Upload
others
View
33
Download
0
Embed Size (px)
Citation preview
1
The normal backwardation of Keynes and Hicks in Crude Oil Futures
Ronald D. Ripple
Mervin Bovaird Professor of Energy Business and Finance
School of Energy Economics, Policy, and Commerce
Collins College of Business
The University of Tulsa
Tulsa, Oklahoma, USA
Extended Abstract
The concepts of normal backwardation and normal contango are presented in terms of theory developed by Keynes (1930) and Hicks
(1946) and compared to current usage for backwardation and contango. The evolution of the forward curves for crude oil futures over
the 1995-2015 period is evaluated against the expectations of the relationships between hedgers’ net short positions and corresponding
risk premia. A series of error correction models based on autoregressive distributed lag (ARDL) specifications are estimated to
evaluate the relationships between the percent of open interest represented by net short hedgers and the slope of the forward curve,
which proxies risk premium. A significant long-run positive relationship is found. Support is also found for Samuelson’s (1965)
prediction that near-maturity futures prices will be more volatile than more distant maturities based on the speed of adjustment
coefficient estimates of the error correction specifications.
Keywords: normal backwardation, normal contango, autoregressive distributed lag, error correction, oil futures, forward curves.
Introduction
Keynes (A Treatise on Money, 1930) is credited with introducing the concept of backwardation into the economist’s lexicon, and
Hicks (Value and Capital, 1946) provides additional clarity. The form of backwardation put forward by these two influential writers is
often referred to as normal backwardation, and this may be distinguished from the current common usage of the stand-alone term,
backwardation.
2
Normal backwardation refers to the relationship between the current futures price for a commodity to be delivered at some future
time, T, and the current expectation for the future spot price at T. This is distinct from the current usage of backwardation which
relates the current futures price for a commodity to be delivered anytime in the future to the current spot price or nearer maturity
futures price.1
The current common usage of the terms backwardation and contango may be represented as backwardation holding when ,t t T
S F and
contango holding when ,t t T
S F , where S is the current spot price, F is the futures price, t is the present time period, and T is the future
date at which the futures contract matures. These relationships imply that the forward curve of futures prices will be downward
sloping for backwardation and upward sloping for contango. Alternatively, normal backwardation and normal contango require these
slope conditions but require further that , ,t T t T
ES F holds for both, where ,t T
ES is the expectation at time t for the future spot price at
date T. Therefore, normal backwardation requires ,t t T
S F and , ,t T t T
ES F , while normal contango requires ,t t T
S F and , ,t T t T
ES F .
Therefore, normal backwardation/contango provides for an explicit representation of a risk premium (or insurance, in the words of
Keynes), while the current usage does not.
Keynes’ theory of normal backwardation relates hedgers’ futures positions to the risk premium (insurance) necessary to attract
sufficient speculative activity to support risk mitigation demand. The fundamental assumption is that hedgers are typically net short in
the futures contract for their commodity and speculators are therefore net long. Being net long, it is argued that the current futures
price for delivery at time T will be below the spot price expected to hold at time T, so that the speculators who are providing a risk
mitigation service may expect to earn a positive return. Moreover, there is a positive relationship between hedgers’ net short positions
and the necessary risk premium.
The current paper diverges from the typical literature related to backwardation,2 and its alternative contango, and seeks to determine
whether or not this fundamental condition and its relationship to normal backwardation and normal contango, as stated by Keynes and
Hicks, does or does not hold. That is, the paper examines the relationships between hedgers’ net short positions and the slopes of
forward curves, and tests for the posited relationships between net short positions and required risk premium.
The specific commodity to be examined is crude oil, and the forward curves to be examined are those for the futures contracts traded
on the New York Mercantile Exchange (NYMEX). The data to be employed for this analysis are drawn from both NYMEX futures
1 This distinction also holds when one constructs a discounted forward curve, taking into consideration the time value of money over the term structure. The
relationship between the discounted forward curve and the current spot price falls into the same category as the current interpretation of backwardation, since the
spot price of interest is the current spot rather than the expected future spot. 2 See for example, Considine and Larson (2001) and Charupat and Deaves (2002).
3
price data and the Commitment of Traders (COT) database collected and published by the Commodity Futures Trading Commission
(CFTC). The COT database contains information on the open interest (OI) positions of large traders, which have been partitioned into
commercial and non-commercial trader categories. The CFTC data are reported weekly, and they are date-matched to the forward
curves for each of the NYMEX crude oil futures contracts. The analysis evaluates the relationship between the net short open interest
positions of the large commercials (hedgers) and the shape of the forward curve over the period from January 1995 through December
2015. This commodity represents a significant input into global economic activity, the futures contract for crude oil is the most heavily
trade commodity contract globally, and it provides a challenging testing ground for this theory given the swings from contango to
backwardation (in current usage), and back, over the period.
The paper examines the net short positions of hedgers, as characterized by those large traders with a commercial interest in the traded
commodity, to determine both whether or not the fundamental assumption held by Keynes and Hicks has been met over the 1995-
2015 period and how the distributions of open interest positions relates to the observed evolving shapes of the forward curves for
crude oil.
The paper proceeds by presenting the theory of normal backwardation and contango of Keynes and Hicks. This is followed by a
discussion of the empirical model and testing procedures. The data used are then presented, and this is followed by the empirical
results. The paper closes with a summary and conclusions.
