The COVID-19 pandemic and speculation in energy, precious metals, and agricultural futures

Abstract

We report new evidence that speculation in energy and precious metal futures are more prevalent in crisis periods and even more so during the COVID-19 pandemic. In contrast, agricultural futures attract more hedging pressure. Post-GFC patterns mirror the 1980s’ recessions. Using quantile regression on a long-horizon sample we also find that speculative pressure generally coincides with abnormal returns in normal circumstances but not in the current pandemic. Instead, volatility is strongly and often non-linearly associated with speculation across instruments.

Graphical abstract

1. Introduction

We contribute concurrently to behavioral finance, crisis, and futures market literature by unraveling new insights on the divergent magnitude of speculative intent of futures market participants for 10 major instruments. We connect the speculative trading metric to profitability measures and qualify their associations through market states represented by liquidity and volatility. A crucial practical merit of this exercise lies in disentangling similarities and dissimilarities of the aforesaid phenomena for the ongoing COVID-19 pandemic and previous crises.

Onset of the COVID-19 pandemic jolted the commodity futures markets through three channels: demand, supply, and heightened uncertainty (Baldwin and di Mauro, 2020). Different commodity classes incurred varied levels of shocks due to fundamental exposure and investor sentiments. For instance, oil plummeted steeper than other commodities as demand depressed globally, dipping to the lowest since 1995. On the supply side, a glut ensued due to an OPEC+ political impasse. Precious metal futures reached all-time or decade-high peaks. Some agricultural (rice and soybean) futures soared while others were steadier (e.g., coffee and corn), arguably due to consistent real demand, weakening dollar, and lower edible oil production. These pandemic-induced economic shocks distend a long-standing conundrum concerning futures. The original purpose of futures markets was to allow economic agents insure against price risk. Uncertainties over the when, how, and shape of the economic recovery imperiled both consumers and producers of commodities—resulting in extreme volatility. Meanwhile, speculation, both hailed as a liquidity conduit and excoriated as a de-stabilizer, is known to spike in volatile times. If it indeed occurs in excess during crises and leads to alpha, real economic stakeholders may become more vulnerable. Furthermore, since some commodities have shown crisis-time resilience, many investors treat them as safe havens. It is controversial whether such flight-to-safety counts as speculation. Technicality aside, for those with economic exposure to the underlying, the net effect is the same as speculation. Thus, if non-hedging futures trades concentrate in crises and lead to profitability, its verification and quantification merit scrutiny (Aziz et al., 2020, Baur and Lucey, 2010). Contributing to this void is this paper’s focus. Specifically, for 10 futures representing energy, precious metals, and agriculture, we pursue answers to three questions: (a) do crisis periods attract more speculation? and (b) is speculation more profitable in crisis periods? (c) what role does the market state play in deciding the speculation–profitability​ association?

Our questions have consequences for the broader commodity financialization debate. It is widely acknowledged that the rise in commodity financialization has led to its acceptance as a core asset class (Adams and Glück, 2015, Cheng and Xiong, 2014, Henderson et al., 2015). This paradigm shift is an inexorable by-product of exploitation of commodities’ erstwhile low correlation with stocks and reliable alpha generation (Bessembinder, 1992, Gorton and Geert Rouwenhorst, 2006, Silvennoinen and Thorp, 2013, Zhang et al., 2017, Aït-Youcef, 2019, Algieri and Leccadito, 2020). Though financialization has its merits (Arezki et al., 2014), it can be both a cause and an effect of noise and speculation (Hamilton and Wu, 2014). Particularly, futures are identified to facilitate the activity of uninformed speculation (Kurov, 2008, Pindyck and Rotemberg, 1990). This is because futures traders likely possess dissimilar investment incentives from traditional investors (Domanski and Heath, 2007, Tang and Xiong, 2012). They are also far more reliant on leverage. In fact, post-GFC studies consider speculation a foremost catalyst of volatility for its interference with commodities’ fundamental tether to supply, demand, and production (Robles et al., 2009). Given that enough evidence exists suggesting that crisis periods invite market participation disconnected from fundamentals (Davidson, 2008, Johnson, 1960, Parsons, 2010), we set out to investigate whether the same persists during the COVID-19 pandemic. One novelty of our paper is that we formalize test this hypothesis across multiple recessions as identified by NBER. Additionally, the lack of conclusive empirical findings on the role of futures in price stability perpetuates the perception of futures as a catalyst for market uncertainty (Cox, 1976, Hou and Li, 2014, Kim, 2015). To account for this, we qualify our speculation–profitability results with volatility and illiquidity covariates, two features with significant track-record of impacting on intra-futures market dynamics.

