The Response of Global Oil Inventories to Supply Shocks

Abstract

Oil inventories are essential in alleviating realized and anticipated supply shocks and represent a key market indicator. This study examines the responses of global and country oil inventories to supply shocks under tight and loose market conditions. We utilize an expanded version of the GVAR model, adding the OECD oil inventories variable, incorporating major oil-producing countries: Iran, Russia, and Venezuela, and extending the coverage period. Our simulations indicate that a negative global supply shock significantly affects oil inventories under “tight” market conditions. The model correctly predicts the trajectory of changes to oil inventories in South Korea following a supply shock to Russian production in tight markets and Iranian output in loose markets. This case also shows that commercial players, using their inventories as a buffer, can negate government attempts to maintain constant levels of reserves. Overall, the response to the oil inventory tends to vary across producing and importing countries and market conditions. Such dynamics highlight potential problems with specific policies, such as using inventories as a buffer to alleviate price fluctuations or disrupting the oil production of individual countries through sanctions, as these measures oftentimes result in unintended consequences due to complex interconnections of the global oil market.

1. Introduction

External shocks can arise from various sources—including geopolitical, macroeconomic, and force majeure events—and have the potential to place considerable pressure on oil markets. These pressures, in turn, shine a spotlight on potential adverse consequences to the world oil market, affecting prices, volatility, and oil inventories. The Abadan crisis of 1950–1954, the oil price shocks of 1973 and 1979, the 1990 oil price shock, and the 2020 Russia–Saudi Arabia price war are among the examples of oil shocks that had a significant impact on the global oil market and the global economy in general. During such events, commercial and strategic inventories can act as a first line of defense, contributing to two major components of energy security—availability and affordability of supply.

Oil inventories have played an important role in alleviating realized and anticipated shocks to the global oil supply. The release or buildup of these stocks can decrease volatility in price [1] or signal the availability of supply to the market [2]. Herrera [3] describes the basic transmission mechanism. Industry-level inventories and sales often respond faster to an oil price shock than gross domestic product (GDP). The initial price shock acts as a negative demand shock, which catches firms by surprise, resulting in higher-than-anticipated inventory builds. Firms desire to smooth their output and use existing inventories to spread production out over time, thereby delaying the reduction in GDP. The intuition underlying the basic theory is that, given a non-zero and negative price elasticity of the demand for oil, information concerning a future shortfall of oil relative to demand will cause an increase in the demand for oil inventories and, hence, the real price of oil [2]. At the same time, knowledge of oil supplies heading into inventories today will reduce the rational expectations of oil prices next year [4]. The relationship is so well established that a simple announcement of a change in inventories can be shown to have significant and lasting implications for world oil prices. Bu [5] investigated the effect of EIA inventory reports on crude oil futures prices and found that negative shocks to inventory information lead to an increase in crude oil prices.

As a matter of fact, one of the International Energy Agency’s (IEA) initial raison d’etre was to ensure the availability of oil inventories sufficient to cover several weeks of demand of their respective economies and coordinate these inventories against intended or unintended disruptions in global oil supply. Many other non-IEA members also actively monitor and manage their oil inventories. China, for instance, initiated the development of the SPR strategy in 2003. Oil exporting countries have also been developing their oil storage capacities—oftentimes in collaboration with the importer via joint oil stockpiling agreements.

The first trial use of the US SPR was held in September 1990—fifteen years after its launch in December 1975—when the Iraqi invasion of Kuwait triggered an oil market shock and sent crude prices soaring. The sale of five million barrels demonstrated the readiness of the US SPR to respond to emergencies. Months later, the SPR released twenty-one million barrels to calm markets during Desert Shield and Desert Storm operations in Iraq [6]. More recently, sanctions against Russia following the Ukraine-Russia conflict have been reflected in soaring spot and future prices of oil and gas, rising 29% and 27%, respectively, from 24 February to 1 June 2022 [7]. Led by the USA, the IEA responded immediately to the crisis, announcing the release of 240 million barrels of oil from April to August 2022 [8]. This was the largest stock release in IEA history to date.

Generally, oil inventories tend to exhibit two distinct responses to supply shocks. In the case of a negative shock, or disruption to oil supplies, oil inventories can be released onto world oil markets to cover shortages and help to reduce prices, creating a buffer effect. Alternatively, they can be hoarded now and sold at higher prices later, creating a speculative margin. The latter, hoarding, is more likely to occur for speculative players (and commercial storage) when the forward curve is in Contango or markets are tight [2]. Certain inventories, such as SPRs, may exhibit both responses as both emergency and commercial-purpose stocks in the form of crude oil and refined products [9]. Such a heterogeneous structure of inventories, in turn, may lead to varied responses of oil inventories to supply shock. Accordingly, Caffarra [10] identified two historical behavior patterns among the IEA countries in response to a supply shock. Hoarders, like Germany and Japan, tended to hold significant reserves but preferred to hold the inventories as a last line of defense. On the other hand, buffers, like the United States, tended to rely on speculative behavior and commercial stock drawdowns but were compelled to take the lead on SPR drawdowns, lobbying the other parties involved to follow suit.

Consequently, the way importers have utilized their oil inventories to counter supply shocks has varied depending on the current conditions of the global oil market, the ownership structure of their oil inventories, and the nature and magnitude of the shock. To illustrate, the net effect of the July 1990 supply shock differed significantly across regions. Japan lost approximately 10% of its oil imports and took immediate steps to increase imports from Iran, augmenting its stockpile ceiling from 300,000 to 580,000 b/d in the first few months following the shock. The United States relied upon inventories to cover the gap, drawing down 10% of its commercial crude oil inventory immediately following the shock. The US drawdowns reduced oil prices in Europe, facilitating hoarding behavior and a net increase in stockpiles in the EU of 2,250,000 b/d over the same time period.

Recently, the occurrence and magnitude of economic and geopolitical shocks impacting the global oil market, such as the aftermath of the COVID-19 pandemic, the conflict in Ukraine, and the sanction pressure on oil-producing countries, have intensified, putting a spotlight on subsequent reactions of market indicators. In the case of crude oil stocks, historical patterns not seen since the 1990s have begun to repeat themselves, carrying valuable lessons for world oil analysts. These disruptions, however, occur in the context of a significantly evolved global oil market, which has become more globalized, arguably more efficient and frictionless, and much more driven by the “paper” or financial component. Thus, despite the repeating nature of economic and geopolitical shocks, the response patterns can vary significantly, both geographically—across different countries and market participants, and historically—different reactions to similar events by the same market players.

In this uncertain environment, inventory level is seen as one of the major indicators of the global oil market by analysts and policymakers [11,12]. A number of studies have explored the impacts of changes in oil inventories on market performance and macroeconomic indicators. Ye et al. [13] apply low- and high-inventory variables to improve forecasting short-term oil prices. Gong et al. [14] find that oil inventories have more significant effects on oil price fluctuation than the aggregate supply and demand. Another strand of research uses changes in oil stocks as a proxy for supply or demand shocks, exploring their impacts on oil prices [2,15], their volatility [16], and their spillover effects across other commodity markets [17,18]. On the macroeconomic level, the shocks represented by the changes in oil inventories have been discovered to affect real housing returns [19], the banking sector [20], the performance of stock markets across the globe [21,22], and macroeconomic uncertainty in general [23]. Even the announcements of changes in oil stocks were found to have significant implications for intraday market moves [24] and volatility of the futures price [5].

