Commodity markets and the global macroeconomy: evidence from machine learning and GVAR

Commodity market prices are in strong connection to the global macroeconomic cycles and crises, as demonstrated, e.g., by the 2008–2009 global financial crisis and the COVID-19 pandemic, not to mention the most recent crisis due to the Russian attack to Ukraine. During these crises, for example at the beginning of pandemic the commodity prices initially declined but eventually rose as the global economic activity recovered after the recovery. As another possibility for the causal relationship between macroconditions and commodity price development, the war in Ukraine highlights the significant impact of changes in macroeconomic conditions (destroying of productive capacity) on commodity markets. Furthermore, it is widely acknowledged that certain commodities play a more crucial role in economic development, at both a national and global level. Recent evidence indicates that crises have significant effects on supply chains especially in food and energy-related commodity markets. On the other hand, for instance, Stuermer (2018) suggests that the relationship between commodity markets and macroeconomic performance is strongly influenced by demand shocks resulting from the changes in global economic activity.

While some researchers have investigated the impact of global economic activity on commodity price changes in general terms, others have focused on specific commodities and their relevance to the global economic performance (Duarte et al. 2021; Liu et al. 2020; Fasanya and Awodimila 2020; Abbas and Lan 2020; Chen et al. 2010; Kilian 2009; etc.). However, these studies have yet to reach a consensus on which commodities have the most significant price effects on the global economy. Consequently, it is still unclear which commodities, among the many traded globally, are the most important ones for driving global economic activities and promoting economic development in different regions. The present study aims to address this research gap by not only identifying the most significant commodities for the global economy but also examining their roles within different countries and groups.

We propose a novel, completely empirically oriented approach to address these questions. First, we employ machine learning techniques, specifically the Least Absolute Shrinkage and Selection Operator (LASSO) estimation procedure, to determine the time-varying significance of the most important commodity markets' price developments in relation to various macroeconomic indicators. Unlike previous studies that have pre-selected specific commodities based on their perceived relevance to macroeconomic indicators, we avoid this bias by letting the data-driven LASSO technique identify the key commodities for the global economy. Second, we utilize a global vector autoregressive (GVAR) analysis to examine the dynamic interactions and the speed at which economies adjust to the most relevant commodity market exposures identified through the LASSO estimation procedure. This analysis provides a deeper insight into the interdependence between the commodity market developments and macroeconomic performance across different countries and groups.

By combining these two techniques, we aim to provide a comprehensive understanding of the significance of various commodities for the global economy, without any a priori assumptions or pre-selection biases. This approach allows for a more robust and objective assessment of the relationship between commodity prices and macroeconomic indicators.

Empirical relationships between specific commodity prices and global economic activities have received considerable scholarly attention. For example, numerous studies have revealed the significance of oil price changes in affecting the real output changes at both country and global levels (Ge and Tang 2020; Cunado et al. 2015; Boschi and Girardi 2011). Additionally, previous research has focused on examining whether the commodity prices act as leading indicators for exchange rate movements in commodity-dependent economies, commonly known as commodity currencies. This set of countries includes, e.g., Australia, Canada, New Zealand, Norway, and South Africa (Beckmann et al. 2020; Liu et al. 2020; Baghestani et al. 2019; Chen et al. 2010; Ferraro et al. 2015). Furthermore, extensive analysis has been conducted on the relationship between commodity market exposure and aggregate inflation rates (Fasanya and Awodimila 2020; Abbas and Lan 2020; Gelos and Ustyugova 2017; Chen et al. 2014). These studies generally agree that the global commodity price changes can serve as leading indicators for inflation, especially in countries heavily dependent on commodity exports.

The GVAR framework has also been used to explore the impact of commodity prices on the global macroeconomy. Within this framework, the previous research has highlighted the significant role of food prices, including wheat, in the global cycles (Gutierrez and Piras 2013; Galesi and Lombardi 2009). Furthermore, extensive amount of research has been conducted on the global effects of commodity prices with particular emphasis on energy prices, especially oil, within the GVAR framework. Studies such as Bettendorf (2017), Chudik and Fidora (2012), Boschi and Girardi (2011), Pesaran, Schuermann, and Weiner (2004), Dées et al. (2007), and Cashin et al. (2014) have focused on oil prices. For instance, Boschi and Girardi (2011) identified oil prices as a global indicator in explaining output variability in the euro area, Latin America, and several major individual economies. Chudik and Fidora (2012) used oil prices to analyze the effects of a strong oil supply shock in a GVAR model comparing the real output developments of various emerging economies to those of advanced economies. They observed a negative impact on real GDP growth in oil-exporting economies, as well as changes on the real exchange rates for oil exporters and importers.