We proceed by detailing our selection of variables and instruments, followed by detailed methodology, and results and analyses. The final section concludes with a recap of salient findings and recommendation for future research.

2. Estimation

2.1. Variable choice

We operationalize two core variables: speculation and profitability and treat all trading activity devoid of hedging intent as speculation. Conceptually, we adopt Garcia et al.’s (1986) specification that open interest information, linked with aggregate volume, suggests the proportion of speculative and hedging trades. This Speculation Ratio (SR) builds on finance literature that open interest implies hedging intent, while total volume reflects speculation Bessembinder and Seguin (1993). To measure profitability, we apply the capital asset pricing model with MSCI World Index as a benchmark.

Recent research shows that speculation is concurrently influenced by the state of liquidity and volatility in the market. For instance, Ludwig (2019) demonstrates that speculative activity exhausts liquidity temporarily but replenishes over the long run. Market liquidity itself implicates the appeal of a particular instrument to speculators. Alpha seekers can thus potentially trigger a feedback effect. To capture liquidity, we employ a well-known measure: Amihud (2002) Illiquidity. Meanwhile, a mature stream of literature postulates speculation to be both an agent and a result of volatility. For agro-commodities, volatility, and liquidity are consequential. Traditional theory of speculations posits that speculation does not lead to any price changes but instead reduces market volatility and illiquidity (Brunetti et al., 2013). Conversely, non-traditional theory suggests that increased participation of traditional speculators in futures markets is deleterious. For instance, Hamilton (2009) shows that fluctuations in oil markets are driven by speculation. Non-oil commodity volatilities show the same (Tang and Xiong, 2012). Results for precious metals are mixed. To capture price fluctuations, we use Garman–Klass Volatility. Recent papers are recognizing that popular volatility measures (e.g., GARCH) are noisy and disregard information encoded in intraday metrics. Horse-racing range-based (OHLC) estimators, Molnár (2012) shows Garman–Klass’s estimator is the most efficient. This volatility is often aggravated by external uncertainties Bakas and Triantafyllou (2018), triggering liquidity risk for a group of traders who fail to meet margin calls. In fact, a high correlation between volatility and liquidity is well-documented Ramos and Righi (2020). With these justifications for our choice variables, we outline their formulae, range, and sources in Table 1.

Table 1.

Variables and Instruments.

Panel A: Variable Specification
Variable	Quantification	Remarks
Speculation Ratio	$SR=\frac{TradeVolume}{OpenInterest}$	Garcia et al. (1986)
Amihud Illiquidity Measure	$ILLI{Q}_t^f=\frac{1}{day{s}_t^f}{\sum }_{d=1}^{days}\frac{\left|ln\frac{p_t}{p_{t-1}}\right|}{Volume}$	Amihud (2002)
Volatility	${\sigma }^2=\frac{1}{n}({\sum }_{i=1}^n\frac{1}{2}{ln}^2\left(\frac{H_i}{L_i}\right)+2Ln\left(2\right)-1){ln}^2\left(\frac{C_i}{O_i}\right)$	Garman and Klass (1980)
Abnormal Return	${\mu }_{it}=R_{it}-[{\alpha }_i-{\beta }_i\left(R_{mt}\right)]$

Panel B: Instruments of Study
Instrument	Type	Unit of Trading	Period studied
Coffee	Index	Cents and hundredths of a cent	Dec 1979–Dec 2020
Corn	Index	U.S. cents per bushel	Jan 1978–Dec 2020
Crude Oil	Index	$0.01 per barrel	Mar 1930–Dec 2020
RBOB Gasoline	Index	U.S. dollars and cents per gallon	Oct 2005–Dec 2020
Gold	Index	$0.10 per troy ounce	Oct 1979–Dec 2020
Natural Gas	Index	U.S. dollars and cents per MMBtu	Apr 1990–Dec 2020
Platinum	Index	U.S. dollars and cents per troy ounce	May 1978–Dec 2020
Rough Rice	Index	Cents per hundredweight	Aug 1986–Dec 2020
Silver	Index	U.S. dollars and cents per troy ounce	Mar 1978–Dec 2020
Soybean	Index	U.S. cents per bushel	Nov 1977–Dec 2020