Despite a significant body of research dedicated to the impacts of oil inventories on market and macroeconomic indicators, our knowledge of how inventories themselves reflect changes in market fundamentals remains scant. Alsahlawi [25] analyzed the shift of oil inventories’ role from smoothing seasonal variations in supply and demand to a more active market management tool, which was driven by political, financial, economic, and speculative actions. Pirrong [26] and Gao et al. [27] explored the impact of oil price volatility on changes in inventories. Kim et al. [28] used a GVAR framework to establish that the buffer responses of inventories—globally, in the US, and in Cushing, Oklahoma—tend to be immediate for demand shocks but gradual for supply shocks. Zhao et al. [29] also utilized the US oil stocks as a benchmark for the inventory variable but found that the buffering effect of inventory tends to appear in the long term, while the short-term dynamics are primarily driven by the speculation effect. The authors used a hybrid Wavelet-ARDL-SVR (WAS) model to demonstrate that speculation on the supply side is more likely to cause market risk. Besides the lack of consensus on the drivers and time horizon of the buffer and speculative patterns of oil inventories, contemporary modeling literature on the subject is limited in geographical or country coverage—primarily focusing on the US or global stocks. However, as shown in earlier studies, countries have demonstrated varied approaches to managing their inventories, especially in response to oil market shocks [10].

Hence, understanding how oil inventories respond to various shocks remains a timely and relevant question for scholars, policymakers, and market players alike. This study aims to promote a systematic understanding of the problem by examining the potential responses of oil inventory levels to supply shocks by modeling such impacts for specific economies and under various market conditions. In doing so, we specifically ask how oil inventories of IEA-member countries respond to various supply (economic and geopolitical) shocks and whether these responses are conditioned by the state of global oil markets and the location of the supply source being shocked.

Our intended contribution to the existing literature on the subject includes the application of the KAPSARC GOVAR (Global Oil and Inventory Vector Autoregression) model to investigate the response of oil inventories to supply shocks. This GOVAR extends Mohaddes and Pesaran’s [30] canonical GVAR model, allowing us to model the complex relationship between oil supply shocks and country-specific oil inventories, taking global macroeconomic dynamics and interdependencies in global oil production into account. To the best of our knowledge, the model we construct for this study is the first one that includes both global and country-specific OECD oil inventories in a GVAR model. We expand the geographical coverage of Mohaddes and Pesaran [30] by adding three additional oil producers at the country level: Russia, Iran, and Venezuela. We extend the temporal coverage to 2022Q2, making our model one of the most up-to-date open-access GVAR models at the time of writing (The open access GVAR Modelling site has an updated data set to 2023Q3). Finally, the model runs from 1979:Q2 to 2022:Q2, allowing us to compare the impact of oil supply shocks on inventories under different market states and conditions. These extensions allow us to simulate global and supplier-specific crude oil shocks under tight and loose oil market conditions and assess anticipated global inventory responses as well as specific responses of selected OECD countries.

This paper proceeds in four sections. Section 2 presents the modeling methodology, the GOVAR framework—a global (panel) vector autoregression model that underpins our analysis—with expanded variables (including oil inventories) and geographic scope. Section 3 contains the simulation results of oil supply shocks to global production and three major producers: Iran, Mexico, and Russia. The implications of these shocks to global inventories and oil stocks held by individual IEA members are analyzed and compared for tight and loose market conditions. Section 4 summarizes our conclusions and a discussion of the policy implications of the study.

2. Methodology

2.1. Incorporating Oil Inventories into Global Oil Models: The Promise of Vector-Autoregression Models

The challenge of modeling the complex interrelationships in world oil markets has been a perennial problem for world oil analysts. Crude oil supplies, demand, and inventories are located in a variety of different regions with distinct political and legislative regimes. As the industry evolves towards a transition to net zero, potential substitutes for energy sources are quickly becoming more and more relevant to forecasts. The problem becomes increasingly intractable. A wide variety of models have been used for economic analysis. These include statistical methods such as maximum likelihood and Bayesian estimation, vector autoregression models (VAR), dynamic stochastic general equilibrium models (DSGE), hybrid wavelet and artificial intelligence systems, and dynamic panel models (DPE) [31,32,33,34,35,36].

Of these, the VAR and self-exciting threshold auto-regressive (SETAR) models have been shown to perform exceedingly well. The VAR models have been shown to outperform futures and random walk forecasts for horizons of up to two years [37,38]. The SETAR model has been shown to outperform the VAR, ARIMA, and random walk forecasts for in-sample analysis and out-of-sample forecasting [39]. The SETAR model has also been found to be the most accurate model for in-sample forecasting when compared to univariate models but only second best when compared to mixed-data sampling models (MIDAS), which allow for the including of high-frequency variables when forecasting a lower frequency dependent variable [40]. Another notable refinement of the standard VAR models is the combination of standard factor-augmented VAR (FVAR) and error correction models (FECM). The combination lends more forecasting precision to forecasting models with large datasets [41]. Ratti and Vespignani [42] investigate the relationship between oil prices, global industrial production, central bank policy and interest rates with a global factor-augmented error correction model.

While a VAR model measures interdependencies across multiple time series within a given region or country, a global vector autoregression (GVAR) can estimate the co-movements of variables in and across multiple countries. To be precise, a cointegrating global vector autoregression structure allows for the cointegration of common variables within and between countries. In the GVAR modeling framework, individual country-specific models are combined to create a complex system in which most variables are treated as endogenous, and interactions between economies with different political systems are all accounted for and quantified. The curse of dimensionality is handled in a theoretically consistent manner by estimating the individual country-specific error-correcting models separately, conditional on foreign variables. The model was introduced in 2004 by Pesaran et al. [43] to model regional interdependencies and to analyze credit risk [44,45]. The model was adapted for forecasting one year later, in 2009 [46]. The GVAR has an appealing quality in that, under reasonable conditions, the country-specific models can be derived as large N approximations to global factor augmented models of different forms and can be estimated consistently [47].

The GVAR methodology has been criticized for its failure to give system-wide shocks a simple economic interpretation, such as demand, monetary policy, or supply shocks. This methodological inadequacy has already been overcome. Smith [48] illustrates that a GVAR system can be combined with a dynamic stochastic general equilibrium to separate the contributions of demand, supply and monetary policy and produce identified shocks.

The modeling framework of the GVAR enables the analysis of potential spillover effects from economic shocks and sanctions [49,50]. To cite only a few examples, Sznajderska [51] employs a GVAR model to estimate the spillover effects of a negative demand shock in China on global GDP growth. Kempa and Khan [52] analyzed the spillover effects of public debt and economic growth in the Euro area. They find that debt shocks do not impede growth trajectories but tend to raise debt levels across the euro area. Salisu et al. [53] investigated the spillover effects of financial uncertainty in the United States using a GVAR framework in which uncertainty shocks to developed and emerging economies are conditional on the state of the Global Financial Cycle.

