You are on page 1of 15

European Journal of Operational Research 155 (2004) 284–298

www.elsevier.com/locate/dsw

The efficiency of financial futures markets: Tests


of prediction accuracy
Jason Laws, John Thompson *

CIBEF, Liverpool Business School, Liverpool John Moores University, John Foster Building, 98 Mount Pleasant, Liverpool L3 5UZ, UK

Abstract

The prices of financial futures contracts can be interpreted as forecasts of the spot rates, which will apply at the final
delivery date of that contract. Financial futures contracts have been traded daily since the early 1980s and provide a
substantial bank of data to test the forecasting efficiency of such contracts. Tests are carried out to examine whether the
interest rates implied by the futures price for eurodollar and short sterling contracts are cointegrated with the final
settlement price over forecasting horizons of 1, 2 and 3 months. Similar analysis is carried out for the yen/dollar ex-
change rate futures contract. The paper then examines the forecasting performance of the three contracts over the
forecasting horizons of 1, 2 and 3 months and in particular whether the forecasts implied by the futures contract
provide better predictions than the na€ıve no-change (i.e. random walk), a vector error correction model (VECM) or an
ARIMA model.
An examination of the relative efficiency of the markets for the three markets over the three time horizons is carried
out and finally trading strategies are simulated to see whether excess profits can be achieved. In fact the results suggest
that both profits and losses would be attracted.
Ó 2003 Elsevier B.V. All rights reserved.

Keywords: Market efficiency; Financial futures; Forecasting

1. Introduction prices will reflect all available information up to


the point where the cost of acquiring additional
A market may be defined as efficient in the in- information is equal to the benefits derived from
formational sense 1 if the prices of the assets tra- that information. Efficiency will be achieved
ded on that market instantaneously reflect all through arbitrage between traders. Note these
available information. This is a strong definition definitions make no mention of the possible exis-
and a weaker definition contains the caveat that tence of risk premiums to compensate for the un-
doubted risk in this type of arbitrage.
Futures markets serve at least four functions:
*
Corresponding author. 1. competitive price discovery,
E-mail addresses: j.laws@livjm.ac.uk (J. Laws), j.l.thomp-
son@livjm.ac.uk (J. Thompson).
2. management of risk,
1
The term ÔefficiencyÕ will be used as short hand for 3. facilitating financing,
informational efficiency in the remainder of the paper. 4. promotion of efficient resource allocation.

0377-2217/$ - see front matter Ó 2003 Elsevier B.V. All rights reserved.
doi:10.1016/S0377-2217(03)00087-0
J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298 285

To serve these functions efficiently, the futures price for the June 2000 contract was shown in the
markets themselves must conform to the efficient Financial Times on 14/4/00 to be 93.590. This
markets hypothesis; i.e. the prices must reflect all implied an interest rate of 6.41% per annum. All
available information. In particular futures prices the contracts are subject to a quarterly cycle with
should provide unbiased predictions of the subse- specific delivery dates 4 in March, June, September
quently observed spot price. This paper looks at and December each year. Each night a settlement
one aspect of the EMH, i.e. how far futures prices price is quoted. 5 A number of contracts are
predict the subsequent spot price and, as such, is available for trading at the same time. For exam-
an extension of earlier work by Holden and ple there are twelve short sterling contracts avail-
Thompson (1996) and Laws and Thompson (1997, able for trading at any one particular time at
2000). LIFFE. The main dollar/yen exchange rate con-
Our analysis 2 is based on a number of different tract is similar but the price quoted is for 100 yen.
approaches including: Two main, but not mutually exclusive, models
of futures prices exist, i.e. (i) the cost-of-carry
1. cointegration between forward and the realized model (see for example Brennan, 1958; or Telser,
spot price over a number of different horizons–– 1958) and (ii) the expectations model (see for ex-
long-term efficiency (Section 3); ample Breeden, 1980; or Hazuka, 1984). Futures
2. short-term efficiency (Section 4); prices according to the cost-of-carry model depend
3. variance bound tests (Section 5); on the current spot price of the ÔunderlyingÕ and
4. the relative efficiency of the various markets the cost of carrying stocks until the delivery period
(Section 6); less any marginal convenience yield obtained
5. comparison of the predictive power of the fu- through holding stocks. Typically the ÔcarryingÕ
tures price as compared with alternative predic- costs would include elements such as warehousing,
tions provided by (a) ARIMA, (b) VECM and insurance and interest rate charges. Arbitrage 6
(c) random walk (Section 7); should ensure that the following identity 7 holds:
6. examination whether simple-trading rules can
generate profits (Section 8); F0;t ¼ S0 ð1 þ CÞ ð1Þ
7. conclusions (Section 9).
where F0;t refers to the futures price quoted at time
The futures markets examined include the 3- 0 for delivery at time t, S0 is the spot price at time 0
month interest rate contracts for sterling (LIFFE) and C is the total cost for carrying the underlying
and the eurodollar (CME) 3 and the dollar/yen
exchange rate contract (CME). The characteristics
of the two interest-rate contracts are very similar.
They are for 3-month interest rates and contract 4
For example, the final trading day for the short sterling
prices are quoted as 100––the implied rate of in- contract is two business days prior to the third Wednesday of
terest for delivery on a specific date. As an example the delivery month.
5
of the method of quoting the relevant prices, the If no trades take place, an end-of-day settlement price is
price of the 3-month sterling futures settlement fixed by the market authorities with reference to the same
contract but different delivery dates. Consequently the end-of-
day settlement price provides a good indication of the market
position irrespective of whether trading has taken place. This
caveat is not likely to be applicable to the end-of-day settlement
2
For the sake of brevity, some statistical detail is not prices used in this study since we mainly concentrate on the
contained in the paper. Additional statistics can be obtained nearest contract which is normally the most heavily traded.
6
from the authors on request. A fuller version of the paper Note such arbitrage is relatively riskless since all prices are
containing additional statistical detail is available on the web known at time 0. There is, of course, the risk of counterparty
site, CIBEF.com. default.
3 7
The terms ÔsterlingÕ and ÔeurodollarÕ will be used in the rest Note for ease of exposition the formula is for one period
of this paper to indicate interest rates not currencies. only and hence no discounting is involved.
286 J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298

