The impact of ESG ratings on the systemic risk of European blue-chip firms

There are diverging results in the literature on whether engaging in ESG related activities increases or decreases the financial and systemic risks of firms. In this paper we explore whether maintaining higher ESG ratings would reduce the systemic risks of firms in a stock market context. For this purpose we analyse the systemic risk indicators of the constituent stocks of S&P Europe 350 for the period of January 2016 - September 2020, which also partly covers the Covid-19 period. We apply a VAR-MGARCH model to extract the volatilities and correlations of the return shocks of these stocks. Then we obtain the systemic risk indicators by applying a principle components approach to the estimated volatilities and correlations. Our focus is on the impact of ESG ratings on systemic risk indicators, while we consider network centralities, volatilities and financial performance ratios as control variables. We use fixed effects and OLS methods for our regressions. Our results indicate that (1) the volatility of a stock’s returns and its centrality measures in the stock network are the main sources contributing to the systemic risk measure (2) firms with higher ESG ratings face up to 7.3% less systemic risk contribution and exposure compared to firms with lower ESG ratings, (3) Covid-19 augmented the partial effects of volatility, centrality measures and some financial performance ratios. When considering only the Covid-19 period, we found that social and governance factors have statistically significant impacts on systemic risk. respectively. We did not find a significant relation to the environment factor. Similar results can be observed for the sub-sample of stocks from northern European countries, but not for the southern ones. These findings are in line with Ionescu et al. (2019), who analysed the impact of ESG factors on the market values of travel and tourism firms. They found that the governance factor had the highest positive impact on the market values and the social factor had a negative impact, while the environment factor had no significant impact. It is very likely that investors value the governance factor since it is a sign of stability for the firm. As Ionescu et al. (2019) also argue, the investors probably see social investments as risky.


Introduction
Since the 2008 financial crisis, there has been ever-growing interest in understanding the systemic risk concept. The term itself refers to the probability or the risk of a large number of financial institutions defaulting simultaneously (Lehar, 2005). Many central banks and other institutions, such as the Systemic Risk Council formed in 2012 and the Systemic Risk Centre created in 2013, look into measuring systemic risk locally and globally. There has been an extensive amount of research on the topic. SRISK of Brownlees and Engle (2017) and CoVaR of Adrian and Brunnermeier (2011b) are two of the many prominent works in the literature, while survey papers such as De Bandt and Hartmann (2000), Benoit et al. (2017) Eratalay et al. (2021) cover many of the prevalent approaches.
As much as it is important to measure the systemic risk of a certain economy, it is also important to find out the key players in this economy: which firms are "too big to fail"? 3 For example the works of Billio et al. (2012) and Adrian and Brunnermeier (2011b) among many others look into the systemic risk contribution and exposure of firms. One interesting line of research that extends from here is to analyse how sustainability influences systemic risk.
Sustainable firms exert effort in making their investments better in environmental, social and governance (ESG) terms, under which there are many subcategories. Cerqueti et al. (2020) mentions that ESG investment could help reduce systemic risk and if firms comply with ESG requirements they would be less vulnerable to systemic shocks. His argument is that the firms with higher ESG ratings have less problems with their stakeholders, possibly due to more transparent governance. Second, he mentions that ESG-related investments rely on the longer term; therefore, the investors of ESG assets are not likely to sell off even in crisis periods. Lastly, he states that ESG related assets are not yet commonly preferred; therefore, they are less vulnerable to shocks. Leterme and Nguyen (2020) found some evidence that ESG factors can be considered a systemic risk factor. There are also studies which found that there may be a negative or neutral relationship between ESG-ratings and the financial performance of firms, while some others found a positive relationship. 4 .
In this paper we aim to study the impact of the ESG-ratings of firms on their systemic risk contribution and exposure. For this analysis we use the daily returns data on the stocks constituting the S&P Europe 350 index, which represents the blue chip firms over 16 developed European countries and the ESG ratings data from S&P Global. We focus on the period of January 2016 -September 2020, which covers days under the Covid-19 situation. If a firm's stock is central, has high volatility and this firm is performing poorly financially, it is likely that this firm is threatening the financial system it is in, or being threatened by a shock from this financial system, and even more so during the Covid-19 period. Hence, as control variables we consider financial performance ratios, and two network centrality measures of these firms, volatility and a Covid-19 dummy variable. We would like to investigate whether, after controlling for the effect of the stock volatilities, financial ratios and the importance of the firms in the S&P Europe 350 network, we can still find statistical evidence that the ESG ratings increase or decrease the systemic risk contribution or exposure of a firm.
The analysis in this paper brings together different tools from several fields. First of all, we estimate an econometric model following Eratalay and Vladimirov (2020) to extract the time-varying conditional correlation matrix. Using the Gaussian graphical model, we derive the dynamic partial correlation network of the stocks and calculate the local and global network parameters as in CortésÁngel and Eratalay (2021). Then we proceed to derive the systemic risk contribution and exposure of the stocks via the principal components method of Billio et al. (2012). Finally, we conduct a panel data analysis regressing systemic risk measures on volatility, ESG ratings, financial ratios and network metrics. The first contribution of this paper is empirical, since we find the relation between systemic risk and ESG ratings, controlling for other factors that affect systemic risk, such as financial ratios and network parameters. Omitting these control variables could have misled previous research results. The second contribution of this paper is in its methodology in combining different fields to extract these control variables. As mentioned above, there are many works studying the effect of ESG ratings on financial performance, and some relating it to systemic risk. However, to our knowledge there is no work which has analysed the systemic risk contribution and exposures of the stocks in a stock market in relation to the ESG ratings and network centralities of these stocks.
Our results suggest that ESG-ratings have a negative effect on the systemic risk contribution and exposure. However, this effect is marginal for small improvements in the ESG-ratings. A firm that has an ESG-rating that is 40 points higher benefits by reducing its systemic risk contribution and exposure by about 5%, reaching up to 7.3% for southern European countries. 5 We also find that the main factors determining the systemic risk contribution and exposure of a firm are the volatilities and network centralities. For the year 2020, we found that while the "social" factor in ESG ratings is positively related to systemic risk contribution and exposure, the "governance" factor was negatively affecting it. We did not find a significant effect from the "environmental" factor. Finally, during Covid-19, the partial effect of volatilities and network centralities increased.
The paper is structured as follows. Section 2 gives a literature review on systemic risk and sustainability. Section 3 discusses the econometric model to extract the partial correlations. Section 4 explains network construction and centralities. Section 5 describes how the systemic risk measures are computed. Section 6 presents the data. Section 7 discusses the results of the OLS and panel data regressions. Section 8 concludes.
2 Literature Review

