Modelling futures price volatility in energy markets: Is there a role for
financial speculation?
☆
Matteo Manera
a, Marcella Nicolini
b,⁎
, Ilaria Vignati
caUniversity of Milan–Bicocca, Fondazione Eni Enrico Mattei, Milan, Italy bUniversity of Pavia, Pavia, Fondazione Eni Enrico Mattei, Milan, Italy c
Fondazione Eni Enrico Mattei, Milan, Italy
a b s t r a c t
a r t i c l e i n f o
Article history:
Received 27 September 2013 Received in revised form 5 June 2014 Accepted 3 July 2014
Available online 13 July 2014 JEL classification:
C32 G13 Q43 Keywords:
Commodities futures markets Speculation
Working's T GARCH models
This paper models volatility in four energy futures markets, adopting GARCH models. The variance equation is enriched with alternative measures of speculation, based on CFTC data: the market share of non-commercial traders, the Working's T index, and the percentage of net long positions of non-commercials over total open interest in future markets. It also includes a control for market liquidity. We consider four energy commodities (light sweet crude oil, heating oil, gasoline and natural gas) over the period 2000–2014, analysed at weekly fre-quency. Wefind that speculation presents a negative and significant sign. The robustness exercise shows that: i) results remain unchanged through different model specifications (GARCH-in-mean, EGARCH, and TARCH); ii) results are robust to different specifications of the mean and variance equation.
© 2014 Elsevier B.V. All rights reserved.
1. Introduction
Financial markets have faced a number of significant changes in the last decade. Commodities' prices grew dramatically during thefirst years of the 2000s and speculators often have been alleged to influence their levels and drive their increases. After the testimonies by hedge fund manager Michael W. Masters before the U.S. Congress and the U.S. Commodity Futures Trading Commission (Masters, 2008, 2009), this argument has been labelled the“Masters Hypothesis”. Recent em-pirical analysis has widely tested this hypothesis,finding that financial speculation generally does not influence the returns of commodities (e.g.Aulerich et al., 2013; Irwin and Sanders, 2012; Sanders and Irwin, 2011). A related issue is whether speculators' activity affects the
volatility of futures prices. On the one hand, speculators increase market liquidity thus reducing price volatility. On the other hand, critics argue that an increasing trading volume, especially by speculators, positively affects volatility.
This paper belongs to this second area of research, as it models vol-atility of four energy commodities over the period 2000–2014, including speculation among the explanatory variables. In particular, we contrib-ute to the literature in two different directions. First, we include in the variance equation different measures of speculation. Second, we adopt alternative specifications for the volatility of futures returns.
Traditionally, the literature has measure speculative activity by means of the Working's T (1960) index, which is based on the relative weight of speculators and hedgers in the market. Alternative measures are the mar-ket share of commercial traders and the percentage of net long non-commercial positions over total open interest. All these measures require a classification of agents between the two categories, which is provided by the U.S. Commodity Futures Trading Commission (CFTC).
We model the volatility of commodities' prices using Generalized Autoregressive Conditional Heteroskedasticity (GARCH). We include macroeconomic controls in the mean equation, such as returns on the T-bill, the Standard & Poor' s 500 (S&P) returns and the junk bond yield, as well as a speculative index in the variance equation. Wefind that speculation is significantly related to price volatility in the period
☆ The authors would like to thank Bahattin Buyuksahin, Marzio Galeotti, Alessandro Lanza, Eduardo Rossi, Giovanni Urga and seminar participants at FEEM, University of Milan-Bicocca, 51st Meeting of the EWGCFM and 1st Conference of the Research Centre for Energy Management (London, 2013), FEEM Workshop on“Oil price forecasts and trends” (Milan, 2013), 6th International Workshop on Empirical Methods in Energy Economics EMEE (Ottawa, 2013) and 54th Italian Economic Association Conference (Bologna, 2013) for helpful comments.
⁎ Corresponding author.
E-mail addresses:[email protected](M. Manera),[email protected] (M. Nicolini),[email protected](I. Vignati).
http://dx.doi.org/10.1016/j.eneco.2014.07.001 0140-9883/© 2014 Elsevier B.V. All rights reserved.
Contents lists available atScienceDirect
Energy Economics
2000–2014. More precisely, the three indexes have a negative coeffi-cient (when significant), thus suggesting that speculation does not de-stabilize prices. Our results are in line with recent evidence (Aulerich et al., 2013; Brunetti et al., 2011; Irwin and Sanders, 2012; Sanders and Irwin, 2011).
We test the robustness of these results moving along two dimen-sions. First, ourfindings remain unchanged through more refined model specifications, such as GARCH-in-mean, threshold GARCH and asymmetric exponential GARCH. Second, we investigate if results are affected somehow by the correct specification of the mean equation. Focussing on crude oil, wefind that the inclusion of controls for the demand, production and stocks does not affect the mainfindings.
The reminder of the paper is structured as follows:Section 2
discusses the relevant literature,Section 3presents the data,Section 4
illustrates the econometric specification, while the results are presented and discussed inSection 5. Finally,Section 6concludes.
2. Literature review
FollowingBrümmer et al. (2013), the existing literature on the analysis of the volatility drivers in commodity futures markets can be classified into three main approaches: descriptive, based on modelling approaches and based on econometric tests.
Thefirst group includes papers that provide descriptive evidence and do not present explicitly quantitative estimates.Regnier (2007)
compares realized volatilities for energy and non-energy commodities from 1946 to 2005,finding that energy price volatility is not as unusual as commonly perceived.Nissanke (2012)discusses price dynamics in agricultural and oil futures markets over the last decade with reference to two conditions— structural changes in market fundamentals and increasing interactive activities betweenfinancial and commodity mar-kets. She suggests that the recent heightened price instability common across commodities is at least partially attributable to“destabilising” trading byfinancial investors.
Among the studies based on mathematical modelling,Babcock (2012)uses a stochastic partial equilibrium model to show that US biofuel policies increased price volatility in agricultural markets.
The third group of econometric works comprises different approaches.
