Some Determinants of the

Volatility of Futures Prices"

Ronald W. Anderson

T his article empirically investigates the determinants of the volatility of futures

prices. We find that in a variety of American markets the dispersion of futures
prices is not constant and that it changes over time in a systematic manner. The
pattern is particularly strong for annually-harvested, storable goods, but it is also
present for other commodity futures.
While these regularities are interesting in their own right and are established
without imposing strong assumptions concerning the class of possible processes
generating observed futures prices, they can be interpreted as testing theories of
futures price volatility. Two theories are examined. The first, ascribed to Samuelson
(1965), states that the volatility of futures price changes per unit of time increases
as the time to maturity decreases. The second has emerged from recent multiperiod
analyses of the simultaneous determination of equilibrium prices in cash and futures
markets. It states that the volatility of futures prices will be relatively high during
periods when significant amounts of supply or demand uncertainty are resolved.
The two views are not necessarily incompatible. The maturity effect may he
interpreted as the hypothesis that progressively more information is revealed as the
delivery date approaches. Stated in this way, it is seen as a rather special result,
and indeed, it has been shown to rest on strong assumptions. In spite of this, the
simplicity of the result makes it appealing. The fact that past empirical studies have
been interpreted as supporting the hypothesis of an inverse relation of time to
maturity and volatility means that some may take it to be a strong and general
characteristic of futures markets.
The results of the present study suggest that the possible relations of futures price
volatility and time to maturity may have received undue emphasis in the past. Other

*An early version of this article was presented at the meetings of the Anrerican Economic Association, D e w n h e r
1981. The author would like to thank D. Edelman, R. Lana, H. Levene, K. Koenker, and S. Sundaresan for
useful discussions, and P. Samuelson for comments on an earlier version. All errors are the author's responsi1)ility.

Ronald W. Anderson is a Professor of Economics at the

Graduate School and University Center, City University oj
New York.

The Journal of Futures Markets, Vol. 5, No. 3, 331-348 (1985)

0 1985 by John Wiley & Sons, Inc. CCC 0270-73 14/85/03O:%31- I8$04.OO
forms of nonstationarity of futures prices are seen to be enipirically more important.
Since these phenomena are reflected in a multiperiod model of simultaneous spot
anti futures equilibrium, the results of the paper are suggestive of the class of models
that might provide a useful framework for further work on futures price determination.
The paper is organized as follows. Section I is devoted to a more careful statement
of the two theoretical hypotheses, and Section I1 reviews the relevant empirical
literature. Section 111 describes the data and explains the choice of statistical
procedures. Section IV presents the results of a nonparametric test treating the
variance of futures prices in a two-way analysis of variance. Section V presents
parametric tests based on classical and robust statistical methods. Section V1 dis-
cusses tlie implications of the results for the formulation of intertemporal pricing
models.

1. THEORIES OF FUTURES PRICE VOLATILITY

‘ h e possibility that the volatility of the price of a futures contract may increase as
the delivery date approaches appears to have been recognized by many observers
o f futures markets and discussed by Telser (1956). Samuelson (1965)is the first to
give a theoretical argument in support of this phenomenon. Assuming the price of
the good for immediate delivery (the cash price) follows a stationary first-order
autoregressive process and the price of t h e good for deferred delivery (the futures
price) is a n unbiased predictor of the price at delivery date, then the variance of
the daily (or weekly, etc.) price changes increases as the delivery date approaches.
For convenience and with some injustice to Samuelson (see below) we will refer to
this as the Samuelson hypothesis.
r ,
I h e Samuelson hypothesis relies on strong assumptions which a r e open to a
number of questions. First, the assumption of a first-order autoregressive process
is very restrictive. Samuelson (1!?76) finds that for higher-order stationary autore-
gressive processes the variance is not generally monotonically decreasing in the
titne to maturity. Instead, h e obtains the weaker result that the variance of a futures
when delivery is sufficiently disl.ant will necessarily be less than the variance of
the same contract very near to delivery.’ Second, the result is not valid when the
cash price is nonstationary. This would be violated if there were trends to cash prices,
as would the case if the first-order autoregressive parameter exceeded unity (see
Kutledge, 1976 and Miller, 1979). It would also be violated if the underlying shocks
to tlie rash price exhibited nonconstant variances. Finally, the assumption that the
futures price equals the expected value of the price at delivery d a t r goes against
IIie fin(1irigs of much recent work on asset pricing. For example, the general equi-
lihrium analysis of Cox, Ingersoll, and Koss (1978) shows that for a wide class of
assets the expeclations hypothesis does not generally hold.
Notwithstanding these criticisms, the Sarnuelson hypothesis was for a time the
onl!- theory of futures price volatility. I<c:cently, the variance of futures price has
been analyzed in models that allow the simultaneous deterniination the equilibrium
cash and futures prices in a multiperiod setting (Anderson and Danthine, 1983,
Richard and Sundaresan, 1980, and Stein, 1979). These models overcome at least
some of the criticisms listed above. In them, the variance of futures prices may
change over time depending upon the distribution of underlying state variables;
however, with the exception of Stein’s model, there is no necessity for the variance
to increase as the delivery date approaches.
Anderson and Danthine (1983) are most explicit and provide an economic inter-
pretation of changes in the variance of futures prices. Theirs is a multiperiod, partial
equilibrium model where producers make production decisions and hedge their
goods i n process by selling futures so as to maximize the expected utility o f end-
of-holding-period wealth. During the production period, production and demand
uncertainty are progressively resolved as random variables are realized and publicly
observed. In this context it is shown that the ex ante variance of futures price is
high (low) in periods when relatively large (small) amounts of supply arid demand
uncertainty are resolved. A n example of large amounts of uncertainty being resolved
is a time when production shocks hare large variances or are highly correlated across
individual producers. We may refer to this theory as the state variuble hypothesis.
The specific form of the state variable hypothesis depends upon the nature of the
supply and demand uncertainty for a particular good. If demand uncertainty is
dominant and if factors affecting ultimate dernand are subject to progressively larger
shocks as the demand date approaches, the futures price variance would increase
over time. This would be a case where the state variable hypothesis aiid the Sam-
uelson hypothesis would give the same predictions. In contrast, agricultural com-
modities, in particular the grains, offer a case where the state variahle hypothesis
is distinguished from the Samuelson hypothesis aiid may readily be tested. For
grains, total annual production is determined by acreage planted and yields. Yields
in turn are heavily dependent upon weather conditions at certain times of the growth
process. These crucial phases tend to occur at approximately the same times during
the calendar year. Consequently, we would expect the resolution of production
uncertainty to follow a strong seasonal pattern. Given the price of a grain, fluctuations
in the final demand for grain depend on the variations of prices of related goods
and income. Since in the near terni income tends to be comparatively stable, shork-
term fluctuations are effectively determined by the prices and thus the supplies of
substitutes. For the reasons just reviewed, t h e supplies of sulistitute grains will also
be heavily dependent on the seasonal determinant of yield. Thus on both the supply
and demand sides, the resolution of uncertainty in grain markets may be expected
to follow a strong seasonal pattern.
Since the crop that is harvested is consumed over the entire crop year, the new-
crop-year futures contracts tend to obey normal carrying charge relationships. New
information on supply necessarily tends to affect all new-crop contracts approxi-
mately equally. Consequently, the same seasonal pattern of price volatility should
apply to all delivery months. For example, if rnost corn production information is
revealed during July and very little information is revealed during Novemlier, we
would expect the volatility of both the December and March corn futures to be
higher in July than in November.
Some notation will be useful in precisely stating the hypotheses and in setting
the stage for statistical testing. Let F,, be the price (or the logarithm of the price)
on date t of the futures calling for delivery at date i ( 2 t ) , and let E,, be the expected
value of F,,conditional on information through date t-1. The current study as well
as past empirical tests of the Samuelson hypothesis investigate the distribution of
the random variables defined by