Systemic Risk
The global financial crisis that occurred in 2007-2008 has encouraged researchers to apply an interdisciplinary approach to studying systemic risk in the financial sector, with the purpose of predicting and controlling it.
In its simplest form, systemic risk can be understood as the risk of fracturing a system that can be triggered by the internal failure of any of its components or other external factors. It occurs much like a domino effect; if each component of the system represents one domino, it only takes one to fail (or fall in this case) in order to force all the components to collapse. In our analysis, the system is a stock market. The assumption that relates systemic risk in a stock market with the systemic risk in an economy is that the stock market represents a significant part of an economy. This could be the case if the stock market has many stocks, large market capitalizations, and has large coverage of different industries. There are other papers that have used stock markets for systemic risk analysis. For example Liu et al. (2020) analyses stock market indices of 43 countries to represent global financial markets, while Zhao et al. (2019) analysed the systemic risk of the Chinese stock market and Eratalay and Vladimirov (2020) focused on the Russian stock market.
There are a lot of papers that have proposed methods of measuring systemic risk. To start with, Gray et al. (2007) uses the risk-adjusted balance sheet and Contingent Claims Analysis method to gauge the asset-liabilities mismatches between sovereign, corporate, household and financial sectors, and through stress-testing depicts systemic instability due to an external factor. Tarashev et al. (2010) used a game-theoretic model, the Shapley value method, where the risk contributed by a bank is measured using the aggregate of the marginal contributions of the banking system. Additionally, Adrian and Brunnermeier (2011a) defined the conditional value-at-risk measures to appraise the individual and cumulative risk that an entity adds to the system. Similarly, Kritzman et al. (2011) applied the absorption ratio to asset prices to gauge the systemic risk in the US stock market, and Acharya et al. (2017) not only measured the systemic risk but also proposed an optimal taxation policy to manage it. Some papers went further to distinguish the systemic risk contribution and exposure of firms. Billio et al. (2012) used the principal components method, which uses the covariance matrix of returns (or return shocks) to capture the commonality between the returns, which would increase in turbulent times. Their systemic risk measure can identify the systemic risk contribution and exposure of firms, which are the same by construction. We use this methodology in our paper, since it is straightforward and easily applicable using stock return shocks derived from our econometric model. Another paper which discusses systemic risk contribution and exposure separately is by Tobias and Brunnermeier (2016), who base their methodology on value-at-risk.
For further reading we recommend Bougheas and Kirman (2015), who gives a detailed review of more non-network examples. On the other hand Caccioli et al. (2018) delve into the topic of systemic risk utilizing network analysis as their primary tool. Please also see Bisias et al. (2012), Benoit et al. (2017), Silva et al. (2017) and Eratalay et al. (2021) among others.

Sustainability and systemic risk
One of the main concerns of humanity lies on the uncertainty of our future, due to all damage caused to the planet. Entrepreneurs, investors and people in general have begun to become aware of this and have become more sensitive when making decisions. This has also had an impact on investors, who seek to contribute by investing in socially responsible and sustainable firms, seeking to be true to their values.
Socially Responsible Investing (SRI) and Environmental, Social and Governance (ESG) investing are two of the most usual value-based investing strategies. In the case of the former, investors avoid investing in tobacco, weapons and gambling stocks Capelle-Blancard and Monjon (2012). In the case of the latter, for a firm to be qualified as ESG, its line of business (excluding tobacco firms, firms involved in any way with chemical or biological weapons, as well as thermal coal generators) is considered along with the management of the risk inherent to it, such as management of human capital, business ethics, product and product governance, among others, are characteristics that are taken into account to obtain ESG certification (See Drempetic et al. (2020), Dorfleitner et al. (2015), Friede et al. (2015), Escrig-Olmedo et al. (2019)). It is worth mentioning here that there seems to be a question of the reliability of the ESG ratings by different firms. Berg et al. (2019) discusses that the ESG ratings of different sources tend to diverge.
When we search the literature, we find different views on whether investing in ESG related activities is beneficial for firms or not. Balcilar et al. (2017) show how socially responsible investment benefits reducing the volatility of conventional equity portfolios worldwide, using daily data from Dow Jones sustainable and conventional indices from around the world -North America, Europe and Asia-Pacific. Cortez et al. (2012) reveal that the performance of conventional and sustainable investments are quite similar for the US and European global socially responsible funds. Cortez et al. (2009) examine the performance of European socially responsible funds in greater depth and establish that their performance matches the performance of conventional and socially responsible standards, agreeing with Jain et al. (2019). There are also meta-analyses which argue in favour of ESG investing. Based on 2000 previous studies, Friede et al. (2015) documents that there is evidence that ESG investing has a positive impact on financial performance. Clark et al. (2015) analyses 200 previous studies and report that 88% of them conclude that ESG practices affect stock prices positively. On the other hand Revelli and Viviani (2015) report, based on 85 studies and 190 experiments, that socially responsible investments do not yield better financial performance than conventional investments.
From the systemic risk perspective, Cerqueti et al. (2020) shows that ESG investments could help reduce systemic risk and the funds that follow ESG requirements would be less vulnerable to systemic shocks. Boubaker et al. (2020) suggests that firms with higher ESG ratings have lower financial distress risk and are less likely to crash. Giese et al. (2019) mentions that the ESG factor could mitigate tail risk and there may be a long-term ESG risk premium.
Notwithstanding the above, Lundgren et al. (2018), using a network approach and the Granger causality test, show that investing in European renewable energy stock is more risky compared with non-renewable. By network connectedness analysis using a wavelet method and a multivariate vector autoregression model, Reboredo et al. (2020) found that green bonds are significantly affected by corporate and treasure bond spillovers, although their transmission is unnoticeable besides the high connectivity among them in Europe and USA. Friede et al. (2015) notes that there are portfolio studies which find negative or neutral relations between ESG and financial performance. Maiti (2021), Jin (2018) and Leterme and Nguyen (2020) mention ESG related factors as a systematic risk of mutual funds in the Eurozone.
Given this diverging view on whether higher ESG ratings could be beneficial for firms in terms of mitigating systemic risk or not, our paper finds a good place in the literature by providing evidence that ESG related investments could indeed reduce systemic risk contribution and exposures of firm stocks. Although the focus of the paper is similar to that of Cerqueti et al. (2020) and Boubaker et al. (2020), we approach to the problem from a different angle, relating ESG ratings with the systemic risk measured in a stock market, where we can derive the importance of the firm's stock in this stock market through network centrality.

