Long memory and fractional integration in high frequency data on the US dollar/British pound spot exchange rate

https://doi.org/10.1016/j.irfa.2013.03.011Get rights and content

Highlights

  • We examine high frequency financial data.

  • We focus on temporal aggregation and fractional integration.

  • Lower degrees of integration are associated with lower frequencies.

  • Data collected every ten minutes display mean reversion.

  • For higher frequencies the unit root cannot be rejected.

Abstract

This paper analyses the long-memory properties of a high-frequency financial time series dataset. It focuses on temporal aggregation and other features of the data, and how they might affect the degree of dependence of the series. Fractional integration or I(d) models are estimated with a variety of specifications for the error term. In brief, we find evidence that a lower degree of integration is associated with lower data frequencies. In particular, when the data are collected every 10 min there are several cases with values of d strictly smaller than 1, implying a mean-reverting behaviour; however, for higher data frequencies the unit root null cannot be rejected. This holds for all four series examined, namely Open, High, Low and Last observations for the US dollar/British pound spot exchange rate and for different sample periods.

Introduction

The Efficient Market Hypothesis (EMH) in its weak form rules out the possibility of abnormal systematic profits over and above transaction costs and risk premia, as prices should fully reflect available information (see Fama, 1970). The implication is that stock prices should follow a random walk process, which implies unpredictable returns (see Summers, 1986). Therefore, a finding of mean reversion in stock prices is seen as inconsistent with equilibrium asset pricing models (see, e.g., Fama and French, 1988, Poterba and Summers, 1988). A large number of studies have been carried out to establish whether (log-)prices are indeed I(1) and, consequently, stock market returns I(0) series, although business cycle variation and short-range dependence might also lead to a rejection of long memory in stock prices (see Lo, 1991). However, as we argued in Caporale and Gil-Alana (2002), the assumptions imposed by standard unit root tests might be too restrictive, and the possibility of fractional orders of integration with a slow rate of decay should be considered. Therefore that study performed tests allowing for fractional alternatives and incorporating the I(0) and the I(1) models as particular cases of interest, and found that US real stock returns are close to being I(0) (note that if shocks are weakly autocorrelated, markets will not be efficient).

A subsequent contribution (see Caporale & Gil-Alana, 2007) decomposed the stochastic process followed by US stock prices into a long-run component described by the fractional differencing parameter (d) and a short-run (ARMA) structure. Finally, in Caporale and Gil-Alana (2008) we introduced a more general model which, instead of considering exclusively the component affecting the long-run or zero frequency, also takes into account the cyclical structure. Specifically, a procedure was applied which allows to test simultaneously for unit roots with possibly fractional orders of integration at both the zero and the cyclical frequencies. Modelling simultaneously with long range dependence the zero and the cyclical frequencies can solve at least to some extent the problem of misspecification that might arise with respect to these two frequencies.

However, the fractional differencing parameter may be very sensitive to the data frequency used in the analysis. In fact, it has often been claimed that aggregation is behind fractional integration: Robinson (1978) and Granger (1980) showed that the aggregation of heterogeneous individual AR processes may produce fractional integration. On the other hand, it is well known that temporal aggregation leads to finite sample biases in the estimates of the fractional differencing parameter (see, e.g. Souza & Smith, 2002).1This is the main issue that will be investigated in the present study by using a high frequency dataset on the US dollar–British pound spot exchange rate collected every 1, 2, 3, 5, and 10 min. As in Caporale and Gil-Alana (2008), we start the analysis by using a general long memory model that incorporates poles at both the zero and a cyclical frequency; however, since the evidence clearly suggested an order of integration not significantly different from zero for the cyclical frequency, we then focus exclusively on the long run or zero frequency.