2.2. Methodology

We compare the statistical significance of the difference between the mean intra-instrument SR for various crisis periods using Welch’s version of unequal variance t-test. The second and computationally intensive part of our investigation relies on a quantile regression model estimated via vine copula. The quantile approach taken in this paper is a continuation of a growing trend in financial and economics literature that tries to predict the quantiles of a regressand conditioned on covariates of interest. A seminal financial application of this is Tobias and Brunnermeier (2016) conditional value at risk (CoVaR), which has become the industry standard for measuring systemic risk. These methods are derivations of Koenker and Bassett’s (1978) linear quantile regression which has undergone a multitude of modifications and extensions since the new millennium. Some extensions include additive models, single-index extensions, distribution-free nonparametric approaches, etc. What is more, Kraus and Czado point out that, in a machine learning context, Support Vector Machine (SVM) and Random Forest (RF) models have employed the quantile approach. Our rationale for choosing Kraus and Czado’s version of quantile regression is that it is unbounded by the strict assumptions imposed on the shape of regression quantiles by the classical quantile regression and its many derivatives. Indeed, Bernard and Czado (2015) demonstrate that if either of the regressor or the regressand deviates from a normal distribution, the model suffers from serious misspecifications. Other problems include quantile crossing, multicollinearity, lack of significance of regressors, difficulty in inferring interactions when variables are transformed, etc. By contrast, the D-Vine based quantile regression employed by us avoids imposing strict assumptions on the shape of conditional quantiles. Rather, we use a nonparametric subclass of vine copulas, which enables flexible modeling of marginals, dependence, and prediction of conditional quantiles.

For an existing estimate of joint density, we can express the conditional distribution function as:

$$q_{\alpha }\left(x_1,\dots ,x_d\right)≔F_{Y\mid X_1,\dots ,x_d}^{-1}\left(\alpha \mid x_1,\dots ,x_d\right)$$ (1)

Applying probability integral transforms with their respective values we get:

$$F_{Y\mid X_1,\dots ,X_d}\left(y\mid x_1,\dots ,x_d\right)=P\left(Y\le y\mid X_1=x_1,\dots ,X_d=x_d\right)=P\left(F_Y\left(Y\right)\le v\mid F_1\left(X_1\right)=u_1,\dots ,F_d\left(X_d\right)=u_d\right)=C_{V\mid U_1,\dots ,U_d}\left(v\mid u_1,\dots ,u_d\right)$$ (2)

Inverting Eq. (2) we get:

$$q_{\alpha }\left(x_1,\dots ,x_d\right)≔F_{Y\mid X_1,\dots ,x_d}^{-1}\left(\alpha \mid x_1,\dots ,x_d\right)$$ (3)

As per Kraus and Czado (2017), we are able to express the cdf through the inverse marginal distribution function of our dependent variable – speculation ratio – and the conditional copula quantile function of our covariates. At this stage, once we compute the marginals and the copula by inputting them jointly into Eq. (3), we arrive at:

$${\hat{q}}_{\alpha }\left(x_1,\dots ,x_d\right)≔{\hat{F}}_Y^{-1}\left({\hat{C}}_{V\mid U_1,\dots ,U_d}^{-1}\left(\alpha \mid {\hat{u}}_1,\dots ,{\hat{u}}_d\right)\right)$$ (4)

In modeling the marginals, we adopt a nonparametric approach, heeding Noh et al. (2013) caution that model misspecification can occur if the model and the marginals are fully parametric. At this stage, a continuous kernel smoother is used:

$$\hat{F}\left(x\right)=\frac{1}{n}\sum _{i=1}^{n}K\left(\frac{x-x^{\left(i\right)}}{h}\right),x\in \mathbb{R}$$ (5)

Thereafter, ${\hat{C}}^{-1}$ V $|$ U ₁, $\dots$, U ${}_{d=}$ ($\alpha |{\hat{u}}_1$, $\dots$, ${\hat{u}}_d$), and k(.) is a symmetric probability density function, and h is a positive valued parameter for bandwidth. Using the technique shown by Duong (2015) and Nagler and Kraus (2017), we obtain estimates of the marginal distribution functions and use them to transform the dependent and independent variables to pseudo-copula data. We define the conditional log-likelihood of the D-vine copula with I ordering to estimate pair copula families and respective copula parameters based on pseudo copula data as follows:

$${\hat{C}}_{V\mid U_1,\dots ,U_d}^{-1}\left(\alpha \mid {\hat{u}}_1,\dots ,{\hat{u}}_d\right)$$ (6)

$$cll(\mathit{l},\hat{\mathcal{F}},\hat{\mathit{\theta }};\hat{\mathbf{v}},\hat{\mathcal{U}})≔\sum _{i=1}^{n}ln{c}_{V\mid \mathbf{U}}\left({\hat{v}}^{\left(i\right)}\mid {\hat{\mathbf{u}}}^{\left(i\right)};\mathit{l},\hat{\mathcal{F}},\hat{\mathit{\theta }}\right)$$ (7)