Pindyck (2004)develops a structural model of inventories, spot and future prices that explicitly accounts for volatility and estimates a VAR model for crude oil, heating oil and gasoline. Market variables are found to poorly explain volatility. Most papers adopt generalized conditional heteroskedasticity models.Sadorsky (2006)compares the forecasting performance of a large number of models in the four energy markets considered in the present analysis,finding that GARCH family models perform better.Kumar and Narayan (2007)model oil price volatility adopting an exponential GARCH (EGARCH) to account for asymmetry and persistence of shocks.Hayo et al. (2012)investigate the impact of US monetary policy on the price volatility of energy, agricultural, livestock and metal commodities over the period 1998–2009. They find that unexpected interest rate changes do increase volatility, while expected ones decrease it.Creti et al. (2013)contribute to the emerging literature on the relationships between commodity and stock markets. Adopting a dynamic conditional correlation (DCC) GARCH model, they show that correlations between 25 commodities and stock markets over the period 2001–2011 are time-varying and highly volatile.
Vo (2009)instead models the volatility of oil as a stochastic volatility process which incorporates regime-switching. The Bayesian Markov Chain Monte Carlo procedure implemented provides evidence of regime-switching in the oil market.
The way speculators can influence markets is the object of a vast literature. In principle, the presence of speculators (i.e. agents that buy or sell an asset because its price is expected to change) is fundamental to the efficient operation and stability of markets. AsSmith (2009)
points out,“Speculation is not price manipulation, but is sometimes used to exploit efforts to manipulate prices by other means. In such
cases, it is the manipulation of prices that is objectionable, not specula-tion, per se” (p. 26). Thus, the role of speculators might be stabilizing or destabilising and understanding their behaviour and how it affects returns and volatility is extremely important. This issue has been debat-ed extensively in literature. On one side, some authors suggest that the participation of speculators, which are considered uninformed traders, lowers the quality of information in the futures market, and might have a destabilizing effect on prices, thus increasing volatility (Stein, 1987).Hart and Kreps (1986)show that, even in a general equilibrium with optimizing speculators, prices can be destabilized. On the other, speculators are supposed to bring efficiency to price predictions, lower-ing volatility. In particular,Friedman (1953)suggests that rational spec-ulation stabilizes prices,Powers (1970)shows that speculative activity of futures traders reduces the random component of price variation, whileCox (1976)suggests that speculation increases the information content of prices.
More recently, several authors have investigated the relationship be-tween volatility and speculative activity in commodity futures markets.
Brunetti et al. (2011)focus, among others, on crude oil and natural gas futures markets in the 2005–2009 period. They find that an unexpected positive shock to swap dealers positions is associated with a significant reduction of volatility in these two markets. Using data on 14 commod-ities from the Disaggregated Commitment of Traders (DCOT),Sanders and Irwin (2011)find that larger net long positions by swap dealers generally lead to lower volatility in a Granger sense. With respect to en-ergy markets, theyfind a significant negative impact of swap dealer net positions on implied volatility in natural gas and a positive and signi fi-cant impact on realized volatility in the crude oil market.Irwin and Sanders (2012)find little evidence that index positions influence vola-tility in 19 commodity futures markets using cross-sectional Fama– MacBeth regressions. Additionally, they provide time series tests for oil and natural gas,finding no casual links between volatility in these two markets and the positions of two large exchange traded funds (ETFs).Alquist and Gervais (2013)find that changes in financial firms' positions do not predict oil price changes, andAulerich et al. (2013)
find no impact of commodity index trader (CIT) positions on volatility in 12 agricultural commodities markets.
We model volatility in four energy markets and we contribute to this literature by including among the explanatory variables in the variance equation some measures of volatility.
Working's T index is the most widely adopted measure for speculation in the literature. It quantifies the excess of speculation relative to hedging based on position data provided by commitments of traders (COT) data from the U.S. Commodity Futures Trading Commission (CFTC). Recently,Till (2009),Sanders et al. (2010)and
Sanders and Irwin (2013)have shown that speculative positions in energy and agriculture U.S. futures markets are not excessive relative to hedging activity.
The market share of non-commercial traders on total open interest is used inBüyükşahin and Robe (2014)to show that the composition of traders in futures markets helps explain the linkages between equity and commodity returns. The authorsfind that hedge funds increase the equity-commodity return correlations, while swap dealers, index traders, commercial traders, etc., do not influence the correlations.
As net long positions of traders are concerned, some authors (Brunetti et al., 2011; Büyükşahin and Harris, 2011; Irwin and Sanders, 2012; Medlock and Jaffe, 2009) adopt the difference between long and short positions held by non-commercial traders. Others adopt this difference relative to total open interest (Hedegaard, 2011) or relative to open interest held by non-commercial traders (Brunnermeier et al., 2008; Sanders et al., 2010).1Net long positions are usually employed because
1
speculators go mostly long on futures contracts (they buy the risk of hedgers traders), hence this measure is considered a good proxy to detect speculative activity. Moreover, net long positions of speculators have increased in commodity markets after 2004, especially in the oil market (Irwin and Sanders, 2010; Khan, 2009; Medlock and Jaffe, 2009), leading to allegations that these positions have pushed prices up (Medlock and Jaffe, 2009).Panella et al. (2013)analyse Brent and WTI prices in the 2000–2010 period using four different neural network models and argue that the growth of assets under management of money managers both caused and not caused prices to rise.
Overall, previous researchfinds that financial speculation has a negative impact on price variability. Our work contributes to this literature in two respects. First, it uses different measures of speculation. Second, it adopts different specifications for the volatility of futures returns.
3. Data description
We collect data of futures prices for four energy commodities (light sweet crude oil, heating oil, gasoline and natural gas).2Daily (5 days) data on futures prices3for each commodity are obtained from Datastream for the period 2000–2014.4
The U.S. Commodity Futures Trading Commission publicly reports, at weekly frequency, data on positions of two groups of traders: commercials and non-commercials. To build measures of speculative activity, the two groups are generally treated in the literature as representing hedgers and speculators, respectively. While this is the only viable possibility to approximate speculative activity in these markets using publicly available data, it has to be acknowledged that this matching might be imprecise: commercials can take speculative positions, and non-commercials can hedge exposure to price risk. AsAlquist and Gervais (2013)suggest, recognizing the inherent imprecision of such distinction, it is nonetheless meaningful to adopt it for pragmatic reasons.Ederington and Lee (2002)show that, with re-spect to the heating oil futures market, it is accurate to treat non-commercials as speculators, while among non-commercials are many firms which appear to be speculators as well. Dewally et al. (2013)use CFTC's Large Trader Reporting System (LTRS) data for crude oil, heating oil and gasoline and conclude that many of the energy traders, classified as commercials, were actually specula-tors. This evidence would suggest that our measure, while being imprecise, would under-estimate the presence of speculators in these markets.