u,, = F,, - E,,. (1)

Let VLfbe a measure of the dispersion of the distribution of I!J!~. The Samuelson
hypothesis states that

VLt =f(mJ (2)

where f’is decreasing function and m,, is the time to maturity of futures i at date t.
The state variable hypothesis for the case of grains or other annually produced goods
is

Vu = g (5,) (3)
where s, is the season associated with date t. Note that in Eq. ( 3 ) the delivery date
i does not enter a s a determinant.

11. PREVIOUS EMPIRICAL STIJDIES

The relevant past empirical investigations of the volatility of futures prices have
concentrated on testing the Samuelson hypothesis. They differ with respect to mar-
kets studied, frequency of observatiotiu, the measurements of V,,, arid the choice of
statistical tests.
Kutledge (1976) studied the COMEX March 1970 silver contract, the N e w York
Cocoa Exchange December 1970 cocoa contract, the Kansas City Board of Trade
September 1969 wheat contract, and the Chicago Board of Trade May 1971 soybean
oil contract. Working with the daily price observations expressed in logarithms, he
assumed the (log) prices followed the Martingale, E,, = Fif- ,. V,, was measured as
the absolute value of (log) price differences. He employed a goodness of fit test for
a three-way contingency table with data grouped h y volatility of cash price, volatility
of futures price, arid time to maturity. Kutledge rejected the Samuelson hypothesis
for wheat and soybean oil but accepted it for silver and cocoa.
Miller ( 1979)investigated the Samuelson hypothesis using the June and Decemher
live beef contracts of the Chicago Mercantile Exchange for the period 1964-1972.
A s did Rutledge, she employed the changes of the logarithm of the daily closing
price; however, the 0‘28-0.72 interfractile range was used in addition to variance
a s a measure of dispersion. She employed the classical noi-ma1 test for the simple
correlation coefficients between dispersion and time to maturity. She concluded that
there is a significant inverse relationship; thus, she supported the Samuelson hypoth-
esis.
Grauer ( 1977) examined the Satnuelson hypothesis using month-end futures prices.
H e pooled the observations for the Decemher contracts of a range of the 10 futures
tnarkets, taking as measures of volatility the variance and the intercluartile range.
He computed Friedman’s nonparametric test statistic for a two-way analysis of
variance and found no support for the Samuelson hypothesis.
Castelino (1982) tested the Samuelsori hypothesis using daily data from 1960
through 1971 for the wheat, corn, soybeans, soybean meal, and soybean oil contracts
a t the Chicago Board ofTrade and for copper at COMEX. ‘The results were interpreted
as supporting the Samuelson hypothesis. Test procedures were based on the sample
variance of daily price changes of a given futures contract within a given calendar
month. As a control for other possible sources of nonstationarity, these sample
variances were standardized by dividing by the geometric mean of the sample
variances of all contracts within the same month of observation. A first test of the
Samuelson hypothesis was obtained by averaging the standardized variances for a
given time to maturity and then using the F-statistic for testing the equality of the
resulting average standardized variances. A second test was based on the ordinary
least squares regression of the logarithm of the standardized variances on time to
maturity.
These procedures rely heavily on the implicit assumption that the standardized
sample variances are distributed normally or lognormally. This is a doubtful assump-
tion which is separate from the maturity effect, and should not be part of the
maintained hypothesis. It leaves open the issue of whether the results of the tests
of the Samuelson hypothesis are robust to alternative distributional assumptions.
Furthermore, these procedures do not permit an assessment of the relative impor-
tance of the maturity effect versus other forms of nonstationarity, and therefore leave
the state variable hypothesis untested.2
To summarize the past empirical research we may say that the Samuelson hypoth-
esis has been supported for daily observations. However, the evidence seems to
vary from market to market and may be sensitive to untested distributional assump-
tions. Furthermore, the evidence drawn from monthly observations does not support
the Samuelson hypothesis.

