Applied Financial Economics, 2005, 15, 1219–1225

A re-examination of the predicting

power of forward premia
Peijie Wang
Business School, The University of Hull, Cottingham Road,
Hull HU6 7RX, UK
E-mail: p.wang@hull.ac.uk

The paper proposes that the spot exchange rate consist of two parts.
Important information content is with its underlying movement, in
accordance with the development in the economy and the adjustment in
economic activity. The paper then extracts the underlying movement from
the spot exchange rate using the state space method and the frequency
domain method. The extracted component is persistent as expected,
catching trend movement and matching the statistical characteristics of
the forward premium to some degree. Based on the results, the paper is
able to reject the finding in many previous studies that the forward
premium predicts the future spot rate in a completely wrong way. The
paper concludes that the forward premium does not help explain the future
spot rate as its most feasible result.

I. Introduction
in economic activity. It is this component that is
The forward premium puzzle in foreign exchange of significance in the foreign exchange market, and
rate determination has been the focus of studies may be explained by the forward premium. This
into foreign exchange market efficiency over the component is persistent as it catches the trend,
last two decades. The issue remains unsolved despite which shares a similar pattern, statistically, with the
tremendous efforts having been made by researchers forward premium. In this spirit, we extract the
using various econometric methods. The present underlying movement from the spot exchange rate
paper differs from all previous studies in its funda- in a fundamental way.
mental perception of the forward premium. The idea of the present paper is motivated by the
Consequently, the modelling strategy for the relation- failure of many previous studies to identify a reason-
ship between changes in the spot rate and forward ably acceptable correspondence between the change
premia differs fundamentally from previous studies. in the spot rate and the forward premium. These
We believe that the forward premium puzzle studies range from early years, such as Geweke and
will never be solved technically unless academics Feige (1979), Hansen and Hodrick (1980), Bilson
change their view of the underlying process of (1981), Longworth (1981), MacDonald (1983),
exchange rates. Fama (1984), Gregory and McCurdy (1984),
The paper proposes that the spot exchange rate Hsieh (1984), MacDonald and Torrance (1988),
or the change in the spot rate consist of two parts. and Copeland (1989); and stretch into the 1990s,
Much of the change in the spot rate is trading noise, for example, Cavaglia et al. (1994), and Peel and
random walk like, cannot and need not be modelled. Pope (1995). More recent efforts have been made
Any important information content is contained by Schotman et al. (1997) and Huisman et al.
within its underlying movements, in accordance (1998). Many possible irregularities and factors
with developments in the economy and adjustments have been considered, e.g., time-varying risk
Applied Financial Economics ISSN 0960–3107 print/ISSN 1466–4305 online # 2005 Taylor & Francis 1219
http://www.tandf.co.uk/journals
DOI: 10.1080/09603100500360920
1220 P. Wang
premium, time-varying volatility, and the news, but State space estimation
they do not achieve fundamentally different results.
The state space model is as follows:
A regression intercept of zero and slope of unity
probably should not be simply interpreted as a st ¼ sT,t þ sH,t ð1aÞ
criterion for market efficiency, which previous
studies have striven hard to achieve, or been sT,t ¼ rt þ sT,t
1 þ wt ð1bÞ
puzzled if is not met. Probably a slope of the
statistically insignificant unity is a better indication rt ¼ c þ
rt
1 þ ut ð1cÞ
of market efficiency. So investors cannot consis-
ðLÞsH,t ¼ vt ð1dÞ
tently make abnormal profits on the one hand,
and they will not be guaranteed a sure outcome where st is the logarithm of the spot exchange rate st;
on the other hand. That is, on average the forward sT,t is the trend component of st; sH,t is the remaining
rate equals the future spot rate, but the former can part in st that is highly variable, or the higher
be higher or lower than the latter and the deviation frequency components of the spot rate; and rt is the
can be large. In fact, the market mechanism for rate of change in the trend component of the spot
forward exchange will eliminate the possibility of rate. sT,t, sH,t and rt are unobserved state variables.
zero deviation but there ought to be no upper Unlike Clark (1993), where the rate of change is a
limit for the deviation. If the above reasoning is pure random walk, our model has a stochastic rate
sensible, then any insignificant slope, preferably of change that can be stationary or non-stationary
small, will not be seen as violating market depending on the value of
. Specifically, if
is
efficiency conditions. Further, a non-zero intercept, smaller than but close to one, the rate of change is
instead of being interpreted as an indication of persistent in its behaviour, similar to the forward
market inefficiency, should be seen as a correct premium. Estimation of the state variables can be
expectation of market movement, when it is not achieved by using the Kalman filter.
statistically different from the value of the forward Equation 1 can be expressed in the matrix form:
premium while the slope is insignificant.
The rest of the paper is organized as follows. st ¼ Ht ð2aÞ
Section II introduces the methodology of this study. t ¼ m þ Ft
1 þ vt ð2bÞ