We measure speculation using three different indexes: Working's T index, the market share of non-commercial traders, and the ratio of net long non-commercial positions over total open interest.Working (1960)proposed the T index to measure the adequacy of speculative positions to balance the hedging positions. This index reflects the view that speculation can only be considered excessive relative the hedging activity in the market. It is calculated as the ratio of speculative positions to total hedgers positions:
1þHSSSþ HL if HS≥HL 1þHSSLþ HL if HSbHL 8 > < > : ð1Þ
where SS is the number of positions held by speculators who are short, SL is speculation long, HS is hedging short and HL is hedging
long. The calculation of the Working's T index crucially depends on the classification of the market operators between hedgers and speculators. While commercials might be considered as hedgers, and non-commercials as speculators, CFTC also provides data for “Non-Reportable” agents,5which are not classified into any of the two categories. It has been noted by several authors (Leuthold, 1983; Sanders et al., 2010) that this index suffers from the problem of how to classify the positions held by these subjects. Several rules to treat non-reportables are at hand. One could consider them as being all hedgers or all speculators. We follow an intermediate approach, assum-ing that 50% of them are speculators and 50% are hedgers.6
While being widely criticized, this index is still the benchmark to measure excess speculation in futures markets.Sanders et al. (2010)
consider nine agricultural markets and provide descriptive evidence that Working's T in the 1996–2008 period is within the historical ranges reported in the literature dating to the 1950s.Till (2009)finds no
exces-sive speculation in oil futures and options markets in the 2006–2009 pe-riod. We provide some descriptive statistics on this and the other measures of speculation considered in the present analysis (see more below). Moreover, we include these indexes in a GARCH specification to model the volatility present in those markets.
As proposed byBüyükşahin and Robe (2014), we compute the market share of non-commercial traders as the average of the long and short positions of all non-commercial (or speculators) traders on the total open interest in that market:
SLþ SS
2 OI ð2Þ
where OI is the total open interest. In treating non-reportable positions, we follow the same approach used for Working's T.
The last measure of speculation is the ratio of net long non-commercial positions over total open interest. As inHedegaard (2011), it is defined as:
SL−SS
OI ð3Þ
where non-reportable are treated as discussed above. This is a measure of the extent to which speculators are long or short in aggregate: if it is positive (negative), speculators go long (short) in futures markets. We adopt index (3) for two reasons. First, it is a “relative” measure, hence it is directly comparable with the other indexes. Second, it is highly correlated (0.92) with the measure of “speculative pressure” (seefootnote 1).
As for the measure of market liquidity we adopt the ratio of volume to open interest, using daily data, sourced from Datastream:7
VO
OI ð4Þ
To control for macroeconomic factors we follow, among others,
Chevallier (2009) andManera et al. (2013) and we collect daily (5 days) data on Moody's Aaa and Baa corporate bond yield, 3-month
2
All commodities are traded on the New York Mercantile Exchange.
3 We use the continuous futures price series, calculated by Thomson Financial. Those
se-ries start at the nearest contract month, which forms thefirst value for the continuous se-ries and switches over on 1st day of new trading month.
4
The detailed description of the variables is presented inTable A.1in theStatistical Appendix.
5 CFTC defines this category as follows: “The long and short open interest shown as Non
Reportable Positions is derived by subtracting total long and short Reportable Positions from the total open interest. Accordingly, for Non Reportable Positions the number of traders in-volved and the commercial/non-commercial classification of each trader are unknown.” (see http://www.cftc.gov/MarketReports/CommitmentsofTraders/ExplanatoryNotes/ index.htm).
6
The rule chosen, in absence of any precise information on non-reportables, is the most neutral one. However, we repeat our econometric exercise attributing non-reportables to commercial and non-commercial in a 30%-70% split, i.e. assuming that they are mostly speculators. Additionally, we attribute them entirely to either the speculators or the hedgers group. Results are unaffected and are available upon request.
7
Treasury bill and S&P 500 index, as these variables have been previously shown to possess forecast power in commodity markets. Data are for the period 01/01/2000–30/04/2014 from Federal Reserve Economic Data (FRED), provided by the Federal Reserve of St. Louis.8For all these series we consider weekly averages of the daily data, as the CFTC data are publicly available at this frequency.
Descriptive statistics of speculation indexes for weekly data are re-ported inTable 1. Since futures prices contain a unit root,9to obtain sta-tionarity we consider the return rit, which is defined as log(Pit/Pit− 1),
where Pitand Pit–1are the prices of commodity i at time t and t–1,
re-spectively. The same transformation is applied to the macroeconomic variables, while thefinancial speculation indexes are stationary in levels and are not transformed.
Thefirst panel ofTable 1shows that the commodity with the aver-age highest values of the share of non-commercial traders is natural gas (0.198), while the lowest mean value is that of crude oil (0.165). Working's T index, reported in the second panel, ranges, on average, from 1.064 (gasoline) to 1.182 (natural gas), while its maximum value is larger than 1.5 (natural gas). In the third panel ofTable 1we observe that net long positions of non-commercial traders are always positive, apart from natural gas. Although the Augmented Dickey Fuller test al-ways does not accept the null hypothesis of a unit root in the series, somefluctuations over time are present.Fig. 1reports the speculative indexes over the 2000–2014 period.
The share of non-commercial traders, reported in panel A, pre-sents the minimum values in the 2000–2004 interval. Subsequently, the series present larger average values. In 2008 we observe a spike in the gasoline market, which is shortly absorbed. From 2013 onwards all different markets show an increasing share of non-commercial positions. This trend however is recent and it is too early to understand whether we are observing a substantial change in market conditions or we are reverting to the mean. Panel B
displays the Working's T index, which has a similar behaviour over time. Again natural gas presents a spike in 2008\2009 of the specu-lative index and in the last months all indexes are increasing. Panel C reports instead the net long positions of non-commercial traders. We find that for heating oil and crude oil positions are on average in equilibrium (indeed the mean values observed are 0.056 and 0.051, respectively). Starting from 2003, net long positions in gasoline are almost always positive, while in natural gas positions are net short from 2007 onwards, which might suggest that markets are expecting a price reduction. In the last months also this index is increasing in all markets.
