The systemic risk of US oil and natural gas companies

We analyse the evolution of the systemic risk impact of oil and natural gas companies since 2000. This period is characterised by several events that affected energy source markets: the real effect of the global financial crisis, the explosion of shale production and the diffusion of the Covid-19 pandemic. The price of oil and natural gas showed extreme swings, impacting companies’ financial situations, which, accompanied by technological developments in shale production, had an impact on the debt issuance and on the overall risk level of the oil and natural gas sector. By studying the systemic impact of oil and natural gas companies on risk in the financial market, measured by the ∆CoVaR, we observe that in the most recent decade, their role is sensibly increasing compared to 2000–2010, even accounting for the possible effect associated with the increase in companies’ sizes. In addition, our results show evidence of a decreasing relevance of traditional drivers of systemic risk, suggesting that additional factors might be present. Finally, when focusing on the impact of Covid-19, we document its relevant role in fueling the increase in the oil and natural gas companies’ systemic impact.


Introduction
The oil and gas upstream production sector in the US has undergone a notable change in the last decade, propelled by the shale oil1 and shale gas production boom.From 2000 to 2021, US crude oil yearly production doubled, from 2.13 to 4.08 gigabarrel per year (reaching a maximum of 4.48 in 2019), driven by the exploitation of shale fields, whose share in production shifted from less than 8% at the beginning of the sample to 65% in 2021.In the same period, natural gas production changed from 19.18 to 33.49 trillion cubic feet, and the share of shale gas moved from 30% to 86%.Production companies financed their growth, and the drilling activities engaging in bond issuance were rated with low grades by credit agencies.According to Moody's, at the beginning of 2020, North American oil and gas exploration and production companies had $86 billion in debt, which will mature between 2020 and 2024, and pipeline companies have an additional $123 billion in debt coming due over the same period.2Oil and gas companies are price takers in the global market for crude oil.This means that oil and gas companies that do not pursue risk management via hedging or other financial engineering activities can experience erratic or significant energy price volatility, which, in turn could impact their cash flows (Fusaro, 1998).Both oil and gas prices have experienced large variations over the past decades.For oil, the WTI spot price at Cushing (OK) (weekly, FOB, dollar per barrel -Source: Refinitiv) was 24.23 in January 2000, spiked up to 133.88 in August 2008, then went down to 16.55 in April 2020, and come back to 75.21 at the end of 2021.For natural gas, the Henry Hub spot price (weekly, dollars per MMbtu -Source: Refinitiv) started in our sample as low as 2.42, increased up to 13.42 (a 544% rise) in October 2005, then moved down to the minimum of 1.63 in June 2020, and again up to 3.76 at the end of December 2021, with at least 15 price spikes throughout the whole sample.Several local and worldwide factors have contributed to this, including geopolitical upheavals, the worldwide financial crisis of 2007/08, the Oil Glut of 2014, the increasing concern about climate change and importance of decarbonisation, and the COVID-19 outbreak, to name just a few.In addition, the shale boom has played a crucial role, shortening the payback time of upstream investments but also increasing companies' risk exposure.This has strongly impacted the financial stability of oil and gas companies: from 2015, the number of fallen angels has been increasing. 3The lesson of the financial crisis of 2007/08 is that idiosyncratic shocks can aggregate each other and become systemic under precise conditions.For instance, Acemoglu et al. (2012) show that aggregate fluctuations may originate from microeconomic shocks to firms, while Gabaix (2011) shows that individual firm shocks do not average out if the distribution of the firm size is fat-tailed.Therefore, it is of the utmost importance to study the extent to which turmoil in the oil and gas sector can fuel a new financial crisis and threaten the stability of the financial system.In this paper, we try to provide an answer to this research question.
We investigate the systemic risk of oil and natural gas sector by looking at its evolution over time and at the determinants of the indicator put forward in Adrian and Brunnermeier (2016), the Delta Conditional Value-at-Risk (∆CoV aR).The advantage of the ∆CoV aR is that it enables us to determine the impact of each oil and gas company on a proxy of the financial market (the system) conditional on the distress in the company.Moreover, this methodological approach allows us to control for possible state variables when building the indicator and to correlate the estimated indicators in the cross-section of oil and natural gas companies with possible risk and impact drivers.This second step is of particular interest, as it enables us to identify the drivers of the rise in the oil and gas sector's systemic risk.
The study focuses on a panel of US companies active from 2000 to 2021 in the oil and gas production sector4 and accounts for the role played by a selection of risk drivers, including both market-wide and company-specific variables.Two sub-periods are specified, 2000-2010 and 2011-2021, taking into account the evolution of the shale extraction and the structural break identified in the oil prices by Caporin et al. (2019).The analysis shows that company size plays a relevant non-linear role in shaping the company's impact on the systemic risk of the market: in the first sub-period, only large companies had an impact on the systemic risk, but in the second period, the impact was expanded to small companies as well.Robustness checks confirm the findings.This result confirms the role played by shale extraction in changing the risk structure of oil and gas companies and, through this, its importance as a driver of systemic risk.On the contrary, the role of debt issuance is negligible.
The paper is structured as follows.In Section 2, we review the related literature and we present an overview of the US corporate debt market by sector.In Section 3, we review the methodology for computing the systemic risk measure (∆CoV aR), while in Section 4, we de-scribe the data we use in the empirical analysis.In Section 5, we investigate the drivers used for predicting the ∆CoV aR measure, and in Section 6, we gather the empirical results and our inferences.We close the paper with a robustness analysis, in Section 7, and with final remarks in Section 8.

