a number of research papers on the latter effect, including [Ding, Granger, 1996; Andersen, Bollerslev, 1997; Granger, Hyung, 2004; Diebold, Inoue, 2001; Choi, Zivot, 2007].

A new research area emerged with the introduction of Markov Switching processes by Hamilton (1988). At each point of time the process can be in one of the specified states of nature, and not only do we estimate each regime parameters, but also the probabilities to be in corresponding states at each moment of time. This idea was used in other specifications of the regime switching models, such as Diebold and Inoue (2001), or Granger and Hyung (2004). These processes can perfectly explain the autocorrelation patterns usually prescribed to long memory evidence.

It is still an open question, how to differentiate between the long memory and regime switching processes. While we do not suggest a formal procedure, we think that in case of financial time series predictive power of the models can be regarded as a valuable criterion. In particular, the main implication of long memory estimations consists in predicting stock returns and testing Efficient Market Hypothesis. Hence, a true long memory process should exhibit a substantial predictive power, which we can test against that of the regime switching process.

4. Methodology

4.1 ARFIMA-model

With a fractional integration parameter d, the ARFIMA (p,d,q) model is written as

where L stands for the lag operator, Ц (L) and И (L) are polynomials of order p and q correspondingly.

The simplest specification of this model if the so-called fractional Brownian motion (FBM):


It is easy to note through a simple factorization why this process exhibits long memory:

where Г stands for gamma-function.

The autocorrelation function for this process demonstrates hyperbolic decay and looks as follows:

An algorithm for the computation of the autocovariance function of a general ARFIMA process is derived in [Sowell, 1992].

We consider three methods to estimate the parameters of ARFIMA:

· Exact Maximum Likelihood, as in [Ibid],

· Modified Profile Likelihood, as in [An, Bloomfield, 1993], and

· Nonlinear Least Squares.

4.2 Markov switching (MS) model

The MS model has been developed by Hamilton (1988). Consider the following AR (1) model

At every period of time the process belongs to one of the regimes, which are characterized by different coefficients and possibly different variance of the disturbances. The state variable y (t) is assumed to follow an ergodic first-order Markov process and is characterized by the matrix П of transition probabilities from state i to state j.

Once the coefficients of the model and the transition matrix have been estimated, the probability of being in state j, based on the knowledge of complete series can be calculated for each date (the so-called smoothed pro-babilities), as described in [Kim, Nelson, 1999]. When only information up to a current date is used, the probability estimation is called the filter probability.

To estimate the model we use (quasi) maximum likelihood method [Hamilton, 1988].

4.3 Tests for predictive power

In order to test the predictive powers of one model against the other we use Granger - Newbold (1976) and Diebold - Mariano (1995) tests for 1-step ahead predictions.

Consider 2 models estimated at the same data sample, generating the forecasts Forecast errors are defined as

Granger - Newbold test

Let us form new sequences and

Provided that the forecast errors are unbiased and normally distributed, if both models produce equally good forecasts, then x (t) and z (t) should be uncorrelated, and the statistics looks as follows:

where is a sample correlation of x (t) and z (t), H is the number of forecasts generated by each model.

Diabold - Mariano test

This procedure uses a more general setting to test the predictive power of the models, for it does not require the forecast errors to be unbiased, normally distributed or serially uncorrelated.

The mean loss is defined as

If the loss time series is autocorrelated of order q, then provided that both forecasts are equally adequate,

As far as the functional form of g () is concerned, a common approach is to use the square.

5. Data

We conduct our analysis using the data from national indexes of BRIC countries.

· BOVESPA of Brazil. The Bovespa Index is the main indicator of the Brazilian stock market's average performance. The issuing companies of the stocks that compose the Bovespa Index theoretical portfolio are responsible, in average, for approximately 70% of the sum of all BOVESPA's companies' capitalization.

· MICEX of Russia. The MICEX Index is a capitalization-weighted index of the market of most liquid 19 stocks of Russian issuers admitted to circulation on the MICEX (Moscow Interbank Currency Exchange).