Econometric method
In this subsection, we explain the econometric methodology, from which we derive the dynamic volatility and correlation estimates.

Conditional returns
Following a similar approach as in Eratalay and Vladimirov (2020), we model the conditional mean of the stock returns as a vector autoregressive model of order 1, VAR(1), with a common factor: where r t is a kx1 vector of returns. µ is a kx1 vectors of intercept coefficients. β is a kxk non-diagonal matrix containing the vector autoregressive model coefficients, which allows for return spillovers. c is a diagonal vector of coefficients of the common observable factor. The error term, ε t is assumed to be normally distributed with zero mean and a conditional variance-covariance matrix H t .
Our approach differs here from Eratalay and Vladimirov (2020), as we consider an observable common factor, namely r M SW I t , which is the returns from the Morgan Stanley World Index (MSWI). 6 Considering MSWI allows us to take into account the common trends in the world that may affect all the stocks in a similar manner. As Barigozzi and Brownlees (2019) states, the consideration of a common factor is essential. If ignored, it could yield a spuriously connected network. The typical stationarity restrictions apply on the coefficients β, such that all eigenvalues of the β matrix should be positive.

Conditional variances
The conditional variance-covariance matrix of the error term ε t is denoted by H t such that: 6 Given the number of series in consideration including an unobservable factor a la Eratalay and Vladimirov (2020) would not be feasible due to the number of parameters to estimate.
In equation 2, the conditional variance-covariance matrix H t is constructed by the diagonal matrix, D t , of conditional variances of each error term, multiplied by the correlation matrix, R t . v t denotes the standardized errors, and h t is the vector of conditional volatilities. By this construction, each element of the variance-covariance matrix is equal to H t,ij = R t,ij h 1/2 t,i h 1/2 t,j , which is the well-known relation between covariance and correlation. W is a kx1 vector and A and B are kxk diagonal matrices of coefficients. This model therefore does not allow for volatility spillovers for simplicity. In fact, estimating a model with volatility spillovers with the data considered in this paper would not be feasible. Under equation 2, the volatility process for each series is given by: The conditional variances, h t,i are stationary under the usual assumption that a i +b i < 1. Moreover, they are positive as long as w i > 0, a i ≥ 0 and b i ≥ 0.

Conditional correlations
The conditional correlations, R t , follow the consistent dynamic conditional correlation GARCH model of Aielli (2013): where Q t is the covariance matrix of the v * t and Q is the long run covariance matrix. We use the correlation targeting approach of Engle (2002), where we replaceQ with the sample covariance matrix of the v * t during estimation. The scalar parameters, δ 1 and δ 2 , of this model are restricted to be non-negative such that δ 1 +δ 2 < 1. To avoid the attenuation biases that occur when the cross-sectional dimension of the data is large, we used the composite likelihood approach of Pakel et al. (2020).
For the estimation of this model, we follow the three-step estimation procedure discussed in Eratalay and Vladimirov (2020), which is consistent and asymptotically normal (See Bollerslev and Wooldridge (1992), Carnero and Eratalay (2014)).