4. The econometric specification
After testing the stationarity of all the series, we estimate a model where the returns of each commodity i at time t depend on two sets of explanatory variables, namely macroeconomic and speculative factors:
rit¼ α0þ α1int ratetþ α2junk bond yieldtþ α3S&Ptþ α4speculationitþ
þα5market liquidityitþ εit
ð5Þ In Eq.(5)the macroeconomic factors are represented by the returns of 3-month Treasury bill (int_ratet), the junk bond premium
(junk_bond_yieldt), defined as the difference between Baa and Aaa
corporate bond yield, and the returns of S&P 500 index (S & Pt). The
speculation variable (speculationit) is represented in turn by the
share of non-commercials, the Working's T or the net long positions of non-commercial traders for the market i at time t. We also control for market liquidity (market_liquidityit). The estimation period for
the four markets spans from 2000:w1 to 2014:w17.
Wefirst estimate the model using Ordinary Least Squares (OLS) and test for autoregressive conditional heteroskedasticity (ARCH) effects in the residuals. As these effects are present, we move to a GARCH specification, including an autoregressive term of order p, AR(p), when necessary. We aim at testing if there is a role for spec-ulation, thus we consider speculation variables as exogenous regressors in the variance equation of the GARCH models. There-fore, we end up estimating a model where the conditional mean equation is:
rit¼ γ0þ γ1int ratetþ γ2junk bond yieldtþ γ3S&Ptþγ4rit−1þ εitð6:aÞ
with an AR(p) error term if the null hypothesis of absence of resid-ual autocorrelation is rejected by the data. The conditional variance is defined as: σ2 it¼ α0þ Xpi j¼1βjε 2 it− jþ Xqi j¼1ωjσ 2
it− jþ δspeculationitþ ϕmarket liquidityit
ð6:bÞ where the varianceσit2of the regression model's disturbances is a
linear function of lagged values of the squared regression distur-bances, of its past value and of measures of speculation, p defines the order of the ARCH term, and q of the GARCH term. Values for p and q are chosen depending on the outcome of residual tests (ARCH-LM test and correlogram on squared residuals).10Financial
8
From the Federal Reserve of Philadelphia we also have retrieved the Aruoba-Diebold-Scotti (ADS) index, which is a measure of real business condition (seehttp://www. philadelphiafed.org/research-and-data/real-time-center/business-conditions-indexfor further details).
9Fig. A.1in theStatistical Appendixreports the behaviour of future prices at daily
fre-quency (the highest frefre-quency available in data) over the time period considered. In each graph, the series show a non-stationary behaviour, as well as an evident spike in prices in 2008.
Table 1
Summary statistics for speculation measures.
Obs Mean Std. dev. Min Max Unit root
test Share non-commercial Gasoline 745 0.180 0.047 0.071 0.322 −3.948*** Heating oil 745 0.170 0.040 0.096 0.304 −3.506*** Natural gas 745 0.198 0.082 0.040 0.378 −3.817** Crude oil 745 0.165 0.030 0.078 0.226 −3.808*** Working's T Gasoline 745 1.064 0.027 1.021 1.177 −4.629*** Heating oil 745 1.099 0.041 1.025 1.270 −2.630* Natural gas 745 1.182 0.118 1.019 1.559 −3.300* Crude oil 745 1.108 0.035 1.034 1.212 −3.725***
Net long positions of non-commercial over open interest
Gasoline 745 0.174 0.097 −0.118 0.386 −4.568***
Heating oil 745 0.057 0.067 −0.133 0.250 −5.617***
Natural gas 745 −0.056 0.089 −0.255 0.155 −3.334*
Crude oil 745 0.051 0.076 −0.173 0.258 −5.188***
Notes: Column“Unit root test” reports the Augmented Dickey–Fuller statistic for the null hypothesis that there is a unit root in the series. *, ** and *** denote significance at 10%, 5% and 1% levels, respectively.
10
speculation is proxied by the market share of non-commercial traders, the Working's T index and the net long positions of non-commercials. Additionally, we control for market liquidity. In
Section 5.1we present the results at weekly frequency.
We choose not to consider a multivariate GARCH approach as finan-cial speculation is likely to be mostly relevant on the price volatility of its own market only. Indeed, a preliminary analysis of the conditional correlations shows that these are generally low and suggests to keep an univariate approach.11
5. Results 5.1. Main results
Table 2shows the results obtained when the market share of non-commercial traders is included in the variance equation.12
11
The constant conditional correlation matrix is available upon request.
12 We also estimate the models without controlling for market liquidity. The results on
thefinancial speculation variable are unaffected. We show one of these robustness checks inTable 6, column 8, while the full set of results is not reported but is available upon request.
We estimate the model using OLS and then test for ARCH effects using a standard Lagrange multiplier test (not reported). For all commodities reported inTable 2, this test suggests the presence of
ARCH effects in the residuals of the estimated model. Thus, we move to a GARCH(p,q) specification: p = q = 1 is the preferred lag order. Additionally, the Ljung–Box test (not reported) on the GARCH(p,q) model shows that the residuals contain autocorrelation up to order 1. Introducing an AR(1) term in the models generally removes autocorrelation, while for heating oil we include also an AR(2) term. The variance equation shows that the ARCH (β) and GARCH (ω) terms are always statistically significant. In particular, the ARCH estimates are generally small (between 0.068 for heating oil and 0.131 for natural gas) and the GARCH estimates are generally high and close to one (see for example 0.904 in the heating oil equation). This indicates that a shock in the volatility series impacts on futures volatility over a long horizon. The only exception
Table 4
Estimates of univariate GARCH models— Net long positions of non-commercial traders over open interest and market liquidity as exogenous variables in the variance equation.