from time 0 to time t net of any marginal conve- 3 months (65 days) prior to the expiration of that
nience yield. contract so as to provide forecasts of the final
In the case of financial futures contracts, settlement price; i.e. the price ruling at the expi-
warehousing costs are virtually non-existent and ration of that particular contract. For all three
financing costs low since the contracts are highly contracts, the data runs from March 1987 to De-
leveraged. This leads on to the speculative model cember 2000 (a total of 56 non-overlapping ob-
which suggests that the price today represents the servations for each forecasting horizon).
price of the underlying speculators expect to rule The spot interest rates are the 3-month London
on the delivery date. Consequently the model can Inter-Bank Offered rates for sterling and eurodol-
be described by lars. The spot dollar/yen exchange rate is the mid
F0;t ¼ E0 ðSt Þ ð2Þ price of the series provided by Global Treasury
Information Service. As is customary in this type
where E0 ðSt Þ represents the spot price expected at of analysis the data is in logarithmic form. The
time 0 to rule at the time of the maturity of the data is obtained from Datastream.
contract, i.e. time t. The definition of F0;t is the Summary statistics of the final settlement price
same as for Eq. (1). for the three contracts are shown in Table 1. As far
Note now, however, that the arbitrage under- as interest rates are concerned, the sterling interest
lying Eq. (2) is not risk free since the final settle- rates were higher on average and, as evidenced by
ment price is not known at time 0. Thus the futures the standard deviation as a measure of dispersion,
price may be expected to incorporate not only the displayed both greater absolute and relative vari-
expected spot price but also a risk premium. 8 Also ation than eurodollar interest rates. The dollar/yen
it is possible to link the two models by specifying series showed lower variability than the interest
the marginal convenience yield in (1) to incorpo- rate series as measured by the coefficient of vari-
rate any expected profits. Nevertheless we would ation. Both the interest rate series displayed posi-
expect the relative importance of carrying costs to tive skewness and excess kurtosis, 10 as did the
decline as the intrinsic value of the underlying in- dollar/yen contract suggesting that all three series
creases. are not normally distributed. 11 This conclusion is
We now move on to describe the data in Section only partly supported by the Jarque–Bera statistic
2. reported in Table 1. A caveat may be appropriate
here. This is that the Jarque–Bera test is a large
sample test and the number of observations (56)
2. The data is quite low. Finally it should be noted that the
eurodollar contract is by far most heavily traded
As noted above we have concentrated on just contract with the dollar/yen contract attracting
three futures contracts, which provide a series of low trading volumes as compared with the other
observations of the end-of-day and final settlement two contracts.
prices running over a sufficient length of time. For A multivariate analysis of the degree of co-
each contract the end-of-day settlement price is integration between the futures prices and the final
sampled 1 month (20 days), 9 2 months (40 days), settlement price follows in Section 3.

8 10
With a risk premium (2) would be transformed to The short sterling series exhibits positive excess kurtosis
F0;t ¼ E0 ðSt Þ  RP0 where RP0 represents the relevant risk (i.e. it is leptokurtic) with slimmer or longer tails of the
premium at time 0. Thus the futures price can be thought of distribution and the eurodollar and dollar/yen series both
as a discounted expected spot price. Note also RP can be time possess negative excess kurtosis (i.e. it is platykurtic) with
varying. shorter or fatter tails.
9 11
The data is recorded for a 5-day week. This is a standard result for financial time series.
J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298 287

Table 1
Descriptive statistics of final settlement price (percentage per annum)
Final settlement price
Sterling Eurodollar Dollar/yen
Mean 8.78 6.13 0.0084
Standard deviation 3.52 1.73 0.0012
Coefficient of variation 0.40 0.28 0.14
Skewness 1.09 0.29 0.50
Kurtosis ()3) 0.39 )0.40 )0.13
Maximum 20.06 10.29 0.0118
Minimum 5.12 3.13 0.0065
Sample period March 1987 to December 2000 March 1987 to December 2000 March 1987 to December 2000
Number of observations 56 56 56
Volumea 5.26 m 21.85 m 0.82 m
JBb 11.44 1.16 2.37
a
Measured by total number of transactions for the contracts ending December 2000.
b
The Jarque–Bera statistic.
*
Indicates significance at the 5% level.

3. Long-run efficiency: Cointegration analysis be differentiated to become stationary. 12 The


usual method of testing for stationarity is standard
3.1. Introduction Dickey–Fuller methodology (see for example
Dickey, 1984; Dickey and Fuller, 1979, 1981).
Efficiency tests can be carried out by trans- Briefly the basic equation takes the following
forming Eq. (2) into form:
St ¼ a þ bF0;t þ et : ð3Þ Yt ¼ / þ hYt1 þ Ut : ð4Þ
Long-term efficiency of a market requires the fol- Subtracting Yt1 from both sides of (4) gives
lowing four conditions to be satisfied: Yt  Yt1 ¼ / þ ðh  1ÞYt1 þ Ut ð5Þ
1. S and F must be integrated to the same order or
(this requirement is tested in Section 3.2). DYt ¼ / þ dYt1 þ Ut :
2. S and F must be cointegrated otherwise S and F
will tend to drift apart over time so that inaccu- Tests of stationarity take the form of examining
rate forecasts are obtained (this requirement is whether the coefficient d is <0, which implies that
checked in Section 3.3 together with require- h < 1 and the series is stationary. 13 The distribu-
ment 3). tion of d is non-standard but critical values for the
3. For the cointegrating equation, the estimated test have been provided by MacKinnon (1991).
slope coefficient (i.e. b) should equal 1 and the The test is also sensitive to the presence of auto-
constant (i.e. a) 0. correlation. Two methods of dealing with this
4. The forecast errors should not be correlated problem have been suggested. The first alters the
with any information known at the time the
forecast is made (this requirement is examined
12
in Section 7). A time series is said to be ÔweaklyÕ stationary if itÕs mean,
variance and covariances remain the same over time. In this
3.2. Order of integration paper the term stationary refers to Ôweakly stationaryÕ.
13
The equation was also tested to see if it was necessary to
incorporate a time trend. In all instances it was not possible to
As is well known the order of integration is reject the null hypothesis (i.e. that the coefficient on the time
determined by the number of times a series has to trend was equal to 0) at the 5% level.
288 J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298