111. DATA AND STATISTICAL CONSIDERATIONS

The markets that are the subject of the present study are listed in Table 1. This is
a more extensive data set than those involved in previous studies; in all, some
160,000 price observations are included." The principal grain futures are included
since they are the markets for which the state variable hypothesis as expressed in
Section I is intended. The two wheat contracts are included since they differ with
respect to delivery location and the category of wheat that is typically delivered.
Commodities 6 through 9 are included to allow the comparison of our conclusions
with those of Rutledge and Miller. All available delivery months are included. Most
contracts are followed for the 12 consecutive months leading up to the delivery
date; however, for certain contracts data limitations restrict coverage to less than
12 months.
As for Rutledge and Miller, the basic unit of observation is the change of the
logarithm of the daily closing or, if available, settlement price. The main reason

'In related studies, Castelint) and Francis (1982) and Castelino and Vora (1984) examined the volatility of basis
(the difference between futures and spot prices) and of futures spreads. Using methodologies similar to Castelino
(19821, they concluded that volat es tended to decline as the time to maturity of the basis or the length of the
spread decreased. These findings are consistent with the Samuelson hypothesis in the sense that using the
assumptions of Samuelson (1965), such behavior can be theoretically deduced. However. they are consistent with
other hypotheses also; therefore, they do not provide direct evidence cvncerning the behavior of futures price
volatilities. Finally, i n a recent thesis completed after the results of the present study had he disseminated, Miiorias
(1984)examined the behavior of futures price volatilities for eleven agriculiural, metal, and interest rate contracts
covering the period 1972-1982. Using methods similar to Castelino, Milonas found the data supporied the
Samuelson hypothesis in all markets except corn.
'The data were generously provided by the Center for the Study of Futures Markets at Columbia University.

DETERMINANTS OF VOLATILITY 1335

 Table I
DESCRIPTION OF DATA

Commodity Exchange" Time Period

I . Wheat CBOT May 1966-July 1980

2. Wheat KC May 1966-Jdy 1980
3. Corn CBOT March 1966-June 1980
4. Oats CBOT May 1966-June 1980
5. Soybeans CBOT March 1966-June 1980
6. Soybean Oil CBOT May 1966-June 1980
7 . Live Cattle Merc February 1966-June 1980
8. Silver COMEX January 1967-June 1980
9. Cocoa Cocoa January 1966-June 1980

,'CBOT: Chicago Hoard of Trade; KC: Kansas City Board of Trade; Merc: Chicago Mercantile
Exchange; COMEX: Commodity E x c h a n g e Inc., New York; Cocoa: New York Cocoa E x c h a n g e (pres-
ently the New York Coffce, Sugar, and Coc:oa Exchange).

for working with the log differences is that as the price level would change we would
expect the dispersion of prices to change in the same direction; using percentage
changes or log differences corrects for this obvious source of nonstationarity. A s a
measure of volatility we employ the variance. I t is recognized that if the futures
prices are not (log) normally distributed, other measures of dispersion may be more
useful for certain purposes. However, we use variances since that is how the Sam-
urlson and state variable hypotheses w e e r e originally stated. This choice is probably
not a critical one; Miller arid Grauer found largely the same patterns for variances
as t h e y did for the interfactile ranges. Furthermore, our statistical tests do not
depend upon the assumption that the futures prices are distributed lognormally.
The sample variances are calculated pooling observations of a single contact for
a single calendar month of a single year. Thus it is assumed that t h e log changes
of futures prices follow the process,

where 1,is the set of time indices associated with the jth rnonthlyear pair. That is,
the daily log changes of futures prices follow a stationary distribution within a month
hut are nonstationary over time periods of greater than one month. The process (4)
is quite gerieral. It is consistent with (but does not impose) futures prices following
a martingale. It is also consistent with other models of asset price determination to
the extent that they do not imply the mean and dispersion of returns change drast-
ically over short time intervals.'' It should he stressed that, unlike Miller, w e do
not pool across delivery years.
The nature of the distribution of cornrnodity futures prices i s an open question.

'It r r i i g h appcai iriore appealing to allou t h e distribution i l l t t i r log changes of daily prii.rb to \ary g r a d ~ ~ a l l y
( i . ? . , daily) over time. However, to implement this appi-oach w e would need to assume a specific functional form
for the paths of k, and u,. Since these are unknown, making such an assumption would he a n important source of
possilde misspecification. By assuming the drstrihutions fixed over a short interval we ean approximate a n arbitrary
fun(~tkitia1 form fur the evolution of k, and u(. The interval of one nionth was r.hosrii so as to a f f d decent ~ I T I ' ~ S I O ~
i n c.stirnating IJ., and u,, on the o n r hand, arid to yield a natural unit of seasonality, on the other.