Partial correlation network
Following Anufriev and Panchenko (2015) and Eratalay and Vladimirov (2020), we use the Gaussian graphical model (GGM) algorithm. The GGM algorithm helps calculate the partial correlation matrices from the correlation matrices, which measure the conditional relation between any nodes in a network. We use partial correlations to isolate the correlation between two specific series eliminating the indirect effect of other series, obtaining the true relationship between every two series. The matrix of partial correlations, P, can be obtained using the correlation matrix R: where K = R −1 , and D K = diag(K) is the diagonal matrix that has the same leading diagonal as the K matrix. The details for the derivation of this equality can be found in Anufriev and Panchenko (2015).
In the model we are constructing, the cDCC-GARCH approach from Section 3.2 provides us with the time varying conditional correlations. Therefore, we are able to construct a partial correlation network for each day in the time interval of our data. This gives us a dynamic network which takes each firm's stock as a node. The strength of the connections between these nodes are obtained using the adjacency matrix, which is derived based on the partial correlations between the stock returns (see Jackson (2010)). A correlation matrix and the partial correlation matrix it implies are always symmetrical. Therefore, the adjacency matrix derived from the partial correlation matrix are also symmetrical. Consequently, this network's connections are bi-directional, meaning that there is no causal relationship. The adjacency matrix is defined as: where I is the identity matrix. The identity matrix is added to the partial correlation matrix P, since the leading diagonal elements of P are equal to -1. Hence, now the leading diagonal elements of A matrix consist of zeros, implying that nodes are connected to each other but not to themselves. Another interesting point to note about this network is that, when there is an external shock to this network, all the nodes receive the shock simultaneously and the strength of the shock is defined through the partial correlations.
In our paper, we are interested in two centrality measures that relate to systemic risk. The first is the eigenvector centrality which states that a node's centrality is proportional to its neighbours' centrality. In other words, a node's eigenvector centrality is high if its neighbours' eigenvector centralities are high. As Anufriev and Panchenko (2015) state, eigenvector centrality shows the extent to which a shock can propagate in a system. Second, we are interested in the closeness centrality, which focuses on the relative distance among nodes. To be more precise, it is the inverse of the total length of the shortest paths from this node to the other nodes. In this sense, closeness centrality relates to how fast and strongly the nodes react to a shock. As Eratalay and Vladimirov (2020) argues, in the GGM approach some partial correlations may turn out to be negative, and therefore may imply that some entries of the adjacency matrix are negative. For this network, eigenvector centrality can be calculated even with negative partial correlations, although with closeness centrality, we cannot; therefore, we considered the absolute values of the partial correlations. More details can be found in Eratalay and Vladimirov (2020), CortésÁngel and Eratalay (2021).

Systemic risk measure
After obtaining the conditional correlation estimates that change over time, we derive the systemic risk measure using the principal components method from Billio et al. (2012). This approach detects the commonality between the stock returns through the correlations between them. When the commonality between the stock returns is large, the system is more connected. In turbulent times, the commonality between the stock returns, and therefore the connectedness between the stocks, increase. Therefore, there is a one-to-one relation between the systemic risk and commonality between the returns. The principal components analysis decomposes the original return vectors to orthogonal uncorrelated factors. These factors are ordered in decreasing explanatory power. Following the same notation above: let r i t be kx1 the vector of the returns of stock i. The system's aggregated return, r S t , therefore is given by: and the variance of the system's return, σ 2 t,S is given by: where h t,i and v t,i are the volatility and standardized residuals that correspond to stock return i as defined in equations 3 and 4, respectively. The uncorrelated factors of the principal components method, ζ i , have zero mean and have variance equal to λ i such that: In fact, the λ k is the k'th eigenvalue of the correlation matrix. In the context of our paper, this correlation matrix is the conditional correlation matrix obtained from equation 4. The principal components approach therefore decomposes the standardized residuals v t,i as: where L ik is the loading vector which is the eigenvector corresponding to the eigenvalue λ k . Hence, the conditional correlation matrix can be written as: = k L ik L jk λ k and the variance of the system becomes: The principal components approach tries to explain a large percentage of the variation in the system with a few components. Hence, if we have k returns, we have n principal components such that n < k. In periods of crisis, the n principal components can explain a large proportion of the total variation, since in the commonality or correlation of these periods is expected to be high. Consequently, if the principal components can explain more than fraction H of the total variation, this indicates increased connectedness in the system. If the total risk of the system is defined as Ω = N k=1 λ k and the risk captured by the first n principal components is measured by ω n = n k=1 λ k then the ratio h n ≡ ωn Ω shows the cumulative risk fraction. If this fraction is larger than the threshold H, then the system is highly connected and a few principal components can explain most of the variation in the system. Billio et al. (2012) derives the contribution of stock i to the risk of the system, when h n > H: The authors also discuss that by construction, systemic risk exposure is the same as the systemic risk contribution of stock i: In our paper, the time varying conditional correlation matrix allows us to extract the systemic risk exposure of each stock i for each day.
Overall, the flow of the methodology is as follows. First, we apply the econometric model to the stock returns and obtain volatilities and dynamic conditional correlations. Then from the volatilities and correlations we derive the systemic risk measures. From the conditional correlations, we derive the partial correlations which help to construct the network of the stocks and to obtain network centrality measures. The obtained volatilities and network centralities along with financial performance ratios, ESG ratings and the Covid-19 dummy variable are used as regressors in fixed effects regressions, where the dependent variable is the systemic risk measures.

Data sources
For this paper we collected the data from three sources. We collected the historical stock market data for the constituents of the S&P Europe 350 index 7 and for the Morgan and Stanley World Index (MSWI) from Yahoo Finance. For the constituents list, we made a formal request to SPGlobal 8 . We were provided with the list of all 362 constituents of S&P Europe 350 index as of December 2019. Afterwards, we collected daily closing values for these constituent stocks for the period 05.01.2016 -15.09.2020 from Yahoo Finance. Some stocks did not have data for the whole data period; therefore, we had to refine our data. The final list of stocks we consider is given in Tables 19-26 in the appendix. After pre-treating the data, we had 1,202 observations for the prices of 331 stocks and the MSWI index. We detected the outliers following the Hampel filter as discussed in Pearson et al. (2015). We replaced the outliers with the local median in the 20 working days window. When detecting the outliers, we set the parameters of the Hampel filter such that the probability of observing an outlier is very small. 9 Our second data source is the S&P Global website 10 . For the constituent stocks, we collected the yearly overall ESG ratings from 2016 to 2020. Moreover, we collected the dimension scores for environmental, social, and governance and economic for 2020. Unfortunately, for some of the constituent stocks, the ESG data was not provided. We were able to collect the data for 308 stocks. 11 Finally, our third data-set is firm level data of financial performance ratios obtained from the Orbis Europe system. We collected the data on current ratios, solvency ratios and profit margins as indicators of firm level financial performance. The data is annual and for years 2016-2020. The stock market performance of the firms not only depend on the trading behaviour of the investors, but also on the firms' profitability and riskiness. Hence, we can assume that the systemic risk contribution and exposure measures derived from the stock market relations should depend on the financial performance ratios. Unfortunately, the data on all these ratios was available for only 200 of the constituent stocks. We summarize the description of these three panels in the Table 1 below.