Literature background and institutional framework
The global challenges faced by the US oil and natural gas sector are becoming increasingly complex.In this respect, it is crucial to investigate the drivers affecting oil and natural gas companies and whether the sector is resilient enough to avoid threatening financial stability.Sadorsky (2001) shows that exchange rates, crude oil prices and interest rates have large and significant impacts on stock price returns in the Canadian Oil and Natural Gas industry.Faff and Brailsford (1999) document a positive and significant impact of oil prices on Australian oil and gas industry equity returns, as well as El-Sharif et al. (2005) who find that gas and oil price have positive effects on UK oil and gas companies.The rig counts have also been studied, taking into account their technology as well as their importance as indicators of the industry's health (Apergis et al. (2021)).Nevertheless, drilling is capital-intensive since companies must finance their investment, which inevitably brings about other elements of risk.Howard and Harp Jr. (2009) suggest that, for a complete evaluation of the risk, company characteristic such as ratios and debt should be taken into account with the drivers mentioned before.
The transition to a lower-carbon economy could make the future of oil and natural gas firms more gloomy and uncertain.From this angle, according to Diaz-Rainey et al. (2021), transition risks could affect the oil and natural gas sectors.These include the following: (i) higher costs in finding and extracting new oil and gas reserves; (ii) low oil prices, as found by Basher et al. (2018); (iii) the falling cost of renewable energy generation; (iv) the switch to electric vehicles, underpinned by technological improvements to batteries and decreasing vehicle and battery costs; (v) environmental-minded investors affecting demand for petrochemicals (see Fama and French (2007)) and (vi) carbon pricing and taxation schemes.Diaz-Rainey et al. (2021) found that the signing of the Paris Agreement had a large negative impact on the oil and gas sector.Monasterolo and De Angelis (2020) found that the after the Paris Agreement, high carbon stocks became less appealing for investors.An eventual and realistic contraction of oil and natural gas demand could reduce the cash flows and the margin of profit for these firms, which are already in distress.
The characterisation of systemic risk, its definition and the identification of methodologies for its estimation is one of the most debated topics in the financial economics and econometrics literature since the advent of the global financial crisis.Following part of the literature, we define systemic risk as an event or a circumstance that could threaten the stability of the financial system.From the measurement point of view, the econometric literature includes several different methodologies that might be considered for the purpose of monitoring the systemic riskiness of a sector by looking at market data.We refer the reader to the survey by (Benoit et al., 2017) for a review of the most relevant systemic risk measures.In this vein, Adrian and Brunnermeier (2016) introduced ∆CoV aR to evaluate the systemic risk of single companies when they enter into a distress state.We chose this measure for our analyses as it allows for the introduction of covariates that, in our setting, will have a relevant role, as we will show in the following sections.Lupu et al. (2021) found that European energy companies enhanced systemic risk spillovers during 2008, early 2009, and 2020.Within the oil and gas sector, the ∆CoV aR has been used in Khalifa et al. (2021) to evaluate the role of oil in driving the systemic risk of the Gulf Cooperation Countries' financial markets, while Tiwari et al. (2020) used ∆CoV aR and MES to show that oil price dynamics contribute significantly more to G7 stock market returns during volatile times than during tranquil times.5 Al-Jarrah et al. (2021) show that traditional models fail to capture the systemic risk of small-middle size banks operating in the Gulf Cooperation Council since they have high levels of economies of scale.In this respect, analysing the systemic risk spillover of the US oil and gas sector is crucial, given the strong dependence on oil prices and the failure of traditional drivers to capture systemic risk.Bond holders could also amplify the spillover on the financial system by withdrawing their position, given the high credit risk of these securities, especially if driven by herding behaviour during market turmoil.
To assess, from a systemic risk perspective, the importance of the US oil and natural gas sector within the US debt market, we retrieve from Dealogic DCM the amount at issuance of US bonds.The most active issuers in the US bond market are governmental issuers and financial corporations. 6In particular, in the period 2000-2010 the amount of bonds issued by financial corporations was 15.3 trillion dollars (T$), followed by the government, with 7.3 T$, and non-financial corporations, reaching a total of 3.4 T$.The debt structure changed after 2010: financial institutions substantially reduced the issuances from 2011-2020.The government led the issuance, with roughly 19 T$, followed by financial corporations at 9.3 T$, while non-financial companies add up to 7.7 T$.The amount at issuance increased from the first (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010) to the second (2011-2020) period as a consequence of the behaviour of the Federal Reserve, which kept down the interest rate.Figure 1 shows the breakdown of the amount at issuance by global industry group for each period.The oil and natural gas companies raised capital of 251 Bn $ in the first period, the fifth sector by issued volume.In the second period, we observe a generalised expansion of bonds' issuance, especially for oil and natural gas firms, which increased the amount of issuance by 134%, reaching 589 Bn $, the fifth sector in terms of amount issued.The financial importance of the oil and gas sector can also be evaluated by looking at its relative rating compared to the other sectors.Figure 2 reports the weighted rating at the sector level.The oil and natural gas sector was less creditworthy in the first period (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010), with a rating averaging at 9.1, which decreases to 8.6 in the second period (2011-2020). 7The evidence presented above about the large amount of debt issued and the low level of creditworthiness of US companies active in the oil and gas sector confirms the importance of undertaking a systemic evaluation.This will allow us to determine how potential risks that originate in the oil and gas sector could spread throughout the entire financial system, affecting its stability.We pursue this line of research by first reviewing the methodology and then we move to the empirical evidence.