· India BSE (100) National for India. India BSE (100) is calculated using the "Free-float Market Capitalization" methodology. The equity shares of 100 companies from the "Specified" and the "Non-Specified" list of the five major stock exchanges, viz. Mumbai, Calcutta, Delhi, Ahmedabad and Madras have been selected for the purpose of compiling the BSE National Index.

· Shanghai SE Composite for China. Constituents for SSE Composite Index are all listed stocks (A shares and B shares) at Shanghai Stock Exchange.

We have chosen the time interval of 01.2000-01.2008, since it was rather stable for the countries of interest (no periods of hyperinflation or financial crisis). Log-returns are then computed and used for the analysis.

6. Estimation

An essential feature of ARFIMA-model consists in the ability to capture both short - and long-range effects of investors' reaction towards news. In addition, autoregressive function for ARFIMA-model demonstrates not exponential, but hyperbolic rate of decay, which can account for significant autocorrelation at high lags.

We chose the model on the basis of information criteria and autocorrelation fit. The estimation was done in OxMetrics 4,0, results are summarized in Table 1.

Table 1.

ARFIMA (p,d,q) estimates



Exact Maximum Likelihood

Modified Profile Likelihood

Nonlinear Least Squares

Coefficient St. error

Coefficient St. error

Coefficient St. error

d** 0,042528 0,01776

d** 0,06291 0,01785

d* - 0,0394184 0,02881

AR-5* - 0,042037 0,02206

AR-5* - 0,041956 0,02207

AR-5* - 0,0411425 0,02248

AR-6 0,003650 0,02195

AR-6 0,003775 0,02196

AR-6 0,0063301 0,02219

AR-7* - 0,036474 0,02194

AR-7* - 0,036313 0,02196

AR-7* - 0,0296034 0,02211

AR-8 - 0,019318 0,02194

AR-8 - 0,019134 0,02196

AR-8 - 0,014776 0,02201

AR-9* - 0,031507 0,02191

AR-9* - 0,031309 0,02193

AR-9* - 0,0262247 0,02194

AR-10** 0,059842 0,02196

AR-10** 0,06006 0,02197

AR-10** 0,0620966,02201


MA-1* 0,0393780 0,03630


AIC - 5, 20955043

AIC - 5, 20955043

AIC - 5, 20955043

Coefficient St. error

Coefficient St. error

Coefficient St. error

d** - 0,164521 0,08484

d* - 0,137394 0,07922

d** - 0,167749 0,08379

AR-1** 0,178968 0,08876

AR-1* 0,152034 0,08294

AR-1* 0,184778 0,08780

AR-2* 0,054598 0,03608

AR-2* 0,0456001 0,03594

AR-2* 0,0519264 0,03550

AR-3 - 0,0079348 0,02835

AR-3 - 0,0134459 0,02841

AR-3 - 0,0055267 0,02823

AR-4** 0,0629145 0,02894

AR-4** 0,0571543 0,02865

AR-4** 0,0601306 0,02866


AR-9** 0,0515149 0,02401

AR-9** 0,0486693 0,02374

AR-9** 0,0520668 0,02385


AIC - 4,73139209

AIC - 4,71767168

AIC - 4,752945

Coefficient St. error

Coefficient St. error

Coefficient St. error

d* - 0,0232635 0,02193

d - 0,0165652 0,03583

d - 0,0077324 0,04165

AR-1** 0,119743 0,04388

AR-1** 0,113240 0,04354

AR-1 0,0008142 0,04352

AR-4** 0,062277 0,02350

AR-4** 0,0612208,02346

AR-4** 0,0562914 0,02340

AR--9** - 0,0465497 0,02225

AR-9** 0,0463685 0,02226

AR-9** 0,0447982 0,02211


AR-10* 0,0360177 0,02198

AR-10* 0,0362668 0,02200

AR-10* 0,0422150 0,02438


AIC - 5,3351933

AIC - 5,3201407

AIC - 5,34626565



Exact Maximum Likelihood

Modified Profile Likelihood

Nonlinear Least Squares

Coefficient St. error

Coefficient St. error

Coefficient St. error

d* 0,118024 0,06802

d* 0,118046 0,06805

d* 0,112709 0,06684

AR-1*** 0,583834 0,1365

AR-1*** 0,583824 0,1365

AR-1*** 0,556809 0,1399

AR-4** 0,057311 0,02594

AR-4** 0,057308 0,02594

AR-4** 0,060619 0,02579

AR-5*** - 0,083117 0,02278

AR-5*** - 0,083118 0,02279

AR-5*** - 0,080423 0,02232

AR-10* 0,031185 0,02237

AR-10* 0,031183 0,02238

AR-10* 0,0290795 0,02198


MA-1 - 0,0690887 0,1074

MA-1 - 0,06909 0,1075

MA-1 - 0,0665491 0,1133


AIC - 5,60483872

AIC - 5,59682986

AIC - 5,62015626