Descriptive statistics
In Figure 1, we plot the returns after being processed through the Hampel filter. The high volatility caused by Covid-19 is visible towards the end of the sample. We marked the date 21/02/2020 with a vertical dashed grid line, which is when a cluster of cases occurred in Lombardy, Italy. 12 It can be seen from the figure that there are many extreme returns which were not eliminated by the Hampel filter. The most extreme negative return belongs to the return series of the company Wirecard, which declared insolvency in June 2020. We discuss more on this series in Section 7.2. In Figure 2, we give the descriptive statistics for the returns of the stocks in a Box plot form.The descriptive statistics were calculated for each series, and then the Box plots of each descriptive statistic are plotted. For example, the Box plot for the means is for the average returns of each of the 331 return series. As we can see, the means of the returns are concentrated around zero for all the stocks, while the standard deviation varies between 1 and 3, but exceeding 3 for some series. For most stocks the returns are negatively skewed, and in some cases exceeding the conventional threshold of unit skewness indicating that the return distribution is highly skewed, implying that there are many negative extreme returns. We also observe that the kurtosis is very high for all the stocks, much above the kurtosis of normal distribution. This means that the sample distribution of the stock returns are leptokurtic and this is one of the stylized facts about financial time series data (Ghysels et al., 1996).

Mean
St. deviation Skewness We now discuss the ESG-ratings data. In Figure 3 we present the histograms of (a) merged ESG-ratings and (b) yearly ESG-ratings. When we look at the figure 3a we see that the distribution is bimodal and the difference between the modes is about 40-50 points. The figure 3b shows that the trend in ESG-ratings over the years is different around these two modes. In particular, on the left side of the distribution, we see that the ESG-ratings are decreasing over the years, while on the right side we see that they are increasing. This implies that over time the firms with lower (higher) ESG-ratings reduced (increased) their ESG-ratings further.  In figure 4 we plot the 5th, 25th, 50th, 75th and 95th quantiles and the mean of the overall ESG ratings of the stocks from the S&P 350 Europe index. Although perhaps the mean and the median have a slightly positive trend, the other quantiles seem stable over time. What is also interesting is that the median was less than the mean before 2018 and more than the mean afterwards. This suggests that the ESG ratings distribution before 2018 was positively skewed with a few firms with high ESG ratings. After 2018, the distribution became negatively skewed, with a few firms with low ESG ratings. This suggests that overall there is an increasing trend in the ESG ratings over the years. As we discussed in Figure 3, however, this increase is not for every quantile of the distribution.
When we look at the averages per country over the years in Table 2, we can see that for many countries the ESG ratings have been decreasing over time, while for some they increased after a slight decrease. It is hard to comment on any country's efforts in creating and maintaining sustainable firms from this table, since only certain firms from each country are in this list. However, even for those countries where the number of stocks is higher, there is a visible decline of ESG ratings in general. The ESG ratings are higher for the Southern European countries, namely Italy, Spain, Portugal and to some extent France. These are all countries which can benefit from solar energy. This provides the motivation for analysing Southern European countries and other countries separately in Section 7.
In Table 10 in the appendix we show as an example 25 stocks that have the highest average ESG rating. It is interesting that there are many firms from electric and gas utilities. In terms of countries, Spain, Italy, Switzerland and the United Kingdom are leading. Interestingly, the United Kingdom, German, France and Switzerland have many firms in the S&P Europe  350 for which ESG-ratings were available, but the average ESG-ratings were not as high for these firms. After obtaining the necessary regressors, we apply a fixed effects regression. However, to avoid the bias that it could introduce, we discard the data related to the company Wirecard. We discuss the reasons more clearly in Section 7. We construct panels considering (1) all 330 stocks for which systemic risk, volatilities, and network centralities are available, (2) 307 of those 330 stocks for which ESG-ratings are also available, (3) 199 of those 307 for which firm-level financial performance ratios were also available. Therefore we have three panels of data to work with. Since some stocks get eliminated due to data limitations through these panels, it makes sense to discuss the content of these panels in terms of the represented countries and industries. In figure 9 in the appendix, we present word clouds to visualize the industries and countries which are dominant in these three panels. In the larger panels of 330 and 307 stocks there are more stocks from the industries such as banking, diversified financial services, machinery and electrical equipment, chemicals and insurance. In terms of countries, there are more stocks from Great Britain, Germany, Switzerland and France. When we look at the smaller panel of 199 stocks, we see that the industries of chemicals, telecommunication services, pharmaceuticals, machinery and electrical equipment, and oil and gas upstream and integrated are more represented. In this panel there are more stocks from Great Britain, Germany and France. Therefore, when discussing the results, we should keep in mind that banks, diversified financial services and insurance industries dominate the bigger panels, while they do not play such a big part in the smaller panel.

Results
In this section, we first explain the findings from the network analysis of the constituent stocks of the S&P Europe 350 index. Afterwards, we discuss the results of the fixed effects and OLS estimations, which study the causal relationship between systemic risk and ESG ratings.