Fractional integration in exchange rates markets has been examined in various papers, many of them testing the Purchasing Power Parity (PPP) condition, which occupies a central place in international economics. Applying R/S techniques to daily rates for the British pound, French franc and Deutsche mark, Booth, Kaen, and Koveos (1982) found positive memory during the flexible exchange rate period (1973–1979) but negative one (i.e., anti-persistence) during the fixed exchange rate period (1965–1971). Later, Cheung (1993) also found evidence of long memory behaviour in foreign exchange markets during the managed floating regime. On the other hand, Baum, Barkoulas, and Caglayan (1999) estimated fractional ARIMA (ARFIMA) models for real exchange rates in the post-Bretton Woods era and found almost no evidence to support long run PPP. Additional studies on exchange rate dynamics using fractional integration are those by Crato and Ray (2000), Wang (2004), Dufrenot, Mathieu, Mignon, and Peguin-Feissolle (2006), Dufrenot, Lardic, Mathieu, Mignon, and Peguin-Feissolle (2008) and Aloy, Booutahar, Gente, and Peguin-Feissolle (2011) among others. All these papers, however, focus on low frequency (mainly quarterly) data, and do not examine the case of high frequency (intra-day) data.

The present study focuses on the case of spot exchange rates with the aim of gaining some insights into the interaction between fractional integration and high frequency data. The results suggest that lower degrees of memory are associated with lower data frequencies. The layout of the paper is as follows. Section 2 describes the econometric methodology used. Section 3 provides details of the data and discusses the empirical results. Section 4 summarises the main findings and offers some concluding remarks.

Section snippets

Methodology

There are two definitions of long memory, one in the frequency domain and the other in the time domain. Let us consider a zero-mean covariance stationary process {xt, t = 0, ± 1, …} with autocovariance function γu = E(xtxt + u). The time domain definition of long memory states that:u=γu=.

Assume that xt has an absolutely continuous spectral distribution, so that it has a spectral density function, f(λ); according to the frequency domain definition of long memory the spectral density function is

Data and empirical results

The data used for the analysis are taken from Reuters, and are intraday data at the 1, 2, 3, 5, 10-minute frequency. Specifically, the series examined is the spot nominal exchange rate of the US dollar pound vis-à-vis the British pound, for different sample periods with a duration of one and a half days. We only report the results for the period 13/05/2010 (11:47)–14/05/2010 (21:07), since those for other samples were very similar.3

Conclusions

Despite the existence of a very extensive literature, there is still lack of consensus on what is the most appropriate specification for many financial series. For instance, whether asset returns of asset prices are predictable or not is still controversial: the efficiency market hypothesis suggests that they should follow a random walk (see Fama, 1970), but mean reversion is often found (see, e.g., Poterba & Summers, 1988). More recently, it has become clear that it is essential to consider

References (45)

  • C. Wang

    Futures trading activity and predictable foreign exchange movements

    Journal of Banking and Finance

    (2004)
  • C.F. Baum et al.

    Persistence in the international inflation rates

    Southern Economic Journal

    (1999)
  • P. Bloomfield

    An exponential model in the spectrum of a scalar time series

    Biometrika

    (1973)
  • T. Bollerslev et al.

    High frequency data, frequency domain inference and volatility forecasting

    The Review of Economics and Statistics

    (2000)
  • G.M. Caporale et al.

    Long run and cyclical dynamics in the US stock market

  • G.M. Caporale et al.

    Modelling long-run trends and cycles in financial time series data

  • G.M. Caporale et al.

    Fractional integration and data frequency

    Journal of Statistical Computation and Simulation

    (2010)
  • Y.W. Cheung

    Long memory in foreign exchange rates

    Journal of Business and Economic Statistics

    (1993)
  • J. Cotter

    Uncovering long memory in high frequency UK futures

    European Journal of Finance

    (2005)
  • N. Crato et al.

    Memory in returns and volatilities of future's contracts

    Journal of Futures Markets

    (2000)
  • R. Dahlhaus

    Efficient parameter estimation for self-similar process

    The Annals of Statistics

    (1989)
  • M. Demetrescu et al.