Methodology
From a methodological point of view, the ∆CoV aR is estimated with a two-step procedure.Let us denote by X i t the losses at time t for company i, where losses are obtained as the negative of stock price returns.Moreover, X system t is the system loss at time t, and M t a vector of state variables at time t.The first step in the ∆CoV aR estimation is the estimation by means of quantile regression (Koenker, 2005) of the two linear models where the subscript q identifies the estimated quantile, and the superscript system | i highlights the coefficients account for the dependence of the system on oil company i.Following the standard practice, for the system, we consider the quantiles q = 0.99 and q = 0.95, i.e., the 1% and 5% upper tails of losses, while for the company, we consider the same quantiles of the system and we also add the median one, i.e., q = 0.50.
Finally, building on the estimated risk measures, we compute the impact on the system of the distress in company i.This distress is associated with the company return moving from the median level of losses to an upper quantile of losses.∆CoV aR q,t equals ∆CoV aR with the usual choices for the quantile level, i.e. q = 0.95 or q = 0.99.
The ∆CoV aR system|i q,t is expressed in the unit of measure adopted for losses, i.e. percentages, but it is also interesting to show its monetary value.For this purpose, the risk measure is multiplied by a proxy of the company size, expressed in US Dollars, leading to the monetary Delta CoVar: ∆ $ CoV aR system|i q,t .We report the results of ∆ $ CoV aR system|i q,t in Section 7.2.The evaluation of ∆CoV aR system|i q,t provides insights into the company-specific systemic risk, and by comparing this risk measure over time and in the cross-section in a panel of companies, we identify changes in the level and dispersion of risk.Moreover, we specify the drivers of the oil and gas company's systemic risk and infer potential sources of risk that can lead to an increase in the overall systemic risk.