estimating Eq. (5) by adding lagged values of DY hansen and Juselius, 1990). 16 The VAR can be
until such time as autocorrelation is eliminated written in general matrix form as
(i.e. the Augmented Dickey–Fuller equation). The X
DYt ¼ l þ CDYtkþ1 þ PYt1 þ et ð6Þ
second, due to Phillips and Perron (1988), alters
the test statistic. where k is the number of variables, 2 in the case of
As far as ADF approach is concerned the pro- the tests carried out in this paper and the CDYtkþ1
cedure adopted by us was to determine the lag represents the error correction terms.
structure according to the Schwarz–Bayesian in- The P matrix can be factorized into
formation criteria (Schwarz, 1978). 14 The results
P ¼ ab
from using the equation with the estimating
equation were then checked for autocorrelation. If where a specifies the speed of adjustment to the
the LM 15 test indicated the presence of autocor- steady state and b the cointegrating vectors.
relation then further lags of the dependent variable The rank of P defines the number of cointe-
were added until such time as the autocorrelation grating vectors. In the case of two variables the
was eliminated. maximum rank of the matrix P is 2. If the rank is
The results of this procedure provide unam- 2, then St and F0;t are jointly stationary. If the rank
biguous evidence that all the series are Ið1Þ and is 1, a single cointegrating vector exists between
their first differences Ið0Þ. The sole exception is for the two variables. On the other hand, if the rank of
the 2-month futures forecast of the eurodollar rate P is 0, then cointegration between will not be
where the ADF indicates non-stationarity for the present. Johansen suggests two tests to determine
first difference whereas the Phillips–Perron test the rank of P and therefore the number of coin-
indicates stationarity. In this case we have as- tegrating vectors. These are based on the maximal
sumed that the first difference series is stationary. eigenvalue and the trace of the stochastic matrix.
These results are typical for financial/commodity These cointegration tests require the specification
price time series. A notable exception was Heaney of the number of lags in the model as represented
(1998) where the hypothesis of a unit root was by Eq. (6) above. Too few lags seriously biases the
rejected for weekly data on lead commodity prices tests. On the other hand too many lags may result
on the London Metal Exchange. in tests with low power to discriminate between the
Thus the first criterion for efficiency is satisfied presence and absence of cointegration (Gonzalo,
and in the following section we consider whether 1994). 17 As before the order of the VAR was se-
the second and third criteria are met. lected by the Schwarz Bayesian information crite-
rion. It was also necessary to test for the existence
3.3. Cointegration of a time trend in the VAR. In all cases it was not
possible to reject the hypothesis of the absence of a
In this section we investigate whether the final time trend at the 5% level. The results are shown in
settlement price and the futures prices are cointe- Table 2 and suggest the existence of a single co-
grated. Our approach is based on the vector au- integrating vector over all the forecasting horizons
toregression (VAR) framework suggested by for the three series.
Johansen (see for example Johansen, 1988; Jo- The results of estimation of the cointegration
vectors by the Johansen method suggested that,

14
The statistical package provided two other information
16
criteria: Akaike (1973) and Hannan and Quinn (1979). Gener- Gonzalo (1994) concludes that the Johansen method of
ally the three criteria indicated the same optimum lag length but estimating long-run relationships is more satisfactory than
in the majority of cases where disagreement occurred the outlier OLS, NLS, principal components and canonical correlations
was the Akaike criterion, which is prone to overestimate the and that also this conclusion holds for finite samples.
17
number of parameters (see Bascßi and Zaman, 1998). Gonzalo (1994) suggests the consequences of overestimat-
15
Given that all the series are quarterly, the order of LM test ing the lag length are less serious than those associated with an
was 4. underestimate.
J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298 289

Table 2 thesis that the futures price and final settlement


Eigenvalue and trace tests for number of cointegrating vectors price were cointegrated. This is, however, a nec-
Sterling Eurodollar Yen essary but not sufficient condition for long-run
One-month forecasting horizon market efficiency. It is also necessary that a ¼ 0
Eigenvalue test and b ¼ 1 in the cointegrating equation. It was not
H0 : r ¼ 0 39.36 22.48 68.99
possible to reject the hypothesis that a ¼ 0 and
H1 : r ¼ 1 (15.87) (15.87) (15.87)
H0 : r ¼ 1 3.71 5.27 5.88 b ¼ 1 for:
H1 : r ¼ 2 (9.16) (9.16) (9.16)
(a) the sterling contract––all three forecasting ho-
Trace test
H0 : r ¼ 0 43.03 27.75 74.84 rizons,
H1 : r ¼ 1 (20.18) (20.18) (20.18) (b) the eurodollar contract––2-month forecasting
H0 : r ¼ 1 3.71 5.27 5.88 horizon only,
H1 : r ¼ 2 (9.16) (9.16) (9.16) (c) the dollar/yen contract for 1- and 2-month
forecasting horizons.
Two-month forecasting horizon
Eigenvalue test
H0 : r ¼ 0 30.35 53.42 81.62 The hypotheses that a ¼ 0 and b ¼ 1 was re-
H1 : r ¼ 1 (15.87) (15.87) (15.87) jected both separately and jointly for the other two
H0 : r ¼ 1 3.28 1.75 5.33 eurodollar contracts. In the case of the 3-month
H1 : r ¼ 2 (9.16) (9.16) (9.16) dollar/yen contract the hypotheses were accepted
Trace test separately but not jointly. 18 The rejection of effi-
H0 : r ¼ 0 33.63 55.17 86.95 ciency for the 1-month eurodollar contract is in-
H1 : r ¼ 1 (20.18) (20.18) (20.18) teresting given the acceptance of efficiency for the
H0 : r ¼ 1 3.28 1.75 5.33
2-month contract. This may be due to a significant
H1 : r ¼ 2 (9.16) (9.16) (9.16)
proportion of the transactions in the last month
Three-month forecasting horizon consisting of hedging contracts being rolled over.
Eigenvalue test Our results are roughly consistent with Kellard
H0 : r ¼ 0 30.67 76.75 19.86 et al. (1999) 19 who found that the twin hypothesis
H1 : r ¼ 1 (15.87) (15.87) (15.87) that a ¼ 0 and b ¼ 1 was not rejected for live cattle
H0 : r ¼ 1 1.70 1.99 4.11
and the dollar/DM markets and that b ¼ 1 for
H1 : r ¼ 2 (9.16) (9.16) (9.16)
Brent Crude, gas oil and soybeans over 28- and 56-
Trace test day horizons. Similarly Kavussanos and Nomikos
H0 : r ¼ 0 31.85 78.92 23.99
(1999), using similar methodology, found that
H1 : r ¼ 1 (20.18) (20.18) (20.18)
H0 : r ¼ 1 1.70 1.99 4.11 freight future prices 1 and 2 months prior to ma-
H1 : r ¼ 2 (9.16) (9.16) (9.16) turity provided unbiased forecasts (i.e. a ¼ 0 and
Note: r ¼ the number of cointegrating vectors. b ¼ 1) predictions of the realized spot price. This
The figures in brackets are the 95% critical values relevant to was not the case for future prices 3 months prior to
the test. expectation. In a similar vein, Graham-Higgs et al.
The eigenvalue and trace tests rely on asymptotic critical values (1999) found that the Australian wool futures
and therefore are not fully reliable in small samples. Reimers
market is efficient up to 6 months but no further
(1992) suggests adjusting the tests statistics by multiplying by
ðn  pkÞ=n where n is the number of observations, p is the into the future.
number of variables in the var and k is the lag order of the var.
This correction has been applied to the test statistics recorded
above. 18
Note that in all cases where the null hypotheses were
*
Indicates significance at the 5% level. rejected the values of the estimated coefficients a and b were
quite close to 0 and 1 respectively.
19
There have been a large number of studies of the
relationship between futures and spot prices with varying
for all the contracts over the three forecasting results. We concentrate on one or two for comparative
horizons, it was not possible to reject the hypo- purposes.
290 J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298