These studies provide valuable insights into the role of commodity prices, particularly oil prices, in the global macroeconomy within the GVAR framework. However, a closer scrutiny of the existing empirical studies reveals that only a limited number of individual commodities have been examined, based on subjective judgments of their importance. These studies often generalize their findings to global practical analyses or forecasting purposes. While we do not dispute the significance of the commodities that have been investigated, it is crucial to acknowledge the numerous traded commodities worldwide, many of which might have been overlooked in the previous studies despite their potential importance. Assuming that only a few of these commodities are globally significant, as suggested by Duarte et al. (2021) and Baghestani et al. (2019), without employing an appropriate model to determine their actual roles, seems unrealistic. Therefore, we consider already as a starting point for our analysis a comprehensive range of commodities as potentially influential and aim to empirically identify the most significant ones in relation to global macrovariables, such as real GDP, real exchange rates, and inflation.

We fully acknowledge that many previous studies have examined the connections between commodity market returns and, e.g., GDP growth using data based on a much more limited number of prominent commodities, such as 27 in Ge and Tang (2020), 17 in Liu et al. (2020), or 12 in Jacks and Stuermer (2020), who analyzed the agricultural goods, metals, and soft commodities from 1870 to 2013. However, in addition to analyzing the spot market price (indices) of 55 commodities, compared to these papers, our study employs different data and methodologies. By scrutinizing a large number of individual commodities and utilizing specific methodological choices, we aim to present novel findings on the roles that different commodities play in relation to the global economic indicators.

This paper makes several significant contributions to the existing literature, especially empirically. First, unlike previous studies that made ad hoc selections or assumptions regarding the importance of specific commodities, we consider all 55 individual commodities traded on a daily basis to be potentially equally important from the start. This approach ensures that no prior judgments are made, allowing for a more comprehensive analysis. Second, we utilize machine learning techniques to identify the most important commodities among the starting large set of globally traded individual commodities. This is in contrast to all existing studies that have often hand-picked only a limited number of commodities based on subjective reasoning. Hence, by employing machine learning, we enhance the objectivity and accuracy of our analysis. Finally, we employ the global vector autoregression (GVAR) model, which combines country-level time series panel data and factor analytic techniques. This model enables us to assess the impact of unit shocks on the identified globally important commodity markets and examine how these shocks are transmitted among different countries and groups of countries. By analyzing the reactions of various regions such as Africa, Asia, the euro area, Latin America, the Middle East, as well as individual large countries like the UK, China, and the USA, we gain valuable insights into the transmission mechanisms and dynamics of these shocks.

Based on global data from 1990Q1 to 2019Q4, our analysis reveals that among the 55 commodity markets considered, the price change, i.e., returns from copper, crude oil, gold, and lead markets are the most important for the development of global macroeconomic variables in general. More specifically, we find that the changes in copper and crude oil prices have a significant impact on especially the global output changes. Furthermore, changes in gold and lead prices exhibit a strong correlation with the real exchange rate changes. Our results also support the traditional view that the global oil market plays a crucial role in transmitting the inflationary pressures across the global economy. This is in line with many previous studies (e.g., Ha, Kose, and Ohnsorge, 2022; Herwartz and Plödt 2016). Also the importance of copper and oil price changes on the aggregate output development is consistent with findings from other studies (e.g., Wen et al. 2019; Boschi and Girardi, 2011), suggesting that a shock to these commodity prices leads to a significant increase in the real GDP for both the advanced and emerging economies. Furthermore, we observe that a positive shock to the gold and lead price changes results in a significant depreciation of the real exchange rates. Considering the current global economic conditions, particularly influenced by the war in Ukraine, our analysis suggests that the oil price shocks will likely continue to transmit the inflationary pressures worldwide for an extended period of time.

