Academy of Economic Studies - Bucharest Doctoral School of Finance and Banking DOFIN Long Memory in Stock Returns: Research over Markets Supervisor: Professor Dr. Moisă Altăr MSc Student: Silvia Bardoş Bucharest, July 2008
Contents Long memory & Motivation Long memory & Motivation Literature review Literature review Steps & data used: Steps & data used: Testing stationarity and long memory Testing stationarity and long memory ADF & KPSS ADF & KPSS Hurst exponent through R/S test & Hurst exponent through wavelet estimator Hurst exponent through R/S test & Hurst exponent through wavelet estimator Determining long memory by estimating fractional differencing parameter Determining long memory by estimating fractional differencing parameter Geweke and Porter-Hudak test & Maximum Likelihood Estimate for an ARFIMA process Geweke and Porter-Hudak test & Maximum Likelihood Estimate for an ARFIMA process Conclusions Conclusions
Long Memory & Motivation Long memory has important implications in financial markets because if it is discovered it can be used to construct trading strategies. Long memory has important implications in financial markets because if it is discovered it can be used to construct trading strategies. Long memory or long range dependence means that the information from “today” is not immediately absorbed by the prices in the market and investors react with delay to any such information. Long memory or long range dependence means that the information from “today” is not immediately absorbed by the prices in the market and investors react with delay to any such information.So: A long memory process is a process where a past event has a decaying effect on future events A long memory process is a process where a past event has a decaying effect on future eventsAND Memory is the series property to depend on its own past realizations Memory is the series property to depend on its own past realizations
Mathematic view: Long memory processes relates to autocorrelation Mathematic view: Long memory processes relates to autocorrelation If a time series of data exhibits autocorrelation, a value from the data set x s at time t s is correlated with another value x s+z at time t s+z. For a long memory process autocorrelation decays over time and the decay is slower than in a stationary process ( I(0) process ) If a time series of data exhibits autocorrelation, a value from the data set x s at time t s is correlated with another value x s+z at time t s+z. For a long memory process autocorrelation decays over time and the decay is slower than in a stationary process ( I(0) process ) So, if a long memory process exhibits an autocorrelation function that is not consistent for a I(1) process (a process integrated of order 1) nor for an I(0) process (a pure stationary process) we can consider a long memory process as being the layer separating the non-stationary process from the stationary ones – namely a fractionally integrated process. So, if a long memory process exhibits an autocorrelation function that is not consistent for a I(1) process (a process integrated of order 1) nor for an I(0) process (a pure stationary process) we can consider a long memory process as being the layer separating the non-stationary process from the stationary ones – namely a fractionally integrated process. Long Memory & Motivation
Literature review Evidence of long memory was first brought up by E. Hurst in 1951 when, testing the behavior of water levels in the Nile river, he observed that the flow of the river was not random, but patterned Evidence of long memory was first brought up by E. Hurst in 1951 when, testing the behavior of water levels in the Nile river, he observed that the flow of the river was not random, but patterned Mandelbrot (1971) was among the first to consider the possibility of long range dependence in asset returns Mandelbrot (1971) was among the first to consider the possibility of long range dependence in asset returns Wright, J. (1999) is detecting evidence of long memory in emerging markets stock returns (Korea, Philippines, Greece, Chile and Colombia) Wright, J. (1999) is detecting evidence of long memory in emerging markets stock returns (Korea, Philippines, Greece, Chile and Colombia) Caporale and Gil-Alana (2002), studying S&P 500 daily returns found results indicating that the degree of dependence remains relatively constant over time, with the order of integration of stock returns fluctuating slightly above or below zero Caporale and Gil-Alana (2002), studying S&P 500 daily returns found results indicating that the degree of dependence remains relatively constant over time, with the order of integration of stock returns fluctuating slightly above or below zero Henry Olan (2002) makes a survey for finding long memory in stock returns from an international perspective. Evidence of long memory is found in the German, Japanese, South Korean and Taiwanese markets against UK, USA, Hong Kong, Singapore and Australia where no sign of long memory appears. Henry Olan (2002) makes a survey for finding long memory in stock returns from an international perspective. Evidence of long memory is found in the German, Japanese, South Korean and Taiwanese markets against UK, USA, Hong Kong, Singapore and Australia where no sign of long memory appears.