Cointegration analysis examines long-run rela- Table 3


tionships and in the following section we look at Short-term efficiency tests
the relationship between the futures price and Horizon Sterling Eurodollar Dollar/yen
subsequently observed spot price over the shorter 1 month Y N N
time horizon. 2 months Y N N
3 months N N N

4. Short-term efficiency rate contract for this reason. In any case it would
not be expected that a market which was not ex-
Short-term efficiency is concerned whether de- hibiting long-run efficiency would also be short-
viations exist from the long-term equilibrium re- run efficient. For this latter reason it would be
lationship. If such deviations exist then speculative assumed that the 3-month horizon eurodollar
profit opportunities might also exist. Short-term contract would also not be short-run efficient. Eq.
efficiency can be tested through estimation of the (7) was estimated for the other contracts and the
following ECM equation: coefficients tested to see if they conform to the
X X criteria noted above. The results are recorded in
DSt ¼ a þ bi DFti;tj þ cj DStj
Table 3 and it can be seen that only the sterling
þ kðSt  Fti;t Þ1 þ Mt ð7Þ contract for 1- and 2-month horizons meets the
criteria listed above and can therefore be consid-
where i is the forecast horizon; i.e. 1, 2 and 3
ered short-run efficient.
months so that DSt is the change in S over the
These results contrast with those obtained by
forecast horizon and (St  Fti;t ) is the error cor-
Antoniou and Holmes (1996) which demonstrated
rection term. 20
(at least for their observation period September
Short-term efficiency requires:
1984 to June 1993) that the FTSE-100 futures
contract was efficient in the short run for 1 and 2
1. a ¼ 0,
months prior to maturity but inefficient for 3, 4
2. b0 ¼ c0 ¼ 1,
and 5 months prior to maturity.
3. k ¼ 1,
This implies that short-term inefficiency exists
4. all other b and c ¼ 0.
and in Section 8 we examine whether this ineffi-
ciency can be exploited to obtain excess profits.
If these four conditions are not met, then the
market is not efficient in the short run and po-
tential for profitable speculation exists. The order
5. Variance-bound tests
of the VAR dictates the form of the ECM. The
optimum order of the VARs was selected by in-
Recent years have seen extensive use of vola-
formation criteria. If the order of the VAR is 1,
tility to test the efficiency of markets (especially the
then only the error correction term appears in the
stock market) through the use of variance-bound
ECM. If the order is 2 then the variables DFtn;ti
tests. Shiller (1981) provides an exposition of the
and DStj appear in the ECM with a lag of 1. If the
simplest of these tests. The EMH implies that the
order of the VAR is more than 2, then condition 4
futures price is the optimum forecast of the sub-
above is violated and it can be seen that it was
sequently realized spot price. Hence, since the
possible to reject the hypothesis of short-run effi-
forecast error is uncorrelated with the forecast,
ciency for the 1-month horizon eurodollar interest
VarðSt Þ ¼ VarðFti;t Þ þ Varðet Þ ð8Þ
where t  i is the forecast horizon.
20
Where long-term efficiency (i.e. a ¼ 0 and b ¼ 1) was
Eq. (8) implies that VarðFti;t Þ must be less or
rejected the actual estimated values for a and b were used to equal to VarðSt Þ thus providing a test of market
estimate the error correction term. efficiency. Rejection of this inequality suggests that
J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298 291

Table 4 entails a comparison of the forecast provided by


Variance bounds tests the futures price with that of the best regression
Horizon Sterling Eurodollar Dollar/yen equation. The ratio of these two measures then
1 month Y N Y provides the coefficient of efficiency. More for-
2 months Y Y Y mally the regression equation takes the following
3 months Y Y Y
general form:
n 56 56 56
Y indicates that the variance bound test indicated by equation St  Sti ¼ a þ bðFti  Sti Þ
is satisfied. X
þ cj ðStj  Stij Þ
either the market is inefficient or a risk premium X
exists. Antoniou and Holmes (1996) point out that þ dj ðFtj  Ftij Þ þ et ð9Þ
testing this inequality requires the two variances to
be stationary. There is a considerable volume of where i is the forecast horizon.
evidence, apart from that discussed in this paper, The lag lengths were selected through the gen-
that financial time series are not stationary so the eral-to-specific testing procedure dropping lags,
inequality is transformed by deducting the ap- which were insignificant at the 10% level. The R2
propriate lagged spot (i.e. 1, 2 or 3 months) value values seemed quite reasonable and neither auto-
of the interest/exchange rate from both sides of Eq. correlation nor heteroscedasticity provided any
(8). 21 The inequality now being tested is cause for concern save for the 3-month eurodollar
equation where the presence of heteroscedasticity
VarðFti;t  S0 Þ 6 VarðSt  S0 Þ: ð8aÞ
was indicated.
The results are reported in Table 4. In all cases The resulting efficiency measure is then given by
except for the 1-month forecasting horizon for the
eurodollar the variance-bound tests are supported p ¼ VarðeÞ=Varðfutures price forecastsÞ:
by the data.
We now turn to examine the relative efficiency This is quite a stringent test because it compares
of the two markets in Section 6. the variance of the error term of an equation es-
timated over the whole period with the benefit of
6. Relative efficiency hindsight with the forward looking futures price.
The p values are shown in Table 5 and these are
Perfect efficiency with no departures from the relatively high and compare with figures of figures
definitions quoted in Section 1 is an illusory or recorded by Kellard et al. ranging 0.53 for live
ephemeral concept. The real question is whether a cattle to 1 for soybeans. It is also noticeable that in
market approximates efficiency. It would therefore general the sterling and dollar/yen contracts are
be interesting to derive a measure to compare the more efficient relative to the eurodollar contract.
degrees of efficiency of the two markets. One such Again we see a certain degree of inefficiency and
method of making this comparison has been sug- we examine in Section 8 whether it is possible to
gested by Kellard et al. (1999). 22 Basically this exploit this to obtain speculative profits.
In the following section we discuss another facet
of efficiency, i.e. the ability to predict future prices.
21
We tested for stationarity of the variables contained in Eq.
(8a) and found that all the variables were in fact stationary at
the 5% level with the possible exception of (F0;t  S0 ) with the Table 5
forecast horizon of 2 periods. Efficiency measures of implied futures forecasts
22
The formula given by Eq. (9) differs slightly from that used Horizon Sterling Eurodollar Dollar/yen
by Kellard et al. The maximum lag in their paper was 10
1 month 0.99 0.84 0.93
whereas ours is truncated
P to 6. AlsoP Kellard et al. specified the
2 months 0.94 0.96 0.95
lag structure on the cI and the dI to be equal to avoid
3 months 0.97 0.91 0.96
autocorrelation. This was not necessary for our data.
292 J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298