To assess the effect of external economic and political shocks on oil markets, we adopt a counterfactual analysis approach. We utilized the KAPSRAC global vector autoregression oil and inventory model (GOVAR), a specific GVAR model developed to assess the effects of hypothetical economic and political shocks on the global oil market [54].

The model builds upon the GVAR model developed by Mohaddes and Pesaran [30], which covers 33 countries quarterly from 1979Q4 to 2015Q4. Our additions to Mohaddes and Pesaran include (i) the modification of the price equation and country-specific models to incorporate two new variables—OECD oil inventories and the Critical Minerals price index, (ii) the extension of the dataset to 2022Q1, (iii) the addition of three major oil-producing countries, Iran, Russia, and Venezuela, and (iv) the expansion of the trade weights and linking matrices to 2022Q1 and 2021, respectively.

The resulting model characteristics make it particularly suited to this analysis: First, the expanded GVAR framework captures the interactions between many countries. Second, world oil supplies and inventories are modeled jointly with key global and country-level macroeconomic variables. The ability of the GOVAR model to assess the implications of geopolitical shocks on oil inventories has been established by a number of publications, including the evaluation of the time sensitivity of crude oil shocks under tight and loose market conditions [54] and the assessment of regional spillover effects of trade and/or financial sanctions on an oil producing nation [55].

2.2. The KAPSARC GVAR Model

The GVAR modeling approach consists of two main stages. Stage 1 involves the specification of country-specific models, and Stage 2 combines the individual models into a single global representation. For Stage 1, we assume a panel structure with $i=\mathrm{1,2},3,\dots .36$ countries (see Table 1). Each individual country model is made up of a $k_i$ dimensional vector of endogenous variables ${\mathit{x}}_{\mathit{i}\mathit{t}}$ that follows a simple VAR model structure with $p_i$ lags. The model is complemented with a set of contemporaneous lagged exogenous regressors ${\mathit{x}}_{\mathit{i}\mathit{t}}^{\mathrm{*}}$ that follows a simple VAR model structure with $q_i$ lags. These star or foreign variables are assumed to be weekly, exogenous, and represent the effects of foreign variables or influences on the country-specific domestic economy. Finally, two global variables, the oil price, and a critical mineral price index, are added to the conditional country models and follow a simple VAR structure with ${\mathit{s}}_{\mathit{i}}$ lags. The number of lags ($p_i,q_i,{\mathit{s}}_{\mathit{i}}$) are all allowed to vary and can be selected by the Akaike information criterion (AIC), or Schwarz information criterion (SBC) methodologies [56].

Table 1. Countries, regions, and variables in the GVAR model.

Countries Utilized in the World Oil and CM Model
Argentina	Indonesia	Russia
Australia	Iran	South Africa
Austria	Italy	Saudi Arabia
Belgium	Japan	Singapore
Brazil	Korea (South)	Spain
Canada	Malaysia	Sweden
China	Mexico	Switzerland
Chile	Netherlands	Thailand
Finland	Norway	Turkey
France	New Zealand	United Kingdom
Germany	Peru	USA
India	Philippines	Venezuela
Regions and subregions accounted for in the world oil and rare earth metal model
Net oil exporters	Europe	Latin America
Euro area	Net oil importers	Asia Pacific
Rest of world
Global variables in the world oil and rare earth metal model
World oil price	Rare earth metal price
Country-specific variables in the world oil and rare earth metal model
Real GDP	Oil inventories	Real exchange rates
Inflation	Crude oil production	Short-term interest rates
Real equity prices		Long-term interest rates

Table 1. Countries, regions, and variables in the GVAR model.

Countries Utilized in the World Oil and CM Model
Argentina	Indonesia	Russia
Australia	Iran	South Africa
Austria	Italy	Saudi Arabia
Belgium	Japan	Singapore
Brazil	Korea (South)	Spain
Canada	Malaysia	Sweden
China	Mexico	Switzerland
Chile	Netherlands	Thailand
Finland	Norway	Turkey
France	New Zealand	United Kingdom
Germany	Peru	USA
India	Philippines	Venezuela
Regions and subregions accounted for in the world oil and rare earth metal model
Net oil exporters	Europe	Latin America
Euro area	Net oil importers	Asia Pacific
Rest of world
Global variables in the world oil and rare earth metal model
World oil price	Rare earth metal price
Country-specific variables in the world oil and rare earth metal model
Real GDP	Oil inventories	Real exchange rates
Inflation	Crude oil production	Short-term interest rates
Real equity prices		Long-term interest rates

$${{\mathit{\Phi }}_{\mathit{i}}\left(\mathit{L},{\mathit{p}}_{\mathit{i}}\right)\mathit{x}}_{\mathit{i}\mathit{t}}={\mathit{a}}_{\mathit{i}\mathit{o}}+{\mathit{a}}_{\mathit{i}1}\mathit{t}+{\Lambda }_i\left(L,q_i\right){\mathit{x}}_{\mathit{i},\mathit{t}}^{*}+{\mathit{\Psi }\left(\mathit{L},{\mathit{s}}_{\mathit{i}}\right)\mathit{\varpi }}_{\mathit{i}\mathit{t}}+{\mathit{u}}_{\mathit{i}\mathit{t}}$$

(1)

where:

• $\begin{array}{l}{\mathit{a}}_{\mathit{i}\mathit{o}},{\mathit{a}}_{\mathit{i}\mathbf{1}}=K\times \\ 1 vectors of fixed intercepts and coefficients on the deterministic time trends.\end{array}$
• ${\mathit{x}}_{\mathit{i}\mathit{t}}=k_i\times 1 vector of country specific domestic variables$
• $$\begin{gathered} {\mathit{x}}_{\mathit{i}\mathit{t}}^{\mathbf{*}}=k_i\times \\ 1 vector of country specific weekly exogeonous star \left(foreign\right) variables \end{gathered}$$
• ${\mathsf{\varpi }}_{\mathbf{t}}$ $=2\times 1$ vector of global variables
• $$\begin{gathered} {\varphi }_i,\dots {\varphi }_{ip},A_{io},\dots ,A_{iq}\equiv \\ {k}_i\times 1 vectors or matrices of fixed coefficients that vary across countries \end{gathered}$$
• ${\mathit{u}}_{\mathit{i}\mathit{t}}=k_i\times 1 vector of country-specific supply shocks$
• $$\begin{gathered} {\mathit{u}}_{\mathit{i}\mathit{t}}~\mathit{i}.\mathit{i}.\mathit{d}.\left(\mathbf{0},{\mathsf{\Sigma }}_{\mathit{i}\mathit{i}}\right)\equiv \\ the shocks are serially uncorrelated with zero mean and non singual covariance matrix \end{gathered}$$
• $$\begin{gathered} {\mathsf{\Phi }}_{\mathit{i}}\left(\mathit{L},{\mathit{p}}_{\mathit{i}}\right)=\mathit{I}-{\sum }_{\mathit{i}=\mathbf{1}}^{{\mathit{p}}_{\mathit{i}}}{\mathsf{\Phi }}_{\mathit{i}}{\mathit{L}}^{\mathit{i}}\equiv \\ the matrix lag polynomial of the domestic variable coefficients \end{gathered}$$
• $$\begin{gathered} {\mathit{\Lambda }}_{\mathbf{i}}\left(\mathbf{L},{\mathbf{q}}_{\mathbf{i}}\right)={\sum }_{\mathit{i}=\mathbf{1}}^{{\mathbf{q}}_{\mathbf{i}}}{\mathit{\Lambda }}_{\mathit{i}}{\mathit{L}}^{\mathit{i}}\equiv \\ the matrix lag polynomial of the foreign variablecoefficients \end{gathered}$$
• $$\begin{gathered} \mathit{\Psi }(\mathit{L},{\mathit{s}}_{\mathit{i}})={\sum }_{\mathit{i}=\mathbf{1}}^{{\mathit{s}}_{\mathit{i}}}{\mathit{\Psi }}_{\mathit{i}}{\mathit{L}}^{\mathit{i}}\equiv \\ the matrix lag polynomial of the global variable coefficients \end{gathered}$$