Overall, our results indicate that commodity market exposure is a significant and prevalent phenomenon in the markets and overall economies. Therefore, given this significance, implementing policies that mitigate price volatility in these specific commodities can help in smoothing global economic performance. We strongly recommend the adoption of such policies.

The paper is organized as follows. Section 2 introduces the empirical models used in this study and provides a description of the data employed. In Sect. 3, a comprehensive discussion of the empirical results is presented. Finally, Sect. 4 concludes the study with a summary of the findings and potential implications for future research.

We utilized quarterly data on primary commodity market indices and three global macroeconomic variables: output (real GDP), inflation, and real exchange rate. The commodity market dataset contains time series observations on 55 global commodity indices obtained from the IMF primary commodities database (refer to Table 2). The macrodata were sourced from the global VAR modeling database (Smith and Galesi 2014),³ originally compiled by Dées et al. (2007) and extended by Mohaddes and Raissi (2020).⁴ This dataset encompasses 33 countries (as shown in Table 1), which collectively account for over 90% of the global GDP. Due to data limitations,⁵ our analysis was conducted for the period ranging from 1990Q1 to 2019Q4. To focus on regional effects and shocks, we grouped the countries in the GVAR model into specific regions (refer to Table 1). In the VARX* model, the euro area is treated as a single region and aggregated using GDP-weighted averages of the country-specific variables, including real GDP, inflation, and real exchange rate. Table 2 provides an overview of the primary commodities based on the IMF primary commodity groupings. For consistency, all observations in the original dataset were transformed into logarithmic values. For a more detailed summary of these transformed series, please refer to Appendixes 9, 10, 11.

To address the issue of multicollinearity in LASSO estimation and to avoid erroneous identification of relevant commodities in the regression, we initiated our empirical analysis by examining the price correlations among different commodity sectors.⁶ Given the strong correlation typically observed in global energy prices, especially in the oil and gas sectors, we anticipated finding high price correlations in these markets. Analysis of our data confirms that energy price series, including EU and US natural gas prices, as well as various oil price series such as Brent, Dubai, and WTI, exhibit significant correlation. For instance, static correlations between the US and EU natural gas prices and WTI crude oil prices are 0.91 and 0.90, respectively (shown in Table 3). Multicollinearity in the energy market sectors is visualized in Fig. 1. The upper panels A and B of Fig. 1 illustrate collinearity in price and return movements within the crude oil market (Brent, Dubai, WTI). Similar patterns can be observed in the lower panels C and D for the natural gas market (USA and EU).

In order to gain a better understanding of the energy market price dependencies, the analysis was expanded to include the second moments for returns in the oil and natural gas markets. This analysis showed that the volatilities in these two markets are quite similar, with the US natural gas shocks appearing to dominate over time (refer to Fig. 2 panel A). Similar to Batten et al. (2017), the dynamic correlations between returns are not stable but rather vary over time⁷ (refer to Fig. 2 panel B). It is also worth noting that the dynamic co-movement between the US natural gas and WTI crude oil markets is time-varying, but the correlation trend has remained relatively stable since the year 2000. Due to the high price co-movement and the similar variance development, it is reasonable to use WTI crude market data to understand the effects of natural gas and other crude oil markets in further estimations. This choice helps to also mitigate the issue of multicollinearity in the LASSO regressions. Therefore, at this stage, we employed the LASSO model using a set of 55 individual commodity return series, instead of the originally extracted 59 time series.

In many cases, machine learning algorithms are employed in empirical analysis to utilize all available information and predict the response variable, without assuming causality for any particular predictor. Therefore, at the initial stage of our analysis, our focus was solely on identifying the set of most relevant commodities associated with macroeconomic changes. The structural impact analysis will commence from Sect. 3.3, where we will examine the macroeconomic responses to the structural shocks originating from a set of the most significant commodity markets, employing a purely empirical perspective.