The theory
The general backwardation or contango character of the forward curve is often related to the balance of current demand and supply
relative to that expected in the future. If supply is relatively limited currently, spot prices will tend to be higher than those expected to
hold in the future, and backwardation will be expected. The reverse holds for current surplus of supply, leading to contango. However,
normal backwardation, according to Keynes and Hicks, is driven by one fundamental relationship; the normality of hedgers being net
short in their commodity contracts. This sense of normal activity in the market for commodities drives the expected relationship
between the forward/futures price for delivery date T and the spot price expected to hold at that date. Since the commercial
participants in the commodity market, with market risk exposure, will normally be out of balance in terms of the desired levels of long
and short hedging positions, speculators must be attracted to the market. Such speculators are said to be attracted to the market by an
expectation of profit just sufficient to provide “fill the gap” between long and short hedgers. The expectation of profit is determined by
the relationship between the current market price for the futures contract for delivery at date T and the speculators’ (and perhaps the
entire market’s) expectation for the spot price that will hold at time T.
4
Also, Keynes (1930, p. 143) states, “[b]ut it is not necessary that there should be an abnormal shortage of supply in order that a
backwardation should be established. If supply and demand are balanced, the spot price must exceed the forward price by the amount
which the producer is ready to sacrifice in order to ‘hedge’ himself, i.e. to avoid the risk of price fluctuations during his production
period. Thus in normal conditions the spot price exceeds the forward price, i.e. there is backwardation.”3
The relationship also must hold between the futures price and the associated expected future spot price for an observed contango to be
termed a normal contango. Indeed, Keynes (1930, p. 144) states, “[b]ut the existence of a contango does not mean that a producer can
hedge himself without paying the usual insurance against price changes. On the contrary, the additional element of uncertainty
introduced by the existence of stocks and the additional supply of risk-bearing which they require mean that he must pay more than
usual. In other words, the quoted forward price, though above the present spot price, must fall below the anticipated future spot price
by at least the amount of normal backwardation; and the present spot price, since it is lower than the quoted forward price, must be
much lower than the anticipated future spot price.”4
Keynes implies that hedgers will be net short in their forward markets, as is evident from the requirement to attract speculators by
“paying the usual insurance”. Hicks (1946, p. 137) provides more context for this situation to be deemed normal by noting, “[n]ow
there are quite sufficient technical rigidities in the process of production to make it certain that a number of entrepreneurs will want to
hedge their sales for this reason…But although the same thing sometimes happens with planned purchases as well, it is almost
inevitably rarer…” In other words, we expect short hedgers to dominate long hedgers, creating a void to be filled by speculators.
Hicks (1946, p. 138) states further that “[b]ut for this very reason forward markets rarely consist entirely of hedgers.”
Moreover, speculators are expected to reduce the divergence between the expected future spot price and the futures price for the same
future date. In the absence of speculators, the larger the initial imbalance between short and long hedgers, the larger will be this
divergence. Since in futures markets there cannot exist an imbalance between short and long positions, the futures price that could be
asked by short hedgers, for any given expected spot price, will have to fall to attract sufficient long hedgers, who according to Hicks
are less motivated to hedge. Effectively, supply (of futures positions) is initially exceeding demand, so the futures price must fall to
reach equilibrium. In terms of hedging services, this situation implies that the demand (coming from shorts) for risk mitigation
services exceeds the supply (coming from longs), again in the absence of speculators, so the price paid for risk mitigation services will
3 It is occasionally noted that Keynes frequently requires some careful interpretation; a role that Hicks played on many occasions. Here some clarity also must be
introduced, given the distinctions between normal backwardation and the current backwardation usage, and for consistency across Keynes discussions of these
topics. When Keynes says “backwardation” he means his normal backwardation. And when he says supply and demand are balanced he is talking about the
supply and demand for futures/forward positions (longs = shorts). And, when he says “the spot price” must exceed the forward price, he is talking about the
expected spot price at the maturity date of the forward, which is central to his normal backwardation. 4 This is contrary to the popular interpretation reflected in Hull (2002, p. 28): “…when the futures price is above the expected future spot price, the situation is
known as contango. (Bold emphasis added.)
5
be pushed upward, i.e., there will be a larger risk premium as captured by the difference between the expected future spot price and
the futures price for that date.
However, if speculators may enter the market to fill this initial void, they will bid up the price of the futures, or bid down the risk
premium, associated with a given expected future spot price. This may be viewed as either the speculators increase the demand for the
commodity or increase the supply of risk mitigation services.5 An alternative way to think about these relationships is that we typically
expect speculators to be less risk averse than hedgers, which would imply that speculators will be willing to enter the market for a
lower risk premium than will long hedgers. Therefore, the divergence between the expected future spot price and the futures price will
tend to be less when speculators are present. This conclusion follows from Hicks (1946, p. 138) where he states that, “…their
[speculators] action tends to raise the futures price to a more reasonable level.”
Nevertheless, while the introduction of speculators will tend to reduce the risk premium that would otherwise hold, it will still be
necessary for hedgers to increase the risk premium to attract more speculative activity to the market to provide more risk mitigation
service. That is, an increased demand for risk mitigation services will only be satisfied if there is a willingness to pay a higher risk
premium.
As a result, it should be expected that the larger the proportion of the total market that is represented by net-short hedgers, i.e., the
greater the hedging pressure, the larger the insurance that must be paid to attract the required speculators. An increase in the share of
the open interest attributed to net short hedgers implies an increase in the demand for the supply of risk mitigation services provided
by speculators, and this should only be expected to be provided at a price, i.e., an increase in the aforementioned differential, albeit at
a lower price than by long hedgers alone.
The model developed here is informed by these relationship expectations put forward by Keynes and Hicks, and also elaborated on by
Kaldor (1939, where references to Hicks are in the context of Hicks’ first edition, 1939, of Value and Capital). In brief, the normal
economic relationship that underpins the concepts of normal backwardation and normal contango is the tendency for hedgers in the
forward/futures markets for commodities to be net short. Moreover, the larger the proportion of the futures market represented by net
short positions by hedgers the larger should be the expected differential between the futures price for delivery at time T and the
expected future spot price at time T. For a given expected future spot price, this will result in steeper backwardation or flatter
contango.