where ${\mathsf{\varpi }}_{\mathbf{t}}$ is a vector of global or dominant variables and its lagged values. The model is augmented to allow for feedback effects from the domestic variables as follows:

$${\mathsf{\varpi }}_{\mathbf{t}}=\sum _{\mathit{l}=\mathbf{1}}^{{\mathit{p}}_{\mathit{w}}}{\mathsf{\Phi }}_{\mathit{w}\mathit{l}}{\mathsf{\varpi }}_{\mathbf{i},\mathbf{t}-\mathbf{l}}+\sum _{\mathit{l}=\mathbf{1}}^{{\mathit{p}}_{\mathit{w}}}{\mathsf{\Lambda }}_{\mathit{w}\mathit{l}}{\mathit{x}}_{\mathit{i},\mathit{t}-\mathbf{1}}^{*}+{\mathit{\eta }}_{\mathit{w}\mathit{t}}$$

(2)

where:

• $$\begin{gathered} {\mathsf{\Phi }}_{\mathit{i}}\left(\mathit{L},{\mathit{p}}_{\mathit{i}}\right)=\mathit{I}-{\sum }_{\mathit{i}=1}^{\mathit{p}\mathit{i}}{\mathsf{\Phi }}_{\mathit{i}}{\mathit{L}}^{\mathit{i}}\equiv \\ the matrix lag polynomial of the global variable coefficients \end{gathered}$$

Once again, the lag ${\mathbf{p}}_{\mathsf{\omega }}$ is allowed to vary and can be selected by the Akaike information criterion (AIC) or Schwarz information criterion (SBC) methodologies [56]. The specification does allow for different lag orders for the dominant and foreign variables, while the global variables ${\mathsf{\varpi }}_{\mathbf{t}}$ are treated as a foreign variable for the purposes of modeling and share the (q) lag order.

In order to allow for country-by-country estimation, we assume that ${\mathit{u}}_{\mathit{i}\mathit{t}}$ and ${\mathit{u}}_{\mathit{j}\mathit{t}}$ are independent $\forall$ i and j, so that ${\mathsf{\Sigma }}_{\mathit{i}}=\mathit{d}\mathit{i}\mathit{a}\mathit{g}\left({\mathsf{\Sigma }}_{1\mathit{t}},{\mathsf{\Sigma }}_{2\mathit{t}},\dots {\mathsf{\Sigma }}_{36\mathit{t}}\right)$ is a block diagonal matrix. As a result, the co-movements between ${\mathit{x}}_{\mathit{i}\mathit{t}}$ and ${\mathit{x}}_{\mathit{j}\mathit{t}}$ are determined by the foreign or star variables ${\mathit{x}}_{\mathit{i}\mathit{t}}^{\mathrm{*}}$. The ${\mathit{x}}_{\mathit{i}\mathit{t}}^{\mathrm{*}}$ in turn are deterministic functions of a set of pre-specified weights used to construct the foreign variables:

$${\mathit{x}}_{\mathit{i}\mathit{t}}^{\mathit{*}}=\sum _{\mathit{j}=\mathbf{1}}^{\mathbf{36}}{\mathit{w}}_{\mathit{i}\mathit{j}}{\mathit{x}}_{\mathit{j}\mathit{t}}$$

(3)

where the weights satisfy the following conditions: ${\sum }_{j=1}^{36}{\mathit{w}}_{\mathit{i}\mathit{j}}=\mathbf{1}$ and ${\mathit{w}}_{\mathit{i}\mathit{i}}=\mathbf{0}$. The use of country-specific weights allows the model to differentiate the number of country-specific variables in different countries by simply allocating a weight of zero on missing variables and rebalancing the rest of the weights to sum to 1. Given this specification, the contemporaneous relations between ${\mathit{x}}_{\mathit{i}\mathit{t}}$ and ${\mathit{x}}_{\mathit{j}\mathit{t}}$ are determined by the weights used to construct the foreign variables (${\mathit{w}}_{\mathit{i}\mathbf{1}},\dots {\mathit{w}}_{\mathit{i}\mathbf{36}}$) and the coefficient on the foreign variables themselves ${\mathit{\Lambda }}_{\mathit{i}}$.

The covariance matrix ${\mathsf{\Sigma }}_{\mathit{i}}$ can be decomposed as follows:

$${\mathsf{\Sigma }}_{\mathit{i}\mathit{t}}=A_0{H}_t{A}_0^{\prime }$$

(4)

where:

• $A_0\equiv a lower uni-triangular matric$
• $H_t=diag\left(e^{h_{i1,t}},e^{h_{i2,t}},\dots e^{h_{i{k}_i,t}}\right)$
• $h_{i2,t}$ $\equiv$ follows an AR(1) process

Given this specification, the GVAR model allows for interactions among country-specific economies through three channels [44]:

${\mathit{x}}_{\mathit{i}\mathit{t}}$ is contemporaneously dependent on ${\mathit{x}}_{\mathit{i}\mathit{t}}^{\mathrm{*}}$ and lagged values.

The country-specific variables depend on the global variables, oil, and CM prices

Shocks in country i are contemporaneously dependent on shocks in country j as determined by the cross-country covariance matrix ${\mathsf{\Sigma }}_{\mathit{i}\mathit{j}}$, where ${\mathsf{\Sigma }}_{\mathit{i}\mathit{j}}$= $cov\left(u_{it},u_{jt}\right)=E\left(u_{it},u_{jt}^{\prime }\right) for i\ne j$.

In Stage 2 of the GVAR modeling process, we use the weights given in Equation (3) to combine the individual models to a single global representation.