Empirical investigations are carried out in Section III, 0 Q for t ¼ s
searching for fundamental links between the change Eðvt vs Þ ¼ ð2cÞ
0 for t 6¼ s
in the spot rate and the forward premium using the
US dollar vis-à-vis the British pound. Section IV where
summarizes the study with concluding remarks.  
H¼ 1 1 0 0
2 3
1 0 0 1
6 7
II. Methodology 60 ’ ’ 07
6 1 2 7
F¼6
6
7
7
We use two approaches to decompose the spot 60 1 0 07
4 5
exchange rate and to extract the trend component
from the spot rate. One is the state space approach 0 0 0 l
and the other is in the frequency domain. Estimating 0  
t ¼ sT,t sH,t sH,t
1 rt
economic time series in state space, realized by a
Kalman filter, was proposed by Clark (1987) in the  
m0 ¼ 0 0 0 c
decomposition of trend and cycle components in
2 3
US GNP. Our modelling strategy is similar but we w2 0 0 0
have a different specification for the growth rate in 6 7
6 0 v2 0 0 7
the trend, or more precisely for our case, changes 6 7
in the underlying trend in the spot rate. The Q¼6
6
7
7
6 0 0 0 0 7
frequency domain method was used by Copeland 4 5
and Wang (1993) to separate high and low frequency 0 0 0 u2
components in the spot exchange rate. The two
approaches are discussed in the following two Equation 2a is the observation equation or measure-
sub-sections. ment equation; Equation 2b is the state equation
A re-examination of the predicting power of forward premia
or transition equation; and t is the state vector.
The Kalman filter procedure is as follows:
t,t
1 ¼ m þ Ft
1jt
1
Pt,t
1 ¼ FPt
1jt
1 F 0 þ Q
"t ¼ st
Htjt
1
1221
is the lower frequency trend component of spot
exchange rate changes. The simplest high pass filter
is a rectangular window with its value being one for
ð3Þ k  kP, where kP is chosen to be the cut-off frequency,
ð4Þ and zero at other frequency points. As sharp edged
filters do not work well, we use a high pass filter that
ð5Þ ‘cosinely’ decreases with frequency.

t ¼ HPtjt
1 H 0 ð6Þ
III. Empirical Results
t,t ¼ tjt
1 þ Ptjt
1 H 0 ð t Þ
1 "t ð7Þ
Monthly spot exchange rates and 30 day forward
Pt,t ¼ Ptjt
1
Ptjt
1 H 0 ð t Þ
1 HPtjt
1 ð8Þ exchange rates of the British pound vis-à-vis the US
dollar are used. The data set runs from January 1976
t can be regarded the system wide variance.
The conditional density function is and ends in August 2000. Table 1 presents the basic
! statistics of the spot exchange rate and the forward
1 "2t premium. Figure 1 shows the spot rate during
f ðst jIt
1 Þ ¼ pffiffiffiffiffiffiffiffiffiffi exp
ð9Þ this period, and Fig. 2 and 3 show and compare
2 t t
changes in the spot rate and the forward premium
where It
1 is the information set at time t
1. in the time domain and the frequency domain,
The Kalman filter can be estimated by maximizing respectively. It can be seen that the standard devia-
the log likelihood of the above density function: tion of the spot rate is more than 12 times larger than
that of the forward premium; or the variance of the
X
T
former is almost 150 times larger than that of the
Max : log f ðst jIt
1 Þ
t¼1
latter. Figure 2 demonstrates the difference strikingly.
( !) It is clearly not sensible to run regression analysis
1X T
"2t between these two variables with an expectation
¼ Max :
logð t Þ þ ð10Þ
2 t¼1 t that the slope can be unity. Figure 3 further reveals
the differences between spot rate changes and the
sT,t can be obtained as one of the state variables forward premium in the frequency domain. First,
accordingly. changes in the spot rate have much more power
or energy than changes in the forward premium.
Frequency domain decomposition Second, the forward premium consists of many
more low frequency components relative to high
Frequency domain decomposition is accomplished frequency components, while changes in the spot
by transforming the exchange rate time series into rate behave like white noise. The log spectra of the
its corresponding frequency elements – a ’series’ two series are graphed on Fig. 3 to highlight these
with the frequency subscript. Then the higher features.
frequency components are filtered out and the lower The Kalman filter decomposition results are
frequency trend components are extracted. reported in Table 2. The table shows that rt follows
The Fourier transform of the time series st is a very persistent process with l being 0.98, suggesting
defined as: sT,t behaves in a way similar to the forward premium
X
T
1 in this regard. The standard deviation of sH,t is much
Sk ¼ st e
jtð2k=N Þ k ¼ 0,1, . . . , T ð11Þ larger that that of sT,t. In fact,  w is close to, and
t¼0 slightly smaller than, the standard deviation of the
Define Fk as a low pass filter, i.e. it filters out the
high frequency components and keeps the low
Table 1. Basic statistics
frequency components. Let st pass through the filter:
Standard
ST,k ¼ Fk Sk ð12Þ
Mean deviation
The inverse Fourier transform of ST,k: st
0.1013 e
2 (0.1872 e
2) 0.3216 e
1
f t
st
0.2009 e
2 * (0.1530 e
3) 0.2635 e
2
X
T
1
sT,t ¼ ST,k e jkð2t=N Þ ð13Þ Notes: *Significant at the 1% level. Standard errors in
k¼0 parentheses.
1222 P. Wang
3

