
 Start

3039
 Prefix

The standard tool in analyzing the stochastic behavior of exchange rates has been
the autoregressive integrated moving average (ARIMA) model. Along these lines, a large
body of past research (see, for example,
 Exact

Baillie and Bollerslev (1989))
 Suffix

documents that
foreign currency rates are best characterized as pure unitroot (randomwalk or
martingale) processes, thus making predictability of exchange rate movements
impossible.
 (check this in PDF content)

 Start

3746
 Prefix

For example, size and limited
maturity of forward contracts may make it difficult for firms to hedge their foreign exchange risk.
improve upon the predictive accuracy of randomwalk forecasts motivated many
researchers to turn to nonlinear models.2
 Exact

Engel and Hamilton (1990)
 Suffix

use a Markov
switching model for exchange rate changes while Diebold and Nason (1990) and Meese
and Rose (1990) use variants of local regression, a nearestneighbor nonparametric
technique.
 (check this in PDF content)

 Start

3838
 Prefix

For example, size and limited
maturity of forward contracts may make it difficult for firms to hedge their foreign exchange risk.
improve upon the predictive accuracy of randomwalk forecasts motivated many
researchers to turn to nonlinear models.2 Engel and Hamilton (1990) use a Markov
switching model for exchange rate changes while
 Exact

Diebold and Nason (1990) and Meese and Rose (1990)
 Suffix

use variants of local regression, a nearestneighbor nonparametric
technique. Kuan and Liu (1995) use feedforward and recurrent artificial neural networks
to produce conditional mean forecasts of foreign exchange rates.
 (check this in PDF content)

 Start

3975
 Prefix

their foreign exchange risk.
improve upon the predictive accuracy of randomwalk forecasts motivated many
researchers to turn to nonlinear models.2 Engel and Hamilton (1990) use a Markov
switching model for exchange rate changes while Diebold and Nason (1990) and Meese and Rose (1990) use variants of local regression, a nearestneighbor nonparametric
technique.
 Exact

Kuan and Liu (1995)
 Suffix

use feedforward and recurrent artificial neural networks
to produce conditional mean forecasts of foreign exchange rates. Meese and Rose (1991)
estimate structural models of exchange rate determination using a variety of
nonparametric techniques.
 (check this in PDF content)

 Start

4116
 Prefix

to nonlinear models.2 Engel and Hamilton (1990) use a Markov
switching model for exchange rate changes while Diebold and Nason (1990) and Meese and Rose (1990) use variants of local regression, a nearestneighbor nonparametric
technique. Kuan and Liu (1995) use feedforward and recurrent artificial neural networks
to produce conditional mean forecasts of foreign exchange rates.
 Exact

Meese and Rose (1991)
 Suffix

estimate structural models of exchange rate determination using a variety of
nonparametric techniques. The success of these studies to explain exchange rate
movements has been very limited, thus leaving the martingale model as the leading
characterization of the data generating process of exchange rates.
 (check this in PDF content)

 Start

5972
 Prefix

2 Absence of riskneutral behavior, the nature of the policy regime, time deformation, and
misspecification of the functional form of the structural exchange rate model are some of the plausible
sources of nonlinearity in foreign exchange rates.
3 Longmemory methods have been applied extensively to financial asset price series: for stock prices
 Exact

(Lo (1991)),
 Suffix

futures prices (Barkoulas, Labys, and Onochie (1999)), commodity prices (Barkoulas, Labys, and
Onochie (1997)), Eurocurrency deposit rates (Barkoulas and Baum (1997)), for example.
 (check this in PDF content)

 Start

6139
 Prefix

of the structural exchange rate model are some of the plausible
sources of nonlinearity in foreign exchange rates.
3 Longmemory methods have been applied extensively to financial asset price series: for stock prices (Lo (1991)), futures prices (Barkoulas, Labys, and Onochie (1999)), commodity prices (Barkoulas, Labys, and
Onochie (1997)), Eurocurrency deposit rates
 Exact

(Barkoulas and Baum (1997)),
 Suffix

for example. See Baillie (1996)
for a survey of fractional integration methods and applications in economics and finance.
distant observations, its presence entails that past returns can help predict future returns,
and the possibility of consistent speculative profits arises.4
Applying rescaledrange (R/S) analysis to daily exchange rates for the British
pound, French franc, and Deu
 (check this in PDF content)

 Start

6192
 Prefix

of the plausible
sources of nonlinearity in foreign exchange rates.
3 Longmemory methods have been applied extensively to financial asset price series: for stock prices (Lo (1991)), futures prices (Barkoulas, Labys, and Onochie (1999)), commodity prices (Barkoulas, Labys, and
Onochie (1997)), Eurocurrency deposit rates (Barkoulas and Baum (1997)), for example. See
 Exact

Baillie (1996)
 Suffix

for a survey of fractional integration methods and applications in economics and finance.
distant observations, its presence entails that past returns can help predict future returns,
and the possibility of consistent speculative profits arises.4
Applying rescaledrange (R/S) analysis to daily exchange rates for the British
pound, French franc, and Deutsche mark, Booth, Kaen, and Koveos
 (check this in PDF content)

 Start

8463
 Prefix

As long memory represents a special form of nonlinear dynamics, it calls into question linear modeling
and invites the development of nonlinear pricing models at the theoretical level to account for longmemory behavior.
 Exact

Mandelbrot (1971)
 Suffix

observes that in the presence of long memory, the arrival of new
market information cannot be fully arbitraged away and martingale models of asset prices cannot be
obtained from arbitrage.
 (check this in PDF content)

 Start

9120
 Prefix

Statistical inferences
concerning asset pricing models based on standard testing procedures may not be appropriate in the
presence of longmemory series.
5 The classical rescaledrange (R/S) method has a number of drawbacks (see
 Exact