From Eq. (6) we proceed to estimate our D-Vine regression model by serially constructing pair-copulas and maximizing the conditional log likelihood at each step. Our total number of steps here is six since in addition to abnormal returns, volatility, and illiquidity, we also consider their surprise values extracted from the residuals of an AR(1) model. The D-Vine algorithm of Kraus and Czado (2017) considers all six covariates at once through an automatic covariate selection procedure and stops when maximizing the conditional log-likelihood is no longer possible. One key advantage of this method is that the most impactful variables are automatically selected and ranked by the algorithm according to their strength in predicting the regressand (speculation). As such, redundant variables are discarded, and a parsimonious picture of the dependence structure between the regressor and regressand are revealed. To check the validity of our results, we check them against traditional quantile regression proposed by Koenker and Bassett (1978).

3. Results & Analyses

To conserve space, we attach descriptive statistics and unit root testing of our variables in a supplementary file and turn our attention to the meat of our findings. First, Table 2 reports results from Welch’s unequal variance t-test to determine if SR across subsamples differs. We compute this for the whole sample as well as all periods identified as a recession by NBER.

Table 2.

Unequal variance Welch t-test results.

Overall, gasoline, crude oil, and gold emerge as the most active speculation magnets. Soybean’s whole sample mean is in the same vicinity. However, soybean’s metrics are inflated by higher values in the 1980s and 1990s. Remarkably, SR during the COVID-19 pandemic is both lower than historical average and other recession periods. A similar trend is visible for Coffee, where speculation shrank in recent decades. Reverting to Crude Oil, markedly higher SR values during the global financial crisis and the COVID-19 pandemic boost the credibility of increased financialization hypotheses. Although of similarly high magnitude, the same interpretation is not extensible to gasoline since RBOB futures debuted only in October 2005. Contrarily, we find agricultural futures far less prone to speculation. Especially for COVID-19, our results are supported by earlier results on government intervention as a significant food-price stabilizer (Glauber et al., 2020, Schmidhuber et al., 2020).

Meanwhile, isolated analyses of crisis (recession periods) reveal a clear pattern. Speculation in all energy futures, gold, and silver is not only higher during recessions than normal times, it leaped substantially in the COVID-19 pandemic compared to preceding recessions. The precious metal results match with standard financial theory. Platinum, interestingly, does not conform to this trend, which is likely suggestive of its relatively diminutive appeal to investors as a safe haven. Speculation in food staples like rice and corn register modest SR values all throughout. In a nutshell, across all samples, the recessions in the 1980s – the earliest days of the futures’ public trading – attracted more speculation. Hedging demand appears to have risen since then, barring the recent COVID-19 pandemic, except for the agriculture futures.

We now proceed to answering our second question of whether speculative trading coincides with alpha generation via D-Vine quantile regression. The model specifications and estimations for three quantiles (0.1, 0.5, and 0.9) are presented in Table 3, and supplementary files SFig1 and SFig2 report the plots from marginal effects of the covariates. In both tables and plots the exclusion of an independent variable suggests that it no longer contributes to reducing the conditional log likelihood of the model. In other words, the chosen covariate(s) subsume the marginal effects of omitted variables. For agricultural futures, only Soybean registers a near-unity relationship between speculation and abnormal returns. Predictive quantile coefficients further reveal that this relationship is marginally mediated by the liquidity and volatility states of the market. The relationships are largely linear, and the quantile-specific variation is low. Much of these results persist during the COVID-19 pandemic, with the exception of coffee, where significant and non-linear associations of SR are detected against returns and volatility, volatility surprise, and unexpected liquidity vacuum. Absence of significant results for corn and rice imply a strong influence of hedging in these commodities. We also discover a near-ubiquitous significance of contemporaneous volatility and insignificance of alpha for the pandemic sample, suggesting that speculative trading causes or is caused by extreme market volatility amid pandemic uncertainties but rarely equates to abnormal returns. Moreover, non-linearity is more pronounced in COVID-19 circumstances, with high abnormal returns attracting very high speculative activities and vice versa. This is the case regardless of commodity class.