Partial correlations network
In this part, we use the partial correlations obtained from the estimation of the econometric model in Section 3 and calculated via equation 5. As can be seen from the kernel density estimate in Figure 5, the partial correlations are primarily positive; however, there are also negative values. Therefore, some relationships among stocks have a negative sign. In other words, while some stocks react similarly (positive edges) to external news, others respond in the opposite way (negative edges). The positive and negative weights exist in the networks of each day since each day's network is constructed using the partial correlation matrices as the adjacency matrices. In fact, 51.45% of all correlations of all times were positive. Considering all positive and negative partial correlations, we calculate the normalized number of edges over time in Figure 6, which suggests that the normalized number of edges stayed more or less the same over time. In Figure 7 we see that the maximum eigenvalues reach an all time high just after the first news of Covid-19 patients and deaths appear in Europe around 21 February 2020. The maximum eigenvalue is related to the eigenvector centrality, and its high values can be seen as an indicator of systemically risky times. In particular, when the maximum eigenvalues exceed one, it indicates that the system is unstable. (Eratalay and Vladimirov, 2020) In this paper, we calculated the eigenvector and closeness centrality measures based on the dynamic partial correlations networks of S&P Europe 350 for the years 2016-2020. 13 We calculate the eigenvector and closeness centralities considering whole daily partial correlation matrices. The eigenvector centrality considers the importance of a node's neighbours and those neighbours' connections. A node has a high eigenvector centrality if its neighbours have a high eigenvector centrality. A node's closeness centrality measures its distance to the rest of the nodes on the network. We can say that, as a node is closer to the rest of the nodes, it has a higher closeness centrality. Therefore, if the node has a high closeness centrality, then in the case of a shock, the rest of the network will have a quicker response to the shock. In terms of shock propagation, the closeness and eigenvector centralities help us measure the impact of a shock by considering the distance among stocks and the possible implications for the neigbouring nodes. This is why we selected these centrality measures.
When calculating the distances among nodes, we found negative cycles. Therefore, it was impossible to calculate any relative distance parameter for net partial correlations. Consequently, the closeness centrality was only calculated for absolute and positive partial correlations. Independently and additionally, positive and negative weights would offset each other when calculating closeness centralities. Therefore, we only consider the absolute value of the closeness centrality.
In tables 11 and 12 in the appendix, we present the top 25 central firms for which the ESG ratings were available for 2016-2019 and 2020, respectively. The most central firms were mostly the same in both periods. These most central firms were mostly from France and Germany and, from the Financials sector, namely from Banking and Insurance industries. We can also note that there is a clear correlation between the centrality measures and ESG ratings or systemic risk measures.

Systemic risk measure
Following the methodology in Section 5, we calculate the total systemic risk of the S&P Europe 350 stocks, given by equation 8. In Figure 8 we plot this PCA-based total systemic risk along with the composite indicator of systemic stress of the European Systemic Risk Board, and the stress sub-indices for financial and non-financial equities. These latter indices are calculated from the realized volatilities of the corresponding stock market indices. The data was obtained from the Statistical Data Warehouse of the European Central Bank. 14 This index is calculated for all the countries in the Euro area and uses the methodology of Hollo et al. (2012), which combines 15 raw mainly market-based financial stress measures. We find that the correlation of PCA-based systemic risk has a medium high correlation of approximately 0.65 with the ESRB composite indicator of systemic stress. Moreover, it is highly correlated with the stress sub-indices: approximately 0.78 with non-financial stocks and approximately 0.75 with financial stocks. It seems that the PCA systemic risk measure reacted more than the other measures when the systemic risk increased in the market in July 2016, and more clearly in early March 2020.
Tables 15 and 16 in the appendix show 25 firms for which the systemic risk was very high in 2016-2019, and 2020, respectively. It can be seen that Wirecard AG from Germany had the highest risk and this risk is calculated as about 9 times higher than the next company in line in 2020. This was probably related to the Wirecard scandal in 2019 and their declaration of insolvency in 2020. Interestingly, Wirecard AG's centrality measures were not very high. In our regression analyses, we removed Wirecard AG from our data set. According to tables 15 and 16, Anglo American Plc, ArcelorMittal Inc, Bank of Ireland Group, Glencore Plc and Unicredit SpA Ord also had high systemic risk measures for 2016-2019. In 2020, Anglo American Plc, Glencore Plc and Unicredit SpA Ord improved their systemic risk measures, while Bank of Ireland Group, ArcelorMittal Inc. suffered in that respect.
Tables 17 and 18 in the appendix show 25 firms for which the systemic risk was the lowest in 2016-2019 and 2020, respectively. We could easily see that most of these low risk firms are from Switzerland and there are many firms from the Communication Services and Consumer Staples sectors.

Systemic risk and ESG ratings
In this subsection we use the variables we obtained from the previous parts and from the datasets. We use the natural logarithm of systemic risk contribution and exposure as the dependent variable. As regressors, we use the eigenvector and closeness centralities, natural logarithm of volatility, ESG-ratings and firm level financial performance ratios. In our regression analyses, we eliminated Wirecard AG from our list since it was an obvious outlier in terms of systemic risk.
A preliminary analysis of scatter plots of average systemic risk exposures in logarithm and ESG ratings of the remaining 307 firms for which the ESG data was available are given in Figure 10 in the appendix. For each year and for the whole sample the slope of the linear relation is negative but small in magnitude. We can also note that in 2018 and 2019 the magnitude of the slope is relatively higher. Hence, in general we can talk about some negative correlation between systemic risk exposure (and contribution) and the ESG ratings.