Let:

$${\mathit{z}}_{\mathit{i}\mathit{t}}=\left(\begin{array}{c}{\mathit{x}}_{\mathit{i}\mathit{t}}\\ {\mathit{x}}_{\mathit{i}\mathit{t}}^{*}\end{array}\right)$$

(5)

Equation (1) can be written for each region as:

$${\mathit{A}}_{\mathit{i}\mathit{o}}{\mathit{z}}_{\mathit{i}\mathit{t}}={\mathit{a}}_{\mathit{i}\mathit{o}}+{\mathit{a}}_{\mathit{i}\mathbf{1}}\mathit{t}+{\mathit{A}}_{\mathit{i}\mathbf{1}}{\mathit{z}}_{\mathit{i}\mathit{t}-\mathbf{1}}+\dots {+\mathit{A}}_{\mathit{i}{\mathit{p}}_{\mathit{i}}}{\mathit{z}}_{\mathit{i}\mathit{t}-{\mathit{p}}_{\mathit{i}}}+{\mathit{u}}_{\mathit{i}\mathit{t}}$$

(6)

where:

$${\mathit{A}}_{\mathit{i}\mathit{o}}=\left({\mathit{I}}_{{\mathit{k}}_{\mathit{i}}}-{\mathit{\Lambda }}_{\mathit{i}\mathit{o}}\right),{\mathit{A}}_{\mathit{i}\mathit{j}}=\left({\mathit{\theta }\mathit{\Phi }}_{\mathit{i}\mathit{j}},{\mathit{\Lambda }}_{\mathit{i}\mathit{j}}\right) \mathit{f}\mathit{o}\mathit{r} \mathit{j}=\mathbf{1},\dots {\mathit{p}}_{\mathit{i}}$$

(7)

Then, the trading weights, or aggregation weights, can be defined so that:

$${\mathit{z}}_{\mathit{i}\mathit{o}\mathit{i}\mathit{t}}={\mathit{W}}_{\mathit{i}}{\mathit{x}}_{\mathit{t}}$$

(8)

where $x_t=\left(x_{ot}^{\prime },x_{1t}^{\prime }\dots ,x_{Nt}^{\prime }\right)$ is the k × 1 vector which includes all of the endogenous variables of the system, and $W_i$ is a $\left(k_i\times k_i^{\mathrm{*}}\right)\times k$ matrix.

$${\mathit{A}}_{\mathit{i}\mathit{o}}{\mathit{W}}_{\mathit{i}}{\mathit{x}}_{\mathit{t}}={\mathit{a}}_{\mathit{i}\mathit{o}}+{\mathit{a}}_{\mathit{i}\mathbf{1}}\mathit{t}+{\mathit{A}}_{\mathit{i}\mathbf{1}}{\mathit{W}}_{\mathit{i}}{\mathit{x}}_{\mathit{t}-\mathbf{1}}+\dots {+\mathit{A}}_{\mathit{i}{\mathit{p}}_{\mathit{i}}}{{\mathit{W}}_{\mathit{i}}\mathit{x}}_{\mathit{t}-{\mathit{p}}_{\mathit{i}}}+{\mathit{u}}_{\mathit{i}\mathit{t}} \mathit{f}\mathit{o}\mathit{r} \mathit{I}=\mathbf{0},\mathbf{1},\mathbf{2}\dots ,\mathit{N}$$

(9)

The final GVAR representation is obtained by pre-multiplying Equation (9) by ${\left({\mathit{A}}_{\mathit{i}\mathit{o}}{\mathit{W}}_{\mathit{i}}\right)}^{\mathbf{-}\mathbf{1}}$:

$${\mathit{x}}_{\mathit{t}}={\mathit{b}}_{\mathbf{0}}+{\mathit{b}}_{\mathbf{1}}\mathit{t}+{\mathit{F}}_{\mathbf{1}}{\mathit{x}}_{\mathit{t}-\mathbf{1}}+\dots +{{\mathit{F}}_{\mathit{p}}\mathit{x}}_{\mathit{t}-\mathit{p}}+{\mathit{\epsilon }}_{\mathit{t}}$$

(10)

where ${\mathit{\epsilon }}_{\mathit{t}}~\mathit{N}\left(\mathbf{0},{\mathit{\Sigma }}_{\mathit{e}\mathit{t}}\right)$is a full matrix, and Equation (10) is similar to a large VAR model with ${\mathit{p}}^{*}=\mathit{m}\mathit{a}\mathit{x}\left(\mathit{p},\mathit{q},\mathit{s}\right)$ lags. The algebraic calculations and dynamic properties of the system, including the derivation of persistence profiles and Generalized Impulse Response Functions are shown in Considine et al. [57].

2.3. Empirical Analysis and Model Specifications

The countries and variables included in the GVAR specification are listed in Table 1. The data sources are summarized in Supplementary A and described in detail in Considine et al. [57] and are available on request.

2.3.1. Trade and Aggregation Weights

The foreign or star variables for the GVAR model are calculated using a combination of country-specific trade shares as illustrated by Equations (11) and (12).

$${\mathit{x}}_{\mathit{i}\mathit{t}}^{\mathbf{*}}=\sum _{\mathit{j}=\mathbf{1}}^{\mathit{N}}{\mathit{w}}_{\mathit{i}\mathit{j}}{\mathit{x}}_{\mathit{j}\mathit{t}}$$

(11)

where ${\mathit{w}}_{\mathit{i}\mathit{j}}$, i,j = 1,2…N, are bilateral trade weights with ${\mathit{w}}_{\mathit{i}\mathit{i}}=0, \mathrm{and} {\sum }_{\mathit{j}=\mathbf{1}}^{\mathit{N}}{\mathit{w}}_{\mathit{i}\mathit{j}}=\mathbf{1}$. The trade weights, ${\mathit{w}}_{\mathit{i}\mathit{j}}$, are computed as a three-year moving average to reduce the impact of extreme annual movements on the trade weights.

Specifically:

$${\mathit{w}}_{\mathit{i}\mathit{j}}={\displaystyle \frac{{\mathit{T}}_{\mathit{i}\mathit{j},\mathbf{2019}}+{\mathit{T}}_{\mathit{i}\mathit{j},\mathbf{2020}}+{\mathit{T}}_{\mathit{i}\mathit{j},\mathbf{2021}}}{{\mathit{T}}_{\mathit{i},\mathbf{2019}}+{\mathit{T}}_{\mathit{i},\mathbf{2020}}+{\mathit{T}}_{\mathit{i},\mathbf{2021}}}}$$

(12)

where ${\mathit{T}}_{\mathit{i}\mathit{j}\mathit{t}}$, i, is the bilateral trade of country i with country j during a given year t, and is equal to the average of exports and imports of country i with country j, and ${\mathit{T}}_{\mathit{i}\mathit{t}}={\sum }_{\mathit{j}=\mathbf{1}}^{\mathit{N}}{\mathit{T}}_{\mathit{i}\mathit{j}\mathit{t}}$ (the total trade of country i) for t = 2019, 2020, 2021, and j = 1,2,…N.

In addition, the KAPSARC GVAR specification allows for the analysis of regional responses to shocks. We define the following regions—Europe, the Euro Area, net oil exporters and importers, Latin America, Asia Pacific, and the rest of the world (ROW). Following Dees et al. [58], we use weights based on the PPP valuation of the individual countries’ real GDP for regional aggregation and the derivation of aggregate impulse response functions. The PPP method has been shown to be more reliable than weights based solely on US dollar valuations [58].