Table 3.

Quantile regression results.

Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Coffee	SR—Garcia	5.31	494.85	−979.08	960.76	NA	0.10	0.50	0.90
	Volatility	0.00	37.43	−74.86	−74.86	0.00	0.81	0.84	0.89
	Abnormal Return	0.00	17.81	−35.62	−35.62	0.00	0.83	0.86	0.89
	Illiquidity Surprise	0.00	10.14	−20.29	−20.29	0.00	0.81	0.84	0.89
	Volatility Surprise	0.00	4.47	−8.94	−8.94	0.00	0.82	0.85	0.89
	AR Surprise	X	X	X	X	X	0.81	0.84	0.89
	Illiquidity	X	X	X	X	X	0.81	0.84	0.90
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Corn	SR—Garcia	2.40	549.75	−1094.71	−1086.42	NA	0.10	0.50	0.90
	Volatility	0.00	56.76	−113.51	−113.51	0.00	0.87	0.89	0.92
	Abnormal Return	X	X	X	X	X	0.86	0.88	0.91
	Illiquidity Surprise	X	X	X	X	X	0.83	0.86	0.89
	Volatility Surprise	X	X	X	X	X	0.84	0.87	0.90
	AR Surprise	X	X	X	X	X	0.85	0.88	0.91
	Illiquidity	X	X	X	X	X	0.88	0.90	0.93

Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Gasoline	SR—Garcia	2.56	576.81	−1148.51	−1139.67	NA	0.10	0.50	0.90
	Volatility	X	X	X	X	X	0.92	0.94	0.97
	Abnormal Return	0.00	31.95	−63.90	−63.90	0.00	0.91	0.94	0.96
	Illiquidity Surprise	X	X	X	X	X	0.91	0.94	0.97
	Volatility Surprise	X	X	X	X	X	0.91	0.94	0.96
	AR Surprise	X	X	X	X	X	0.91	0.94	0.96
	Illiquidity	X	X	X	X	X	0.91	0.94	0.97
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Gold	SR—Garcia	2.47	523.64	−1042.33	−1033.78	NA	0.10	0.50	0.90
	Volatility	0.00	55.20	−110.39	110.39	0.00	0.92	0.95	0.97
	Abnormal Return	X	X	X	X	X	0.92	0.94	0.97
	Illiquidity Surprise	X	X	X	X	X	0.91	0.94	0.97
	Volatility Surprise	X	X	X	X	X	0.94	0.97	1.00
	AR Surprise	X	X	X	X	X	0.91	0.94	0.96
	Illiquidity	X	X	X	X	X	0.91	0.94	0.96
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Natural Gas	SR—Garcia	3.22	591.35	−1176.26	−1165.12	NA	0.10	0.50	0.90
	Volatility	0.00	45.83	−91.67	−91.67	0.00	0.89	0.92	0.94
	Abnormal Return	X	X	X	X	X	0.89	0.92	0.94
	Illiquidity Surprise	X	X	X	X	X	0.89	0.92	0.94
	Volatility Surprise	X	X	X	X	X	0.89	0.92	0.94
	AR Surprise	X	X	X	X	X	0.89	0.92	0.94
	Illiquidity	X	X	X	X	X	0.90	0.92	0.94
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Crude Oil	SR—Garcia	2.82	525.73	−1045.83	−1036.09	NA	0.10	0.50	0.90
	Volatility	0.00	75.67	−151.34	−151.34	0.00	0.98	0.99	0.99
	Abnormal Return	X	X	X	X	X	0.90	0.93	0.95
	Illiquidity Surprise	0.00	17.21	−34.42	−34.42	0.00	0.88	0.92	0.95
	Volatility Surprise	0.00	29.37	−58.75	−58.75	0.00	0.92	0.96	1.00
	AR Surprise	X	X	X	X	X	0.92	0.97	0.97
	Illiquidity	0.00	12.82	−25.64	−25.64	0.00	0.00	0.91	0.95
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Platinum	SR—Garcia	2.94	442.43	−878.99	−868.83	NA	0.10	0.50	0.90
	Volatility	0.00	84.72	−169.45	−169.45	0.00	0.82	0.85	0.89
	Abnormal Return	0.00	19.33	−38.65	−38.65	0.00	0.84	0.87	0.90
	Illiquidity Surprise	X	X	X	X	X	0.82	0.86	0.89
	Volatility Surprise	X	X	X	X	X	0.82	0.86	0.89
	AR Surprise	0.00	9.07	−18.13	−18.13	0.00	0.82	0.86	0.89
	Illiquidity	X	X	X	X	X	0.82	0.86	0.89
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Rice	SR—Garcia	7.04	259.64	−505.22	−480.87	NA	0.10	0.50	0.90
	Volatility	X	X	X	X	X	0.59	0.69	0.81
	Abnormal Return	X	X	X	X	X	0.60	0.71	0.82
	Illiquidity Surprise	X	X	X	X	X	0.59	0.70	0.82
	Volatility Surprise	0.00	24.95	−49.90	−49.90	0.00	0.59	0.70	0.82
	AR Surprise	0.00	16.81	−33.62	−33.62	0.00	0.68	0.76	0.84
	Illiquidity	X	X	X	X	X	0.61	0.71	0.83
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Silver	SR—Garcia	2.83	460.37	−915.09	−905.32	NA	0.10	0.50	0.90
	Volatility	0.00	82.68	−165.37	−165.37	0.00	0.90	0.92	0.96
	Abnormal Return	X	X	X	X	X	0.90	0.93	0.96
	Illiquidity Surprise	X	X	X	X	X	0.90	0.93	0.96
	Volatility Surprise	X	X	X	X	X	0.92	0.95	0.98
	AR Surprise	X	X	X	X	X	0.90	0.93	0.96
	Illiquidity	X	X	X	X	X	0.90	0.92	0.96
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Soybean	SR—Garcia	2.53	579.99	−1154.93	−1146.19	NA	0.10	0.50	0.90
	Volatility	0.00	55.76	−111.53	−111.53	0.00	0.90	0.92	0.94
	Abnormal Return	X	X	X	X	X	0.87	0.89	0.92
	Illiquidity Surprise	X	X	X	X	X	0.87	0.89	0.91
	Volatility Surprise	X	X	X	X	X	0.88	0.90	0.92
	AR Surprise	X	X	X	X	X	0.88	0.90	0.92
	Illiquidity	X	X	X	X	X	0.87	0.89	0.92
Panel B: Full Sample
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Coffee	SR—Garcia	20.98	17323.34	−34604.71	−34452.33	NA	0.10	0.50	0.90
	Volatility	0.00	1614.50	−3228.99	−3228.99	0.00	0.80	0.90	0.94
	Illiquidity	0.00	119.27	−238.54	−238.54	0.00	0.39	0.79	0.83
	Illiquidity Surprise	0.00	56.68	−113.36	−113.36	0.00	0.36	0.73	0.81
	Volatility Surprise	0.00	94.23	−188.47	−188.47	0.00	0.56	0.82	0.91
	AR Surprise	0.00	46.60	−93.21	−93.21	0.00	0.41	0.82	0.84
	AR Surprise	0.00	46.84	−93.68	−93.68	0.00	0.59	0.81	0.89
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Corn	SR—Garcia	7.07	21018.10	−42022.07	−41970.56	NA	0.10	0.50	0.90
	Volatility	0.00	1471.46	−2942.92	−2942.92	0.00	0.79	0.83	0.87
	Illiquidity	0.00	210.96	−421.91	−421.91	0.00	0.80	0.84	0.88
	Illiquidity Surprise	x	x	x	x	x	0.80	0.85	0.89
	Volatility Surprise	x	x	x	x	x	0.78	0.83	0.87
	AR Surprise	x	x	x	x	x	0.79	0.83	0.87
	AR Surprise	x	x	x	x	x	0.79	0.86	0.90
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Gasoline	SR—Garcia	21.75	8234.81	−16426.12	−16290.23	NA	0.10	0.50	0.90
	Volatility	x	x	x	x	x	0.89	0.93	0.96
	Illiquidity	x	x	x	x	x	0.89	0.93	0.96
	Illiquidity Surprise	x	x	x	x	x	0.89	0.93	0.96
	Volatility Surprise	x	x	x	x	x	0.89	0.93	0.96
	AR Surprise	0.00	48.49	96.99	96.99	0.00	0.89	0.93	0.96
	Abnormal Return	x	x	x	x	x	0.89	0.93	0.96
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Gold	SR—Garcia	2.47	523.64	−1042.33	−1033.78	NA	0.10	0.50	0.90
	Volatility	0.00	55.20	−110.39	−110.39	0.00	0.92	0.95	0.97
	Illiquidity	x	x	x	x	x	0.92	0.94	0.97
	Illiquidity Surprise	x	x	x	x	x	0.91	0.94	0.97
	Volatility Surprise	x	x	x	x	x	0.94	0.97	1.00
	AR Surprise	x	x	x	x	x	0.91	0.94	0.96
	Abnormal Return	x	x	x	x	x	0.91	0.94	0.96
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Natural Gas	SR—Garcia	3.02	12516.59	−25027.14	−25006.16	NA	0.10	0.50	0.90
	Volatility	0.00	763.31	−1526.61	−1526.61	0.00	0.83	0.88	0.98
	Illiquidity	0.00	353.14	−706.29	−706.29	0.00	0.80	0.86	0.91
	Illiquidity Surprise	0.00	58.12	−116.24	−116.24	0.00	0.79	0.87	0.91
	Volatility Surprise	0.00	115.81	−231.62	−231.62	0.00	0.77	0.85	0.90
	AR Surprise	x	x	x	x	x	0.78	0.86	0.90
	Abnormal Return	x	x	x	x	x	0.72	0.82	0.89
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Crude Oil	SR—Garcia	17.17	17915.47	−35796.60	−35673.73	NA	0.10	0.50	0.90
	Volatility	x	x	x	x	x	0.78	0.86	0.94
	Illiquidity	0.00	362.99	−725.99	−725.99	0.00	0.84	0.90	0.95
	Illiquidity Surprise	0.00	198.27	−396.54	−396.54	0.00	0.68	0.79	0.85
	Volatility Surprise	0.00	661.25	−1322.51	−1322.51	0.00	0.87	0.91	0.95
	AR Surprise	0.00	79.58	−159.15	−159.15	0.00	0.85	0.90	0.95
	Abnormal Return	0.00	94.49	−188.98	−188.98	0.00	0.85	0.90	0.95
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Platinum	SR—Garcia	19.84	12772.77	−25505.87	−25361.52	NA	0.10	0.50	0.90
	Volatility	0.00	1498.14	−2996.29	−2996.29	0.00	0.71	0.83	0.89
	Illiquidity	0.00	451.71	−903.41	−903.41	0.00	0.51	0.73	0.89
	Illiquidity Surprise	0.00	116.21	−232.42	−232.42	0.00	0.12	0.41	0.64
	Volatility Surprise	0.00	103.22	−206.45	−206.45	0.00	0.41	0.76	0.88
	AR Surprise	0.00	61.82	−123.64	−123.64	0.00	0.70	0.85	0.92
	Abnormal Return	x	x	x	x	x	0.00	0.78	0.85
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Rice	SR—Garcia	18.56	8088.19	−16139.26	−16008.16	NA	0.10	0.50	0.90
	Volatility	x	x	x	x	x	0.61	0.71	0.81
	Illiquidity	0.00	87.82	−175.64	−175.64	0.00	0.56	0.68	0.79
	Illiquidity Surprise	x	x	x	x	x	0.56	0.68	0.78
	Volatility Surprise	0.00	440.05	−880.10	−880.10	0.00	0.54	0.67	0.78
	AR Surprise	x	x	x	x	x	0.56	0.68	0.78
	Abnormal Return	x	x	x	x	x	0.51	0.67	0.78
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Silver	SR—Garcia	17.66	13699.67	−27364.02	−27235.48	NA	0.10	0.50	0.90
	Volatility	0.00	1867.17	−3734.34	−3734.34	0.00	0.69	0.81	0.91
	Illiquidity	0.00	309.74	−619.48	−619.48	0.00	0.72	0.79	0.89
	Illiquidity Surprise	0.00	220.31	−440.62	−440.62	0.00	0.68	0.79	0.86
	Volatility Surprise	0.00	88.58	−177.16	−177.16	0.00	0.70	0.79	0.85
	AR Surprise	0.00	66.12	−132.25	−132.25	0.00	0.72	0.78	0.85
	Abnormal Return	x	x	x	x	x	0.71	0.79	0.86
Commodity	Variable	EDF	CLL	CAIS	CBIC	p-Value	Prediction by Quantiles
Soybean	SR—Garcia	23.34	22317.60	−44588.51	−44418.31	NA	0.10	0.50	0.90
	Volatility	0.00	1511.25	−3022.51	−3022.51	0.00	0.89	0.92	0.95
	Illiquidity	0.00	260.50	−521.00	−521.00	0.00	0.81	0.92	0.95
	Illiquidity Surprise	0.00	76.37	−152.75	−152.75	0.00	0.77	0.91	0.95
	Volatility Surprise	0.00	42.81	−85.62	−85.62	0.00	0.83	0.93	0.96
	AR Surprise	x	x	x	x	x	0.83	0.93	0.97
	Abnormal Return	0.00	55.73	−111.47	−111.47	0.00	0.88	0.92	0.95