Steps – Modeling long memory A series x t follows an ARFIMA (p,d,q) process if: where Φ(L), θ(L) are the autoregressive and moving average polynomials, L is the lag, d is the fractional differencing parameter, εt is white noise. For d within (0,0.5), the ARFIMA process is said to exhibit long memory or long range positive dependence For d within (-0.5, 0), the process exhibits intermediate memory or long range negative dependence For d within [0.5, 1) the process is mean reverting and there is no long run impact to future values of the process The process is short memory for d=0 corresponding to a standard ARMA process
Testing stationarity Memory is closely related to the order of integration Memory is closely related to the order of integration In the context of non-fractionally integration is equivalent to establish whether the series is I(0) or I(1) and the commonly used tests are ADF and KPSS In the context of non-fractionally integration is equivalent to establish whether the series is I(0) or I(1) and the commonly used tests are ADF and KPSSADF Null hypothesis: H 0 : d = 1 (returns series are containing a unit root) Hassler and Wolter (1994) find that this test of unit root is not consistent against fractional alternatives so the ADF can be inappropriate if we are trying to decide whether a set of data is fractionally integrated or not. KPSS Null hypothesis: H0: d = 0 (return series are stationary) Lee and Schmidt (1996) find that KPSS test can be used to distinguish short memory and long memory stationary processes
Testing stationarity KPSS Consider x t ( t = 1, 2, …, N), as the observed return series for which we wish to test stationarity The test decomposes the series into the sum of a random walk, a deterministic trend and a stationary error with the following linear regression model: The KPSS statistics: andis the residual from regressing the series against a constant or a constant and a trend Under the null hypothesis of trend stationary, the residuals e t ( t = 1, 2, …, N ) are from the regression of x on an intercept and time trend. Under the null hypothesis of level stationarity, the residuals e t are from a regression of x on intercept only. Rejection of ADF and KPSS indicates that the process is described by neither I(0) and I(1) processes and that is probable better described by the fractional integrated alternative (d is a non-integer).
Estimating long memory using R/S test R/S test Mandelbrot & Wallis (1969) method allows computing parameter H, which measures the intensity of long range dependence in a time series Return time series of length T is divided into n sub-series of length m. For each sub-series m = 1,..., n, we: a) find the mean (E m ) and standard deviation (S m ); b) we subtract the sample mean Z i,m = X i,m − E m for i = 1,..,m; c) produce a time series taking form of W i,m = j,m where i = 1,…,m d) find the range R m = max{W 1,m,…., W n,m } – min{ W 1,m,…., W n,m } e) rescale the range R m by How does this procedure relates to the Hurst exponent?
Einstein discovered that the distance covered by a random variable is close related to the square root of time (Brownian motion), where R is the distance covered by the variable, k is a constant and T is the length of the time. Using R/S analysis, Hurst suggested that:, where R/S is the rescaled range, m is the number of observations, k is the constant and H is the Hurst exponent, can be applied to a bigger class of time series (generalized Brownian motion) The Hurst exponent can be than found as: log(R/S)m= log k + H log m H valueReturn time series = 0.5follow a random walk and are independent (0,0.5) are anti-persistent, process covers only a small distance than in the random walk case (0.5,1) are persistent series, process covers a bigger distance than a random walk (long memory) Estimating long memory using R/S test
Hurst exponent using wavelet spectral density For computing the Hurst Exponent, the R wavelet estimator uses a discrete wavelet transform then: averages the squares of the coefficients of the transform, performs a linear regression on the logarithm of the average, versus the log of the parameter of the transform The result provides an estimate for the Hurst exponent. Wavelet transform behaves as a microscope that decomposes our return series into components of different frequency so this is why we tend to consider that results obtained for H through the wavelet estimator are being more accurate.