5 It should be emphasized that speculators do not come to the futures markets for the purpose of providing insurance services to hedgers. They come to these
market with the expectation/hope of making a profit, and it is purely an ancillary benefit to the hedgers that there is sufficient interest from speculators to be on
the opposite side of the deals where they are looking to hedge.
6
Normal backwardation/contango is therefore a function of the net short positions of hedgers. And, the degree of normal
backwardation/contango is a function of the proportion of the futures market represented by net-short hedgers. The “degree” of normal
backwardation/contango should be understood to mean the relative divergence between the expected future spot price and the lower
futures price, which, ceteris paribus, may be captured as the steepness of the slope of the futures-price forward curve.
According to the theory, hedgers being net short is consistent with both normal backwardation and normal contango, so a relationship
must be identified that relates the hedgers’ net positions to the slope of the forward curve that captures the theoretical link between net
short hedging and unobserved risk premium.
Assume a risk premium exists, such that 𝐸𝑅𝑃𝑡𝑇 = ln(𝐸𝑆𝑡
𝑇) − ln(𝐹𝑡𝑇) > 0. For ∆𝐸𝑅𝑃𝑡
𝑇 > 0, we will have ∆[ln(𝐸𝑆𝑡𝑇) − ln(𝐹𝑡
𝑇) > 0,
which can occur under five conditions:
(a) ∆𝐸𝑆𝑡𝑇 > 0 and ∆𝐹𝑡
𝑇 < 0,
(b) ∆𝐸𝑆𝑡𝑇 > 0 and ∆𝐹𝑡
𝑇 = 0,
(c) ∆𝐸𝑆𝑡𝑇 > 0 and ∆𝐹𝑡
𝑇 > 0, with ∆𝐸𝑆𝑡𝑇 > ∆𝐹𝑡
𝑇,
(d) ∆𝐸𝑆𝑡𝑇 = 0 and ∆𝐹𝑡
𝑇 < 0, and
(e) ∆𝐸𝑆𝑡𝑇 < 0 and ∆𝐹𝑡
𝑇 < 0, with |∆𝐸𝑆𝑡𝑇| < |∆𝐹𝑡
𝑇|
Only condition (e) is likely to be relevant to an increase in the net short positions of hedgers. To see this, consider the following.
Assume the market is in equilibrium such that the risk premium being charged/earned clears the market for hedging services. Now, if
hedgers increase their net short positions, they are effectively bringing more of the commodity onto the market for sale in period T,
which will tend to decrease the expected future spot price. However, the act of shorting more futures contracts requires that more
speculators (and/or long hedgers) be attracted to take the matching long positions, and this will occur only if the equilibrium risk
premium is increased. Therefore, both the expected future spot price and the futures price will fall. However, since the risk premium
must increase to attract more long-position takers, the futures price will have to fall by more than the expected future spot price.
What market conditions are likely to stimulate an increase in net short positions, relative to long positions, by hedgers? First, the
market may conclude that the expected future spot price will decline because of either expected demand declines or supply increases.
So, the short hedgers are faced with changed market expectations that have a negative impact on their expected future profits, and they
react by acquiring more insurance, that is, they short more contracts.6 Alternatively, but likely inter-connectedly, producers find they
will be able to bring more commodity to market than previously thought, and they seek to increasing their hedging operations
6 This is because they may have limited ability to reduce output.
7
accordingly. A third possible driver is a change in hedging preferences, which could be tied to greater comfort with the use of these
derivatives instruments for risk mitigation. So, whether the increase in hedgers’ net short positions is the result of a changed
expectation for future spot prices or changed hedging preferences, the result will be the same.
On the other hand, it is difficult to conceive of a circumstance when we would have both an increase in hedgers’ net short positions
and an increase in the expected future spot price. If hedgers were pure hedgers, those who are insensitive to expectations on future
spot prices, they will only change their hedging position to match their actual change in exposure, and perhaps a change in hedging
preferences. If hedgers’ net positions become more short as a result of pure hedging pressures, this will be most likely associated with
a decline in the expected future spot price, because it implies that more supplies of the commodity will be being delivered to the
market at time T. And, if hedgers do actively consider the expected future spot price when making their hedging decisions, an
expectation of increasing future spot prices will not be associated an increase in short hedging; these hedgers will withhold some share
of their exposure to take advantage of the expected rise in the price of the commodity to be sold. Therefore, conditions, (a), (b), and
(c) are quite unlikely to apply to an analysis of a market where hedgers’ net short positions are increasing.
Finally, condition (d) is also unlikely to be applicable. Pure hedgers will not pay attention to expected future spot price, but if they do
bring more supplies to market, the most likely effect will be for expectations of declining future spot prices. And, if the hedgers do pay
attention to expected future spot prices, the lack of change in expectations will not provide any basis for them to change their hedging
positions from the existing equilibrium levels.
A similar set of five conditions may be developed with reversals in direction or inequality of relative change for a ∆𝐸𝑅𝑃𝑡𝑇 < 0. The
equivalent condition to (e), for a decrease in hedgers’ net short positions7, is: (e´) ∆𝐸𝑆𝑡𝑇 > 0 and ∆𝐹𝑡
𝑇 > 0, with ∆𝐸𝑆𝑡𝑇 < ∆𝐹𝑡
𝑇.
The expected risk premium is assumed to be time sensitive, in that it is expected that 𝐸𝑅𝑃𝑡𝑚 > 𝐸𝑅𝑃𝑡
𝑛 for 𝑚 > 𝑛, where m and n
indicate maturity dates. It is also expected that for a given change in hedging pressure (represented below as the percentage of open
interested account for by net short positions, phns) ∆𝐸𝑅𝑃𝑡𝑚 > ∆𝐸𝑅𝑃𝑡
𝑛 for 𝑚 > 𝑛. Following from these conditions, it is expected that
an increase in the unobserved ERP will tend to lead to a steepening of the forward curve.