Fixed effects regressions
In this subsection, we discuss the fixed effects estimation results. As mentioned above, we have three panels to consider, with cross-section sizes 330, 307 and 199. In the larger panels, we have more stocks from many industries. However, in the smallest panel, although we have the variables for firm-level financial performance ratios, we do not have as many stocks from the banking and insurance industries. We discussed how different industries and countries are represented in these panels in Section 6.2.
The dependent variable in all these regressions is the natural logarithm of the systemic risk. Since it had some outliers and only has positive values, taking a logarithm of this variable helps to bring the distribution closer to normal. The main variables in these regressions are the net eigenvector centrality, absolute closeness centrality, logarithm of volatility, and the dummy variable that takes the value 1 for 2020. We also added certain interactions of the variables. For example, it made sense to include the interaction of centralities with the logarithm of volatility, since a stock's high volatility becomes dangerous for the system if that stock is more central. A similar argument follows for the interaction of centralities with financial performance ratios. We also included interactions with the dummy variable since the partial effects might change during Covid-19. In all the following regressions, we removed some of the interaction terms between regressors due to strong multicollinearity. Since we found that the ESG ratings of the firms from southern countries (Italy, Spain, France and Portugal) are relatively higher in Table 2, we performed the same regressions using sub-samples with respect to geographical location.
In Table 3 we present the fixed effects regression results using the large panel with 330 stocks. The estimation results suggest that both centrality measures are positively linked to the systemic risk of the stock. Similarly, higher volatility of a stock implies higher systemic risk contribution and exposure. As expected, the partial effect of eigenvector centrality and volatility increased in Covid-19 times. The coefficient estimates and their signs are similar for the stocks from southern and northern European countries. One difference can be that for southern European countries, closeness centrality has a higher impact than for northern European countries. On the other hand, eigenvector centrality has a higher impact on northern European countries. One interpretation could be that for the stocks from southern European countries, being "close" to the rest of the stocks has more impact. In contrast, for northern European countries, the centrality of the neighbouring stocks matters more. The correlation between unobserved heterogeneity and the regressors validate that fixed effects is a better approach than the random effects method for these regressions.
In Table 4 we present the regression results with 307 stocks, where ESG ratings are also considered as a regressor. We again see similar relations that centralities and volatility are positively linked to systemic risk. We also notice the same way that the partial effects of centralities and volatility increased in 2020. What is more in these results is that the ESG rating is negatively linked to systemic risk. The coefficient is significant at 10% and is small in magnitude. However, if we consider the approximately 40 point difference between the two modes in the histogram of Figure  3a, we can calculate that a 40-point increase in ESG ratings would decrease systemic risk contribution and exposure by 2.90%. 15 This means that firms with higher ESG ratings are benefitting from a lower systemic risk contribution and exposure compared to firms with lower ESG ratings. When we compare the results for southern and northern European countries, we see that the ESG ratings had no significant impact on systemic risk for southern European countries. For stocks from northern European countries it had a higher impact, which would imply 3.41% decline in systemic risk contribution and exposure for a 40-point increase in ESG ratings.
In Table 5, we further include the financial ratios of the firms to the regression. As we said before, due to lack of data, we end up with 199 stocks among which there are less banks and insurance firms. As before, the coefficients of centrality measures are positive. In addition, we found that the partial effect of eigenvector centrality decreases as profit margin increases, but this does not depend on volatility or other financial performance ratios. This means that a stock becomes systemically less risky if the firm's profit margin is higher. The coefficient of log-volatility is positive, but the partial effect of volatility decreases when profit margin and solvency ratios are higher. This could mean that a stock's high volatility is less of a threat to the market if its profit margin and solvency ratios are higher. Financial performance ratios are positively linked to systemic risk contribution and exposure, but the sign of the partial effects quickly change for higher levels of eigenvector centrality and log-volatility, which implies that having better financial performance reduces systemic risk contribution and exposure further for central and volatile stocks.
The coefficient of the ESG rating is -0.0012 and it is significant at 5%. Following the previous discussion, an increase of 40 points in the ESG rating would mean a decrease of 4.87% in the systemic risk contribution and exposure. This implies that the high ESG-rating firms, in the right mode of the histogram in Figure 3a, are enjoying approximately 5% less systemic risk contribution and exposure compared to the low ESG-rating firms in the left mode of the same histogram. In the extreme case, the difference between the left and right tails of the ESG-rating distribution is over 80 points, and this would imply about 9.5% less systemic risk contribution and exposure for the high ESG-rating firms. Another note is that the partial effects of eigenvector centrality and log-volatility are higher in 2020, but no such effect is seen for ESG rating and financial ratios. Comparing the results for southern and northern European countries, we find that most coefficients are quantitatively and qualitatively very similar. We observe the difference that for southern countries the impact is much larger, yielding a 7.27% decrease in systemic risk contribution and exposure for a 40-point increase in ESG ratings, while for northern countries this impact is about 4.05%. This is a stronger result than that of the second panel, which had 307 stocks and it is most likely due to the change in the stocks we considered. In this small panel, banks and insurance firms are not well represented due to lack of data. The results call for further research considering different industries, which we consider in Section 5.3.3.