Gasoline Heating oil
Natural gas
Crude oil
Mean equation Tbill 0.021*** 0.009* 0.014 0.017***
(0.007) (0.006) (0.010) (0.006)
S&P500 0.452*** 0.413*** 0.135 0.435***
(0.071) (0.065) (0.097) (0.065)
Junk bond yield −0.048 −0.043 −0.018 −0.054
(0.043) (0.037) (0.061) (0.037) AR(1) 0.180*** 0.197*** 0.208*** 0.202*** (0.038) (0.039) (0.038) (0.039) AR(2) −0.124*** (0.037) Constant 0.003* 0.001 0.001 0.002 (0.002) (0.001) (0.002) (0.002) Variance equation ARCH(1) 0.080*** 0.068*** 0.106*** 0.065*** (0.022) (0.016) (0.034) (0.025) GARCH(1) 0.897*** 0.904*** 0.820*** 0.847*** (0.030) (0.027) (0.058) (0.056) Net long positions/OI −0.008 0.004 0.026 −0.055** (0.017) (0.008) (0.050) (0.026) Market liquidity −0.018 −0.017 −0.041 0.011 (0.023) (0.012) (0.070) (0.015) Constant 0.0001 0.0001 0.0003 0.0001 (0.0001) (0.0001) (0.0002) (0.0001) ARCH + GARCH terms 0.977 0.972 0.926 0.912 Test ARCH LM (F-stat) 3.317 0.385 0.172 0.220 No. of obs. 743 742 743 743
Notes: Standard errors in parentheses. * significant at 10% level, ** significant at 5% level, and *** significant at 1% level.
Table 3
Estimates of univariate GARCH models— Working's T and market liquidity as exogenous variables in the variance equation.
Gasoline Heating oil Natural gas Crude oil
Mean equation Tbill 0.021 0.010* 0.012 0.014*
(0.007)*** (0.005) (0.011) (0.008)
S&P500 0.446*** 0.414*** 0.185* 0.460***
(0.071) (0.065) (0.101) (0.067)
Junk bond yield −0.048 −0.043 −0.063 −0.029
(0.042) (0.037) (0.064) (0.035) AR(1) 0.179*** 0.198*** 0.188*** 0.196*** (0.037) (0.039) (0.041) (0.042) AR(2) −0.123*** (0.037) Constant 0.003* 0.001 0.001 0.001 (0.002) (0.001) (0.002) (0.002) Variance equation ARCH(1) 0.078*** 0.068*** 0.134*** 0.098*** (0.022) (0.015) (0.030) (0.020) GARCH(1) 0.903*** 0.904*** 0.756*** 0.875*** (0.030) (0.027) (0.062) (0.028) Working's T −0.040 −0.003 −0.094** −0.048* (0.038) (0.015) (0.041) (0.026) Market liquidity −0.007 −0.016 0.016 0.010 (0.024) (0.012) (0.050) (0.008) Constant 0.001 0.0001 0.001** 0.001* (0.001) (0.0002) (0.001) (0.0003) ARCH + GARCH terms 0.981 0.972 0.89 0.973 Test ARCH LM (F-stat) 3.535 0.410 0.270 0.210 No. of obs. 743 742 743 743
Notes: Standard errors in parentheses. * significant at 10% level, ** significant at 5% level, and *** significant at 1% level.
Table 2
Estimates of univariate GARCH models— market share of non-commercial traders and market liquidity as exogenous variables in the variance equation.
Gasoline Heating oil Natural gas Crude oil Mean equation Tbill 0.021*** 0.010* 0.012 0.016** (0.007) (0.005) (0.010) (0.007) S&P500 0.419*** 0.413*** 0.203** 0.452*** (0.076) (0.065) (0.091) (0.069)
Junk bond yield −0.016 −0.043 −0.061 −0.041
(0.043) (0.037) (0.063) (0.036) AR(1) 0.186*** 0.198*** 0.188*** 0.196*** (0.038) (0.039) (0.041) (0.042) AR(2) −0.123*** (0.037) Constant 0.002 0.001 0.001 0.001 (0.002) (0.001) (0.002) (0.002) Variance equation ARCH(1) 0.087*** 0.068*** 0.131*** 0.088*** (0.017) (0.015) (0.030) (0.020) GARCH(1) 0.889*** 0.904*** 0.746*** 0.861*** (0.023) (0.027) (0.063) (0.032) Share non-commercial −0.052** −0.003 −0.021*** −0.161*** (0.026) (0.015) (0.077) (0.049) Market liquidity −0.002 −0.016 0.053 0.019** (0.016) (0.012) (0.048) (0.010) Constant 0.0001 0.0001 0.001** 0.0003*** (0.0001) (0.0001) (0.0003) (0.0001) ARCH + GARCH terms 0.976 0.972 0.877 0.949 Test ARCH LM (F-stat) 3.356 0.410 0.250 0.239 No. of obs. 743 742 743 743
Notes: Standard errors in parentheses. * significant at 10% level, ** significant at 5% level, and *** significant at 1% level.
Table 5
Speculation measures' coefficients estimated with different GARCH models.
Gasoline Heating
oil
Natural gas
Crude oil
GARCH Share non-commercial −0.052** −0.003 −0.021*** −0.161***
Working's T −0.040 −0.003 −0.094** −0.048*
Net long positions/OI −0.008 0.004 0.026 −0.055**
GARCH-M Share non-commercial −0.051** −0.006 −0.211*** −0.164***
Working's T −0.036 −0.004 −0.092** −0.045*
Net long positions/OI −0.009 0.004 0.016 −0.055**
TARCH Share non-commercial −0.056** −0.004 −0.212** −0.408***
Working's T −0.041 −0.004 −0.110** −0.062**
Net long positions/OI −0.007 0.005 0.007 −0.065**
EGARCH Share non-commercial −0.300* −0.096 −0.529** −0.950***
Working's T −0.600 −0.075 −0.307** −0.340**
Net long positions/OI −0.426 0.054 −0.041 −0.711***
is represented by natural gas, which shows lower GARCH estimates. In the mean equation the S&P 500 index is always positive and significant, suggesting that returns are pro-cyclical. The T-bill is generally significant and positive and junk bond yield is never significant.13The speculation index displays a negative and signi fi-cant coefficient in the variance equation in all markets but heating oil, indicating that an increase of the market share of non-commercials corresponds to a decrease, although small, in the vola-tility of commodity futures returns. This is in line with the rational speculation view (Friedman, 1953) and the strand of literature whichfinds that speculation has the stabilizing effect of smoothing the price discovery process (Aulerich et al., 2013; Brunetti et al., 2011; Irwin and Sanders, 2012; Sanders and Irwin, 2011). The control for market liquidity is generally not significant, apart in the crude oil market, where it is positive and significant at 5% level. This positive correlation between market liquidity and volatility is consistent with previous evidence (Du et al., 2011; Robles et al., 2009).