A specification for the relationship between hedgers’ net short positions and the slope of the forward curve may be employed to test
the Keynes-Hicks theory with an unambiguous expectation of a positive relationship. This may be represented generally as
7 Remember that a decrease in net short positions implies that less product is being brought to market at time T. Furthermore, the specification of net short
positions is robust to conditions where hedgers are net long; the value of the variable will be negative.
8
( )s f phns (1)
where: s is the spread between the near-month futures price and that of a distant maturity
phns is the percentage of open interest implied by hedgers net short positions
For the present date, t, and maturity date, T, ,1 ,
ln( ) ln( )T
t t t Ts F F , phns hns OI , and hns hs hl , where hs and hl are hedgers’
short positions and hedgers’ long positions, respectively. s provides a proxy for the slope of the forward curve. As defined, s>0
implies backwardation, s<0 implies contango, and phns>0 implies hedgers are net short.
A linear representation of this relationship is
T
t ts phns (2)
where: s phns , which is expected to be positive. That is, as the hedgers’ net short positions as a percentage of open interest
increases (i.e., hedging pressure increases), it is expected that the spread, s, will increase, representing the expected increase in risk
premium; i.e., we get a steeper backwardation or flatter contango.
The expectation noted above whereby 𝐸𝑅𝑃𝑡𝑚 > 𝐸𝑅𝑃𝑡
𝑛 for 𝑚 > 𝑛 implies that 𝑠𝑡𝑚 > 𝑠𝑡
𝑛 for 𝑚 > 𝑛 . And further, that ∆𝐸𝑅𝑃𝑡𝑚 >
∆𝐸𝑅𝑃𝑡𝑛 for 𝑚 > 𝑛 implies ∆𝑠𝑡
𝑚 > ∆𝑠𝑡𝑇 for 𝑚 > 𝑛 for a given ∆𝑝ℎ𝑛𝑠. The implication is that a change in hedging pressure is
expected to be revealed by a change in the slope of the forward curve through its effect on the expected risk premium.
The empirical model and testing procedures
A series of error correction models based on autoregressive distributed lag models are specified to test the theoretical relationships
between hedgers’ net short positions and the spreads of the forward curves for different futures maturities. This specification will
provide estimates of both the long-run relationships between the variables and the speed of adjustment of the slope to changes in
hedgers’ net short positions. The models will also allow comparison across maturities.
The relationships between the spread and the proportion of open interest represented by hedgers’ net short positions are modelled as
ARDL(p,q),
0 1( , ) ( , )T
t t tL p s L q phns u (3)
9
where L is the lag operator, p and q are the numbers of lags, s is the spread, phns is the percent of open interest accounted for by
hedgers’ net short positions, and u is a disturbance term.
The ARDL approach8 has been shown to produce cointegration vector estimates that are equivalent to those from error correction
models, see Hassler and Wolters (2006). Pesaran et al. (2001) have demonstrated that the ARDL bounds approach may be employed
to test for the existence of a long-run “forcing” relationship between and among variables regardless of whether they are I(0), I(1), or
fractionally integrated. The long-run relationship of the levels variables is tested employing a non-standard F-statistic evaluated
against a band of critical values. If the test statistic exceeds the upper limit of the critical-values band, a long-run relation cannot be
rejected.9 This approach is used here.
Once the existence of a long-run relationship is established the optimal lag lengths, p and q, for the ARDL(p,q) specifications may be
determined.10 The resulting long-run coefficient estimates may then be employed to estimate an error correction model.
The error correction model associated with Equation 3 may be written as follows:
ˆ ˆ1 1
1 01 1
p qT T T
t t t j t j j t j t tj j
s EC phns s phns w u
(4)
where ECt is the error correction term
T T
t t t tEC s phns w
and w are deterministic variables (in this case the intercept, a trend, and three dummies to capture abnormally large spikes in
the slopes in early 2009).
8 The ARDL methodology has a long history of application to a range of economics questions, including the savings-investment relationship, the trade-GDP
relationship, international reserves accumulation and exchange rate intervention, and recent renewed application to energy-related questions. See for example,
Hendry et al. (1984), Bentzen and Engsted (2001), De Vita and Abbott (2002), Pahlavani et al. (2005), Kollias et al. (2006), Vita, Endresen, and Hunt (2006), and
Ramachandran and Srivinsan (2007). 9 While this approach can identify long-run relationships in levels variables, regardless of the degree of integration, to discuss the long-run “forcing” relationship
inn terms of cointegration it is still required that the dependent variable be I(1). In the case of the variables under study in this paper, all series are I(0).
10 Optional tests include an 2R test, the Akaike Information criterion (AIC), the Schwartz Bayesian criterion (SBC), and the Hannan-Quinn criterion (HQC).
10
Equation (4) implies that when 0T
tEC the spread (slope) will react by decreasing to return to the equilibrium, long-run relationship,
and vice versa.
The results from these testing procedures will provide estimates of the long-run relationship between s and phns, i.e., estimates of ,
as well as estimates of the error correction coefficient, which provides an estimate of the speed with which the variables return to
equilibrium following a shock. As noted, it is expected that 0 . The magnitude of δ is expected to decrease as the time period
represented by the maturity of the contract increases. This expectation derives from Samuelson (1965) where it is proved that we
expect nearer maturity futures prices to be more volatile relative to more distant maturities. This may be represented in the error
correction representation by differing speeds of adjustment to common shocks, with the speed of adjustment for nearer maturities
being faster. This implies m n for m < n, where m and n represent time to maturity.