    Long memory testing in the time domain

    Econometric Theory

    (2008)
  • Cited by (25)

    • Persistence and long run co-movements across stock market prices

      2023, Quarterly Review of Economics and Finance
      Citation Excerpt :

      Evidence of mean reversion (i.e., estimates of d significantly below 1) is only obtained in some cases for the two American indices (S&P500 and Dow Jones) and generally, lower degrees of integration are obtained at lower (monthly) frequencies. These results, showing little evidence of mean reversion are consistent with some other previous works (DePenya & Gil-Alana, 2004; Tabak, 2007; Narayan, 2008; Hasanov, 2009; Gozbasi et al., 2014; Tiwari & Kyophilavon, 2014; etc.), and the fact that there is a reduction in the magnitude of d at lower (monthly) frequencies is also consistent with Caporale et al. (2013) who show that lower degrees of integration were associated with lower frequencies.3 Our results for the individual series differ from Adekoya (2021) who found evidence of mean reversion and long memory in 18 OECD countries using monthly samples and longer periods of analysis than in our work (in that paper the data spans from January, 1973 to August, 2018) and from Caporale et al. (2020a, 2020b) regarding the Russian market that found no-persistence or mean-reversion with daily samples between 2010 and 2018.

    • The economic importance of rare earth elements volatility forecasts

      2020, International Review of Financial Analysis
    • Long memory or regime switching in volatility? Evidence from high-frequency returns on the U.S. stock indices

      2020, Pacific Basin Finance Journal
      Citation Excerpt :

      Long-memory persistence describes the property of financial series whose sample autocorrelations are significantly different from zero, even for large lags (Baillie et al., 1996; Bollerslev and Mikkelsen, 1996; Diebold and Inoue, 2001). In many recent studies, long-memory persistence is extensively observed, especially for the high-frequency financial series (Fleming and Kirby, 2011; Degiannakis et al., 2013; Caporale and Gil-Alana, 2013; Ho et al., 2013). However, apart from the actual everlasting effects of autocorrelations, it is well known that regime switching can also cause the long memory (Diebold and Inoue, 2001; Choi et al., 2010; Wang and Vasilakis, 2012).

    • Long memory and regime switching: A simulation study on the Markov regime-switching ARFIMA model

      2015, Journal of Banking and Finance
      Citation Excerpt :

      Long memory describes the property of financial series, whose sample autocorrelations are significantly different from zero, even for large lags (Baillie et al., 1996; Bollerslev and Mikkelsen, 1996; Diebold and Inoue, 2001). In many recent studies, long memory is extensively observed, especially for the high-frequency financial series (Chortareas et al., 2011; Fleming and Kirby, 2011; Degiannakis et al., 2013; Caporale and Gil-Alana, 2013; Ho et al., 2013). In Fig. 3, the autocorrelation functions (ACFs) of hourly log Garman–Klass and realized volatility of London Stock Exchange (FTSE) index ranging from 1/1/2001 to 31/12/2012 are plotted.

    • Structural breaks and long memory in modeling and forecasting volatility of foreign exchange markets of oil exporters: The importance of scheduled and unscheduled news announcements

      2014, International Review of Economics and Finance
      Citation Excerpt :

      Aloy, Boutahar, Gente, and Péguin-Feissolle (2012), using the U.S. Dollar, the UK pound and the German Deutsche mark as numeraires, find that only few bilateral real exchange rates exhibit the LM mean-reverting properties. More recently, using intra-day data (1, 2, 3, 5, and 10 min) for the USD/GBP spot exchange rate and for different sample periods, Caporale and Gil-Alana (2013) examine the LM properties and find evidence that a lower degree of integration is associated with lower data frequencies. When the data are extracted every 10 min, there are several cases with values of the LM parameter d strictly smaller than 1, implying a mean-reverting behavior.

    View all citing articles on Scopus

    The second-named author gratefully acknowledges the financial support from the Ministry of Education of Spain (ECO2011-2014–28196-ECON Y FINANZAS, Spain) and from a Jeronimo de Ayanz project of the Government of Navarra. We are also grateful to the Editor and the two anonymous referees as well as Walter Kramer and the other participants in the Symposium on high frequency data in empirical finance, Technische Universitat Dortmund, Dortmund, Germany, 1–2 July 2010.

    View full text