336 I ANDERSON
However, there is substantial evidence that futures prices are not distributed nor-
mally or log normally. (See Mann and Heifner, 1976 for a survey of the literature,
as well as some new results.) As a consequence, the distribution of t h e sample
variances (or other measures of dispersion) of futures price (or log price) changes
must be viewed as unknown. In light of this, inferences based on classical statistical
methods should be viewed with suspicion until they are corroborated using methods
that are not sensitive to departures from normality.
There are several possible approaches to testing when classical methods fail. One
is to employ a nonparametric test, usually based on ranks, which is valid under
weak assumptions. Typically, this requires only that ranks are assigned randomly
under the null hypothesis. A criticism of nonparametric procedures is that they
often have low power relative to the parametric methods appropriate for the true
distribution.
An alternative to nonparametric methods is to employ parametric methods that
have high relative efficiency for a range of possible distributions and are in this
sense robust to departures from any single distribution. A nonparametric test suited
to the present study is employed in Section IV. Section V reports parametric tests
based on classical and robust methods.

IV. NONPARAMETRIC TESTS

The Samuelson hypothesis and the state variable hypothesis are not necessarily
mutually exclusive; instead, they may indicate two separale determinants of the
volatility of futures prices. In this section we investigate the composite model

where E;, is an independent, unobservable random variable drawn from a fixed

distribution. If we do not specify functional forms forf( ) and g( ), this is a two-
way layout without interaction. Under the normality of E,,, we would treat this as a
standard two-way analysis of variance. For the reason discussed in Section 111 the
distribution of E;, is unknown. An approach which makes no distributional assump-
tions involves converting the observations of the V,, to their corresponding ranks
and estimating a two-way model for these ranks.
Since the observations of the variances are monthly, the seasonal effect can take
12 levels. As described in Section 111, the data were chosen so that the maturity
effect also takes on twelve levels. If we had one observation for each of the resulting
144 cells, an appropriate rank test would be the well-known Friedman (1937)test.
Because the futures markets listed in Table I do not have deliveries for all the
months of the year, some combinations of time to maturity and seasons of the year
are never observed. Our data sets cover many years, so the remaining combinations
are observed more than once. Consequently, we use the generalized rank test for
the case of an arbitrary number of observations per cell which is derived by Henard
and Van Elteren (1953). Since this test is unfamiliar, we describe it briefly in
Appendix 1. For each market we compute two Henard and Van Elteren statistics.
One tests the null hypothesis of no seasonal effects against the alternative of arbitrary
seasonal effects. ‘The second tests the n u l l of no maturity effects against the alter-
native of arbitrary maturity effects. Under the null either of these statistics is
asymptotically distributed Chi-square with 11 degrees of freedom. Table I1 lists
these statistics for each of the markets of the study.

DETERMINANTS 01. \OIL4TILITY / 337

 Table I1
BENARD AND VAN ELTEREN STATISTICS
- ~~

Market Seasonal Effect Maturity Effect

1. Wheat (CBOT) 30.73" 11.62

2. Wheat (KC) 35.3P 3.21
3. Corn 119.7" 4.24
4. Oats 59.09" 33.60"
5. Soybeans 71.94" 8.11
6. Soybean Oil 56.29" 36.45"
7. Live Cattle 19.73" 27.6"
8. Silver 9.71 12.21
9. Cocoa 18.17 22.33"

"Significant at the 0.99 level.

"Significant at the 0.95 level.

An inspection of the statistics of column 1 from the table reveals seasonal effects
are highly significant for the five grain markets in the sample. Clearly, seasonality
emerges as an important determinant of the variation of the volatility of futures
prices over time. An examination of the ranks averaged over years suggests that
the pattern of seasonality is similar in all the grains considered. These average
ranks for corn, soybeans, and the two wheat contracts are displayed in Figure 1.
They have been scaled so as to equal zero in January and unity at the maximum.5
Volatility is at its minimum in mid-winter (usually February) and rises steadily until
it reaches a peak in the summer (June for Kansas City wheat, July for the other
grains). Subsequently, volatility decreases monotonically until it reaches the winter
lows. Soybean oil also shows a strong seasonal pattern of volatility as is indicated
by the large Benard and Van Elteren statistic. For the remaining markets the test
statistics indicate slight or no seasonality. This is not surprising. The form of the
state variable hypothesis developed in Section I may not be relevant to these markets.
Thus the data considered give unambiguous support to the state variable hypothesis.
The evidence is less clear for the Samuelson hypothesis. The statistics in Table
I1 indicate significant maturity effects for oats, soybean oil, live cattle, and cocoa.
There is no indication of a maturity effect for the grains other than oats. The results
are somewhat at variance with those of Rutledge in that we find a significant maturity
effect for soybean oil whereas he did not. On the other hand, we find no support
for a maturity effect in silver where he did. However, it should be recalled that the
sample used in Table I1 included more years and more delivery dates than were
included in his study. Otherwise our results agree with those of Rutledge and Miller.
The hypothesis of increasing futures price variability as time to maturity decreases
was proposed as a general characteristic of futures markets. The evidence presented
in Table I1 casts doubt on its general validity. Some markets show strong maturity
effects; others show little or no maturity effects. However, there is reason for not
taking this evidence as being entirely conclusive. The nonparametric test was chosen

,r .
Ihe wits contrac.t ib omitted from the figure foi- h e sake. of clarity. The patteyn of arasonalily fot oats is very
3itnilar to Chi(.ago wheat.