Without making any prior assumptions in our machine learning estimation, the LASSO regression revealed that out of the 55 commodity price changes analyzed, only four—crude oil, copper, gold, and lead prices—emerged as relevant predictors for the global macroeconomic performance. These findings hold true even after accounting for the multicollinearity issues and conducting robustness checks. To validate these results, we conducted principal component analysis (PCA)⁸ and factor analyses as the robustness checks. The results from PCA and factor analysis support the notion that these commodities (crude oil, copper, gold, and lead) are of utmost importance when it comes to the global macroeconomic performance. These findings are consistent with various previous studies that have also identified at least the crude oil, copper, and gold market price developments as indicators of global economic performance (Bildirici and Gokmenoglu 2020; Stuermer 2018; Arora and Cai 2014; Jaunky 2013; Boschi and Girardi, 2011).

The results reported in Table 5 indicate that changes in copper and crude oil prices are the most significant indicators for assessing the global output (real GDP) changes. Furthermore, our empirical estimates from this stage suggest that when the prices in these markets increase, there is a general improvement in the macroeconomic performance. These estimates further support the notion that the positive developments in the copper market prices, often referred to as 'Dr Copper,' serve as significant indicators of global economic health, corroborating previous findings such as those of Stuermer (2018).

In addition, our estimation results consistently demonstrate that the crude oil market has the utmost significant role among the primary commodity markets when it comes to the global transmission of inflationary pressures. This finding aligns with several recent studies that have come to the same conclusion (Fasanya and Awodimila 2020; Bettendorf 2017; Chudik and Fidora 2012; Kilian 2009). When examining the economic impact, Table 5 illustrates that the changes in copper and oil market prices have similar effects on the real economic activity across different countries. In most cases, the coefficient ranges between 0.1 and 0.2, rounded to one decimal point. However, it is worth noting that the primary driver of aggregate inflation appears to be the changes in crude oil market prices, and this effect seems to be consistent across all countries. Furthermore, a comparison at the country level indicates that macroeconomic indicators in the European countries are particularly vulnerable to the commodity market price risks.

The results presented in Table 5 emphasize the significant impact of gold and lead market price changes on the development of the real exchange rate. Notably, an increase in gold prices is strongly associated with a depreciation of the real exchange rate. This outcome is not surprising, considering the gold market's sensitivity to global economic fluctuations, making its price a reliable indicator of global economic well-being.⁹ Our findings align with previous studies (e.g., Capie et al. 2005), which have consistently shown a negative relationship between gold price movements and dominant currencies like the US dollar. Moreover, the research of Giannellis and Koukouritakis (2019) and Wang and Lee (2016), among others, provides support for the idea that the gold market offers protection against currency risks. Table 5 also reveals the diverse effects of various commodity markets on the real exchange rates of different countries. In the majority of cases, an increase in commodity market prices correlates with a depreciation in the domestic real exchange rate, implying a decline in the currency's value.

Based on our LASSO analysis, one of the four most important global commodities was detected to be the lead metal. Hence, our research revealed completely novel findings regarding the role of lead market prices, which bear similarity to the impact observed in the gold market. From our GVAR results it is now also evident that the lead market price changes significantly influence the real exchange rates of nearly every individual country. Digging deeper into the economic significance of the lead market, we found that lead, like the other base metals such as zinc, silver, and copper, is extracted from galena and sourced from ore. At the global level, over 86% of refined lead and lead-related products are utilized in various industries including automobile manufacturing, batteries, pigments, and ammunition (U.S. Geological Survey, 2019). China, Australia, and the USA are the primary producers of lead, followed by Peru, Canada, Mexico, Sweden, and South Africa (U.S. Geological Survey, 2019). It is important to note that all these countries were included in our study to ensure comprehensive analysis.

Our newly revealed prominent role of lead in international trade may explain why the development of lead market prices is considered crucial for analyzing exchange rate movements using the LASSO method. Furthermore, according to the results reported in Table 5, the impact of lead market price development on economies varies across countries. For instance, when the global lead market prices increase, countries like Australia and Japan experience a significant appreciation of their real exchange rates, while others witness a depreciation. Recently, lead prices have surged to their highest level since 2018 amid the post-Covid-19 global economic recovery and the energy crisis. This increase is mainly attributed to the anticipation of supply disruptions in Europe and the growing global demand for the traditional lead-acid car batteries from China, the USA, and Europe (World Bank Group 2021). The World Bank's commodity market outlook for 2021 projected that lead prices would remain stable at pre-pandemic levels in both the medium and long terms (World Bank Group 2021). Nonetheless, this outlook might have changed due to the onset of the war in Ukraine in early 2022.