2.5

1.5

0.5

0
76M2 78M2 80M2 82M2 84M2 86M2 88M2 90M2 92M2 94M2 96M2 98M2 00M2

$/£ spot

Fig. 1. Spot exchange rate of £ vis-à-vis $

0.15
0.1
0.05
0
−0.05
−0.1
−0.15
−0.2
76M2 78M2 80M2 82M2 84M2 86M2 88M2 90M2 92M2 94M2 96M2 98M2 00M2

spot rate changes forward premium

Fig. 2. Spot rate changes and forward premium

−6
−8
−10
−12
−14
−16
−18
−20
1 21 41 61 81 101 121 141 161 181

spot rate changes forward premium

Fig. 3. Log spectra of spot rate changes and forward premium

forward premium; and  v is close to, and slightly
smaller than, the standard deviation of the change
in the spot rate. We know from Equation 1 that the
total variance of spot rate changes is the sum of the
variances of sT,t and sH,t, and the variance sT,t is
w2 plus a function of u2 . Therefore the Kalman filter
results are sound. Figure 4 plots changes in sT,t
against the forward premium. It can be observed
that the two series have similar sizes of change,
in sharp contrast to Fig. 2. They move similarly
sometimes and in rather different ways at other
times. Figure 5 demonstrates their spectra, together
with the spectrum of spot rate changes. Now
the spectrum of changes in sT,t and the spectrum
A re-examination of the predicting power of forward premia 1223
of the forward premium have approximately the same a small degree. Therefore, little information has been
power or energy. At the very low frequency end of the extracted, as the spot rate itself is white noise like,
spectrum, changes in sT,t decrease with the frequency, containing little information. After sT,t has been
matching that of the forward premium, but only to extracted, the remaining part of st becomes even
whiter.1
We progress to Table 3 without reporting the
technical detail of frequency domain decomposition
Table 2. Kalman filter decomposition results of the spot exchange rate, which can be obtained
’1 1.0979* (0.1111 e
1) upon request. The first row is the regression of
’2
0.1297* (0.1161 e
1) changes in the spot rate on the forward premium.
c 0.1162 e
4 (0.1036 e
4) The slope is
1.62 and is significant at the 5%
l 0.9800* (0.1329 e
1) level, which is typical for many other currencies in
w 0.2487 e
2 (0.1059 e
1)
v 0.3168
1 * (0.1013 e
2) previous studies. The second row of the Kalman
u 0.3022 e
5 (0.3198 e
2) filter analysis reveals a smaller (magnitude) slope
that is insignificant too. The last row of frequency
Notes: *Significant at the 1% level. Standard errors in domain filtering confirms that the slope is smaller in
parentheses.
magnitude and insignificant. Moreover, the Kalman

0.01

0.005

−0.005

−0.01

−0.015
76M2 78M2 80M2 82M2 84M2 86M2 88M2 90M2 92M2 94M2 96M2 98M2 00M2

changes in underlying s forward premium

Fig. 4. Underlying spot rate changes and forward premium

−6
−8
−10
−12
−14
−16
−18
−20
1 21 41 61 81 101 121 141 161 181

changes in s forward premium changes in underlying s

Fig. 5. Log spectra of spot rate changes, underlying spot rate changes and forward premium

1
The spectrum of sT,t is not that of st being made smaller as it may appear so. If, for example, 0.1 st is used in
the regression, then the corresponding intercept in Table 3 simply becomes
0.4292 e
3 and the slope becomes
0.16195.
Their t-statistics remain exactly the same.
1224
Table 3. Test results
Constant