338 1 ANDERSON
 Jan Jul Oct

0 CBT Wheat + K C Whcot 0 Soybaune A Corn

Figure 1
Volatility seasonals-grains.

because of the weak assumptions necessary to assure its validity. The cost is that
the Benard and Van Elteren test may not be powerful enough to reveal a maturity
effect even if it is present. The possible loss of' power is generally a worry when
using nonparametric tests. Furthermore, the test entails a loss of power since the
alternative hypothesis to no maturity effects was arbitrary maturity effects. This is
too broad an alternative since the Samuelsoii hypothesis requires that the maturity
effects be monotonic decreasing.
When the ranks of the variances for a given maturity are averaged across years,
the results appear to be monotonic or nearly so for all the markets considered. I n
particular the highest average rank of variance is attained for the smallest time to
maturity in all the markets. Thus the data appear to be consistent with the Samuelson
hypothesis even though some of the rank statistics in Table I1 were not significant.
The next section investigates parametric methods which allow the imposition of the
monotonicity of the maturity effects.

V. PARAMETRIC TESTS
The principal reasons for adopting a parametric approach are that we are able to
impose the monotonicity of the maturity effects and that we employ statistical tests
that may have greater power than rank tests. The approach has the additional
advantage in that it gives us a somewhat freer hand in introducing factors in addition
to season and maturity affecting the variance of futures prices. That there are such
factors is clear. Our data set spans a period of time during which the American
economy was submitted to a number of severe shocks, and virtually all markets saw
periods of enormous price volatility. While we do not seek to explain these additional
factors affecting volatility of futures prices, we wish to control them so that our
estimates of seasonality and maturity effects are not obscured.
In this section we report results based on the model

lnVzf = P m,, + g(s,) + MY,) + E,f, (6)

wherta Y , is the calendar year within which date t occurs. This variable is intended
to control for possibly nonrecurrent factors affecting futures price volatility. We
leave t h e function h( 1 unspecified; instead, we estimate a qualitative year effect.
The Samuelson hypothesis states that the parameter p < 0. In Eq. (6)the dependent
variable is the natural logarithm of the variance.” This is the preferred specification
since in this nonlinear relation for P < 0 the maturity effect is greater the nearer
the delivery date; this agrees with the Samuelson hypothesis (see Rutledge, 1976
or Miller, 1979).
As discussed above in Section Ill, we must view the shape of the distribution of
the dependent variable in Eq. (6) as unknown; if the distribution departs markedly
from the normal distribution there is a risk that tests based on least squares estimates
nray have low power. The guard against this possibility we adopt an estimation
strategy suggested by recent work on robust statistics (see Koenker, 1981). Spe-
cifically, we estimate model (6) and test hypotheses both by the least squares (L2)
and the least absolute deviations (Ll) methods. For some time, L 1 estimation has
Iwen recognized as being robust to departures of the distribution from normality in
the sense that it produces parameter estiniators that are ntore efficient than least
squares for a variety of non-normal error distributions. The major obstacle to the
application of least absolute deviations was the absence of a theory of hypothesis
testing based on 1,1 estimators; this recently has been overcome (see Koenker and
Hassett, 1982). In what follows we employ both likelihood ratio tests based on least
squares estimates and Lagrange multiplier tests based on least absolute deviations
estimates. The latter test may b e unfamiliar and is briefly described in Appendix
11.
‘Table I11 lists the results of tests of the seasonal effect and the maturity effect
for each market. In addition, we report the results of a test of no “year” effect which
gives some indication of other factors affecting futures price volatility. All test
statistics should he compared to the F distribution with the indicated degrees of
freedom.
As w e would expect from the results of Section IV, the statistics testing the
hypothesis of no seasonality art: very highly significant for all the grain markets.
More surprising is the fact that t h e seasonality effects are seen to be highly significant
for all the other markets as well. This is true for both the LL and L2 results. On
the basis of the results in Table 111, a seasonal pattern to the variance of futures
prices appears to b e a strong and general phenomenon. The pattern of seasonality
reflected in the individual parameter estimates is similar for all the grain markets.
Starting from a period of relatively low price volatility in January and February,
volatility rises through the spring, reaching a peak in July or June (for Kansas City
wheat) after which time volatility falls. The fact that peak volatility tends to come
in June or July is probably a reflection of the fact that crop yields are particularly
sensitive to weather conditions at that time. This pattern plus the finding of sig-
riificant seasonality in all markets is strong evidence i n favor of the state variable
hypothesis as formulated in Section i . 7
As with the rank tests of Section IV, the evidence in support of the Samuelson
hypothesis appears to be somewhat weaker. The maturity effect is highly significant
for oats, soybeans, soybean oil, live cattle, and cocoa. It is irisigriificant for Kansas
City wheat, corn, and silver. The maturity effect appears significant for Chicago
wheat when estimated by least squares but not when estimated by least absolute
deviations. Thus judging by the size of the test statistics, the evidence concerning
the maturity effect again seems to be mixed. It is striking, however, that all the
estimated coefficients of the maturity variable are negative, as suggested by the
Samuelson hypothesis. The fact that the maturity effect was of the right sign in all
nine markets and that it was at least marginally Significant in six of them is stronger
evidence than w e have had until now that the inverse relation between volatility
and time to maturity may be a general property of futures markets.
The year effects are highly significant in all markets as judged by either test.
Clearly, the volatility of futures prices underwent significant changes during the
sample period beyond those which can be attributed to seasonality and changing
time to maturity. It is interesting to note, however, that when the year effects are
omitted, the results for seasonality and maturity are essentially unchanged.
If we accept that the data support the maturity effect, we are in the position of
saying that the results support both the state variable hypothesis and the Samuelson
hypothesis, and that these suggest two distinct determinants of futures price vola-
tility. The question then becomes, which is quantitatively more important? Here
the results leave little question that the seasonality effects generally dominate the
maturity effects. To demonstrate this we have calculated the 12-month variation of
the log-variances of futures prices that can be attributed to seasonality alone and
to the maturity effect alone. As measures of variation we have used the sample
variances; these are listed in the last two rows of Table 111.' With the single exc,eption
of the live cattle market, the amount of variation due to seasonality is greater than
that due to changing time to maturity. In most cases the seasonality effect exceeds
the maturity effect by a large margin; in four cases, by a factor of at least 10 to 1.
Consequently, we conclude that while the maturity effect may be a general char-
acteristic of futures markets it is usually a comparatively minor factor in accounting
for changing variability when compared to seasonal factors.
The conclusions concerning the relative importance of seasonality and time to
maturity of this section are consistent with those of Section IV. Since qualitative