Finally, the fixed weights used to construct the feedback variables are: (i) the PPP for the financial-economic variables, real GDP, inflation, real equity prices, real exchange rates, and short- and long-term interest rates; (ii) contributions to OECD inventories for the inventory variable; and (iii) contribution to total oil production for the crude oil production variable. The weighting system is derived as follows:

$$\begin{array}{c}{Y_t}^{\mathrm{*}}={\displaystyle \sum _{\mathit{j}=\mathbf{1}}^{\mathit{N}}}{\mathit{\omega }}_{\mathit{j}}^{\mathit{Y}}{\mathit{Y}}_{\mathit{j}\mathit{t}}\\ {Qs}_t^{0\mathrm{*}}={\displaystyle \sum _{\mathit{j}=\mathbf{1}}^{\mathit{N}}}{\mathit{\omega }}_{\mathit{j}}^{\mathit{Q}}{\mathit{Q}\mathit{s}}_{\mathit{j}\mathit{t}}\\ {Dp}_t^{\mathrm{*}}={\displaystyle \sum _{\mathit{j}=\mathbf{1}}^{\mathit{N}}}{\mathit{\omega }}_{\mathit{j}}^{\mathit{Y}}{\mathit{D}\mathit{p}}_{\mathit{j}\mathit{t}}\\ I_t^{0\mathrm{*}}={\displaystyle \sum _{\mathit{j}=\mathbf{1}}^{\mathit{N}}}{\mathit{\omega }}_{\mathit{j}}^{\mathit{I}}{I}_{jt}^{\mathbf{0}}\end{array}$$

(13)

where:

${\mathit{\omega }}_{\mathit{j}}^{\mathit{Y}}$ is calculated as a three-year average of 2019, 2020, and 2021 of the PPP GDP weights of country j, and ${\sum }_{\mathit{j}=\mathbf{1}}^{\mathit{N}}{\mathit{\omega }}_{\mathit{j}}^{\mathit{Y}}=\mathbf{1}$.

${\mathit{\omega }}_{\mathit{j}}^{\mathit{I}}$ is calculated as a three-year average from 2019Q1 to 2021Q1 of quarterly weights of country j in terms of its contribution to OECD inventories, and ${\sum }_{\mathit{j}=\mathbf{1}}^{\mathit{N}}{\mathit{\omega }}_{\mathit{j}}^{\mathit{I}}=\mathbf{1}$.

${\mathit{\omega }}_{\mathit{j}}^{\mathit{Q}}$ is calculated as a three-year average from 2019Q1 to 2021Q1 of quarterly weights of country j in terms of its contribution to total oil production from the producing countries listed in the GVAR model, and ${\sum }_{\mathit{j}=\mathbf{1}}^{\mathit{N}}{\mathit{\omega }}_{\mathit{j}}^{\mathit{Q}}=\mathbf{1}$.

2.3.2. Empirical Estimation, Unit Root, Weak Exogeneity and Stability Tests

Given quarterly estimates of the data for the domestic variables from 1979Q2 to 2022Q1, we estimate the 36 individual country-specific models. The modeling exercise assumes that the country-specific foreign variables are weakly exogenous variables and that the parameters are stable over time. Unit root tests performed on the variables utilized by the GVAR show that the variables utilized in the model are integrated into order one. The unit root tests, weak exogeneity, and structural stability test results are reported in Supplementary Online A.

The stability of the GVAR system can be verified by the eigenvalues. To be specific, the model has 165 endogenous variables with 59 cointegrating relationships, so that at least 165 − 59 = 106 must lie on the unit circle for the system to be stable. In fact, the system has 108 eigenvalues on the unit circle, and all of the remaining values are less than one. This suggests that the system as a whole is stable and that some shocks can be expected to have permanent effects on the endogenous variables. The price shock scenarios are run without restrictions on the equations or parameters. The selection of lag orders, cointegrating relationships, and persistence profiles is described in detail in Supplementary Online B.

2.3.3. Scenarios

To assess the time profile of the effects of supply shocks on oil inventories, we investigate the implications of a shock to the oil output of oil production under different oil market conditions: “tight” and “loose” markets.

We define loose markets as having crude oil inventories above the 5-year moving average and a forward curve in a state of backwardation [54]. These conditions were seen in Q1 2018 when the ICE Brent futures curve traded was trading in backwardation at levels around USD 67.00/bbl., and OECD commercial inventories had reached levels as high as 2.870 mb, a 53 mb surplus over the five-year average in February [59]. At that time, a one standard deviation shock in global oil production was equivalent to a reduction of 1.24% or 1.2 MMB/d on a quarterly basis.

We define tight markets as having crude oil inventories below the 5-year moving average and a forward curve in a state of backwardation [54]. The tight market scenario is represented by the world oil markets in Q1 2022 when OECD industry stocks were estimated at 2621 mb, 35.6 mb below the five-year moving average. At the same time, ICE Brent oil futures had shifted into contango in January 2022, with the ICE front-month January Brent contract trading at a discount to February [60]. With sanctions on Russian crude oil supplies on the horizon, Brent crude oil prices reached a high of USD 140/bbl on 8 March 2022 [61]. In this scenario, a period of high volatility in world oil markets, a one standard deviation reduction in global oil production, is equal to a reduction of 1.37% or 1.4 MMB/d on a quarterly basis.

In the first set of scenarios, we shock the global oil production by applying a one standard deviation reduction in output under tight and loose market conditions. The second set of scenarios estimates the impacts of country-specific supply shocks using the examples of Iran, Mexico, and Russia under the same market condition assumptions.

The implications of these counterfactual scenarios are examined by means of a GIRF, which considers only the shocks to individual country-specific error terms and integrates the effects of all other shocks using the observed probability distributions of all of the shocks. In the cases where there is no a priori belief set or information concerning the ordering of countries or variables in the GVAR system, the GIRF can provide valuable information about the transmission of shocks from the sanctions countries to the rest of the world.

3. Results and Discussion

3.1. Global Supply Shock Scenario

The first set of scenarios to be considered applies a one standard deviation reduction in world oil production under the tight and loose market conditions described in Section 2.

Figure 1a,b show the simulated impact of the global oil production shock on oil inventories. The Y-axis indicates deviations from the baseline levels of oil inventories for each particular time period represented by quarters on the X-axis. The solid line represents the median simulation within a 90% confidence interval displayed between the dashed lines. The median cumulative changes in inventories arising from the shocks and statistical significance are reported in

Table 2. Tests of statistical significance for a negative shock to global oil production in tight and loose markets.

Location	Tight Market				Loose Market
	Median Cumulative Changes, %			Significance *	Median Cumulative Changes, %			Significance *
	Year 1	Year 2	4 Years		Year 1	Year 2	4 Years
World	0.46%	−0.59%	−1.50%	A	0.22%	0.02%	−0.11%	-
Japan	0.23%	−0.21%	−0.42%	A	0.24%	0.16%	0.57%	-
South Korea	−1.81%	−3.78%	−11.38%	-	−2.07%	−3.44%	−14.05%	D
UK	−0.27%	−1.23%	−3.94%	G	−0.57%	−1.19%	−5.32%	-
USA	0.90%	−0.36%	−0.62%	B	0.55%	0.43%	1.55%	A

Notes: Median cumulative changes after one year in %, * refers to 90% confidence intervals. A—Statistically significant in one quarter. B—Statistically significant in two quarters. C—Statistically significant in three quarters. D—Statistically significant in four quarters. E—Statistically significant in five quarters. F—Statistically significant in six quarters. G—Statistically significant in 7 quarters. H—Statistically significant for at least 2 years.