7. The predictive of power of futures prices indicates a rise in the variable. This is a particu-
larly useful indicator because very often the di-
In this section we examine how well the futures rection of the change is more important than the
price forecasts the subsequent final settlement magnitude of the change. The mean 23 error is also
price. It is useful to have some yardstick(s) against a useful, but often neglected, statistic since it il-
which comparisons can be made. We use three lustrates any bias in the forecast. The mean ab-
series for comparison: (a) an ARIMA forecast solute error overcomes the problem associated
based on Box–Jenkins methodology (see Box and with the mean error of errors being cancelled out
Jenkins, 1970), (b) a random walk and (c) a by errors of the same magnitude but opposite sign.
VECM model. The random (RW) forecast is The root mean square error penalizes large fore-
simply the FSP 1, 2 and 3 months prior to expi- cast errors. The Theil inequality coefficient (Theil,
ration of the contract. The ARIMA forecast was 1966) is useful becomes it is a pure number free of
based on a rolling 260 days prior to expiration. units of measurement. A value of 0 indicates per-
There is a timing problem here. The FSP is timed fect forecasts. Details of the comparisons are
during the day whereas the daily interest rate series shown in Table 6.
refers to the end of the day. Consequently the For the four types of forecast (i.e. futures price,
comparisons which are listed in Table 6a–c refer to ARIMA, RW and VECM) every forecast is un-
the FSP for the random walk and futures forecasts biased on the mean-error criteria for all three
but close-of-trading rates for the ARIMA fore- contracts. As far as the sterling forecasts are con-
casts. This should introduce very little bias. cerned, the futures price provides the best forecast
The ARIMA forecasts are based on the fol- for all the criteria except for the mean error. The
lowing model: CDI is always significantly different from the
random value of 50% and is far superior in this
Yt ¼ / þ a1 Yt1 þ    þ an Ytn þ b0 Ut respect to the ARIMA forecast and also the RW
þ b1 Ut1 þ    þ bnUtn : ð10Þ forecast. 24 This latter is by definition equal to 0
since it forecasts no change. As far as the RW,
It was decided to drop all the Us and concentrate ARIMA and VECM forecasts, there is little to
on the lagged values of Y . Given the order of in- choose between them. This is perhaps not sur-
tegration already defined for FSP the estimates prising given the low level of predictive power of
were carried out in first difference terms. Early the ARIMA equations over and above the random
tests suggested that there was no stable-lag pattern walk.
for the autoregressive component of (10) so the Generally speaking all the forecasts for the eu-
number of lags was standardized as 5. The ex- rodollar contract are slightly superior to those for
planatory power of the differenced series was the sterling interest rate except for the CDI crite-
quite low with an average R2 of 0.05 suggesting rion. Again for most criteria the eurodollar futures
that little explanatory power was obtained over price provides a superior forecast to that generated
that achieved by the random walk. by the ARIMA, RW and VEDM processes. It is
The specification of the VECM was the same as also noticeable that the CDI is only significantly
for Eq. (6) but, in this instance, a time trend was different from 0.5 for the 3-month horizon in the
included. The order of the VECM was 2 in all case of both the futures and ARIMA forecast.
cases.
The criteria used to assess the quality of the
forecasts include the standard statistical measures
for assessing the accuracy for forecasts together 23
with the ÔCorrect Direction IndexÕ (CDI). This Holden and Peel (1990) suggest the mean error is the best
indicator of bias.
latter measure indicates whether the forecast value 24
These results mirror earlier studies of dollar interest rates
indicates the direction of any movement. Thus if (Batchelor, 1994; Belognia, 1987) where predictive accuracy was
the forecast value is above the current value, this of the order of 55%.
J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298 293

Table 6 Table 6 (continued)