The GPH test (1983) Semi-parametric approach to obtain an estimate of the fractional differencing parameter d based on the slope of the spectral density function around the frequency ξ=0 Periodogram (estimator of the spectral density) of x at a frequency ξ I (ξ) = Geweke, J. and S. Porter-Hudak(1983) proposed as an estimate of the OLS estimator of d from the regression:, λ= 1,…..,v the bandwidth v is chosen such that forbut Geweke and Porter-Hudak consider that the power of T has to be within (0.5,0.6). In our test we have considered: V =
Maximum likelihood estimates for ARFIMA model In the present paper we have used the MLE implemented based on the approximate maximum likelihood algorithm of Haslett and Raftery (1989) in R. If the estimated d is significantly greater than zero, we consider it an evidence of the presence of long-memory.
For testing the existence of long memory we have selected indexes around the world trying to compare return series in mature markets (US, UK, Germany, France, Japan) with emerging markets (Romania, Poland and the BRIC countries) For testing the existence of long memory we have selected indexes around the world trying to compare return series in mature markets (US, UK, Germany, France, Japan) with emerging markets (Romania, Poland and the BRIC countries) For the data series (1997 – 2008) we have first established the length as being 2 (for the wavelet transform performed by the soft) and then we have transformed it in return series through: For the data series (1997 – 2008) we have first established the length as being 2 (for the wavelet transform performed by the soft) and then we have transformed it in return series through: For testing and comparing we have selected mainly, daily returns For testing and comparing we have selected mainly, daily returns Stationarity test were run in Eviews and long memory tests and estimation procedures were run in R Stationarity test were run in Eviews and long memory tests and estimation procedures were run in Rn Data used
Is there evidence of long memory in the return time series? S&P 500 daily return series S&P 500 daily return series Null Hypothesis: SP500DAY has a unit root Exogenous: Constant, Linear Trend Lag Length: 0 (Automatic based on SIC, MAXLAG=25) t-Statistic Prob.* Augmented Dickey-Fuller test statistic Test critical values:1% level % level % level GPH d tstat sd (d=0) tstat asd (d=0) ARFIMA (0,d,0)mleValue d4.583E-05 ADFKPSS R/S Hurst Exponent Diagnostic: Wavelet estimator for H: Null Hypothesis: SP500DAY is stationary Exogenous: Constant Bandwidth: 22 (Newey-West using Bartlett kernel) LM-Stat. Kwiatkowski-Phillips-Schmidt-Shin test statistic Asymptotic critical values*:1% level % level % level
FTSE100 daily return series FTSE100 daily return series Null Hypothesis: FTSE100DAY has a unit root Exogenous: Constant, Linear Trend Lag Length: 2 (Automatic based on SIC, MAXLAG=25) t-Statistic Prob.* Augmented Dickey-Fuller test statistic Test critical values:1% level % level % level GPH\d= d tstat sd tstat asd ARFIMA (0,d,0)mleValue d4.583E-05 ADFKPSS R/S Hurst Exponent Diagnostic: Wavelet estimator for H: Null Hypothesis: FTSE100DAY is stationary Exogenous: Constant Bandwidth: 17 (Newey-West using Bartlett kernel) LM-Stat. Kwiatkowski-Phillips-Schmidt-Shin test statistic Asymptotic critical values*:1% level % level % level
Null Hypothesis: BETFIDAYP has a unit root Exogenous: Constant, Linear Trend Lag Length: 0 (Automatic based on SIC, MAXLAG=21) t-Statistic Prob.* Augmented Dickey-Fuller test statistic Test critical values:1% level % level % level GPH\d= d tstat sd tstat asd ARFIMA (0,d,0)mleValue d BET-FI daily return series BET-FI daily return series ADFKPSS R/S Hurst Exponent Diagnostic: Wavelet estimator for H: Null Hypothesis: BETFIDAYP is stationary Exogenous: Constant Bandwidth: 7 (Newey-West using Bartlett kernel) LM-Stat. Kwiatkowski-Phillips-Schmidt-Shin test statistic Asymptotic critical values*:1% level % level % level