Data and preliminary analyses
We downloaded US oil and natural gas companies' stock returns, weekly-based, from Refinitiv/Eikon.  3 exhibits the price of Crude Oil-West Texas Intermediate and the price of natural gas (Henry Hub Spot price) over time.Given their strong dependence with the cash flows of the oil and natural gas companies, we associate Figure 3 with the number of US-listed companies in the oil and natural gas sector in Figure 4. Overall, the number of public oil and natural gas companies raised in 2000, dropped after the global financial crisis in 2008 and reached a maximum of 137 in 2018; finally, the number plummeted to 97 at the end of 2021 (Figure 4).The first drop might be related to the consequences of the global financial crisis; the more recent contraction in active companies is due to the COVID-19 outbreak and its impact on the real economy.The entire sample is thus characterised by two sub-samples including, first a phase of moderate increase in the number of companies, a second phase with a steeper increase in the companies, and then a third phase with a drop.
The evolution of the number of companies can also be associated with the margin that oil and natural gas firms have on their cost, which depends on the oil and gas prices.Periods of drops in oil and natural gas prices are coupled with reductions in the number of firms.By combining the patterns in the number of active companies and the patterns in the oil and gas prices, we decided to analyse separately the sub-samples 2000-2010 and 2011-2021; such a choice is also coherent with the evidence in Caporin et al. (2019) that identity a structural break in the oil price time series, located at the beginning of 2011.
We first proceed to a filtering step on the downloaded stock return series: we consider all the companies that, in the time period considered, are characterised by the presence of non-missing values for at least two years (100 consecutive observations).Such a choice excludes companies active from 2010 to 2010, or from 2011 to 2021, for short periods of time and, at the same time, keeps track of all firms that entered or left the market during the time period considered.Our sample is thus composed by 127 listed companies in 2000-2010 and 165 firms in 2011-2021.9The table reports the number of firms (absolute value and percentage) and the weekly market capitalisation averaged across time expressed in millions of dollars (M$). We compute, over the cross-section, mean, standard deviation, median, and 10% and 90% quantiles.Sadorsky (2001) has shown that exchange rates, crude oil prices and interest rates each have large and significant impacts on stock price returns in the Canadian oil and natural gas industry.
To consider these effects in the analysis of the systemic risk of US oil and natural gas companies, we select the following state variables to be included in Equations ( 3) and ( 5).First, we use Tbill3M, which is the change in the three-month Treasury Bill rates.This variable measures the attractiveness of the risk-free rate in the US economy.Then, we include Term spread 10Y3M, a measure of term spread, computed as the difference between the 10-year bond yield and the threemonth Treasury Bill rate.The term spread measures the slope of the bond yield curve.Further, we include Credit spread, which is the difference between the ICE Bank of America BBB US corporate index and the Treasury 10-year bond yield.The credit spread monitors the additional risk faced by investors when buying corporate debt in place of a safer government debt.In addition, we introduce Ret M KT , the Standard & Poor's 500 market index return, and Ret W T I , the West Texas Intermediate (WTI) crude oil price return.We also add P rice W T I − P rice BRE , the difference between the WTI and the European Brent oil prices.The difference measures the disagreement in oil price between the two most predominant world oil benchmarks.To monitor the market stress, we introduce ∆ VIX, the change in the Russell volatility index, defined as the implied volatility of a synthetic at-the-money option of the Russell 2000 index.Finally, to take into account the demand of urban consumers, we introduce Inflation rate, a measure of the average monthly change in the prices for goods and services paid by urban consumers.A detailed description of the variables and their sources is reported in the Appendix, Table 10.
For the analyses, the sample is split into two time periods: the first spans from 07-01-2000 to 31-12-2010, the second from 07-01-2011 to 31-12-2021.This aims at evaluating the role played by the shale oil and gas boom and is coherent with the findings in Caporin et al. (2019), who identify a structural break located in early 2011 by studying the long-run relationship in the WTI-Brent oil time series.11 Table 2 reports the descriptive analysis of the state variables in the two time periods.
Except for Tbill3M, which is approaching zero, moving from −9.1× −5 in the first period to −1.3 × 10 −6 in the second period, Term spread 10Y3M, Ted spread and Credit spread, on average, register a contraction in the second period.By contrast, Ret M KT increases (from 0.0001 to 0.0027).Ret W T I shrinks its average value from 0.0022 to −0.0003, while ∆ VIX and P rice W T I − P rice BRE increase their average, reaching a value of −0.0009 and 0.0849 during the second period, compared to −0.0120 and −0.0373, respectively, in the first period.The inflation rate drops from 0.0903 in the first period to 0.0722 in the second.All variables are leptokurtic except for Term spread 10Y3M.Table 3 reports the summary statistics for the market equity losses X i t and for the risk measures of the oil and natural gas firms in the two periods.The time series of each firm are first averaged, and then the mean, standard deviation, median and 10% and 90% quantiles are calculated cross-sectionally.Both V aR i q,t and ∆CoV aR i q,t are obtained by running the quantile regressions in Equations 3, 5 and 5, at quantiles 0.99 and 0.95.We also report V aR Sys q,t , the financial system's value-at-risk, again at quantiles 0.99 and 0.95.The last two columns report the number of firms used in the corresponding period and the number of weeks in each period, respectively.
In the first period (2000 − −2010), the oil and natural gas firms register negative losses (gains) equal to −0.002, on average (−2.9% in terms of annualized average percentage).On the contrary, in the second period (2011 − −2021), the losses rise to 0.003 (3.8% in terms of annualized average percentage).On average, V ar i 95 at firm level equals 0.124, which is roughly comparable with the V ar i 95 of the second period, 0.130.Additionally, the V ar i 99 rises from 0.218 to 0.228.The effect is more pronounced looking at the median; both V aR i 99 and V aR i 95 increase over time, from 0.183 to 0.200 and from 0.113 to 0.121, respectively.∆CoV aR 95 does not change across the two periods (0.018 vs. 0.013).The same considerations apply to the results looking at the 0.99 quantile (0.033 vs. 0.025).Finally, looking at the market capitalisation, we note that firms decisively increase their magnitude across two periods, from $3599 million to $4901 million, on average.12 5 ∆CoVaR and its predictors over time In this section, we investigate the drivers affecting the cross-sectional dispersion of the ∆CoV aR in the two periods under study.In the fashion of ( The table reports the summary statistics of the market equity losses X i t and the systemic risk measures V aR i q,t , ∆CoV aR i q,t and the financial system value at risk V aR Sys q,t at quantiles 99 and 95 percent.First we average the weekly time series across time and then we compute the distribution.