The data
The data used to construct the forward curves for this analysis are the futures prices for NYMEX crude oil contracts from January
1995 through December 2015. For each observation date, the forward curve represents the settlement prices for crude oil for each
contract maturity traded. The data for the percentage of net short open interest held by hedgers is taken from the CFTC database for
large traders; the Commitment of Traders (COT). These data are disaggregated into subsets for commercial, non-commercial, and
non-reporting, where the non-reporting are relatively small traders who are not required to report. As is typical when employing these
data, the positions of the large commercial traders are used to proxy hedger activity.
The frequency of observations is limited by the COT, which is reported on a weekly basis for Tuesday open interest. These data
represent an aggregate value for all contracts traded for a given commodity. Currently for crude oil the open interest associated with
any subset of the COT is for all contracts open for trading from the current March 2016 delivery to the December 2024 expiry, with no
means of disaggregating further to the individual contract level.
The analyses in this paper match the Tuesdays reporting of the CFTC, which provides 1,095 observations over the period. The series
are first “spliced” to provide continuous time series of expiring futures contracts. For crude oil on the NYMEX, the last trading day is
three business days prior to the 25th of the month in the trading month prior to delivery. As the last trading day nears, the volume of
trade and the open interest in the near-month contract wanes quickly, and prices frequently become more volatile. As a result there
may be a considerable disconnection between the price for the near-month contract and that for the rest of the forward curve as the
market shifts its attention to the next-to-near-month contract. The analysis in this paper evaluates the spread (or slope of the forward
11
curve) as the relationship between the near-month contract price and successive, later-dated maturities; the spot price is not used, so
the analysis will be sensitive to the unusual volatility on the near-month contract prices.
The splicing method employed aims to produce a series that represents the trading activity of the contract of most market interest on
each reported trading date. The method in this paper incorporates a three-step process. The first step employs an algorithm that shifts
the series from the near-month contract to the next-to-near month contract when both the trading volume and open interest of the next-
to-near month contract exceed those for the near-month. The second step employs an algorithm to determine if there are any instances
of contract reversal; that is once shifting to the next-to-near contract does the first algorithm lead to a return to the near-month
contract. The third step applies only when such a reversal occurs. In such a case, the data are manually inspected to set the roll
between contracts so as to produce a smooth, non-reversing (in the sense of flip-flopping between maturities) series that represents the
contracts upon which the market was focused; this occurred fewer than 20 times over 5,256 daily observations, and only four of these
involved a Tuesday.
The forward curve analysis is conducted for contract maturities of two months, three months, six months, and nine months, i.e., T = 2,
3, 6, and 9. Table 1 shows the concentration of trading activity accounted for by these contracts, along with the near-month contract.
The reported values are averages of the percentage of the daily trading volume and open interest for each grouping. For example, the
value of 0.81 for 3-months (in column 2) under the heading for trading volume implies that on average 81% of all trading volume was
attributable to trades for the nearest three contract maturities.
It is clear that trading volume is more concentrated toward the very near-term contracts than is open interest. Nevertheless, including
in the analysis the first nine traded contracts captures 94% of daily trading volume and nearly 70% of the open interest available for
trading. Another way to think about these results is that the 30% of open interest found in the distant-maturity tail of the forward curve
sees only about 6% of the daily trading activity.
Table 2 reports how frequently the forward curves over the 2-month through 9-month maturities are in backwardation, as defined by
,t t TS F , how frequently hedgers are net short, and how frequently hedgers are net short for the full period, 1995-2004, and 2005-
2015. The Keynes and Hicks assumption that hedgers are typically net short is strongly supported by this evidence, with this class of
trader net short 75 percent of the time.11 Moreover, hedgers are found to be net short under both backwardation and contango
conditions.
11 Crude oil is not unusual in this respect; each of the other four energy futures contracts traded on the NYMEX–gasoline, natural gas, heating oil, and propane–
have hedgers net short over 76% of the time. Many commodities (agricultural and mineral) appear to have hedgers net short well over 50% of the time, while
financial futures are more mixed.
12
Figure 1 shows the growth of open interest over the period along with the evolution of the hedgers’ net short positions. Figures 2 and 3
show the percentage of open interest accounted for by hedger net short positions and the four forward curve spreads, respectively, that
are the data for the analyses carried out in this paper.
Empirical results
Four forward curve models are specified, one for each of the more distant maturity futures contracts considered. There is a separate
model for the 2-month, 3-month, 6-month, and 9-month forward curves, with the corresponding spreads designated as s2L, s3L, s6L,
and s9L, respectively. Each s is defined as ,1 ,
ln( ) ln( )T
t t t Ts F F , so, for example, 3
,1 ,3ln( ) ln( )
t t ts F F is a representation of s3 at time
t. And, positive values imply backwardation in the current usage of the term.
The first step of the ARDL bounds approach12 is to estimate an OLS regression of the short-term difference variables and then
perform a non-nested test for the statistical significance of added variables. The initial OLS regression of short-term variables allowed
for a maximum of twelve lags, representing approximately three months of weekly observations, and include an intercept and trend.
These results were tested for stability using the CUSUM and CUSUMSQ tests; all were visually inspected and passed with confidence
bounds at 0.05.
The lagged levels of s and phns are then added to the initial OLS regression, and the statistical significance of these added variables is
tested via an F-statistic. As noted above, the resulting F-statistic is non-standard, so the test statistic is evaluated against the bounds
estimated and reported in Pesaran et al. (2001)13. The F-tests for all four models pass this test. This implies that there is long-run
forcing relationship between the variables, irrespective of their degree of integration, which permits moving on to the ARDL
specification phase and then on to the error correction representation.