Emerging markets are characterized by long memory (significant value of d at least at 10%), which is regarded by the previous researchers as a sign of market inefficiency. Using similar estimates, they described the current stage of development in countries like BRIC as relatively low and the degree of predictability as rather substantial.

Table 2 summarizes MS-estimation results for the same data.

Table 2.

Markov switching estimates


Model Specification

Transition matrix





Although not all the coefficients in the estimated model are significant, the overall tendency is that the algorithm differentiates between positive and negative dynamics on the market, the latter characterized by a substantial increase in volatility.

Having estimated the models, we divide our sample in 2 parts, taking 7 years for the estimation period and 1 year (250 observations) for the forecast period. Note, that it is rather short compared to the dominating approach in the literature (about 1/3 of the sample). However, robust long memory estimation requires a very large data set; hence, we had to allow only for 250 observations. Each period we estimate models and make 1-step ahead forecast For ARFIMA-model we use EML-estimates., then update the information and repeat the procedure. This yields 250 static forecasts for both models.

Finally, we carry out the tests on the predictive power of our models.

Table 3.

Granger - Newbold test











H1 (1%)



H1 (10%)

For Brazil and China we tend to reject the null hypothesis of equal predictive power. However, for Russia and India the statistics lies within the confidence interval. However, Granger - Newbold test is rather restrictive, for it assumes unbiased normally distributed forecast errors with no serial autocorrelation, while Diabold - Mariano test is free from these drawbacks. To conduct this test, we assumed that the first model is ARFIMA, and the second - MS, so that a strictly positive value of the statistics means worse predictive power of the former.

Table 4.

Diebold - Mariano test for different orders of autocorrelation included






q = 1





q = 2





q = 3





q = 4





q = 5

6, 20797




q = 6










q = 7





q = 8





q = 9





q = 10





Regardless of the number of autocorrelation coefficients included MS models demonstrate superior predictive power to that of the ARFIMA-model for any country. This means that long memory models do not actually have significant predictive power as it was postulated in the literature, since even the simplest regime-switching model outperforms them. Next, the quality of long memory estimation is not likely to lead to strong conclusions, such as market inefficiency. Hence, we are more inclined to think that it is actually regime-switching nature of the process, not a long-range dependence that drives the market dynamics.

Concluding remarks

In this paper we empirically investigate validity of the long memory specification for BRIC countries. It has been noted in the literature that processes with structural breaks are often confused with those exhibiting long memory. We suggest assessing the validity of the model by its predictive power. We estimate ARFIMA and Markov Switching models, and build series of one-step ahead forecasts. Predictive powers of the models are compared using Granger - Newbold and Diebold - Mariano tests. We demonstrate that even the simplest specification of MS-model shows superior predictive power to that of ARFIMA for all the countries. This evidence raises doubts concerning the use of the long memory presence as an indicator of good time series prediction and a sign of market inefficiency, while the regime-switching nature of the process is more likely to be present in the data. However, one should carefully interpret the conclusions of this paper, since fractional integration is only one of the means to capture long-range dependence. Alas, most of other techniques do not allow for predictability of the series, and ARFIMA-model stays one the most popular means of analysis.