The second set of estimates introduces the Working's T index in the variance equation. Results are presented inTable 3. As concerns the macroeconomic variables in the mean equation and the GARCH specification, we obtain similar results. The Working's T index is generally negative, being significant in two markets out of four (natural gas and crude oil). This again supports the view
that larger presence of speculators is associated with reduced vola-tility of commodities futures prices, as predicted by the theory and supported by empirical evidence.
The last set of estimates, reported inTable 4, presents the percentage of net long positions of non-commercials. For the mean equation we get results similar to those reported inTables 2 and 3. In the variance equa-tion, the speculative measure is significant only in the crude oil market, whereas market liquidity is not significant.
To sum up, the evidence shows speculation measures gen-erally have a negative sign in the variance equation, suggesting that a larger presence of non-commercials in the markets smoothes the price process, in line with previous empirical evi-dence. We have a robust result in the crude oil market, where all different measures are significant. In the other markets the coefficient is not always significant, and in the heating oil we neverfind a significant coefficient for the financial speculation variables.
5.2. Robustness analysis
In order to analyse if the main results vary under different conditions, we focus on two types of robustness checks: we in-vestigate whether the results are unaffected adopting alternative GARCH models and we check if different controls in the mean equation impact somehow on the results obtained in the variance equation.
5.2.1. Econometric specification
We repeat the previous analysis adopting alternative GARCH models to see if the results are influenced by the type of models
13
We also estimate the model with the ADS index in the mean equation. It is generally poorly significant, therefore we prefer the specification reported.
Table 6
Focus on crude oil— Estimates of univariate GARCH models — market share of non-commercial traders and market liquidity as exogenous variables in the variance equation. Weekly— crude oil
(1) (2) (3) (4) (5) (6) (7) (8)
Mean equation Demand – −0.011 – – −0.008 – −0.004 −0.004
(0.023) (0.025) (0.023) (0.023) Production – – −0.005 – −0.006 – −0.008 −0.009 (0.018) (0.019) (0.018) (0.020) Ending stock – – – 0.009 0.011 – 0.005 0.005 (0.022) (0.019) (0.021) (0.022) Tbill – – – – – 0.016** 0.016** 0.015** (0.007) (0.007) (0.007) S&P500 – – – – – 0.452*** 0.452*** 0.467*** (0.069) (0.068) (0.067)
Junk bond yield – – – – – −0.041 −0.043 −0.040
(0.036) (0.037) (0.035)
AR(1) 0.185*** 0.185*** 0.184*** 0.184*** 0.184*** 0.196*** 0.196*** 0.197***
(0.041) (0.041) (0.041) (0.041) (0.041) (0.042) (0.042) (0.042)
Constant 0.002 0.020 0.005 −0.007 0.006 0.001 0.006 0.006
(0.002) (0.035) (0.010) (0.022) (0.042) (0.002) (0.040) (0.040)
Variance equation ARCH(1) 0.086*** 0.085*** 0.086*** 0.085*** 0.084*** 0.088*** 0.086*** 0.091***
(0.022) (0.022) (0.022) (0.022) (0.022) (0.020) (0.022) (0.019) GARCH(1) 0.832*** 0.834*** 0.833*** 0.834*** 0.836*** 0.861*** 0.832*** 0.865*** (0.045) (0.045) (0.045) (0.045) (0.044) (0.032) (0.045) (0.031) Share non-commercial −0.184*** −0.184*** −0.184*** −0.181*** −0.179*** −0.161*** −0.184*** −0.121*** (0.066) (0.065) (0.066) (0.067) (0.066) (0.049) (0.066) (0.046) Market liquidity 0.034** 0.034** 0.034** 0.034** 0.033** 0.019** 0.034** – (0.016) (0.016) (0.016) (0.016) (0.016) (0.010) (0.016) Constant 0.0003** 0.0003** 0.0003** 0.0003** 0.0003** 0.0003*** 0.0003** 0.0003*** (0.0001) (0.0001) (0.0001) (0.0001) (0.0001) (0.0001) (0.0001) (0.0001)
ARCH + GARCH terms 0.918 0.919 0.919 0.919 0.92 0.949 0.918 0.956
Test ARCH LM (F-stat) 0.002 0.0001 0.002 0.004 0.002 0.239 0.265 0.158
No. of obs. 743 743 743 743 743 743 743 743
employed. We estimate the GARCH-in-Mean (GARCH-M, see
Engle et al., 1987), which introduces the conditional variance or standard deviation into the mean equation, the threshold ARCH (TARCH, seeZakoïan, 1994), which allows the conditional standard deviation to depend upon the sign of the lagged inno-vations, and the asymmetric exponential GARCH (EGARCH, see
Nelson, 1991), which explicitly allows for asymmetries in the lationship between returns and volatility. We compare the re-sults of the three models with the GARCH model previously presented.
Table 5shows the results across different econometric spec-ifications.14
We can see that speculation variables across GARCH-M, TARCH and EGARCH have the same sign and signi fi-cance of the baseline GARCH model. Moreover, in the GARCH-M estimation, we have found that the conditional variance (or standard deviation) added in the mean equation is generally not significant. This means that the estimated coefficient on the expected risk (the risk premium) has no influence on ex-pected returns of commodities investments, i.e. there is no feed-back from the variance to the mean. The asymmetric EGARCH model obtains larger coefficients but generally gets to the same results. Finally, the asymmetric models, TARCH and EGARCH, do not display significant asymmetric effects on condi-tional variance. Overall, we might say that the leverage effect does not seem to be present.
5.2.2. Focus on crude oil
The macroeconomic controls we have used in our analysis might be not sufficient to model the economic cycle. Hamilton (2009), for example, suggests that economic fundamentals such as demand, supply and storage are more relevant in explaining crude oil returns. Thus, we focus on crude oil15and verify how results on speculative indexes change when the mean equation is other-wise specified.