These results support the primary hypothesis of this paper. An increase in the share of open interest attributable to hedgers is
positively related to the slope of the forward curve for crude oil through the expected risk premium as represented by the spread
between the near-month contract and more distant maturities. An increase in the demand of hedgers for the services of speculators, an
increase in hedging pressure, comes at a price, the expected risk premium, that is revealed as an increase in the spreads; this implies a
steepening forward curve that is in backwardation and a flattening of one that is in contango. The important estimated coefficients that
12 All estimations and tests were conducted in Microfit 5.0 (2009). 13 For this paper, the critical value bounds reported in Pesaran et al. (2001, p.300) in “Table CI(iii) Case III, Unrestricted intercept and no trend” were used. The
applicable critical value bounds are 4.94 – 5.73, for k=1 and confidence level 0.05. These critical values differ slightly from those found in Microfit 5.0, p. 544-
545, due to a larger T and number of repetitions in the stochastic simulations. Nevertheless, the calculated F-statistics exceed the upper bound against both sets of
tables. The results of these tests are available upon request.
13
provides this support is that for phns, which are reported in Table 3. For the four model specifications, each for a progressively more
distant maturity, the estimated coefficients range from 0.1123 to 0.9698, and they are highly statistically significant, with all positive
as expected and with magnitudes increasing with maturity as expected.
Table 4 reports the error correction term estimates. Each estimated coefficient for the four models has the expected negative sign and
is highly statistically significant. The decrease in magnitude of the speed of adjustment estimates, as the maturities constructing the
spreads increase, is also as expected. On the on hand, the relatively small magnitude of the estimates suggests a modest speed of
adjustment to disequilibria, but it also must be recognized that these speed of adjustments are for a week. And even with the modest
magnitudes, the declining characteristic of these results may be interpreted as providing support for Samuelson (1965) in the following
sense. Samuelson (1965) argues that near-term futures prices will be more volatile than those for more distant maturities. The larger
magnitude of the speed of adjustment term for the nearer term contracts suggests that the prices for these contracts will respond more
quickly to divergences from equilibrium, and this response will be observed as greater volatility in those prices. Table 4 shows that for
a given change in the percentage of open interest accounted for by hedgers the speed of adjustment for the spread between the near-
term contract and the next-to-near contract, s2, will be roughly twice as fast as the adjustment of the spread between the near-month
contract and the ninth-month contract, s9. Since both spreads are anchored by the same near-month contract price, the adjustments
must reflect relative changes to the distant month prices. Therefore, for this example, the next-to-near month contract price must adjust
at roughly twice the speed of the ninth-month price, and this will be observed as greater volatility in the nearer contract price.
Table 5 comprise three sections, A, B, and C, and reports a sample the results of the ARDL estimation sequence, in this case for the
s3L variable. Section A provides the estimation results for the ARDL, Section B presents the long-run coefficient estimates (the most
important being that for phns), and Section C reports the error correction estimation results. The results are presented along with the
range of diagnostic test automatically performed within the Microfit software. The estimates for s3L are very representative of those
for the other three, especially in that all estimated coefficients are statistically significant. Moreover, even the short-run dynamics
reflect the posited positive relation between hedgers’ net short positions and the slope of the forward curves, and, hence, the risk
premium. The other results are available upon request.
Summary and conclusions
The paper develops a theoretical framework for the relationship between a measure of hedging pressure (the net short positions of
hedgers) and the backwardation and contango characteristics of futures forward curves. The analysis finds that the expectations of
Keynes and Hicks regarding hedgers normally being net short is revealed in the data for the NYMEX crude oil futures contracts,
14
which could not have pre-conceived. Moreover, hedgers being net short is consistent with both backwardation and contango
conditions, which is also consistent with Keynes’ observations.
Empirical models, based on autoregressive distributed lag, are specified and estimated. The long-run coefficient estimates for the
relationship between hedgers net short positions and the slopes of the forward curve at different maturities are found to be statistically
significant and of the expected positive signs. Thus, increases in hedgers demands for the services of speculators, i.e., increased
hedging pressure, comes at a cost (increased expected risk premium) revealed as an increase in the spreads between the near-month
contract price and those of more distant maturities.
Finally, the speed of adjustment estimates on the error correction term have the expected sign and decrease with more distant maturity.
This result supports the Samuelson position that near-maturity futures prices will be more volatile than more distant maturities.
While we cannot directly observe the relationships between futures prices and their corresponding expected future spot prices, it is
shown that the theoretical basis for normal backwardation and normal contango has the expected influence on the shape of the forward
curve. The results demonstrate that even over the past 21 years, which have exhibited wide ranging shifts in both the physical and
financial markets for crude oil, the Keynes-Hicks theories of normal backwardation linking the demand for risk mitigation services
(hedger net short positions) and the expected risk premium required to attract those services has remained strong.
Future research may evaluate the possibility for asymmetric responses to changes in the percentage of open interest accounted for by
hedger net short positions depending on whether we are in a backwardation or contango environment.
15
References
Bentzen, J. and Engsted, T. 2001. “A revival of the autoregressive distributed lag model in estimating energy demand relationships,”
Energy, vol. 26, pp.45-55.
Charupat N. and Deaves, R. 2002. “Backwardation and normal backwardation in energy futures markets,” ZEW Discussion Paper No.
02-59, pp. 1-19, ftp.zew.de/pub/zew-docs/dp/dp0259.pdf.
Considine, T.J. and Larson, D.F. 2001. “Uncertainty and the convenience yield in crude oil price backwardation,” Energy Economics,
vol. 23, pp. 533-548.
De Vita, G. and Abbott, A. 2002. “Are savings and investment cointegrated? An ARDL bounds testing approach,” Economics Letters,
vol. 77, pp. 293-299.