Despite poor predictive power, long range dependence has a clear advantage of representing the movements in a heterogeneous agents market, which is quite a realistic assumption. On the other hand, it is not clear what causes sudden changes in the regime of the model. We think that the solution might lie in the combination of agents' heterogeneity and changes in that structure. For instance, a substantial negative shock could undermine the trust of long-term investors, and they temporarily abstain from transactions, so that the market is driven mostly by short-term investors. It seems that dynamic structure of market participants could perfectly reconcile heterogeneous agents' framework and regime-switching behavior of the index. We believe that further research will be able to capture these effects.


1. An S., Bloomfield P. Cox and Reid's Modification in Regression Models with Correlated Errors: Discussion paper. Department of Statistics. North Carolina State University. Raleigh, 1993.

2. Andrews D. Tests for Parameter Instability and Structural Change with Unknown Change Point // Econometrica. 1993.61.4 Р.821-856.

3. Andrews D., Ploberger W. Optimal Tests When a Nuisance Parameter Is Present Only under the Alternative // Econometrica. 1994. Vol.62 (6). Р.1383-1414.

4. Bai J., Perron P. Estimating and Testing Linear Models with Multiple Structural Changes // Econometrica. 1998.66. Р.47-78.

5. Brock W. A., Hommes C. H. Heterogeneous Beliefs and Routes to Chaos in a Simple Asset Pricing Model // Journal of Economic Dynamics and Control. 1998. Vol.22 (8-9). Р.1235-1274.

6. Brown R. L., Durbin J., Evans J. M. Techniques for Testing for Constancy of Regression Relationships Over Time // Journal of the Royal Statistical Society B. 1975.37. Р.149-163.

7. Cajueiro D. O., Tabak B. M. Testing for Predictability in Equity Returns for European Transition Markets // Economic Systems. 2006. Vol.30 (1). Р.56-78.

8. Choi K., Zivot E. Long Memory and Structural Changes in the Forward Discount: An Empirical Investigation // Journal of International Money and Finance. 2007.26. Р.342-363.

9. Dacorogna M., Gencay R., Muller U. et al. An Introduction to High Frequency Finance. San Diego, CA: Academic Press, 2001.

10. Dacorogna M., Aste T., Di Matteo T. Long-term Memories of Developed and Emerging Markets: Using the Scaling Analysis to Characterize their Stage of Development // Journal of Banking & Finance. 2005. Vol.29. Р.827-851.

11. Diebold F. X., Inoue A. Long Memory and Regime Switching // Journal of Econometrics. 2001.105. Р.131-159.

12. Diebold F., Mariano R.S. Comparing Predictive Accuracy // Journal of Business & Economic Statistics. 1995. Vol.13 (3). Р.253-263.

13. Ding Z., Granger C. W. J. Varieties of Long Memory Models // Journal of Econometrics. 1996. Vol.73 (1). Р.61-77.

14. Eom C., Oh J., Jung W. - S. Relationship Between Efficiency and Predictability in Stock Price Change // Physica A. 2008.387. Р.5511-5517.

15. Granger C., Hyung N. Occasional Structural Breaks and Long Memory with an Application to the S&P 500 Absolute Stock Returns // Journal of Empirical Finance. 2004.11. Р.399-421.

16. Granger C., Newbold P. Forecasting Transformed Series // Journal of the Royal Statistical Society. 1976. Series B.38. Р.189-203.

17. Hamilton J. D. Rational-Expectations Econometric Analysis of Changes in Regime: An Investigation of the Term Structure of Interest Rates // Journal of Economic Dynamics and Control. 1988.12. Р.385-423.

18. Jamdee S., Los C. Long Memory Options: LM Evidence and Simulations // Research in International Business and Finance. 2007. Vol.21 (2). Р.260-280.