Forecasting performance (panels A–C: 1-, 2- and 3-month Sterling Eurodollar Dollar/yen
forecasting horizon respectively)
Correlation coefficient
Sterling Eurodollar Dollar/yen Futures 0.990 0.974 0.931
Panel A AR 0.985 0.976 0.933
Mean error RW 0.984 0.967 0.933
Futures 0.078 0.036 )0.148 (104 ) VECM 0.983 0.973 0.903
AR 0.073 )0.012 )0.174 (107 )
TheilÕs inequality coefficient (1966)
RW 0.034 )0.023 )0.106 (107 )
Futures 0.055 0.061 0.054
VECM 0.005 0.006 )0.589 (107 )
AR 0.061 0.063 0.054
Mean absolute error RW 0.062 0.077 0.052
Futures 0.238 0.178 0.248 (103 ) VECM 0.066 0.069 0.068
AR 0.278 0.207 0.236 (103 )
CDI
RW 0.249 0.197 0.245 (103 )
Futures 66.07 50.00 30.36
VECM 0.317 0.216 0.265 (103 )
AR 50.00 53.21 57.14
RMS RW 0 0 0
Futures 0.373 0.271 0.311 (103 ) VECM 53.57 58.93 48.21
AR 0.409 0.346 0.304 (103 )
RW 0.397 0.413 0.309 (103 ) Panel C
VECM 0.459 0.303 0.343 (103 ) Mean error
Futures )0.012 )0.061 )0.250 (104 )
Correlation coefficient
AR )0.010 0.031 )0.245 (104 )
Futures 0.995 0.988 0.965
RW )0.116 )0.067 0.441 (104 )
AR 0.992 0.979 0.966
VECM )0.051 0.026 0.249 (106 )
RW 0.992 0.974 0.965
VECM 0.989 0.984 0.958 Mean absolute error
Futures 0.588 0.400 0.435 (103 )
TheilÕs inequality coefficient (1966)
AR 0.746 0.422 0.455 (103 )
Futures 0.039 0.043 0.037
RW 0.650 0.437 0.416 (103 )
AR 0.045 0.055 0.036
VECM 0.719 0.433 0.618 (103 )
RW 0.043 0.065 0.037
VECM 0.050 0.048 0.041 RMS
Futures 0.930 0.570 0.576 (103 )
CDI
AR 1.095 0.566 0.605 (103 )
Futures 69.64 51.79 44.64
RW 1.024 0.656 0.558 (103 )
AR 46.43 48.21 60.71
VECM 1.048 0.590 0.812 (103 )
RW 0 0 0
VECM 25.0 55.36 55.36 Correlation coefficient
Futures 0.964 0.944 0.889
Panel B AR 0.946 0.953 0.886
Mean error RW 0.948 0.937 0.892
Futures 0.056 0.0004 )0.671 (104 ) VECM 0.950 0.948 0.821
AR )0.008 0.015 )0.655 (104 )
TheilÕs inequality coefficient (1966)
RW )0.047 )0.039 )0.219 (104 )
Futures 0.099 0.089 0.068
VECM )0.064 )0.008 )0.709 (107 )
AR 0.120 0.089 0.072
Mean absolute error RW 0.112 0.103 0.066
Futures 0.368 0.278 0.333 (103 ) VECM 0.115 0.093 0.096
AR 0.414 0.293 0.328 (103 )
CDI
RW 0.385 0.288 0.315 (103 )
Futures 67.86 64.29 39.29
VECM 0.441 0.294 0.420 (103 )
AR 53.57 64.29 62.50
RMS RW 0 0 0
Futures 0.518 0.386 0.454 (103 ) VECM 55.36 57.14 46.43
AR 0.558 0.401 0.456 (103 ) *
Indicates value significantly different from zero at the 5%
RW 0.564 0.487 0.439 (103 ) level of significance.
VECM 0.601 0.436 0.571 (103 )
294 J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298

In contrast to the other two contracts, the errors do not contain any information, which is
dollar/yen futures price is not generally the best correlated with the forecast at the time the forecast
forecast. For the 1-month horizon, the ARIMA is made. We therefore regressed the forecast error
forecast generally across most criteria provides the with up-to-4 of its own lagged values. For the
best performance whereas for the 2- and 3-month three futures forecasts over the three time hori-
horizons, the random walk performs the best. It is zons, the F value of the estimating equation was
also noticeable the futures price gives little or no never significant though, in the case of the 2-
indication of the direction of future changes in the month eurodollar futures forecast, the F statistic
exchange rate as the CDI statistic is always lower verged on significance (P value 0.054). It was also
than the random value of 50% at the 5% level and noticeable that only two of the individual coeffi-
significantly so in the case of the 2-month fore- cients were significantly different from zero at the
casting horizon. Possibly this is caused by inter- customary 5% level (two lags for the 1-month
vention in the foreign exchange market by the horizon for the dollar/yen and three lags for the
authorities. Neely (2000) reported that interven- eurodollar 2-month horizon).
tion by the US authorities (mainly in the DM and
Yen markets) over the period 1973:04 to 1998 as
follows: 8. Simulated trading

average (absolute) monthly intervention 355 In the previous sections we have identified areas
million dollars, where the markets are less then fully efficient but
maximum monthly intervention 6735 million the acid test for efficiency is whether excess profits
dollars (purchase of foreign currency), can be made from trading rules. The term excess
minimum monthly intervention )3413 million profit implies profits that are persistently above the
dollars (sales of foreign currency). level associated with the risk associated with that
direction of trading. Whether the level of profits is
On the reasonable assumption that intervention excess or not is a matter of judgement and could be
was leaning against the wind, this would provide the subject of considerable controversy.
an adequate explanation, in the case of the dollar/ In this paper the scope for trading is provided
yen series, of the relatively good performance of by the coexistence of two forecasts of the prices
the random walk for 2- and 3-month forecasting likely to exist in the future. Our trading rule is to
horizons and the ARIMA for the 1-month fore- take a position and hold that to the expiration of
casting horizon as well as the poor performance of the futures contract. Clearly this is more reason-
the futures price as far as the CDI criterion is able for the 1-month forecasts than those for 2 or 3
concerned. months. Nevertheless these are included for com-
Finally it should also be seen that, as would be parison purposes. The precise trading rules 25 are
expected, the performance of all contracts tend to as follows:
deteriorate as the forecasting horizon lengthens. (1) If the ARIMA or VECM interest-rate
Kavussanos and Nomikos (1999) also found forecast exceeds the futures rate forecast plus the
that, for 1-, 2- and 3-month horizons, the freight trading cost then our prediction is that the futures
futures prices provided a better prediction of the rate will rise in the future. The predicted rise in the
spot price than the other models considered (i.e. interest rate is equivalent to predicting a fall in the
vector error correction (VECM), ARIMA, ran-
dom walk and exponential smoothing models)
25
with the sole exception that the VECM provided a This is the same rule adopted by Kellard et al. (1999) but
marginally better forecast for the 1 month horizon. for 1-month forecasts only. This rule could be easily altered to
(a) permit trading only when the profit exceeds trading cost by a
We also examined the forecast errors for the fixed margin and (b) to increase the intensity of trading by
futures contracts. In all cases these were station- permitting more than one trade if the profit margin is above a
ary. For efficiency, it is also necessary that these certain level(s).
J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298 295

price of the futures contract. Hence the trading rate forecast exceeds the futures rate forecast then
rule is to sell one contract. a contract should be bought. Conversely if, after
(2) Conversely if the ARIMA or VECM fore- allowance for the trading cost, the ARIMA or
cast is less than the implied futures rate by more VECM interest-rate forecast is less than the fu-
than the trading cost, then the trading rule gives a tures rate forecast then a contract should be
signal to buy one futures contract. sold.
(3) In the case of the dollar/yen exchange rate (4) The simulated trade takes place at the rates
contract, the rule is quite simple and the inverse of ruling at the close of the day following the fore-
the interest rate contracts. If, after allowance for cast. Hence the trades could have been made at the
the trading cost, the ARIMA or VECM interest- quoted prices.