We contextualize our results within literature as follows. Joëts (2015) documents that energy price fluctuations are mainly governed by fundamentalist expectations when agents in the markets evolve in the context of certainty, whereas both fundamentalist and speculative behaviors are the source of price movements in an uncertain world. Our findings indicate, however, that the weights of the types of traders could be different if we look at extreme situations and perhaps even more so in unprecedented pandemic periods. Looking at safe-haven and non-staple futures, the proportion of uncertainty-averse agents increased during extreme downward movements, leading to situations where the fundamental nature of the markets fades in favor of irrational fluctuations, such as “cascade behavior”. We conjecture that our findings are likely after-effects of such a cascade-like phenomenon. Such cascades, usually in informational form, suggest the possibility of individuals taking decisions purely based on mimicking other investors’ trading activities. For financial instruments, especially those that frequently attract speculators due to a speculation-hedging divide, can be particularly prone to manifestations of cascade-like behavior such as herding. Such behavior increases the fragility of the instrument in question. Also, for futures, if such instruments are of global importance (e.g., oil and gold), it can exacerbate the prevailing fragility of the markets–particularly during crisis periods. This raises the vulnerability of both speculators and hedgers as sudden arrival of regulatory or precise information can alter the course of price and volume and thereby reverse the cascade. Sornette et al. (2009) also provide similar support for higher speculation in energy commodities during the 2006–2008 oil bubble. Our results also support that during time of crisis, investors adopt flight-to-safety behavior and show behavioral prejudices associated with gold’s history as a currency, a store of value and a safe haven Baele et al. (2020). We find that silver closely follows suit but not platinum. It is worth pointing out that, as one anonymous referee aptly points out, it is plausible that the similarity between the role of gold and silver are driven by their conditional quantiles and tail dependence. This phenomenon was demonstrated by Bouri and Jalkh (2019) for an earlier sample period. Investigating whether the same applies to the COVID-19 pandemic is a worthy matter. We leave this as an open conjecture, which future researchers may pursue. Besides this, our results also lend support to the report of Conlon and McGee (2020) for COVID-19 and early 1980s’ recessions.