Table 6presents eight different specifications. The dependent variable in each equation is the crude oil return. The first model presents a simple specification, which includes only an AR(1) term in the mean equation. Then, from columns 2 to 5, we include some controls specific for the oil market, i.e. data on de-mand, production and stocks, which are however not significant. Column 6 corresponds to the crude oil equation inTable 2, while column 7 enriches this specification by adding the oil-specific controls. Across all these different specifications of the mean equation, wefind that the share of non-commercials in the vari-ance equation displays a negative and significant coefficient, whose magnitude is rather stable. The coefficient estimated for market liquidity is also robust, close to zero and significant at 5% level, independently from the choice of variables in the mean equation.
Finally, in column 8 we propose a different specification for the var-iance equation, omitting the market liquidity control. The coefficient at-tached to the share of non-commercial agents is still negative and significant at 1% level.
We can conclude that the results presented inSection 5.1are invariant to changes in the econometric specification and that the choice of controls in the mean equation does not affect the
results in the variance equation, which is the focus of our analysis. Moreover, the coefficient attached to the speculation variable is not influenced by changes in the specification of the variance equation.
6. Conclusions
This paper considers alternative measures of speculative activity to evaluate if there is a role forfinancial speculation in modelling the volatility of commodity futures prices. We test this relationship using data for futures prices in four energy markets (crude oil, heating oil, gasoline and natural gas) over the period 2000–2014 at weekly frequency. Speculation can be proxied by the market share of non-commercial traders, the Working's T index and the percent-age of net long positions held by non-commercials over total open interest.
Our work brings fresh evidence in the literature under different respects. First, we consider different measures of speculation: we employ the most frequently adopted Working's T index, but also the market share of non-commercial traders and the net long posi-tions of non-commercial. Second, we analyze if these different mea-sures of speculation are significant in modelling volatility. Third, we run a robustness exercise to check if the main results are invariant to changes in the econometric specification and control variables in the mean equation.
In the econometric analysis the commodity returns are modelled according to a GARCH(p,q) with an AR(1) term. Our estimation results suggest that, among macroeconomic factors, S&P 500 index and T-bill are generally positive and significant and are relevant controls to explain commodity futures returns. Speculation indexes are included as exogenous variables in the variance equation of the different models. Wefind that the three speculation indexes have a negative coefficient (when significant), confirming evidence that speculation does not destabilize prices (see, among others,
Aulerich et al., 2013; Brunetti et al., 2011; Irwin and Sanders, 2012; Sanders and Irwin, 2011). We have a robust result in the crude oil market, where all different measures are statistically significant. In the other markets thefinding is less robust, while in the heating oil we neverfind a significant coefficient for the financial speculation variables.
We evaluate if and how the main results change moving along several dimensions. In particular, the main results in the variance equa-tion remain unchanged across different econometric models (such as GARCH-M, TARCH and asymmetric EGARCH). Moreover, if we change the specification in the mean equation or in the variance equation, the results concerning thefinancial speculation variables in the variance equation are unaffected.
Statistical Appendix
14The complete set of estimation on weekly data of GARCH-M, TAGARCH and EGARCH
models is available upon request.
15
The focus on oil is motivated by the availability and the frequency of the oil data, which are not generally matched by other commodities.
Table A.1
Variables description.
Commodity Start date End date Mnemonic datastream
Energy Gasoline 01/04/2000 04/30/2014 NHUCS00 and NRBCS00
Heating oil 01/04/2000 04/30/2014 NHOCS00
Natural gas 01/04/2000 04/30/2014 NNGCS00
Crude oil 01/04/2000 04/30/2014 NCLCS00
References
Alquist, Ron, Gervais, Olivier, 2013.The role offinancial speculation in driving the price of crude oil. Energy J. 34 (3), 35–54.
Aulerich, Nicole M., Irwin, Scott H., Garcia, Philip, 2013.Bubbles, food prices and speculation: evidence from the CFTC's daily large trade datafiles. In: Chavas, Hummels, Wright (Eds.), Economics of Food Price Volatility. University of Chicago Press, Chicago, IL.
Babcock, Bruce A., 2012.The impact of US biofuel policies on agricultural price levels and volatility. China Agric. Econ. Rev. 4 (4), 407–426.
Brooks, Chris, Burke, Simon P., 2003.Information criteria for GARCH model selection. Eur. J. Finance 9 (6), 557–580.
Brümmer, Bernhard, Korn, Olaf, Shlüßler, Kristina, Jaghdani, Tinoush Jamali, Saucedo, Alberto, 2013.Volatility in the after crisis period— a literature review of recent em-pirical research. ULYSSES Working Paper No 1.
Brunetti, Celso, Büyükşahin, Bahattin, Harris, Jeffrey H., 2011. Speculators, prices and mar-ket volatility, mimeo (available at SSRN:http://ssrn.com/abstract=1736737). Brunnermeier, Markus K., Nagel, Stefan, Pedersen, Lasse H., 2008.Carry trades and
currency crashes. Macroeconomics Annual 2008, 23. NBER (National Bureau of Eco-nomic Research), pp. 313–347 (Cambridge, MA).
Büyükşahin, Bahattin, Harris, Jeffrey H., 2011.Do speculators drive crude oil futures prices? Energy J. 32 (2), 167–202.
Büyükşahin, Bahattin, Robe, Michel A., 2014.Speculators, commodities and cross-market linkages. J. Int. Money Financ. 42, 38–70.
Chevallier, Julien, 2009.Carbon futures and macroeconomic risk factors: a view from the EU ETS. Energy Econ. 31, 614–625.
Cox, Charles C., 1976.Futures trading and market information. J. Polit. Econ. 84 (6), 1215–1237.
Creti, Anna, Joëts, Marc, Mignon, Valérie, 2013.On the links between stock and commod-ity markets' volatilcommod-ity. Energy Econ. 37, 16–28.
De Roon, Frans A., Nijman, Theo E., Veld, Chris, 2000.Hedging pressure effects in futures markets. J. Financ. 55 (3), 1437–1456.