Table 7
Simulated trade profits (panel A–C: 1-, 2- and 3-month horizon)
Sterling (£) Eurodollar ($) Dollar/yen ($)
Panel A
Total profit )3370 ()5060) 950 (170) 11,740 ()1980)
Maximum profit on one trade 1400 (1090) 1200 (2880) 7980 (7980)
Minimum profit on one trade )2180 ()2180) )1400 ()1400) )7890 ()7890)
Maximum cumulative profit 1990 (140) 3180 (2180) 20,980 (17,580)
Minimum cumulative profit )3370 ()5060) 0 ()3170) )7190 ()7780)
No. No. No.
Transactions 32 (36) 33 (34) 33 (39)
Sales 17 (22) 19 (19) 19 (21)
Profitable trades 15 (15) 16 (13) 21 (22)

Sterling (£) Eurodollar ($) Dollar/yen ($)


Panel B
Total profit 1640 (1820) 4050 (590) )11,250 ()57,050)
Maximum profit on one trade 1610 (1610) 2650 (2650) 11,160 (9700)
Minimum profit on one trade )1160 ()1230) )1750 ()2200) )23,630 ()23,630)
Maximum cumulative profit 5930 (4540) 11,050 (930) 1290 ()6540)
Minimum cumulative profit )830 ()2250) 150 ()2500) )34,360 ()70,810)

No. No. No.


Transactions 42 (43) 42 (39) 43 (46)
Sales 22 (26) 17 (18) 17 (24)
Profitable trades 19 (19) 25 (21) 22 (21)

Sterling (£) Eurodollar ($) Dollar/yen ($)


Panel C
Total profit )3090 ()480) 8330 (5900) 37,790 ()91,560)
Maximum profit on one trade 1810 (1810) 3030 (2980) 11,880 (11,950)
Minimum profit on one trade )3030 ()3030) )4530 ()3930) )29,060 ()29,490)
Maximum cumulative profit 2650 (6450) 8330 (5900) 58,450 ()2740)
Minimum cumulative profit )3230 ()480) )4700 ()5700) 3940 ()101,080)
No. No. No.
Transactions 48 (50) 50 (51) 54 (55)
Sales 22 (26) 25 (20) 19 (24)
Profitable trades 21 (24) 26 (28) 31 (23)
Figures in brackets show corresponding figures for trades based on VECM.
296 J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298

(5) Sutcliffe (1997) identifies the cost of trading forecasts for the dollar/yen contract. The loss
the FTSE-100 futures contract as 0.116% for a was substantial in the latter case.
round trip including two commissions and the (b) Profits were obtained in all other cases and
spread between the bid and offer rate. It is likely these were substantial in the case of the euro-
the costs of trading interest and exchange rate dollar contract (both forecasts) and the dol-
contracts will be of a similar nature. However, lar/yen (ARIMA forecast).
since our strategy is to buy and hold, we have re-
duced the cost so as to include only one commis- Further points about the overall profitability
sion. The trading cost is therefore specified as 0.1% are:
or £500 per transaction for the sterling contract. A
similar cost was applied to the eurodollar interest 1. For all contracts the cumulative profit was neg-
and dollar/yen exchange rate futures contract so ative at some time throughout the sample pe-
that the total cost per transaction came to £500 for riod except in the cases of (a) the 1-month
all contracts. eurodollar forecasting horizon (zero profit),
The results of the trade are shown in Table 7 (b) eurodollar 2-month ARIMA forecast
with the transaction cost being incorporated into (small profit $150) and (c) dollar/yen 3-month
the profit calculations. The trades include both ARIMA forecast (significant profit of $3940).
purchases and sales with sales tending to be more 2. Individual trades showed substantial profits as
numerous than purchases for the 1-month fore- well as losses.
casting horizon and less numerous for the 2- and 3. The ARIMA forecasts generally performed bet-
3-month forecasting horizons. The overall profit- ter than the VECM forecasts.
ability is summarized below:
Kellard et al. (1999) also reached a similar con-
One month forecasting horizon clusion for simulating trade with a 56-day forecast
(a) Losses incurred on the sterling contract for horizon. After allowance for trading costs, profits
both methods of forecasting (i.e. ARIMA were obtained during December 1991 and October
and VECM). 1994. However in one sub-period a loss was in-
(b) Very small profits were obtained on the euro- curred and they also concluded that the Ôsmall size
dollar contract for both methods of forecast- of the average return and the large variance suggest
ing (i.e. ARIMA and VECM). a substantial risk-return trade-off Õ.
(c) For the dollar/yen contract profits could be In conclusion to this section, the outlay for
made if ARIMA forecasts were used as the trading a single contract is quite low on the other
basis of trading but not in the case of VECM hand the total profit at the end of the period also
forecasts. seems quite low given the length of the period and
the degree of risk associated with the trading rule.
Two month forecasting horizon Whether these profits are excessive is, as noted
(a) For the sterling and eurodollar contracts prof- earlier, a matter of judgement.
its would be obtained whichever forecasting
method was used. Only in the case of the euro-
dollar ARIMA forecast would the level of 9. Conclusions
profits be significant.
(b) For the dollar/yen contract a substantial loss We have examined the sterling and eurodollar
was incurred whichever forecasting method futures markets for efficiency. Our conclusions for
was used. this sample period can be summarized as follows:

Three month forecasting horizon 1. For horizons of 1, 2 or 3 months, the futures


(a) Losses were incurred for both methods of fore- price and the subsequently realized spot rate
casting for the sterling contract and VECM are cointegrated for the three contracts.
J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298 297