4. Conclusion

We summarize the key findings of our paper as follows: energy and safe-haven precious metals invite high degrees of speculation. This trend is on the rise since the GFC. Agriculture-based futures have historically attracted comparatively modest speculation. We also document a largely linear relationship between speculation and excess returns since 1980s. During the COVID-19 pandemic, however, the tie is broken. Instead, rise in volatility coincides with rise in speculation. Scope constraints inhibit broadening our scope to engage in a recession-by-recession analysis to unravel further subtleties in the speculation–profitability nexus. Future researchers can consider this as a promising area for extension.

Our study offers several implications for various stakeholders. Given that precious metals and energy futures exhibit potentially inefficient behavior after the outbreak of the pandemic, the markets for these commodities suggest the possibility of irrational speculative intent which may distort the payoff expectations derived from fundamental investment analysis. Consequently, regulation in such exploitable instruments can becomes necessary. Moreover, in contrast to a common belief that precious metals offer safe-haven benefits in the event of a crisis, our result suggest that investors need to be more careful in underestimating the risk exposure to these commodities. Furthermore, the switching behavior (Pre vs. post-crises) of prices for energy and precious metals futures shows that these markets are sensitive to market trends that warrant real-time monitoring by market analysts and regulators. Our results also show that agro-based futures are likeliest to follow the classical valuation methods—assuming Brownian motion of the underlying asset. The results imply that traditional candidates for safe havens such as precious metals and energy might be replaced by other commodities, suggesting that the nature of the event (crisis) determines the market dynamics.

CRediT authorship contribution statement

Imtiaz Sifat: Conceptualization, Methodology, Software, Validation, Formal analysis, Data curation, Writing - original draft, Writing - review & editing, Visualization. Abdul Ghafoor: Formal analysis, Writing - original draft, Writing - review & editing. Abdollah Ah Mand: Data curation, Visualization.

Appendix A. Supplementary data

The following is the Supplementary material related to this article.

MMC S1

The supplementary material contain additional information pertinent to the paper (e.g., summary stats and stationarity tests) in the paper as well as a video abstract.