De Vita, G., Endresen K., and Hunt, L.C. 2006. “An empirical analysis of energy demand in Namibia,” Energy Policy, vol. 34, pp.
3447-3463.
Hassler, U. and Wolters, J. 2006. “Autoregressive distributed lag models and cointegration,” Allgemeines Statistisches Archiv, vol. 90,
no. 1, 59-74.
Hendry, D.F., Pagan, A.R., and Sargan, J.D. 1984. “Dynamic specification,” in Handbook of Econometrics, Vol. II, eds. Z. Griliches
and M.D. Intriligator, Elsevier, Amsterdam, pp. 1023-1100.
Hicks, J.R. 1946. Value and Captial, 2nd Edition, Oxford University Press, London.
Hull, J.C. 2002. Fundamentals of Futures and Options Markets, 4th edition, Prentice Hall, Upper Saddle River, New Jersey.
Kaldor, N. 1939. “Speculation and economic stability,” The Review of Economic Studies, vol. 7, no. 1, pp. 1-27.
Keynes, J.M. 1930. A Treatise on Money, Vol. II, MacMillan and Co., Ltd, London.
Kollias, C., Mylonidis, N., and Paleologou, S-M. 2006. “The Feldstein-Horioka puzzle across EU members: Evidence from the ARDL
bounds approach and panel data,” International Review of Economics and Finance, Article In Press, December, pp.1-8.
16
Pahlavani, M., Wilson, E., and Worthington, A.C. 2005. “Trade-GDP nexus in Iran: An application of the autoregressive distributed
lag (ARDL) model,” American Journal of Applied Sciences, vol. 2, no. 7, pp. 1158-1165.
Pesaran, H.M. and Pesaran, B. 2003. Microfit 4.0, Oxford University Press, London.
Pesaran, H.M, Shin, Y., and Smith, R.J. 2001. “Bounds testing approaches to the analysis of level relationships,” Journal of Applied
Econometrics, vol. 16, no. 3, pp. 289-326.
Ramachandran, M. and Srivinsan, N. 2007. “Asymmetric exchange rate intervention and international reserve accumulation in India,”
Economics Letters, vol. 94, pp. 259-265.
Samuelson, P. 1965. “Proof that Properly Anticipated Prices Fluctuate Randomly,” Industrial Management Review, 6, 41-49.
17
Figure 1
18
Figure 2
19
Figure 3
20
Table 1: Concentration of trading activity by contract spread period: volume and open
interest – Daily January 1995 – December 2015 (5,256 observations)
Cumulative Percentage of Volume
2-months 3-months 6-months 9-months
0.72 0.81 0.90 0.94
Cumulative Percentage of Open Interest
2-months 3-months 6-months 9-months
0.36 0.45 0.60 0.69
Table 2: Frequency of backwardation and hedgers net short: averages for the period
Backwardation – Full period
2-months 3-months 6-months 9-months HNS
0.40 0.44 0.49 0.52 0.75
Backwardation – 1995-2004
2-months 3-months 6-months 9-months HNS
0.63 0.70 0.77 0.79 0.61
Backwardation – 2005-2015
2-months 3-months 6-months 9-months HNS
0.19 0.20 0.24 0.28 0.88
21
Table 3: Estimated long-run coefficients employing the ARDL(p,q) approach and the
Schwarz Bayesian Criterion (1,086 observations)
Dependent ARDL(p,q) Regressor Coefficient - β Probability
s2 (2,2) phns 0.1123 0.000***
c 0.0153 0.000***
s3 (2,2) phns 0.2396 0.000***
c 0.0316 0.000***
s6 (2,2) phns 0.6148 0.000***
c 0.0776 0.000***
s9 (2,2) phns 0.9698 0.000***
c 0.1176 0.000***
*** represent significance at the 99 percent level.
Table 4: Error correction representation based on the ARDL(p,q) model selection and
Schwarz Bayesian Criterion (1,086 observations)
Dependent ARDL(p,q) ECM - δ Probability R-bar-square
ds2 (2,2) -0.1001 0.000*** 0.3383
ds3 (2,2) -0.0756 0.000*** 0.3035
ds6 (2,2) -0.0470 0.000*** 0.2921
ds9 (2,2) -0.0365 0.000*** 0.2901
*** represents the 99 percent level.
22
Table 5: Estimation results for S3L 5/25/2016 1:42:11 AM
A. Autoregressive Distributed Lag Estimates ARDL(2,2) selected based on Schwarz Bayesian Criterion
*******************************************************************************
Dependent variable is S3L
1084 observations used for estimation from 13 to 1096
*******************************************************************************
Regressor Coefficient Standard Error T-Ratio[Prob]
S3L(-1) .80773 .027893 28.9586[.000]
S3L(-2) .11667 .027347 4.2664[.000]
PHNS .11522 .013222 8.7142[.000]
PHNS(-1) -.13628 .019710 -6.9141[.000]
PHNS(-2) .039174 .013453 2.9118[.004]
INPT .0023924 .7307E-3 3.2742[.001]
TREND -.6182E-5 .1440E-5 -4.2926[.000]
D11309 -.12321 .010671 -11.5458[.000]
D12009 .085904 .011228 7.6506[.000]
D21709 -.10286 .010656 -9.6522[.000]
*******************************************************************************
R-Squared .91558 R-Bar-Squared .91487
S.E. of Regression .010582 F-Stat. F(9,1074) 1294.2[.000]
Mean of Dependent Variable -.0010399 S.D. of Dependent Variable .036269
Residual Sum of Squares .12027 Equation Log-likelihood 3397.6
Akaike Info. Criterion 3387.6 Schwarz Bayesian Criterion 3362.6
DW-statistic 1.8586
*******************************************************************************
Testing for existence of a level relationship among the variables in the ARDL model
*******************************************************************************
F-statistic 95% Lower Bound 95% Upper Bound 90% Lower Bound 90% Upper Bound
25.9998 6.6238 7.3285 5.5918 6.2706
W-statistic 95% Lower Bound 95% Upper Bound 90% Lower Bound 90% Upper Bound
51.9997 13.2476 14.6570 11.1836 12.5413
*******************************************************************************
If the statistic lies between the bounds, the test is inconclusive. If it is
above the upper bound, the null hypothesis of no level effect is rejected. If
it is below the lower bound, the null hypothesis of no level effect can't be
rejected. The critical value bounds are computed by stochastic simulations
using 5000 replications.