2. The sterling contract meets the conditions nec- References


essary for efficiency imposed on the constant
and slope coefficients in the case of 1-, 2- and Akaike, H., 1973. Information theory and the extension of the
3-month forecasting horizons. This is not the maximum likelihood principle. In: Petrov, B.N., Csaki, F.
(Eds.), Proceedings of the Second International Symposium
case for the eurodollar market where the condi- on Information Theory, Budapest, pp. 267–281.
tion is only met for the 2-month forecasting ho- Antoniou, A., Holmes, P., 1996. Futures market efficiency, the
rizon or the dollar/yen contract where condition unbiasedness hypothesis and variance bound tests: The case
is only met for the 1- and 2-month forecasting of the FTSE-100 futures contract. Bulletin of Economic
horizons. Research 48, 115–128.
Bascßi, S., Zaman, A., 1998. The effects of skewness and kurtosis
3. Only the sterling futures market exhibits short-
on model selection criteria. Economics Letters 59, 17–22.
term efficiency and then solely for 1- and 2- Batchelor, R., 1994. Is there expertise in interest rate forecast-
month forecasting horizons. ing. Paper presented to the International Symposium on
4. Variance bounds requirements for efficiency are Forecasting, Stockholm, June 1994.
satisfied for all three futures markets with the Belognia, M.T., 1987. Predicting interest rates: A comparison
sole exception of the 1-month forecasting hori- of professional and market-based forecasts. Federal Reserve
Bank of St Louis Review 69, 9–15.
zon for the eurodollar futures market. Box, G.E.P., Jenkins, G.M., 1970. Time Series Analysis:
5. An examination of the relative efficiency of the Forecasting and Control. Holden-Day, San Francisco.
three futures markets, reveals that, whilst the Breeden, D.T., 1980. Consumption risks in futures markets.
level of relative efficiency is high for all markets, Journal of Finance 35, 503–520.
the sterling market is relatively more efficient Brennan, M.J., 1958. The supply of storage. American Eco-
nomic Review 48, 50–72.
than the either eurodollar or dollar/yen futures Dickey, D.A., 1984. Powers of unit root testsÕ. In: Proceedings
markets. of the American Statistical Association, Business and
6. For all the forecasting horizons, the interest- Economics section.
rate futures price tends to provide a superior Dickey, D.A., Fuller, W.A., 1979. Distribution of the estima-
forecast of the subsequent spot price as com- tors of autoregressive time series with a unit root. Journal of
the American Statistical Association 74, 427–431.
pared with the random walk, VECM or AR- Dickey, D.A., Fuller, W.A., 1981. Likelihood ratio statistics for
IMA forecasts. This is not the case for the autoregressive time series with a unit root. Econometrica 49,
dollar/yen exchange rate futures contract. 1057–1072.
7. The futures forecast errors are both stationary Gonzalo, J., 1994. Five alternative methods of estimating long-
and uncorrelated with past errors. run equilibrium relationships. Journal of Econometrics 60,
203–233.
8. During this period, both profits and losses Graham-Higgs, J., Rambaldi, I., Davidson, B., 1999. Is the
could be made through the use of a simple trad- Australian wool futures market efficient as a predictor of
ing rule. The presence of significant losses sug- spot prices? The Journal of Futures Markets 19, 565–582.
gests that the risk/profit trade off is substantial Hannan, E.J., Quinn, B.J., 1979. The determination of the
and, in general, losses would be incurred if al- order of an autoregression. Journal of the Royal Statistical
Society B 41, 190–195.
lowance were made for the existence of a risk Hazuka, T.B., 1984. Consumption betas and backwardation in
premium. commodity markets. Journal of Finance 39, 647–655.
Heaney, R., 1998. A test of the cost-of-carry relationship using
This leaves the efficiency of the three markets as the London metal exchange lead contract. The Journal of
an open question. As usual the empirical evidence Futures Markets 18, 177–200.
Holden, K., Peel, D.A., 1990. On testing for unbiasedness and
is ambiguous but we would argue that nevertheless efficiency of forecasts. Manchester School of Economic and
it is sufficient to indicate ÔapproximateÕ efficiency in Social Studies 58, 120–127.
these three markets. Holden, K., Thompson, J.L., 1996. Forecasting interest rates
from financial futures markets. Applied Financial Econom-
Acknowledgements ics 6, 449–461.
Johansen, S., 1988. Statistical analysis of cointegrating vectors.
Journal of Economic Dynamics and Control 12, 231–254.
The authors are grateful for helpful comments Johansen, S., Juselius, K., 1990. Maximum likelihood estima-
provided by Professors C. Dunis and K. Holden. tion and inference on cointegration––with applications to
298 J. Laws, J. Thompson / European Journal of Operational Research 155 (2004) 284–298

the demand for money. Oxford Bulletin of Economics and Neely, C.J., 2000. Are changes in foreign exchange reserves well
Statistics 52, 169–210. correlated with official intervention. The Federal Reserve
Kavussanos, M.G., Nomikos, N.K., 1999. The forward pricing Bank of St. Louis Review 82, 17–31.
function of the shipping freight futures market. The Journal Phillips, P.C.B., Perron, P., 1988. Testing for a unit root in time
of Futures Markets 19, 353–376. series regression. Biometrica 75, 335–346.
Kellard, N., Newbold, P., Rayner, T., Ennew, C., 1999. The Reimers, H., 1992. Comparisons of tests for multivariate
relative efficiency of commodity futures markets. The cointegration. Statistical Papers 33, 335–339.
Journal of Futures Markets 19, 413–432. Schwarz, G., 1978. Estimating the dimension of a model.
Laws, J., Thompson, J.L., 1997. Forecasting spot exchange Annals of Statistics 6, 461–464.
rates from futures markets. Presented at the International Shiller, R.J., 1981. The use of volatility measures in assessing
Symposium on Forecasting (ISF) in Barbados, June 1997. market efficiency. Journal of Finance 36, 291–304.
Laws, J., Thompson, J.L., 2000. The short-term forecasting Sutcliffe, S., 1997. Stock Index Futures: Theories and Interna-
ability of financial futures contracts. Paper presented to the tional Evidence. Thomson Business Press, London.
26th conference organized by the Euro Working Group on Telser, L.G., 1958. Futures trading and the storage of
Financial Modelling at Trondheim, May 2000. cotton and wheat. Journal of Political Economy 66, 233–
MacKinnon, J.G., 1991. Critical values for cointegration tests. 255.
In: Engle, R.F., Granger, C.W.J. (Eds.), Readings in Theil, H., 1966. Applied Economic Forecasting. North-Hol-
Cointegration. Oxford University Press, Oxford. land, Amsterdam.

You might also like