2
st
0.4292 e
sT,t (Kalman)
0.8237 e
3 *
sT,t (Fourier)
0.1144 e
2 *
Notes: *Significant at the 1% level. **Significant at the 5% level. Standard errors in parentheses.
and Fourier analyses suggest that the mean value of
changes in the spot rate is significant and negative in
the range of 0.001, reflecting consistent depreciation
of the British pound against the US dollar during this
period. While with the basic statistics in Table 1 and
the first row in Table 3, the mean value of changes is
insignificant, losing any rough guideline for future
spot rate movement.
At first glance our results appear to be inconclu-
sive, as they suggest that the future spot exchange
rate is not predictable by the forward premium in
any way. In fact, that the future spot exchange rate
cannot be predicted in either a wrong way or in an
unbiased fashion by the forward premium is a market
reality. Most people do not believe that future spot
rate changes can be explained by the forward
premium in a completely wrong way, as found in
many previous studies; and few, if any, have exploited
the opportunity based on that kind of formulation.
However, why should people expect the forward
premium to be an unbiased predictor of the future
spot rate? If it were so, covered interest parity
would hold and transactions made with and without
a forward contract would yield the same outcome.
This is certainly not true, as with a forward contract,
every possibility, good or bad, may occur. There is
simply no such certainty in a financial market, be it
efficient or not, viewed in the standard definition of
the efficient market or other sensible forms. There
would be no demand for forward exchange if it
were an unbiased predictor of the future spot
exchange rate. Moreover, our analysis suggests that
the intercept in the second and third rows in Table 3
is not statistically different from the mean value of the
forward premium, which appears to be the correct
expectation, and the only reliable indication of
exchange rate movement in this particular case.
In this way, our results are the most feasible that
one can expect to achieve.
IV. Summary
By considering the underlying trend movement of
the exchange rate, the paper proposes that the
P. Wang
Forward premium

2
(0.2350 e )
1.6195** (0.7090)
(0.1771 e
3)
0.6370 e
1 (0.5343 e
1)
(0.4010 e
3)
0.1996 (0.1212)
spot exchange rate consist of two parts, and accord-
ingly, extracts the underlying movement from the
spot exchange rate in a fundamental way, using a
state space method and a frequency domain method.
The extracted component is persistent, catching
the trend movement and sharing some similar statis-
tical characteristics with the forward premium, as
expected. Our conclusion is simple and straight-
forward: the forward premium does not explain the
future spot rate in either a completely wrong way or
in a ‘right’ way. It is this simple mechanism that keeps
the foreign exchange market moving in the way it
is and should be.
References
Bilson, J. F. O. (1981) The speculative efficiency hypothesis,
Journal of Business, 54, 435–51.
Cavaglia, S. M., Verschoor, W. F. and Wolff, C. C. (1994)
On the biasedness of forward foreign exchange rates:
irrationality or risk premia, Journal of Business, 67,
321–43.
Clark, P. K. (1987) The cyclical component of US economic
activity, Quarterly Journal of Economics, 102, 797–814.
Copeland, L. S. (1989) Exchange rates and news: a vector
autoregressive approach, in Exchange Rates and Open
Economy Macroeconomics (Eds) R. MacDonald and
M. P. Taylor, Blackwell, Oxford, and Cambridge, MA.
Copeland, L. S. and Wang, P. (1993) Estimating daily
seasonals in financial time series: the use of high-
pass spectral filters, Economics Letters, 43, 1–4.
Fama, E. F. (1984) Forward and spot exchange rates,
Journal of Monetary Economics, 14, 319–38.
Gregory, A. and McCurdy, T. (1984) Testing the unbiased-
ness in the forward foreign exchange market, Journal
of International Money and Finance, 3, 357–68.
Geweke, J. and Feige, E. (1979) Some joint tests of markets
of forward exchange, Review of Economics and
Statistics, 61, 334–41.
Hansen, L. P. and Hodrick, R. J. (1980) Forward exchange
rate as optimal predictors of future spot rates: an
econometric analysis, Journal of Political Economy,
88, 829–53.
Hsieh, D. (1984) Tests of rational expectations and no risk
premium in forward exchange markets, Journal of
International Economics, 17, 173–84.
Huisman, R., Koedijk, K., Kool, C. and Nissen, F. (1998)
Extreme support for uncovered interest parity, Journal
of International Money and Finance, 17, 211–28.
A re-examination of the predicting power of forward premia
Longworth, D. (1981) Testing the efficiency of the
Canadian-US exchange market under the assumption
of no risk premium, Journal of Finance, 36, 43–9.
MacDonald, R. (1983) Tests of efficiency and the impact of
‘‘news’’ in three foreign exchange markets: the
evidence of the 1970s, Bulletin of Economic Research,
35, 123–44.
MacDonald, R. and Torrance, T. S. (1988) Exchange rates
and the ‘‘news’’: some evidence using UK survey data,
1225
The Manchester School of Economic and Social Studies,
56, 69–76.
Peel, D. A. and Pope, P. F. (1995) Time-varying risk
premia and the term structure of forward exchange
rates, The Manchester School of Economic and
Social Studies, 63, 69–81.
Schotman, P., Straetmans, S. and Vries, C. G. de (1997) Big
news in small samples, Tinbergen Institute Discussion
Paper, No 97-083/2.