'Identifying the sourves of seasonality in a iionagricultural commodity such a3 s i l w r is toore ~ m p l t : x .5iini.e

silver stocks are usually large relative tu production, the sources of volatility are likely to come rnostl) frorn the
demand side. Here there is a complex mixture of factors affecting tleinand for use in photographic materials. for
jewelry and silverplate, and for bullion as a store of value. WP find the net result of these factors tends to be
relatively high volatility in May, October, and November and relatively low \ulatility i n January. F e h a r ! . and
March.
9 n the notation of Eq. (6), the formula for the vanation of variance due to seasonality is

Noting that m, , ~ =
, j, the analogous expIession with respect to maturity is

p' (i j L - 6.5'
1 = p' 11.917.

DETERMINANTS OF VOLATILPI'Y / 341

 Table I11
LINEAR MODEL RESULTS MODEL: In V,, = pm,, + g (s,) + h (Y,)
Commodity Wheat (CBOT) Wheat (CBOT) Wheat (KC) Wheat (KC) Corn Corn Oats Oats Soybeans
Method L1 L2 L1 L2 L1 L2 L1 L2 L1

Seasonality
F-Statistic 9.02 11.324 11.056* 11.615” 21.79” 33.77” 14.69” 18.19” 29.83
D. F. Numerator 11 11 11 11 11 11 11 11 11
D.F. Denominator 831 820 816 805 840 829 781 770 1179
Maturity
Coefficient -0.021 -0.027 -0.006 - 0.007 - 0.008 -0.007 - 0.038 - 0.047 - 0.026
F-Statistic 2.6 19.7” 0.185 0.767 0.39 1.02 13.98^ 58.59” 15.88“
D.F. Numerator 1 1 1 1 1 1 1 1 1
D.F. Denominator 82 1 820 806 805 830 829 771 770 1169
Year
F-Statistic 30.1” 125.7” 28.811‘ 105.00” 24. .W 82.51” 31.738 98.51“ 59.37”
D. F. Numerator 14 14 14 14 14 14 14 14 14
D.F. Denominator 834 820 819 805 843 829 784 770 1182
12-Month Variation
Seasonality 0.44820 0.05225 0.0786 0.0887 0.23963 0.22149 0.08742 0.08308 0.20017
Maturity 0.00546 0.00862 0.0005 0.0006 0.00084 0.00062 0.01728 0.02695 0.00823
 Soybeans Soybean Oil Soybean Oil Live Cattle Live Cattle Silver Silver Cocoa Cocoa
L2 L1 L2 L1 L2 L1 L2 L1 L2

Seasonality
F-Statistic 45.49" 19.862* 21.768" 4.04" 5.00" 5.5073" 7.2383" 4.698' 4.570
D.F. Numerator 11 11 11 11 11 11 11 11 11
D. F. Denominator 1168 1349 1338 1044 1033 1166 1155 848 837
Maturity
Coefficient -0.029 - 0.045 - 0.053 - 0.047 - 0.048 -0.007 - 0.010 - 0.023 - 0.027
F-Statistic 31.14" 50.024a 130.41' 44.48" 81.45* 3.2280 3.1503 7.379b 27.850
D.F. Numerator 1 1 1 1 1 1 1 1 1
D.F. Denominator 1168 1339 1338 1034 1033 1156 1155 838 837
Year
F-Statistic 442.18" 53.82 la 247.85' 46.87' 187.W 40.455" 97.3370 25.347' 64.40
D.F. Numerator 14 14 14 14 14 13 13 14 14
D.F. Denominator 1168 1352 1338 1047 1033 1168 1155 851 837
12-Month Variation
Seasonality 0.16083 0.0723 0.0595 0.01399 0.01809 0.0195 0.0353 0.0136 0.014
Maturity 0.01014 0.0234 0.0334 0.02645 0.02758 0.0006 0.0013 0. oofj6 0.008

"Significant at the 0.995 level.

'Significant at the 0.99 level.
'Significant at the 0.95 level.
results were similar for parametric tests based on either least squares or least
absolute deviations estimates as well as for the rank tests, the results are not
dependent upon any specific assumption concerning the shape of the distribution
of futures prices.