19. Kramer W., Sibbertsen P. Testing for Structural Change in the Presence of Long Memory: Technical Report 31/2000. SFB 471. University of Dortmund, 2000.

20. Kim C., Nelson J. State Space Models with Regime Switching: Classical and Gibbs Sampling Approaches with Application. Massachusetts: MIT Press, 1999.

21. LeBaron B. Agent Based Computational Finance: Suggested Readings and Early Research // Journal of Economic Dynamics and Control. 2000.24. Р.679-702.

22. Mandelbrot B., Wallis J. Noah, Joseph and Operational Hydrology // Water Resources Research. 1968. Vol.4. Р.909-918.

23. Rogers L. Arbitrage with Fractional Brownian Motion // Mathematical Finance. 1997.7 Р.95-105.

24. Sowell F. Maximum Likelihood Estimation of Stationary Univariate Fractionally Integrated Time Series Models // Journal of Econometrics. 1992. Vol.53. Р.165-188.

Размещено на Allbest.ru

Подобные документы

  • The General Economic Conditions for the Use of Money. Money and Money Substitutes. The Global Money Markets. US Money Market. Money Management. Cash Management for Finance Managers. The activity of financial institutions in the money market involves.

    реферат [20,9 K], добавлен 01.12.2006

  • Составление портфеля ценных бумаг. Изменение стоимости портфеля, нахождение его фактической доходности. Оценка эффективности инвестиционного проекта с точки зрения владельца портфеля. Виды финансовых инструментов. Депозитные и сберегательные сертификаты.

    курсовая работа [47,2 K], добавлен 26.01.2015

  • The concept, types and regulation of financial institutions. Their main functions: providing insurance and loans, asset swaps market participants. Activities and basic operations of credit unions, brokerage firms, investment funds and mutual funds.

    реферат [14,0 K], добавлен 01.12.2010

  • Тhe balance sheet company's financial condition is divided into 2 kinds: personal and corporate. Each of these species has some characteristics and detail information about the assets, liabilities and provided shareholders' equity of the company.

    реферат [409,2 K], добавлен 25.12.2008

  • Types and functions exchange. Conjuncture of exchange market in theory. The concept of the exchange. Types of Exchanges and Exchange operations. The concept of market conditions, goals, and methods of analysis. Stages of market research product markets.

    курсовая работа [43,3 K], добавлен 08.02.2014

  • The economic benefits to the recipient countries by providing capital, foreign exchange. The question of potential causality between foreign debt and domestic savings in the context of the Kyrgyz Republic. The problem of tracking new private businesses.

    реферат [26,7 K], добавлен 28.01.2014

  • Economic essence of off-budget funds, the reasons of their occurrence. Pension and insurance funds. National fund of the Republic of Kazakhstan. The analysis of directions and results of activity of off-budget funds. Off-budget funds of local controls.

    курсовая работа [29,4 K], добавлен 21.10.2013

  • Theoretical basis of long-term loans: concept, types. Characteristics of the branch of Sberbank of Russia. Terms and conditions of lending to households in Sberbank of Russia. Financing of investment projects. Risk - the main problem in the credit market.

    реферат [28,0 K], добавлен 17.09.2013

  • The chiral model of graphene based on the order parameter is suggested in the long-wave approximation, the ideal graphene plane being determined by the kink-like solution. Corrugation of the graphene surface is described in the form of ripple and rings.

    статья [211,7 K], добавлен 23.05.2012

  • Short and long run macroeconomic model. Saving and Investment in Italy, small open economy. Government expenditure and saving scatterplot. Loanable market equilibrium in closed economy in the USA. Okun’s Law in the USA and Italy, keynesian cross.

    курсовая работа [1,6 M], добавлен 20.11.2013

Работы в архивах красиво оформлены согласно требованиям ВУЗов и содержат рисунки, диаграммы, формулы и т.д.
PPT, PPTX и PDF-файлы представлены только в архивах.
Рекомендуем скачать работу.