In order to conduct further robustness checks, we excluded the data from the global economic and financial crises of 2008–2009. These data spanned from 2007Q1 to 2009Q4. Subsequently, we performed the LASSO estimation on two sub-samples referred to as 'pre-crisis' (covering the period from 1990Q1 to 2006Q4) and 'post-crisis' (covering the period from 2010Q1 to 2019Q4). The results of these estimations are presented in Tables 9 and 10 in the Appendix. Overall, the set of four significant commodities identified in the full sample estimation remains consistent. Notably, we observed a significant increase in the overall importance of these commodities to the respective macrovariables during the post-crisis period (see Table 10 in the Appendix).

Based on the results obtained from the unit root tests for our data, the second stage of our estimation process began by examining the weak exogeneity of the foreign and global variables in our estimated GVAR systems. The weak exogeneity assumption is critical in the estimation process of the global VAR approach. To test the joint significance of the foreign and global variables in each country-level regression equation, we estimated an error correction model. Following the approach used in studies by Pesaran et al. (2004), Boschi and Girardi (2011), and Gutierrez and Piras (2013), we first grouped the foreign and global variables together in the vector \(\Delta {y}_{{\text{it}}}^{*}\). Subsequently, we constructed the regression model based on Eq. 7, i.e.,

In the given equation, the ECM term, denoted as \({{\text{ECM}}}_{i,t-1}^{j}, j=1, 2, .. {, r}_{i}\) corresponds to the estimated error correction term for each macrovariable (j) of each country. The value of \({r}_{i}\) represents the number of cointegration relations found for that particular macrovariable. The notation ∆ represents the first difference operator, so \(\Delta {\mathbf{y}}_{i,t-k}^{j}\) implies the domestic macrovariables expressed as differences over the lag period of k. Similarly, \(\Delta {\widetilde{\mathbf{y}}}_{i,t-m}^{*}\) represents the foreign and global variables expressed as differences over the lag period of m. Furthermore, \({k=1, \dots ,p}_{i}\) and\(m=1,\dots ,{q}_{i}\), where \({p}_{i}\) and \({q}_{i}\) are the maximum lag orders of the domestic, foreign and global variables for each of the \({i}^{th}\) country models,¹⁰ respectively. The test for weak exogeneity is an F test for the joint significance of the hypothesis \({\gamma }_{ij}=0, j=1, 2, .. {, r}_{i}\) at a 5% risk level of the ith country model using the above regression representation. In particular, the test assumes that both the foreign-specific and global variables enter the model as weakly exogenous. We thus verify the null hypothesis of weak exogeneity for both the foreign-specific and global variables against the alternative hypothesis of no weak exogeneity for both the foreign-specific and global variables.

The results presented in Table 6 indicate that out of the 130 cases analyzed, only 10 cases show a rejection of the null hypothesis regarding the weak exogeneity assumption. This favorable outcome strengthens the support for the weak exogeneity hypothesis within our sample. However, it is important to note that the assumption of weak exogeneity for global variables is challenged in several countries. This implies that, from a purely statistical perspective, the changes in prices observed in these markets cannot be considered weakly exogenous in relation to the macroeconomic developments of those countries.

For instance, our analysis from this stage reveals that the changes in copper prices cannot be considered weakly exogenous for countries such as Indonesia, Norway, the Philippines, and Sweden. Similarly, the exogeneity of crude oil price changes is rejected for Malaysia. On the other hand, the exogeneity of gold price changes is only rejected for Saudi Arabia, while the exogeneity of lead price changes is rejected for both Sweden and the USA.

At the next step, we utilized the error correction model versions of the VARX* representations to analyze the contemporaneous impacts of foreign variables on domestic counterparts. We also assume the weak exogeneity of the foreign variables. The estimation procedure in the GVAR framework maintains consistency, ensuring that the variables of each model interact in the long run. This analysis is particularly valuable at the global scale, as it allows us to investigate the feedback effects from the foreign variables (Pesaran and Smith 2006; Pesaran et al. 2004; Galesi and Lombardi 2009). Specifically, we focus on the impact elasticities, as discussed in Galesi and Lombardi (2009). These elasticities measure the immediate variation of a domestic variable resulting from a 1% change in the corresponding foreign variables.