VI. IMPLICATIONS FOR FURTHER RESEARCH

What are the implications of the findings of this article for further work concerning
commodity futures markets? An immediate implication is that in any empirical study
of futures prices there should be a presumption that the variance of futures prices
is nonconstant; consequently, efficient estimation will generally require a correction
for heteroscedasticity.
While we have avoided the use of any specific functional relation, the results
also have implications for the investigation of intertemporal models of futures pricing.
Current interest is focused on models that can be derived within an intertemporal
multigood equilibrium (see Richard arid Sundaresan, 1981 and French, 1981). It
is convenient to refer the French's discrete-time model of futures prices. I n a very
general setting, equilibrium futures prices must satisfy

F:t = P.: E;,(MRS{,R,,) (7)

where F{t is the price at time t of a futures contract for good j for delivery at time
i, P{ is the cash price of goodj at time t , E, is the expectations operator based on
information through time t , ICZRS{, is the marginal rate of substitution (ratio of
marginal utilities) of g o o d j consumption between date t and date i, and R,, is the
short-term yield (in excess of the subjective rate of interest) compounded from date
1 to date i. According to this relation, fluctuations of futures prices originate from
fluctuations of the current cash price, of expectations concerning interest rates, and
of expectations of intertemporal marginal rates of substitution. In view of our empir-
ical findings, in order for a model based on Eq. (7) to be consistent with the observed
behavior of futures prices, there must be a seasonality in at least one of these factors.
Nominal interest rates do not manifest very strong seasonal fluctuations. Cash price
fluctuations may be seasonal. However, for annually harvested goods where stock
holding can approach zero prior to harvest, fluctuations of distant futures may be
isolated at times from cash price fluctuations. Consequently, it seems plausible that
some futures price fluctuations may be attributable to changing expectations of
intertemporal marginal rates of substitution. We would expect models based on a
constant marginal rate of substitution to be poor choices when attempting to model
commodity futures.
It should be noted that this reasoning is consistent with the interpretation of the
seasonality of futures price fluctuations as reflecting seasonal information patterns
concerning yields. This follows because marginal rates of substitution for a good
will be related to the change in the consumption of that good over t h e relevant time
period. Finally, we note that the argument of this section parallels those of Grossman
and Shiller (1981) concerning the volatility of securities prices within the context
of a single good model.

VII. CONCLUSION
The volatility of daily price changes is investigated in nine American futures markets
over the period 1966 to 1980. This data set is considerably more extensive than
those employed in previous studies of daily price volatility. We make use of a
nonparametric and robust statistical methods; this reduces the chances that our
inferences are conditional on incorrect or nonverifiable assumptions.
We find that the variance of futures price changes is not constant and that the
changes of variance follow a partially regular pattern. The principal predictable
factor in changes of variance is seasonality. A secondary factor is the changing time
to maturity.
The fact that there is seasonality in the volatility of futures prices in markets with
annual harvests will hardly come as a surprise to those familiar with the fundamental
factors of supply and demand in those markets. However, these important seasonal
factors have been overlooked in the previous studies of the volatility of futures prices
which have been concerned with the effect of changing time to m a t ~ r i t yWhile
.~ we
find seasonality is relatively more important than maturity, we do find the inverse
relation between time to maturity and volatility is a general tendency in the markets
studied. In view of the coverage of our data set and the generality of our methods,
this is stronger support for the Samuelson hypothesis than found in previous studies.
In view of their pervasiveness, it is hard to state all the implications of these
empirical findings. They can be summarized by saying that users, regulators, and
students of futures markets should generally allow for the fact that futures price
changes are likely to be nonstationary in second moments. Thus at a practical level,
hedgers guided by the portfolio theory of hedging should adjust hedge ratios (ratios
of the covariance of futures and cash prices to the variance of futures prices)
seasonally. Another practical implication would be that in the pricing of options
on futures contracts allowance should be made for the regular pattern to the volatility
of futures.
An implication for research is that studies of futures prices should generally
correct for heteroscedasticity in order to achieve efficient estimates. Finally, the
results suggest that intertemporal equilibrium pricing models for futures should
allow for the variability of the intertemporal marginal rates of substitution.

Appendix I
For convenience we provide a brief description of the rank test for two-way layouts
without interaction and with arbitrary numbers of observations per cell due to Benard
and Van Elteren (1953). Let the model be
-
v,, = 01; + p, + C'bk i = 1, . . . , I
j = 1 , ...,.I
k = l , . . . , nG
{ebk}are i.i.d. (0, a')

91t is true that Rutledge may have implicitly allowed for seasonal fwtors biiice he controlled tor the v o l i t i i l i ~ y
of cash prices. This is unlikely to be a complete correction since in years when little o r norir of Ihr old crqi is
stored into the new crop year, Ihp cash market may be effectively segmented from the new-c-rop futures.

DETERMINANTS OF VOLATILITY / 345

where v,k is the kth observation of cell ij, a, and p, are unknown constants, and ny
is the number of observations in cell z j . The null hypothesis is
a, = a1 = .. . = a,.
The test statistic can be computed as follows. Let r 6 k be rank of V,, within set
of j' = j}. There are n, =
{VLtlrktl x,
n, such ranks. Under the null the expected
ranks are given by
1
14: r{,k = - nj(n, + 1).
2
Define

the sum of mean adjusted ranks for effect i.

Under the null the variance/covariance matrix of u = (u,,. . . , u,) is 2 =
where
[a,,!]

and

It can be shown that 2 is singular by construction. Define the bordered matrix

Let C* and c,T

be obtained from 2 and Xu by deleting any one of the first I columns
and any one of the first I rows. The test statistic is

x=- Det 2:
Det C*
Under the null hypothesis its asymptotic distribution is Chi square with I - 1
degrees of freedom. It should be noted that:

(1) When ny = 1, i = 1, . . . , I aridj = 1, . . . , J, X is the Friedman (1937)

rank test;
(2) If the same column and row have been deleted and u* is the corresponding
subvector of u, then it is equivalent to compute,
X = u*C*-luT.
9

( 3 ) The above treatment assumes there are no ties in the rankings and that X *
is invertible. See Benard and Val Elteren (1953) for the appropriate modi-
fications when these assumptions are violated.