Lo (1991)).
 Suffix

differencing) parameter for the foreign currency rates. We use the Gaussian
semiparametric method to estimate the longmemory parameter. Contrary to previous
findings of strong persistence, our evidence strongly favors the martingale model
against longmemory alternatives in the foreign currency markets, thus supporting the
market efficiency hypothesis in its weak form.
 (check this in PDF content)

 Start

11017
 Prefix

The arbitrary restriction of d to integer values gives rise to the standard
autoregressive integrated moving average (ARIMA) model. The stochastic process ty is
both stationary and invertible if all roots of Φ(L) and Θ(L) lie outside the unit circle and
d<0. 5. Assuming that d∈0, 0.5() and d≠0,
 Exact

Hosking (1981)
 Suffix

shows that the correlation
function, ρ(⋅), of an ARFIMA process is proportional to 2d−1k as k→∞. Consequently,
the autocorrelations of the ARFIMA process decay hyperbolically to zero as k→∞
which is contrary to the faster, geometric decay of a stationary ARMA process.
 (check this in PDF content)

 Start

11999
 Prefix

For d∈0. 5,1[) the process is nonstationary (having an infinite
variance) but it is mean reverting, as there is no longrun impact of an innovation on
future values of the process.
We estimate the longmemory parameter using Robinson's Gaussian
semiparametric method.
 Exact

Robinson (1995)
 Suffix

proposes a Gaussian semiparametric estimate,
GS hereafter, of the selfsimilarity parameter H, which is not defined in closed form. It is
assumed that the spectral density of the time series, denoted by f⋅(), behaves as
fξ()~G1−2Hξ as ξ→+0(3)
for G∈0,∞() and H∈0,1().
 (check this in PDF content)

 Start

13188
 Prefix

The GS estimator appears to be the most
efficient semiparametric estimator developed so far. It is also consistent and has the
same limiting distribution under conditional heteroscedasticity
 Exact

(Robinson and Henry (1999)).
 Suffix

3. Data and Longmemory Results
3A. Data
The data set consists of U.S. dollar nominal rates of weekly frequency for the
Canadian dollar, Deutsche mark, British pound, French franc, Italian lira, Japanese yen,
Swiss franc, Netherlands guilder, Swedish krona, and Belgian franc and of monthly
frequency for the Austrian schilling, Danish krone, Luxembourg franc, Norw
 (check this in PDF content)

 Start

16549
 Prefix

For the
Italian lira, rejection of the unitroot null hypothesis is obtained but it is restricted to a
particular estimation window size, thus lacking robustness. This evidence for the major
currency rates does not confirm the longmemory evidence reported in
 Exact

Booth et al. (1982) and Cheung (1993).
 Suffix

For the remaining currencies, consistent evidence of long
memory is apparent for the Danish krone, Luxembourg franc, Portugese escudo, and
Spanish peseta. For the Belgian franc, Swedish krona, Finnish markka, and Greek
drachma, weak evidence of fractional dynamic behavior is found, but only for a specific
estimation window size.
 (check this in PDF content)

 Start

18077
 Prefix

The GPH results are not reported here but they are
available upon request from the authors.
7 Monte Carlo simulations have shown that such bandwidth choices provide a good balance in the tradeoff between bias (which tends to increase with bandwidth) and sampling variability (which tends to
decrease with bandwidth).
 Exact

Henry and Robinson (1999)
 Suffix

provide heuristic approximations of the minimum
mean squared error optimal bandwidth. Inferences drawn below remain unaltered for other plausible
bandwidth choices.
8 Alternative estimation methods of longmemory models include Sowell's (1992) exact maximum
likelihood method, Fox and Taqqu's (1986) frequency domain approximate maximum likelihood method,
and the condition
 (check this in PDF content)

 Start

18548
 Prefix

Inferences drawn below remain unaltered for other plausible
bandwidth choices.
8 Alternative estimation methods of longmemory models include Sowell's (1992) exact maximum
likelihood method, Fox and Taqqu's (1986) frequency domain approximate maximum likelihood method,
and the conditional sum of squares (CSS) method (Chung and
 Exact

Baillie (1993)).
 Suffix

The maximum likelihood
methods are computationally burdensome, especially in light of the repeated estimation in the sensitivity
analysis, and rely on the correct specification of the highfrequency (ARMA) structure to obtain consistent
parameter estimates.
 (check this in PDF content)

 Start

24118
 Prefix

A sensitivity analysis suggests that, when evidence
of long memory is obtained, it is sporadic and generally temporally unstable. The
evidence of long memory in major currency rates reported in
 Exact

Booth et al. (1982) and Cheung (1993)
 Suffix

would appear to be an artifact of the sample period and currencies
considered in those studies. For all but three of the broader set of currencies analyzed
here, we find that the unitroot hypothesis is robust to longmemory alternatives,
providing strong support for martingale behavior of currency rates and, therefore, for
foreign exchange market efficiency.
 (check this in PDF content)

 Start

25250
 Prefix

These weekly rates represent Friday noontime bid prices from the New
York foreign exchange market and were obtained from the Federal Reserve Board of
Governors. When Friday prices are not available Thursday prices are used. (The
construction of the data set of weekly observations follows
 Exact

Cheung (1993)).
 Suffix

For the U.S.
dollar nominal rates for the Austrian schilling, Danish krone, Luxembourg franc,
Norwegian krone, Finnish markka, Portugese escudo, and Spanish peseta, the frequency
of observation is monthly and the sample covers the period 08/1973 to 12/1995 for a
total of 268 returns observations.
 (check this in PDF content)