. Following Dees et al. [58], we focus the discussion on the first two years after the shock as a reasonable time horizon over which the model presents credible results.

The simulation results demonstrate a clear statistically significant relationship between oil production shocks and inventory levels under tight market conditions. The observed dynamics shown in Figure 1a, globally and most oil importers, follow the expected reaction pattern: there is an initial increase in inventories in the first two quarters as market players anticipate higher future prices and potential supply shortages. After the initial increase, inventories are drawn down in the third quarter. The impacts are strongest in the second quarter, with a subsequent decline after 4 or 5 quarters, as the shock is absorbed by the system and begins to die out (It is important to notice that the trade weights for the loose market scenario were calculated using the trade flows from 2019–2021 through COVID, US–China trade tensions and deglobalization and the Russian invasion of Ukraine. Trade weights for the loose market scenario, on the other hand, were calculated using trade weights from 2015–2018, which experienced the rise of US Shale and the 2016 price collapse. While the spillovers may be over or underestimated relative to ‘today’s’ market, the weights were specifically designed to emulate conditions in two different states of the market—tight and loose markets—and not to represent or reflect current interdependencies).

Not surprisingly, and as demonstrated by Caffarra [10], the reaction differs in different countries around the globe. Thus, inventories in the United States rose by approximately 1% during the first year following the shock. A similar trend, albeit of a lower magnitude, is observed in Japan. The UK and South Korea, on the other hand, see their inventories depleting as soon as the first year after the shock.

The presumptive trajectory of oil inventories under the loose market conditions follows a similar pattern across the observed locations, although with less pronounced swings (see Figure 1b). However, as shown in Table 2, the simulation results are not statistically significant for the world and most countries under investigation. Interestingly, the results for the US and South Korea are significant at the 90% confidence level. Crude oil inventories in the net oil importing countries rose by 0.41% in the first year following the shock, as the net importers (hoarders) took advantage of ‘relatively’ lower oil prices to increase strategic petroleum reserves. Inventories in the Asia Pacific fell by 0.37% in the first year, primarily due to sharp declines in South Korea. The impulse response of the UK and South Korea to the global supply shock is similar under tight and loose market conditions.

In 1990, during the first test of the US SPR, different patterns emerged in at least three major IEA regions: the US, East Asia, and the Net Importers, which are dominated by Europe. It is interesting to notice that the UK and South Korea demonstrate different reactions to the global supply shock. Under both tight and loose market conditions, their inventories tend to deplete substantially—without the anticipated initial uptick. Moreover, these results are statistically significant for both the tight and loose market scenarios. Both countries react immediately to the loss of suppliers during the supply shock, drawing down inventories as a buffer effect to cover the gap.

Interestingly, such discrepancies cannot be explained by either the geographic location or the state of the oil export-import balance of the observed countries. Japan and Korea—both located in Asia Pacific and heavily dependent on imports—demonstrate varied inventory responses. Likewise, more balanced from the international oil trade perspective, the UK and the US react differently, especially under loose market conditions. Japan, the US, and the Net Importers exhibit hoarding behavior, purchasing new inventories immediately after the shock and drawing them down later in the forecast period. In the case of tight markets and contango, the effect is considerably more pronounced as inventories (speculative or commercial) can be sold at a higher price.

3.2. Impacts of Country-Specific Shocks

The output of scenarios discussed in the previous section shows that the response of oil inventories to supply shocks differs by the location of stocks. The next set of oil market shock scenarios explores the variety of impacts depending on the producing country whose supply is being shocked. To illustrate this thesis, we shock the oil output of three producing nations—Iran, Mexico, and Russia under the same tight and loose market scenario assumptions. The resulting magnitude of simulated production cuts is shown in Figure 2.

In order to assess the impacts of oil supply disruptions in various producing countries, we focus on a single storage location. A one-standard-deviation supply shock in a particular country instead of the global market aggregate results in smaller volumes of oil being taken off the market. This approach can provide insights into a potential market for crude oil and the policy drivers that guide inventory responses in that specific country. South Korea was chosen as a representative storage location for this analysis for a number of reasons. Specifically, the country has a balanced mix of strategic and commercial reserves and has reported significant imports from all three of the supply-shocked countries: Iran, Mexico, and Russia. For the simulation results across other storage locations and their statistical significance, please see Supplementary B.

South Korea holds the largest oil inventories of all of the Asia Pacific countries, with 247 days of forward demand cover of net oil imports in May 2023—over 50 days more than Japan at 193 [9]. Two acts provide the legislative framework for Korea’s oil emergency response strategies: the 2017 Petroleum and Petroleum Substitute Fuel Business Act (PAPSA) and the 2016 Energy Act (EA). Emergency measures and drawdowns can be triggered at any time a supply shock threatens Korea’s public order or national economy [9].

At the same time, the country has a vibrant commercial oil storage industry. Over 65% of Korea’s 400 million barrels of oil reserves are held as commercial storage, leaving about 35% of the stocks in nine government storage facilities and a number of international joint oil stockpiling facilities [62].

South Korea is the world’s fourth-largest oil importer, with the bulk of its refinery feedstock coming from Saudi Arabia: 33.6% or 56.68 million barrels in January and February 2023. The next largest shares come from Kuwait and the United States, over 10% each, the UAE, Kazakhstan, and Mexico [63]. In recent years, Korea has been divesting itself from Russian crude rapidly despite the USD 60 per barrel price cap in an effort to distance itself from complicated logistical and legislative bottlenecks. Russian oil imports have fallen precipitously by 60% year on year to 20.98 million barrels in 2022 and zero in February 2023. To our knowledge, there have been no crude oil imports from Iran in the first half of 2023.

As shown in Figure 3, South Korea’s oil inventory response to supply shocks differs significantly according to the country of origin of the supply shock and the state of the market at the time of the shock. This can be partially explained by the relative size of the shocked sanctioned country on world oil markets. The one standard deviation of the shock to oil output of Mexico is three to five times less in absolute (barrels) terms than similar cuts in the oil production of Iran and Russia. Figure 3 illustrates the size of a one-standard-deviation oil shock in the three countries under investigation: Iran, Mexico, and Russia. The median cumulative changes in inventories arising from the shocks and their statistical significance are reported in

Table 3. South Korea’s inventory response to a negative oil production shocks statistical significance.