346 I ANDERSON
 Appendix I1
We briefly outline the procedure for hypothesis testing with linear models estimated
by least absolute deviations. Let the model be
9, = X,p + C,, t = 1, . . . ,T (All
where p : p X 1, X, : 1 X p , X1,= 1all t (i.e., intercept is included), and {E,} - i.i.d.
(0, d). The L1 estimator 6 solves,

It can be shown that 6 can be found by solving a linear program. Partition p and
X, so that
Vt = XI,@, + X,*Pz + C,
where p1 : ( p - k) X 1, p2 : k X 1, X1,: 1 X (p - k) and X, : 1 X k. The
restriction to be tested is

@z = 0. (A31
The Lagrange multiplier test for L1 regressions are derived and analyzed in Bassett

Let 6' = :.I

and Koenker (1982) and can be computed as follows.

[. be the constrained L1 estimator; that is, 6' solves Eq. (AZ)

subject to Eq. (A3). Let X = [X,: X,] be the matrix of the T vectors X,. Define
the vector of signs of the constrained L1 residuals
s = sign (V - Xlbp).
The Lagrange multiplier statistic is
P = s'x2(x;x2- x;x,(x;x,)-'x;x,)-'x;s.
If (1IT)X'X converges to a positive definite matrix and if the density of {E,} exists
and is continuous at the median, then under the null, S has the limiting distribution
XW.
In practice, 5? is corrected for degrees of freedom
P* = ( 6 / k ) ( T - p + k)/T).
P* is compared to the critical values of the F ( k , T - p + k).
Bassett and Koenker discuss the asymptotic efficiency of 2 compared to the least
squares F-test. It is less efficient for normal (Gaussian) errors. It is more efficient
for fat tailed distributions.
The small sample efficiency of 2* versus the least squares F-test was studied in
a Monte Carlo experiment by Bassett and Koenker. For moderate or large sample
sizes (T 2 60) the power of 2* was almost as large as the power of F when the
underlying error distribution was normal. 6* was distinctly more powerful than F
for Laplace or Cauchy errors.

DETERMINANTS OF VOLATILITY 1 347

 Bibliography
Anderson, R. W., and Danthine, J. P. (1983): “The Time Pattern of‘ Hedging arid the
Volatility of Futures Prices,” Review o j Economic Studies, 50: 249-266.
Benard, A . , and Van Elteren, P. (1953):“A Generalization of the Method of m Rankings,”
Indagationes Mathematime, 15: 358-369.
Castelino, M. G. (1982):“Price Volatility of Futures Contracts: l h e Maturity Effect,” Hofstra
University, mimeo.
Castelino, M. G., and Francis, J. C. (1982): “Basis Speculation in Commodity Futures:
The Maturity Effect,” Journal of Futures Markets, 2: 195-206.
Castelino, M. G., and Vora, A. (1984):“Spread Volatility in Commodity Futures: The Length
Effect,” Journal of Futures Markets, 4: 39-46.
Cox, J., Ingersoll, J., and Ross, S. A. (1978): “A Theory of the Term Structure of Interest
Rates,” Research Paper No. 468, Graduate School of Business, Stanford University.
French, K. (1981): ‘The Pricing of Futures Contracts,” mimeo.
Grauer, F. L. A. (1977): “Equilibrium in Commodity Futures Markets: Theory and Tests,”
unpublished Ph. D. disseration, Stanford University.
Grossman, S. J., and Shiller, R. j.(1981):“Determinants of the Variability of Stock Market
Prices,” American Economic Review, 71 : 222-227.
Koenker, R. (1981):“Robust Methods in Econometrics,” Bell Laboratories Discussion Paper
228.
Koenker, R., and Bassett, G. (1982): “Tests of Linear Hypotheses and L1 Estimation,”
Econometrica 50: 1577-1583.
Mann, J. S., and Heifner, R. B.: “The Distribution of Shortrun Commodity Price Move-
ments,” USDA Economic Research Service Technical Bulletin No. 1536.
Miller, K. D. (1979): “The Relation Between Volatility and Maturity in Futures Contracts,”
in Commodity Murkets and Futures Prices, R. M. Leuthold, (Ed.), Chicago Mercantile
Exchange, Chicago, 11,.
Milonas, N. T. (1984):“Liquidity, Price Variability and Storage Asymmetry: A Study of the
Behavior of Prices in Futures Markets,” unpublished Ph. D. dissertation, City University
of New York.
Richard, S. F., and Sundaresan, M. (1980):“A Continuous Time Equilibrium Model of
Commodity Prices in a Multigood Economy,” Carnegie Mellon University, mimeo.
Kichard, S. F., and Sundaresan, M. (1981): “A Continuous Time Model of Forward Prices
and Futures Prices in a Multigood Economy,” Journal of Financial Economics, 9: 347-
371.
Rutledge, D. J. S. (1976):“A Note on the Variability of Futures Prices,” Review of Economics
and Statistics, 58, 118-120.
Samuelson, P. A. (1965): “Proof that Properly Anticipated Prices Fluctuate Randomly,”
Industrial Management Review, 6: 41-49.
Samuelson, P. A. (1976):“Is Real-World Price a Tale Told by the Idiot of Chance,” Review
of Economics and Statistics, 58: 120-123.
Stein, J. L. (1979): “Spot, Forward and Futures,” Research in Finance, 1: 225-310.
Telser, L. G. (1956): “The Supply of Stocks: Cotton and Wheat,” unpublished Ph.D.
dissertation, University of Chicago.

348 I ANDERSON