We examine the impact of a 1% increase in foreign-specific inflation on Argentina's domestic inflation, based on the research conducted by Pesaran et al. (2004). This analysis is crucial to evaluate the spillover effects originating from the foreign variables. Table 7 presents the results, showing that the estimated coefficients are predominantly positive, except for the inflation effects observed in Brazil, Chile, and Japan. Many of the countries/regions demonstrate statistically significant impact elasticities, indicating a notable influence of foreign variables. Additionally, all countries/regions exhibit statistically significant exchange rate elasticities. The statistical significance of the effects of foreign variables on domestic counterparts is particularly high for both the output and inflation. For output, the estimated coefficients range from 0 to 2, with Argentina showing the lowest value (0.15) and Turkey demonstrating the highest (2.01). Notably, the impact elasticity is statistically significant and exceeds that of Turkey (2.01), Sweden (1.247), Singapore (1.32), and Malaysia (1.12). This implies that the domestic output reacts more strongly to an increase in the output of major trading partners.

The estimated impact elasticity with respect to inflation ranges between -3 and 1. Additionally, the impact elasticity for the output effects is significant for Argentina (1.08) and India (1.02), indicating an overreaction of inflation relative to the increase in inflation of their main trading partners. On the other hand, statistically non-significant estimates suggest that the inflation dynamics would be independent of foreign countries' inflationary pressure. Furthermore, a statistically significant impact elasticity greater than one is observed for the real exchange rate in several countries such as Argentina (1.77), Canada (1.21), Euro area (1.18), Korea (1.05), Malaysia (1.01), Norway (1.08), Sweden (1.23), and Thailand (1.38). This indicates an overreaction or influence of an increase in the real exchange rate of major trading partners on the domestic real exchange rate. Overall, these findings align with the research conducted by e.g. Galesi and Lombardi (2009), which demonstrates that the foreign variables significantly impact the domestic macroeconomic performance in most countries.

In this section, we employed the GIRFs (generalized impulse response functions) to analyze the response of global macroeconomic variables to the commodity price shocks. We obtained the point estimates for the simulation horizon, which covers a period of up to 40 periods (10 years). This allows us to examine the long-term effects of these shocks. Moreover, we also conducted a comparative analysis to highlight the differences between our findings and those of several other studies that have focused on somewhat similar kinds of sets of commodities as ours.

In this study, we have employed the machine learning (LASSO) and GVAR approaches to analyze the relationship between key macroeconomic variables (output, inflation, and real exchange rate) and commodity market price developments in a globally comprehensive dataset. Our novel approach enabled us to assess the statistical significance of various commodities without any preconceived assumptions about their relevance. By considering at the start of our analyses a large set of 55 commodities, we identified the four globally most significant commodity market segments to be the crude oil, gold, copper, and lead markets. These markets seem to play a crucial role in influencing the global economic cycles in the contemporary context. In the second stage of our analysis, we investigated the dynamic impacts of price change shocks in these identified markets on the performance of economic regions in both the advanced and emerging countries.

Overall, our findings support the notion that the traditional commodity markets like copper, oil, and gold significantly contribute to the global economic performance. However, as a somewhat novel finding we revealed the important role of lead metal market in the global economic development, too. These findings have broader implications, suggesting the growing importance of these commodities in shaping the global economic trends.

Given the significant influence of crude oil, gold, copper, and lead markets on global economic cycles, policymakers should encourage the diversification of commodity market exposures in their political decisions. This can include promoting the development of other commodity markets, such as renewable energy sources or emerging technologies, to reduce the dependency on a few key commodities and mitigate the potential risks associated with their price volatility. Additionally, we propose that achieving optimal stabilization policies for output, inflation, and the development of real exchange rates requires meticulous forecasting exercises for the global commodity prices that hold the greatest significance in general. This is especially important for some individual economies in the group of more developed market economies.