Location of Shock	Tight Market				Loose Market
	Median Cumulative Changes, %			Significance *	Median Cumulative Changes, %			Significance *
	Year 1	Year 2	4 Years		Year 1	Year 2	4 Years
Mexico	1.13%	1.44%	5.24%		1.45%	1.24%	5.30%	A
Russia	−0.68%	−0.58%	−1.50%		0.36%	2.32%	6.39%
Iran	−1.33%	−1.49%	−5.50%	A	−0.97%	−0.92%	−2.73%

Notes: Median cumulative changes after one year in %, * refers to 90% confidence intervals. A—statistically significant in one quarter. B—statistically significant in two quarters. C—statistically significant in three quarters. D—statistically significant in four quarters. E − statistically significant in five quarters. F—Statistically significant in six quarters. G—statistically significant in seven quarters. H—Statistically significant for at least 2 years.

.

In the tight market, supply shocks to oil production in Iran and Mexico tend to have a larger impact on South Korean inventories than supply shocks arising from a disruption to Russian oil flows, while in the loose market case, the magnitude of relevant changes in inventories is much more significant. This pattern mirrors the dynamics of actual real-world oil imports from Russia: in 2018 (the base year for the loose market scenario), Russian oil amounted to 6.9% of total South Korean imports, while in 2022, which represents the tight market, this share has dropped to 2.3% [64]. Similar dynamics have been observed for Iran, whose share in total oil imports of South Korea, following the reimposition of the US sanctions, declined from 4.8% in 2018 to zero in 2022. However, the reaction of oil inventories to a decrease in Iranian output shows a similar trajectory for the loose and tight scenarios—with a higher absolute withdrawal rate in the latter case. Mexico’s share of South Korean oil imports remained relatively stable in recent years, fluctuating in the 3–4% range. The simulated response patterns in oil inventories to shocks in Mexican oil production under tight and loose market conditions are similar both in direction and magnitude but contrast with the case of Iran. The South Korean inventory responses contrast with supply shock from Russia and Mexico under a tight market scenario but exhibit similar patterns under a loose one.

It should be noted that the results are statistically significant in only two cases: Iran (tight market scenario) and Mexico (loose market scenario). The model output demonstrates that such varied effects cannot be solely explained by any potential single factor, such as the country’s total output, its share in oil trade flows, or current market conditions. Rather, the output reflects the complex interconnections and interdependencies captured by the GVAR modeling approach.

In retrospect, the GVAR simulation was extremely timely, with the tight market scenario beginning in 2022Q1, months after the start of the Ukrainian conflict in February 2022, which was followed by an abrupt drop in Russian oil output in Q2 2022 and the sanctions placed on Russian oil (from December 2022) and oil products (from February 2023) [65]. In response to the observed oil output shock and in anticipation of sanctions, oil inventories held by South Korea initially declined in Q2 2022 by −1.6% compared to the Q1 2022 baseline. However, in subsequent quarters—Q3 and Q4 2022—they have substantially rebounded at +4.3% and +3.7% compared to the baseline [9].

As illustrated in Figure 3, this pattern mirrors the simulated trajectory of the actual shock to Russian oil production. Interestingly, oil inventories held by the industry participants and those controlled by the government behaved differently in response to this crisis. Industrial oil stocks demonstrated a significant increase or hoarding effect, starting from +2.4% in Q2 2002 and followed by +20.0% in Q3 and Q4 2022 (all compared to the Q1 baseline). The “public” oil inventories, on the other hand, experienced a buffer-like response, decreasing steadily by −4.9% in Q2, −8.7% in Q3, and −10.7% in Q4 2022, as South Korea released a record 11.65 million barrels of strategic petroleum reserves in coordinated IAE effort [66]. The distinctive responses of commercial vs. government-managed oil inventories emerge as another key factor in country-specific responses to supply shocks.

The reference case for an oil market shock in the “loose” market conditions can be considered the reimposition of the Iranian sanctions after the US’s withdrawal from the JCPOA Agreement in May 2018. The effect on Iranian oil production started to manifest itself in Q3 2018, with an initial 25% drop from the 2017–2018 plateau [67]. Following this supply shock, oil inventories in South Korea sharply decreased in Q3 2022 by −4.3% and somewhat leveled off in Q4 at −2.7% (both compared to Q2 2022 baseline) [9]. This pattern matches the simulated trajectory of response to Iranian oil output shock under loose market conditions, as shown in Figure 3. Notably, in this case, the oil inventories of South Korea managed by the government did not change throughout 2018. The changes in the country’s oil stocks only occurred in the industry/commercial sector. These two cases demonstrate different approaches of government and market players to oil supply shocks depending on the producing country and general market conditions—potentially explaining the variance observed in the GOVAR model output.

4. Conclusions

The study demonstrates that oil supply shocks have an impact on oil inventories, and this impact varies significantly depending on the magnitude of the shock, market conditions, and oil producers, as well as on the location and structure of inventories. The model output suggests that it takes a global negative supply shock—measured as one standard deviation of the historical output—to have a statistically significant effect on the global oil inventories. Similarly, one standard reduction in supply from individual sample countries does not produce such an impact. Moreover, this effect is statistically significant only under a particular state of the market (defined in this study as “tight”) global market conditions: when the future oil price curve is in contango and oil inventories are below their historical average.

The state of the market can also affect the response of individual countries to global supply shocks. Depending on whether oil markets are tight or loose, oil importers can choose either a strategic buffer or speculative hoarding strategy in order to profit from future price increases. As shown in the case of South Korea, these response strategies can also coexist in one country or storage location when the government releases its SPR reserves in an attempt to stabilize the market, whereas the industrial and commercial inventories show a steady increase.

Supply shocks from individual oil-producing countries tend to have a less prominent impact on country-specific oil inventories than global shocks. In short, international oil markets would appear to be fully diversified. Country-specific supply shocks, which may have been devastating in the past (pre-1980s), no longer demonstrate statistically significant implications for global oil inventories irrespective of the market conditions and rarely affect individual oil importers.

Given the reaction of South Korea to country-specific oil supply shocks, it is difficult to pinpoint a particular factor that drives the changes in inventories. Rather, the output of these simulations is defined by a complex set of country-specific and international economic ties, which is captured by the GOVAR modeling framework. The response patterns vary significantly, both geographically—across different countries and market participants, and historically—different reactions to similar events by the same market players.

Complexity and the demonstrated variance in responses to oil supply shocks pose a significant challenge to policymakers. From the perspective of oil importers, the efforts to preemptively intervene in anticipation of market shifts or to stabilize the markets after shocks via micro-managing SPRs can be muted by opposing dynamics in commercial stocks. On the supply side, attempts to constrain the oil output of a country by imposing sanctions may succeed in terms of reducing its domestic oil production, but the implications for global and country-specific markets are extremely hard to predict, leading to a variety of unintended and adverse policy consequences. Such complex dynamics suggest that oil inventories as a policy instrument are more applicable to addressing the “supply availability” rather than the “affordability” component of energy security—both on the country and global levels.

The GOVAR model has the capacity to account for these complex interactions and can be applied as a tool for the assessment of policy and market strategies related to oil supply shocks under various scenario assumptions and market conditions. Potential steps to further refine the model and directions for further research may include expanding the geographic coverage of oil inventories with a particular focus on oil-exporting countries, the addition of refined oil product stocks, and the separation of government-managed and commercial oil inventories.