the cross-sectional average of market equity at time t. 13 Finally, we obtain quarterly figures of ∆CoV aR system|i q,t by averaging over the weekly observations within each quarter.
∆CoV aR system|i q,t becomes our dependent variable, which is regressed over lagged state variables, M t−1 , and lagged firm characteristics, X t−1 .Note that the predictors are all lagged by a single quarter.The specification we adopt is the following: The state variables and the firm controls that are included in the regression are summarised in Table 10.The firm characteristics we take into account are the following: the time series of quarterly losses at the q% quantile, obtained by averaging the weekly observation within the quarter.The weekly VaR is estimated by using (3).The time horizon is defined by the quarter QP = Q1,...,Q4 and the year YYYY; Size: computed as the logarithm of the market capitalisation; Debt: computed as the logarithm of the total debt; ROA: the ratio between operating income and total asset income; ROE : the ratio between operating income and common equity; ∆N RIG % : the percentage variation in the number of active oil rigs.
We also define two alternative specifications in Equation 6 by interacting VaR with Size as Equation 7 and interacting VaR with Size 2 as in Equation 8.
13 We use as the dependent variable ∆CoV aR system|i q,t in the baseline model.Table 7 exhibits the results by using ∆ $ CoV aR system|i q,t as the dependent variable.
The two alternative specifications are crucial for spotting possible non-linearities that could affect the dependent variable.Table 4 reports the summary statistics of ∆CoV aR and V aR at the 0.99 and 0.95 quantiles, and of the firm characteristics.We focus on the two periods, Q12000 − Q42010 and Q12011 − Q42021, to identify the impact of the oil and gas sector's structural and technological changes in the relation between predictors and systemic risk measures.
The average value of ∆CoV aR at 95% and 99% remains constant from the first period values, from 0.018 and 0.034 to 0.014 and 0.025 in the second period.The same considerations hold for ∆CoV aR $ at 95% and 99%.The average of VaR at 95% and 99%, slightly rises from 0.120 and 0.213 in the first period to 0.129 and 0.225 in the second period; the effect is stronger for the median.Tbill3M is bounded to zero across the two periods.Ted spread moves from 0.005 in the first period to 0.003 in the second.Similarly, Credit spread reduces to 0.016 in the second period, from 0.02.The oil and natural gas industry portfolio (Oil ) remains constant across the two periods, at 0.005.The first period sees a lower inflation rate compared to the second one, decreasing from 1.16 to 1.33.Ret W T I falls from 0.03 to −0.04, while Ret HenryHub spurred its value from 0.01 to 0.005.If we focus on firm variables, Size remains constant to 20 across the two periods, while the average of Debt rises in the second period to 11.7 from 10.4.ROA raises to 0.136 in the second period, demonstrating an opposite behaviour with respect to ROE, which sinks to 1.4; this reminds us of the importance of using both firms' profitability indicators in the analyses.The number of rig percentage changes ∆N RIGS % slumps in the second period to −0.017.Overall, a preliminary evaluation of the data features of the two samples shows differences, which depend on the changes in the oil and natural gas production structure, and notably the increased shale oil production, but also on the variation in the financial system risk dominated in the first period by the global financial crisis, while the end of the second period suffers from the effect of the first wave of the COVID-19 pandemic.
6 The drivers of systemic risk in oil and natural gas companies  14 We stress that all the drivers are lagged by one quarter, coherently with Equation 6.Since we are dealing with panel estimates, we control for the time-invariant heterogeneity by resorting to a fixed effect estimator.The standard errors are clustered by firm.A shows that the V aR at the 99% and 95% levels has a lower impact on ∆CoV aR in the first period than in the second.In the second period, the V aR at 99%, on average, affects ∆CoV aR with a magnitude equal to 4.9%; the effect is, roughly, the double of the first period.On the contrary, if we look at the 95% V aR, the impact is only slightly increasing.
Looking at the control variables (reported in Table 14 in the Appendix), Ted spread, Credit spread and Inflation rate have a positive and significant impact on ∆CoV aR in all periods.Tbill3M is negative and significant for ∆CoV aR at 95% and 99%, but only in the second period, with coefficients roughly equal to −0.44 in both cases.The impact of the oil and natural gas sector (Oil ) changes across two periods: in the first period, it has a positive effect on ∆CoV aR, but the effect is lower and negative in the second period.Ret W T I and Ret HenryHub have a negative and significant impact on ∆CoV aR at 95% and 99%; in particular Ret HenryHub is significant in all periods, while Ret W T I is relevant only in the second period.
Looking at firm's characteristics, Size has a limited, positive impact on ∆CoV aR, but only at 99% in the second period.Debt is mildly positive and significant only in the first period.
ROE has an oscillating impact: the sign of ROE changes at different quantile levels.Finally, ∆N RIGS % negatively affects ∆CoV aR at 95% and 99%, but only in the first period.Interestingly, a positive trend of ∆N RIGS % shows that the extraction business is expanding.Therefore the cash flow of that activity rewards the shareholders, thus reducing the systemic risk associated with that business.
The system risk is not impacted by the company size.However, the company contribution to the systemic risk might change in the cross-section of companies according to their size, thus impacting on the relation between the company risk measure and the systemic risk measure.
Therefore, the relevance of the company risk measure's impact on the systemic risk might be modulated by the company size.To test this, we report in Panel B the coefficients interacting V aR with Size.The interaction term is positive and significant, while the two variables alone, V aR and Size, have a negative and significant impact on ∆CoV aR.This signals that the size of the companies, per se, reduces the systemic risk, while the company's own risk, weighted by the company size, raises the systemic risk, since the company's risk increases the systemic risk, and the bigger the company, the bigger the effect.However, the overall impact is difficult to identify by simply looking at estimated coefficients, as it also depends on the company's size.Figure 5 helps to visualize the overall impact of the company risk on the systemic risk; the overall effect is due to the interaction between V aR and size for different levels of Size: the figure reports the