Dewally, Michaël, Ederington, Louis, Fernando, Chitru, 26 (10), 2013.Determinants of trader profits in commodity futures markets. Rev. Financ. Stud. 2648–2683. Du, Xiaodong, Yu, Cindy L., Hayes, Dermot J., 2011.Speculation and volatility spillover in
the crude oil and agricultural commodity markets: a Bayesian analysis. Energy Econ. 33, 497–503.
Ederington, Louis, Lee, Jae Ha, 2002.Who trades futures and how: evidence from the heating oil futures markets. J. Bus. 75 (2), 353–373.
Engle, Robert F., Lilien, David M., Robins, Russell P., 1987.Estimating time varying risk premia in the term structure: the Arch-M model. Econometrica 55 (2), 391–407. Friedman, Milton, 1953.The case forflexible exchange rates. Essays in Positive
Econom-ics. University of Chicago Press, Chicago, pp. 157–203.
Hamilton, James D., 2009.Understanding crude oil prices. Energy J. 30, 179–206. Hart, Oliver D., Kreps, David M., 1986.Price destabilizing speculation. J. Polit. Econ. 94 (5),
927–952.
Hayo, Bernd, Kutan, Ali M., Neuenkirch, Matthias, 2012.Communication matters: US monetary policy and commodity price volatility. Econ. Lett. 117 (1), 247–249. Hedegaard, Esben, 2011.How margins are set and affect asset prices. Job Market Paper. Hughes, Anthony W., King, Maxwell L., Kwek, Kian T., 2004.Selecting the order of an
ARCH model. Econ. Lett. 83 (2), 269–275.
Irwin, Scott H., Sanders, Dwight R., 2010. Thefinancialization of commodity futures markets or: How I learned to stop worrying and love the index funds. Working Papers Series (available at:http://ssrn.com/abstract=1699793).
Irwin, Scott H., Sanders, Dwight R., 2012.Testing the Masters hypothesis in commodity futures markets. Energy Econ. 34, 256–269.
Khan, Mohsin S., 2009.The 2008 oil price“bubble”. Peterson Institute for International Economics (Policy Brief 09-19).
Kumar, Narayan Paresh, Narayan, Seema, 2007.Modelling oil price volatility. Energy Pol-icy 35, 6549–6553.
Leuthold, Raymond M., 1983.Commercial use and speculative measures of the livestock commodity futures markets. J. Futur. Mark. 3 (2), 113–135.
Manera, Matteo, Nicolini, Marcella, Vignati, Ilaria, 2013.Financial speculation in energy and agriculture futures markets: a multivariate GARCH approach. Energy J. 34 (3), 55–81.
Masters, Michael W., 2008. Testimony of Michael W. Masters before the Committee on Homeland Security and Governmental Affairs United States Senate. (May 2008. http://hsgac.senate.gov/public/_files/052008Masters.pdf).
Masters, Michael W., 2009.Testimony of Michael W. Masters before the Commodity Futures Trading Commission (August 2009).
Medlock III, Kenneth B., Jaffe, Amy M., 2009.Who is in the oil futures market and how has it changed? James A. Baker III Institute for Public Policy, Rice University.
Mitchell, Heather, McKenzie, Michael, 2003.GARCH models selection criteria. Quant. Finance 3 (4), 262–284.
Nelson, Daniel B., 1991.Conditional heteroskedasticity in asset returns: a new approach. Econometrica 59 (2), 347–370.
Nissanke, Machino, 2012.Commodity market linkages in the globalfinancial crisis: excess volatility and development impacts. J. Dev. Stud. 48 (6), 732–750.
Panella, Massimo, D'Ecclesia, Rita L., Stack, David G., Barcellona, Francesco, 2013.Crude oil prices and kernel-based models. Int. J. Financ. Eng. Risk Manage. 1 (3), 214–238. Pindyck, Robert S., 2004.Volatility and commodity price dynamics. J. Futur. Mark. 24,
1029–1047.
Powers, Mark J., 1970.Does futures trading reduce pricefluctuations in the cash markets? Am. Econ. Rev. 60 (3), 460–464.
Regnier, Eva, 2007.Oil and energy price volatility. Energy Econ. 29, 405–427. Robles, Miguel, Torero, Maximo, von Braun, Joachim, 2009.When speculation matters.
International Food Policy Research Institute (Issue Brief 57).
Sadorsky, Perry, 2006.Modelling and forecasting petroleum futures volatility. Energy Econ. 28, 467–488.
Sanders, Dwight R., Irwin, Scott H., 2011.The impact of index funds in commodity futures markets: a systems approach. J. Altern. Invest. 14, 40–49.
0
1
2
3
4
01jan2000 01jan2005 01jan2010 30apr2014
Gasoline 0 1 2 3 4
01jan2000 01jan2005 01jan2010 30apr2014
Heating oil 0 5 10 15 Natural gas 0 50 100 150
01jan2000 01jan2005 01jan2010 30apr2014
Crude Oil
01jan2000 01jan2005 01jan2010 30apr2014
Sanders, Dwight R., Irwin, Scott H., 2013.Measuring index investment in commodity futures markets. Energy J. 34 (3), 105–127.
Sanders, Dwight R., Boris, Keith, Manfredo, Mark, 2004.Hedgers, funds, and small specu-lators in the energy futures markets: an analysis of the CFTC's Commitments of Traders reports. Energy Econ. 26, 425–445.
Sanders, Dwight R., Irwin, Scott H., Merrin, Robert P., 2010.The adequacy of speculation in agricultural futures markets: Too much of a good thing? Appl. Econ. Perspect. Policy 32 (1), 77–94.
Smith, James L., 2009.World oil: Market or mayhem? J. Econ. Perspect. 23 (3), 145–164.
Stein, Jeremy C., 1987.Informational externalities and welfare-reducing speculation. J. Polit. Econ. 95 (6), 1123–1145.
Till, Hilary, 2009.Has there been excessive speculation in the US oil futures markets? Working Paper. EDHEC— Risk Institute.
Vo, Minh T., 2009.Regime-switching stochastic volatility: evidence from the crude oil market. Energy Econ. 31, 779–788.
Working, Holbrook, 1960.Speculation on hedging markets. Food Res. Inst. Stud. 1, 185–220. Zakoïan, Jean-Michel, 1994.Threshold heteroskedastic models. J. Econ. Dyn. Control. 18,