23
Diagnostic Tests
*******************************************************************************
* Test Statistics * LM Version * F Version *
*******************************************************************************
* * * *
* A:Serial Correlation*CHSQ(1) = 31.6779[.000]*F(1,1073) = 32.3004[.000]*
* * * *
* B:Functional Form *CHSQ(1) = 18.6694[.000]*F(1,1073) = 18.8038[.000]*
* * * *
* C:Normality *CHSQ(2) = 2131.9[.000]* Not applicable *
* * * *
* D:Heteroscedasticity*CHSQ(1) = 145.0200[.000]*F(1,1082) = 167.1087[.000]*
*******************************************************************************
A:Lagrange multiplier test of residual serial correlation
B:Ramsey's RESET test using the square of the fitted values
C:Based on a test of skewness and kurtosis of residuals
D:Based on the regression of squared residuals on squared fitted values
5/25/2016 1:44:08 AM
B. Estimated Long Run Coefficients using the ARDL Approach ARDL(2,2) selected based on Schwarz Bayesian Criterion
*******************************************************************************
Dependent variable is S3L
1084 observations used for estimation from 13 to 1096
*******************************************************************************
Regressor Coefficient Standard Error T-Ratio[Prob]
PHNS .23960 .060051 3.9900[.000]
INPT .031647 .0088836 3.5624[.000]
TREND -.8178E-4 .1672E-4 -4.8897[.000]
D11309 -1.6298 .25522 -6.3858[.000]
D12009 1.1363 .23770 4.7805[.000]
D21709 -1.3606 .22586 -6.0239[.000]
*******************************************************************************
Testing for existence of a level relationship among the variables in the ARDL model
*******************************************************************************
F-statistic 95% Lower Bound 95% Upper Bound 90% Lower Bound 90% Upper Bound
25.9998 6.6238 7.3285 5.5918 6.2706
W-statistic 95% Lower Bound 95% Upper Bound 90% Lower Bound 90% Upper Bound
51.9997 13.2476 14.6570 11.1836 12.5413
*******************************************************************************
24
If the statistic lies between the bounds, the test is inconclusive. If it is
above the upper bound, the null hypothesis of no level effect is rejected. If
it is below the lower bound, the null hypothesis of no level effect can't be
rejected. The critical value bounds are computed by stochastic simulations
using 5000 replications.
5/25/2016 1:44:54 AM
C. Error Correction Representation for the Selected ARDL Model ARDL(2,2) selected based on Schwarz Bayesian Criterion
*******************************************************************************
Dependent variable is dS3L
1084 observations used for estimation from 13 to 1096
*******************************************************************************
Regressor Coefficient Standard Error T-Ratio[Prob]
dS3L1 -.11667 .027347 -4.2664[.000]
dPHNS .11522 .013222 8.7142[.000]
dPHNS1 -.039174 .013453 -2.9118[.004]
dTREND -.6182E-5 .1440E-5 -4.2926[.000]
dD11309 -.12321 .010671 -11.5458[.000]
dD12009 .085904 .011228 7.6506[.000]
dD21709 -.10286 .010656 -9.6522[.000]
ecm(-1) -.075597 .010684 -7.0758[.000]
*******************************************************************************
List of additional temporary variables created:
dS3L = S3L-S3L(-1)
dS3L1 = S3L(-1)-S3L(-2)
dPHNS = PHNS-PHNS(-1)
dPHNS1 = PHNS(-1)-PHNS(-2)
dTREND = TREND-TREND(-1)
dD11309 = D11309-D11309(-1)
dD12009 = D12009-D12009(-1)
dD21709 = D21709-D21709(-1)
ecm = S3L -.23960*PHNS -.031647*INPT + .8178E-4*TREND + 1.6298*D11309 -
1.1363*D12009 + 1.3606*D21709
*******************************************************************************
R-Squared .30927 R-Bar-Squared .30349
S.E. of Regression .010582 F-Stat. F(8,1075) 60.1107[.000]
Mean of Dependent Variable -.6477E-4 S.D. of Dependent Variable .012680
Residual Sum of Squares .12027 Equation Log-likelihood 3397.6
Akaike Info. Criterion 3387.6 Schwarz Bayesian Criterion 3362.6
DW-statistic 1.8586
*******************************************************************************
R-Squared and R-Bar-Squared measures refer to the dependent variable
25
dS3L and in cases where the error correction model is highly
restricted, these measures could become negative.
Testing for existence of a level relationship among the variables in the ARDL model
*******************************************************************************
F-statistic 95% Lower Bound 95% Upper Bound 90% Lower Bound 90% Upper Bound
25.9998 6.6238 7.3285 5.5918 6.2706
W-statistic 95% Lower Bound 95% Upper Bound 90% Lower Bound 90% Upper Bound
51.9997 13.2476 14.6570 11.1836 12.5413
*******************************************************************************
If the statistic lies between the bounds, the test is inconclusive. If it is
above the upper bound, the null hypothesis of no level effect is rejected. If
it is below the lower bound, the null hypothesis of no level effect can't be
rejected. The critical value bounds are computed by stochastic simulations
using 5000 replications.