The impact of speculation on commodity prices: A Meta-Granger analysis

Abstract

This paper uses Meta-Granger analysis to explain and summarize the mixed results in the literature on the impact of financial speculation on commodity prices. The sample covers 2106 manually collected p-values from Granger causality (GC) tests reported in 54 prior studies. Our results show that the heterogeneity in previous findings can be largely explained by the commodity type under examination, the sample period of the data, the measurement of the focus variables (return, volatility, or spread), and the inclusion of control variables in the GC model. Even after accounting for 23 observable differences in study and test design, our results indicate that studies published in higher ranked journals present significantly less evidence for speculation to drive commodity prices. Moreover, we use the Meta-Granger results to predict ‘best choice’ models considering preferred model setups. The results reveal that the hypothesis of Granger non-causality between speculation and commodity prices cannot be rejected at standard significance levels when assuming a best choice study design and various variations of it. We conclude that either there is no genuine overall speculation effect in agricultural, energy and metal markets, or the research design of the frequently applied GC testing is not powerful enough to detect those effects.

Introduction

The market environment of commodity trading has undergone substantial changes over the last decades. Often referred to as financialization, commodities have become an increasingly attractive asset class for investors. Financial speculation, which is amplified by the emerging popularity of index related financial products, is often associated with the increased trading activity in commodity futures markets (Tang and Xiong, 2012). This rise of speculation is frequently claimed to be a driver for surging commodity prices and volatilities, especially in the 2000s, sparking an ongoing public debate (e.g., Greely and Currie, 2008; Masters and White, 2009; U.S. Senate, 2009).

From a theoretical perspective, the influence of speculation on commodity prices can be seen through the lens of two major transmission mechanisms: risk sharing and information discovery (Cheng and Xiong, 2014). According to Keynes (1923), Hicks (1939), and Hirshleifer (1988), futures markets are characterized by hedging pressure as commercial hedgers are usually net short. Speculators acting as counterparty via long positions accommodate this pressure, leading to more efficient risk sharing. On the contrary, price distortions might arise if speculators engage in selloffs when seeking risk reductions in their portfolios. Considering information discovery (Grossman and Stiglitz, 1980; Hellwig, 1980), informational frictions hamper an efficient price discovery process in commodity markets. On the one hand, well-informed speculators might enhance this process by providing information regarding supply and demand through their trading activity. On the other hand, speculators might drive prices away from fundamental values due to imperfect information and heterogeneous beliefs (Singleton, 2014).

The empirical literature picks up the ambiguity in the public debate and academic theory by analyzing pricing mechanisms of commodities. The most common methodological approach is the analysis of the relation between non-commercial traders' or index traders' open interest and commodity futures prices using methods of Granger causality (GC) testing.¹ Looking at the findings of this literature stream, we see that results are rather diverse. One strand finds no or limited evidence for GC from speculation to commodity prices (among others, Alquist and Gervais, 2013; Büyükşahin and Harris, 2011; Sanders et al., 2004). In contrast, other authors detect substantial evidence that speculators Granger-cause changes in commodity prices (among others, Bohl et al., 2018; Mayer, 2012; Obadi and Korecek, 2018). A third group of studies finds evidence for speculation effects depending on the model and data characteristics of the GC test (among others, Ciner, 2002; Fujihara and Mougou, 1997; Huchet and Fam, 2016).

The previous GC studies often vary in terms of their study design, especially their sample composition (data period, frequency and data source), the configuration of empirical testing (variations of the GC model and inclusion of control variables), and the diverging measurement of financial speculation and commodity market behavior. Hence, it is challenging to directly compare the previous evidence without accounting for this heterogeneity.

Driven by the wide range of literature and its inconclusive outcomes, several review articles aggregate the existing research record to find out what we really know about the impact of speculation on commodities. Boyd et al. (2018), Grosche (2014), Haase et al. (2016), Shutes and Meijerink (2012), and Will et al. (2016) conduct literature reviews of articles on commodity (index) speculation. As an overall result, these reviews document mixed evidence for speculation to raise commodity prices or amplify its volatility. The paper closest to this study is Haase et al. (2016). The authors apply a vote counting approach to summarize the distribution and apparent disagreement among 100 studies on the effects of financial speculation. They use an integer scale from −2 to +2 to categorize the studies' results. This scale refers to the direction and strength of the impact of speculation on commodity markets. In addition, they apply this categorization to subgroups of results depending on the examined speculation measure, the response variable, paper quality, as well as the commodity type. The authors find that within their sample of 100 studies, the evidence is equally distributed among weakening effects (−2 and −1), no impact (0), and reinforcing effects (+1 and ​+2).² For specific subsamples, they report more conclusive results. For example, weakening effects dominate in the case of direct speculation measures. However, the vote counting procedure is widely criticized (Borenstein et al., 2009; Friedman, 2001; Hedges and Olkin, 1985; Mann, 1994; Stanley and Doucouliagos, 2012). First, it ignores the sample size of the primary studies and, thus, also the fact that the probability of finding a significant relation increases with the number of observations, i.e., when applying vote counting procedures, small-sample studies receive the same weight as large-sample studies. Second, the subsampling approach divides vote counts into separate categories. However, the examined categories – like paper quality or commodity type – might influence the collected primary studies' findings simultaneously. To avoid spurious and biased aggregation through omitted-variable bias, regression-based meta-methods are widely preferred to the univariate approach (Stanley and Doucouliagos, 2012). Third, the vote counting by Haase et al. (2016) collects one finding per study,³ albeit empirical articles typically report a wide range of test results for different commodity types, lag structures, and time periods. Thus, condensing a study outcome into a single result neglects valuable information about the distribution of findings and the drivers of heterogeneity.

The Meta-GC analysis presented in this paper extends the existing reviews by aggregating 2106 reported p-values from GC tests on the relation between speculation and commodity markets reported in 54 empirical studies. Through this approach, we address three main research questions:

• RQ1: Is there an overall effect of financial speculation on commodity prices?

• RQ2: Is the literature contaminated by publication bias or overfitting?

• RQ3: Which observable differences in study design explain variation across studies?

By collecting and aggregating statistical data from previous studies, meta-analysis provides a powerful method to improve our understanding of why reported study results are so diverse and, thus, explains sources of heterogeneity. The Meta-GC analysis contributes to the literature in several ways: (i) We provide a statistical integration of previous GC tests. By applying the panel GC method according to Dumitrescu and Hurlin (2012), we explore differences between individual commodities. (ii) We apply the meta-analysis method for GC testing by Bruns and Stern (2019) to investigate the presence and impact of publication selection bias and overfitting via lag selection. (iii) We use meta-regression to explicitly test the joint impact of 23 aspects of study design and quantitatively discover heterogeneity drivers such as the commodity type, measurement differences, methodological characteristics, as well as journal quality of the prior studies. (iv) Based on the results from meta-regression analysis, we define a best choice model with preferred study and test characteristics.

The remainder of the article is structured as follows. Section 2 describes the construction of the meta-data set. Section 3 presents the methodology of the Meta-GC analysis. Section 4 explains the sources of heterogeneity among primary study results. Section 5 presents and discusses the empirical findings. Section 6 concludes.