Sensitivity analyses and robustness checks 7.1 Oil, Gas and Supporting Activities
The sample of firms considered in the analysis belongs to three subsectors, as reported in Table 1.
We investigate here the extent to which the effect on ∆CoV aR varies across the sub-groups, to detect whether the results are also impacted by the main companies' activity.Panel A of Table 6 shows the regression coefficients, considering only firms in the extraction of crude oil sector.The results indicate that V aR has a significant and stronger impact on ∆CoV aR in the second period, looking at extreme risks (99% quantile).The effect is the opposite, considering the VaR at 95%.
In both periods, the coefficients are significant.Panels B and C show the regression coefficients for firms belonging to the extraction of natural gas and support activities, respectively.In both cases, the V aR is economically and statistically significant only in the second period, showing that these firms are sensibly more risky than in the first period.Particularly, the VaR at the 99% quantile has the highest magnitude across time and subgroup, with a value of 0.07.

Dollar-valued systemic risk
In this section, we verify that our results do not depend on the scale of the dependent variable, and consider the effect of V aR on ∆CoV aR $ as computed in Adrian and Brunnermeier (2016).
The difference with Equation 6 is that the risk measure is now multiplied by a proxy for of the company size expressed (usually) in US Dollars, leading to ∆ $ CoV aR system|i q,t .Therefore, the dependent variable is weighted by the firm size.The empirical findings in Table 7 report a sharp difference in terms of the impact of V aR on ∆CoV aR $ across two periods.The results corroborate our finding in Table 5.The risk metric is not significant in the first period, while the coefficients become significant, and equal to 0.05 and 0.07, for VaR at 99% and 95%, respectively, in the second period.The coefficients in Table 5 are the sensitivities of ∆CoV aR $ with respect to the characteristics expressed in decimal units.For example, the coefficient of 0.0504 on the V aR 99 in the second period implies that an increase in an institution's VaR (say, from 0.05 to 0.06) is associated with an increase in ∆CoV aR $ of 0.0504 decimal points of quarterly market equity losses at 99%.Second, the variable Size is positively significant in all periods and for all quantiles.

Balanced panel
In this section, we test our findings by replicating them with a more balanced panel.Two alternative specifications are adopted: first, the balanced case, in which we choose the listed firms that, in the time period considered, are characterised by the presence of missing values less than 20% of the time.In this case, we consider firms possessing available observations for more than 80% of the time horizon.In the second specification, termed strong balanced, only listed firms that have been running their business in both periods are included in the sample.
The results are collected in Table 8.The first four columns show that the V aR at 95% and 99% has a positive and significant effect on ∆CoV aR, and that the effect is stronger in the second period.By contrast, in the strong balanced case, the V aR has a positive and significant impact on ∆CoV aR, but the impacts seems reduced from the first to the second period, even though the magnitude is higher than the baseline results in Table 5. Since, in the strong balanced case, the sample of companies does not change over time, it includes only those firms who survived the different shocks across the years.At first glimpse, we could incur in a survivorship bias, since we are selecting those firms with solid characteristics with respect to the others. 15If we look at the descriptive statistics for these firms, they are in line with the baseline case, and the Debt surged from 10.59 to 11.65.The profitability ratios are in line with the baseline case, across two periods.For instance, on average, ROE plummeted from 12.54 to 3.6; the ROA rose to 0.3 from 0.2 in the second period and showed contrasting behaviour with the ROE.V aR at 95% (99%) surged to 0.121 (0.212) from 0.109 (0.189).Only the evolution of Size differs from the baseline, as it increases from 20.48 to 21.65.Finally, the ∆CoV aR at 95% (99%) of these firms declines across the two periods from 0.019 (0.037) to 0.015 (0.026).Since the drivers' patterns are comparable with the baseline case, we suspect that the exclusion of newborn oil and gas firms can provide a different picture of the systemic risk of the US oil and natural gas sector.

Removing the COVID outbreak
To corroborate our results, we replicate the regressions, excluding from the sample the period after the COVID-19 outbreak (from Q1 2020 to Q4 2021).With this further robustness, we control for the possibility that our results may be driven by the turbulence induced by the pandemic.The results are collected in Table 9.The empirical findings show that after the  exclusion of COVID-19, the V aR has a positive and significant effect on ∆CoV aR, and the effect is stronger if compared with the results in Table 5.For the V aR at 99% (95%) (columnn I and IV in Table 9), we note that it has an impact equal to 0.061 (0.056) on ∆CoV aR.Columns II and V report the coefficients when we interact V aR with Size.The variables V aR and Size have a negative and significant impact on ∆CoV aR.In contrast, their interaction is positive and significant, as in Table 14. Figure 7 shows the relation between the composite effect across different Size levels.Graphs (a) and (b) report the coefficients' evolution when V aR is at 99% and 95%.The joint effect is significant for firm sizes grater than M$2.6 and M$1.62, respectively.
The impact is comparable with graphs (b) and (d) in Figure 5.
Panel C reports the regression results when we interact V aR with Size and Size 2 .The left plots of Figure 8 have already been reported in Section 6.The effect of size is significant only for firms greater than M$884 (a) and M$14.7 (c) (99% and 95% quantiles).The right part of the panel reports the joint coefficient for the second period.It is always significant, considering that the risk is measured with the quantile at 99%.Nevertheless, when we look at the quantile at 95%, the effect is significant for those firms with Size values greater than M$1.2 (see panel (d)).
Surprisingly, if we compare the 2011-2021 period in Figures 5 and 6 with 2011-2019 in Figures 7 and 8, the latter graphs have a sharper slope.Summarising the main finding, the results are confirmed even if we exclude the COVID-19 period.Indeed, the sensitivity to the size is further increased.16A possible explanation is that the COVID-19 impacted the relation between companies' risk and systemic risk, leading to a structural break.After the pandemic outbreak, a systemic event, the entire economy was affected, and in relative terms, other sectors have Notes: The table reports the results of the fixed effect panel regression in Equation ( 6) by excluding the first two quarters of 2020 (COVID Outbreak).Clustered standard errors are in parentheses.*** p<0.01, ** p < 0.05, * p < 0.1.
contributed more to the overall systemic risk, thus lowering the impact of the oil and gas sector.
Therefore, by excluding the COVID-19 period from the analysis, our results are confirmed and even stronger than those previously reported.