BOVESPA daily return series BOVESPA daily return series ADFKPSS Null Hypothesis: BOVESPADAY has a unit root Exogenous: Constant, Linear Trend Lag Length: 0 (Automatic based on SIC, MAXLAG=25) t-Statistic Prob.* Augmented Dickey-Fuller test statistic Test critical values:1% level % level % level GPH\d= d tstat sd tstat asd ARFIMA (0,d,0)mleValue d R/S Hurst Exponent Diagnostic: Wavelet estimator for H: Null Hypothesis: BOVESPADAY is stationary Exogenous: Constant Bandwidth: 16 (Newey-West using Bartlett kernel) LM-Stat. Kwiatkowski-Phillips-Schmidt-Shin test statistic Asymptotic critical values*:1% level % level % level
RTS daily return series ADFKPSS Null Hypothesis: RTSDAY has a unit root Exogenous: Constant, Linear Trend Lag Length: 0 (Automatic based on SIC, MAXLAG=25) t-Statistic Prob.* Augmented Dickey-Fuller test statistic Test critical values:1% level % level % level GPH\d= d tstat sd tstat asd ARFIMA (0,d,0)mle\d=0Value d R/S Hurst Exponent Diagnostic: Wavelet estimator for H: Null Hypothesis: RTSDAY is stationary Exogenous: Constant Bandwidth: 1 (Newey-West using Bartlett kernel) LM-Stat. Kwiatkowski-Phillips-Schmidt-Shin test statistic Asymptotic critical values*:1% level % level % level
SENSEX daily return series ADFKPSS Null Hypothesis: SENSEXDAY has a unit root Exogenous: Constant, Linear Trend Lag Length: 0 (Automatic based on SIC, MAXLAG=25) t-Statistic Prob.* Augmented Dickey-Fuller test statistic Test critical values:1% level % level % level GPH\d= d tstat sd tstat asd ARFIMA (0,d,0)mle\d=0Value d R/S Hurst Exponent Diagnostic: Wavelet estimator for H: Null Hypothesis: SENSEXDAY is stationary Exogenous: Constant Bandwidth: 10 (Newey-West using Bartlett kernel) LM-Stat. Kwiatkowski-Phillips-Schmidt-Shin test statistic Asymptotic critical values*:1% level % level % level
Hang Seng daily return series ADFKPSS Null Hypothesis: HANGSENGDAY has a unit root Exogenous: Constant, Linear Trend Lag Length: 0 (Automatic based on SIC, MAXLAG=25) t-Statistic Prob.* Augmented Dickey-Fuller test statistic Test critical values:1% level % level % level GPH\d= d tstat sd tstat asd ARFIMA (0,d,0)mleValue d4.583E-05 R/S Hurst Exponent Diagnostic: Wavelet estimator for H: Null Hypothesis: HANGSENGDAY is stationary Exogenous: Constant Bandwidth: 5 (Newey-West using Bartlett kernel) LM-Stat. Kwiatkowski-Phillips-Schmidt-Shin test statistic Asymptotic critical values*:1% level % level % level
Index dailyH value via R/S S&P FTSE CAC DAX NIKKEI WIG0.593 BET BET C BET FI BOVESPA RTS SENSEX HANG SENG Index dailyH value via Wavelet estimator S&P FTSE CAC DAX NIKKEI WIG BET BET C BET FI BOVESPA RTS SENSEX HANG SENG Comparison between indices - Hurst BRIC
Conclusions Using a range of test and estimation procedures we have investigated whether stock returns exhibit long memory Our results come to increase a bit the idea that emerging markets have a weak form of long memory as resulted in case of Russia and India or a stronger form like discovered in case of Romania (BET-FI), China and Brazil. Mature markets, in which we include US & UK among Germany, France show mixed evidence We have tested for long memory the return series for BRIC countries indicesWhy? Because it is important to see is there is some kind of correlation between distant observations in these markets as emerging markets are of great interest to potential investors first taking into account their returns and second because they can be used in case of portfolio diversification as emerging market countries have low correlation with mature markets.