Alternative specifications
Finally, we define two alternative specifications of Equation 6 by interacting VaR with debt as in Equation 9 and interacting VaR with Debt 2 as in Equation 10.This allows us to test whether the debt structure of the oil and gas companies has an effect similar to that of the companies' size in modulating the impact of the companies' risk on the systemic risk.
Although Debt weakly affects ∆CoV aR only in the first period, we investigate whether the company's contribution to the systemic risk may change according to their debt burden.
The company's sensitivity to systemic risk could change according to its debt structure.For this reason, we test the coefficient interacting V aR and Debt.Table 11 in panel A shows the and increasing their level of risk.We show that the company's losses, as summarised by the V aR at 95% and 99%, have contributed to systemic risk (in the sense that they are ∆CoV aR predictors).This effect has been higher in the second period, when oil and gas supply has been mostly driven by shale.Additional results highlight that the size of the company has indeed played a role in the systemic risk, but mostly through the indirect impact on the companies' own risk.Moreover, this effect depended non-linearly on the size of the company and has become smaller in the second period, showing that small companies in the oil and gas sector have also started contributing to the systemic risk after the shale production boom.
Robustness checks confirm these findings.The effect is stronger and the difference between the two periods is higher when we change the measure to compute ∆CoV aR, expressing it in monetary terms.Looking at each sub-sector individually (namely, extraction of oil, extraction of gas and supporting activities, respectively), we show that the natural gas extraction and the support activities sector are the subgroups for which the impact on ∆CoV aR is higher and for which the difference across the two periods are more relevant, compared to the oil sector.The results are still valid if we take into account a balanced sub-panel of firms composed only of companies with available observations more than 80% of the time and also when we exclude the COVID-19 period from the analysis.
Overall, our findings can be of interest to investors, who can rely on this measure of the contribution to systemic risk stemming form the oil and gas sector, as well as from a policy point of view.They contribute to the debate on the consequences of oil and gas production, in general, and the shale production in particular, allowing us to assess the role that it plays not just in the environment and/or the energy supply but also as a driver of systemic risk in the financial system.

Figure 3 :
Figure 3: The figure reports on the left y-axis, the price of WTI (dollars per barrel), defined by the red line, and on the right y-axis the Henry Hub Spot price of natural gas (dollars per MMbtu) over time from 2000 to December 2021.

Figure 4 :
Figure 4: The figure reports the number of US oil and natural gas firms over time from 2000 to December 2021 (dashed line).

Figure 5 :
Figure 5: The figure shows the pattern of the joint coefficient b 1 + b 3 Size t−1 (red line) with respect to Size.The blue dashed lines represent the confidence interval at 99%.

Figure 6 :
Figure 6: The figure shows the pattern of the joint coefficient b 1 + b 3 Size t−1 + b 5 Size 2 t−1 (red line) with respect to Size.The blue dashed lines represent the confidence interval at 99%.

Figure 8 :
Figure 8: The figure shows the pattern of the joint coefficient b 1 + b 3 Size t−1 + b 5 Size 2 t−1 (red line) with respect to Size.The blue dashed lines represent the confidence interval at 99%.

Figure 7 :
Figure 7: The figure shows the pattern of the joint coefficient b 1 + b 3 Size t−1 (red line) with respect to Size.The blue dashed lines represent the confidence interval at 99%.

Figure 9 :
Figure 9: The figure shows the pattern of the joint coefficient b 1 + b 3 Debt t−1 (red line) with respect to Debt.The blue dashed lines represent the confidence interval at 99%.

Figure 10 :
Figure 10: The figure shows the pattern of the joint coefficient b 1 + b 3 Debt t−1 + b 5 Debt 2 t−1 (red line) with respect to Debt.The blue dashed lines represent the confidence interval at 99%.
fixed effect panel regression relative to the entire sample.Panel A reports the coefficients of interest defined by Equation 11 by introducing in the model the interaction between VaR and Size, VaR and Debt and, Size and Debt.Panel B exhibits the coefficients of interest as in Equation 12 by adding to the interaction in panel A the second-order interaction VaR and Size 2 , VaR and Debt 2 d an, Size 2 and Debt 2 .Standard errors clustered by firm are in parentheses.*** p<0.01, ** p < 0.05, * p < 0.1.
The time horizon spans from January 2000 to December 2021.Figure

Table 1
The high standard deviation and the discrepancy between the average and the mean show evidence of the relevant heterogeneity in the sectors' market value.Firms that belong to the extraction of natural gas, have the highest median in the first period, reaching a value of $2125 million, followed by firms in the supporting activities, with a value of $992 million, and extraction of crude oil firms, with $471 million.If we look at the mean, the ranking of the market value remains unchanged.In the second period, the extraction of the crude oil sector increased its dimension, moving to the second highest market cap, with $5087 million, on average.The extraction of natural gas remains the largest sector, with a median capitalisation equal to $5132 million.Notably, focusing on the difference between the mean and median and on the market value dispersion, we see that companies involved in the extraction of crude oil and in support activities are more heterogeneous that the gas extraction companies.

Table 1 :
Firms descriptive statistics by NACE code 4-digits

Table 2 :
State Variable Summary StatisticsThe table shows the weekly-based descriptive statistics of state variables.First we average the weekly-based time series, across time and then we compute the distribution.The table reports the mean, standard deviation, skewness, kurtosis, minimum and maximum value of the state variable distribution.

Table 3 :
Adrian and Brunnermeier, 2016), we use as predictors the state variables and some institutional characteristics.This larger set of risk pre-Summary statistics for estimated risk measures market capitalisation of the conditioning institution i at time t and then we normalise it by Table 5 reports the panel regression results for two periods, Q1 2000 − Q4 2010 and Q1 2011 −

Table 5 panel
14The table shows only the coefficients for the variables of interest.The complete regressions that include all the control variable coefficients are reported in the Appendix; seeTable 14, 15, and 16, respectively.

Table 4 :
Summary statistics for estimated risk measures, state variables and company characteristics at quarterly level The table reports each variable and the descriptive statistics in two sub-periods: from January 2000 to December 2010, reported in panel A, and from January 2011 to December 2021, reported in panel B. Data are quarterly based.The descriptive statistics comprehend the mean, standard deviation, skewness, kurtosis, quantile at 25%, median, quantile at 75 % , minimum and maximum values of the state variable distribution.

Table 5 :
V aR i as predictor of ∆CoV aR i -Baseline model The table reports the results of the fixed effect panel regression relative to the entire sample.Panel A shows the coefficients of interest defined by Equation6.Panel B reports the coefficients of interest, defined by Equation7by introducing the interaction between VaR and Size.Panel C exhibits the coefficients of interest as in Equation 8 by interacting VaR and Size 2 .The control variable parameters are given, respectively, in Tables14, 15 and 16.Standard errors clustered by firm are in parentheses.*** p<0.01, ** p < 0.05, * p < 0.1.

Table 6 :
V aR i as predictor of ∆CoV aR i -NACE 4-digit breakdown The table reports the results of the fixed effect panel regression.Panel A, in particular frames the analysis for those firms belonging to the extraction of crude petroleum(NACE 06.20).Panel B reports the results for the extraction of natural gas firms (NACE 06.20).Finally, Panel C reports the results for those firms operating as support of the oil and natural gas extraction(NACE 06.20).

Table 7 :
V aR i as a predictor of dollar ∆CoV aR

Table 8 :
V aR i as predictor of ∆CoV aR Notes: The table reports the results of the fixed effect panel regression relative to the entire sample defined by Equation6.In this table, we test two alternative specifications.The columns under 'Balanced' represent the coefficients for those listed firms that have non-missing observations for at least 80% of the period of time considered.The 'strong balanced' case reports the coefficients for the same firms in both periods.Standard errors clustered by firm are in parentheses.*** p<0.01, ** p < 0.05, * p < 0.1.

Table 9 :
V aR i as predictor of ∆CoV aR i -Removal of the COVID-19 outbreak period

Table 11 :
V aR i as predictor of ∆CoV aR i -The role of Debt The table reports the results of the fixed effect panel regression relative to the entire sample.Panel A reports the coefficients of interest defined by Equation 7 by introducing the interaction between VaR and Debt.Panel C exhibits the coefficients of interest as in Equation 8 by interacting VaR and Debt 2 .The control variable parameters given, respectively in Table14, 15 and 16.Standard errors clustered by firm are in parentheses.*** p<0.01, ** p < 0.05, * p < 0.1.

Table 12 :
V aR i as predictor of ∆CoV aR i -The role of Size and Debt.

Table 13 :
V aR i as predictor of ∆CoV aR i , an alternative specification -The role of Size and Debt The table reports the results of the fixed effect panel regression relative to the entire sample.Panel A reports the coefficients of interest defined by Equation 11 by introducing in the model the interaction between VaR and Size and, VaR and Debt.Panel B exhibits the coefficients of interest as in Equation 12 by adding to the interaction in panel A the second-order interaction VaR and Size 2 and, VaR and Debt 2 .Standard errors clustered by firm are in parentheses.*** p<0.01, ** p < 0.05, * p < 0.1.

Table 14 :
The predictors of ∆CoV aR -All drivers VARIABLES Notes: The table reports all the coefficients of the fixed effect panel regression defined by Equation (6).Standard errors clustered by firm are in parentheses.*** p<0.01, ** p < 0.05, * p < 0.1.Results are reported for the sub-samples of the main results as well as for the sub-samples adopted in the robustness check with separate analyses for the COVID-19 period.

Table 15 :
The predictors of ∆CoV aR -Interaction with Size Notes: The table reports all the coefficients of the fixed effect panel regression defined by Equation7.The table reports the coefficients by introducing the interaction between VaR and Size.Standard errors clustered by firm are in parentheses.*** p<0.01, ** p < 0.05, * p < 0.1.

Table 16 :
The predictors of ∆CoV aR -Interaction with Size and Size 2 Notes: The table reports all the coefficients of the fixed effect panel regression defined by Equation8.The table reports the coefficients by introducing the interaction between V aR and Size and V aR and Size 2 .Standard errors clustered by firm are in parentheses.*** p<0.01, ** p < 0.05, * p < 0.1.