OLS regressions for 2020
As explained in Section 6.2., we were able to collect data for the subcategories of the ESG ratings for the 199 firms in our smallest panel in 2020. To have a fair comparison, we run three OLS regressions, one for each cross-section size in our panels: 330, 307 and 199. The stock tickers were used as a clustering variable for calculating the standard errors.
Using the 330 stocks of the first panel, we found similar results as in the fixed effects regression that the centralities and volatility significantly affect the systemic risk contribution and exposure. We present these results in Table 6. However, we should note that the coefficient for eigenvector centrality was negative and larger in magnitude for the stocks from southern European countries compared to the northern ones. For the 307 stocks that have ESG rating data available, we found similar coefficients in Table 7. Interestingly, in these regressions we found that ESG subcategories did not have an affect on the dependent variable. When we move on to include the financial performance ratios to the OLS regressions in Table 8, we see that eigenvector centrality and volatility regressors are significant, while in the sub-samples the former is not significant.
Table 8 also suggests that while the social factor in the ESG ratings is positively linked to systemic risk contribution and exposure, the governance/economic factor is negatively related. The coefficients are not very large, but for a 40-point improvement in these factors, the effect is 3.25% and -3.35% respectively. We did not find a significant relation to the environment factor. Similar results can be observed for the sub-sample of stocks from northern European countries, but not for the southern ones. These findings are in line with Ionescu et al. (2019), who analysed the impact of ESG factors on the market values of travel and tourism firms. They found that the governance factor had the highest positive impact on the market values and the social factor had a negative impact, while the environment factor had no significant impact. It is very likely that investors value the governance factor since it is a sign of stability for the firm. As Ionescu et al. (2019) also argue, the investors probably see social investments as risky.   Notes: For this regression yearly average of systemic risk, network characteristics, volatilities, ESG ratings and firm level financial data are used. Cross section size is 199. Other interaction terms were eliminated due to multicollinearity. Significance: * 10%, ** 5%, *** 1%. Source: S&P Global ESG ratings and authors' calculations.

Further regressions
In Table 9 we present the coefficients of the ESG ratings (ESG Coef) and their interaction with the dummy variable (D*ESG Coef) for 2020 in the fixed effects regressions we ran for each sector. The industries that constitute these sectors are given in Table 29 in the appendix. As Hox et al. (2017) mentions, when a panel data has less than 50 groups and less than 5 cases for each group, the standard errors for the fixed effects regressions might be too small. We need to keep this in mind when interpreting the results of Table 9. That is why we report the number of firms in each sector in the last column of this table.
If we consider the panel of 307 stocks, where the regressors were as in Table 4, we find significant coefficients for ESG ratings for Energy, Financials and Utilities sectors. An increase of 40 points in ESG ratings in these sectors suggests a decrease of 16.60%, 6.07% and 17.56% in systemic risk, respectively. For these sectors, keeping ESG ratings high might have Finally, we ran OLS regressions for each sector for 2020 using the panel with 307 stocks, where we used ESG sub-factors as ESG related regressors as in Section 7.5. In most cases, there were too few stocks in the sectors we wanted to analyse, which rendered these OLS regressions useless. There were 64 stocks in the Industrial sector and we found that the coefficient of the environmental factor was -0.0006, significant at 10%, while the other factors were not significant. On the other hand, for the Financial sector, where there were 59 stocks, we found that the coefficients of social and governance/economic factors were -0.0008 and 0.0010, respectively, which were both significant at 1%. Harrell et al. (2001) suggests that for each regressor, one should have 10-20 observations per regressor, while Green (1991) suggests to have at least 50+8*p number observations where p is the number of regressors. In these regressions we had 7 regressors, which required at least 70 or 106 observations based on the suggestions of Harrell et al. (2001) and Green (1991), respectively. Therefore, it is possible that the results of these OLS regressions were suffering from a small sample size. We do not present the results of these regressions to save space.

Conclusions
In this paper we explore the effect of the ESG ratings of firms on the systemic risk contribution and exposure of their stocks. Our aim was to show that keeping ESG ratings high would benefit the firms by reducing the systemic risk they face. For this purpose we used the daily returns of the stocks constituting the S&P Europe 350 index for the period 05.01.2016 -15.09.2020, and yearly ESG ratings and firm performance ratios for these firms. We employ an interdisciplinary approach that connects financial econometrics, panel data econometrics and social networks. To be more precise, we fit a rigorous model to estimate the daily volatilities and dynamic correlations, and using principal components method we derived the systemic risk contribution and exposure measures. Subsequently, we obtain dynamic partial correlations using Gaussian graphical modelling and construct the daily partial correlation networks of stocks, which provided us with the network centralities. Finally, we employ panel data and OLS regressions, where the systemic risk contribution and exposure of each firm is the dependent variable and the volatility estimates, network centralities, ESG ratings and firm performance ratios are the regressors. We also consider a dummy variable for the year 2020 to keep account of the effect of Covid-19.
Our results indicate that volatilities and network centralities are the main determinants of systemic risk contribution and exposure, and the impact of these variables increased during the Covid-19 period. We also found that the systemic risk contribution and exposure could be reduced by almost 5% through a 40-point increase in ESG ratings. When we consider the southern European countries (Italy, France, Spain and Portugal) alone, this effect rises to about 7.3%. This finding could be interpreted such that the firms to the higher end of the ESG ratings are benefitting from reduced systemic risk contribution and exposure compared to those with lower ESG ratings.
We were also able to analyse the effect of ESG subcategory ratings (environmental, social and governance/economic factors) for 2020, and we found no significant impact of the environmental factor. On the other hand, the results suggest a positive coefficient for the social factors and a negative coefficient for the governance/economic factors on the systemic risk contribution and exposure. Interpreting these results could suggest that investors might see social investments as risky, but value how the firms are governed. The findings of this paper are highly useful for firms. Although firms may find it costly or risky to engage in ESG related activities, our results show that it pays to keep ESG ratings high. In particular, firms should pay attention to the governance/economic factors to satisfy the interests of their shareholders.
This work can be extended in multiple ways. The first would be to expand the dataset further, not only in terms of the number of stocks considered but also the ESG ratings and subcategories. For example, our data did not allow us to estimate regressions per sector, although this would have been a valuable analysis. Another interesting point could be to explore whether the systemic risk measures and firm performance ratios are simultaneously determined. Although it could provide a different insight into the possible relations between the variables, the firm-specific effects would not be captured by such a regression.           Tables related to stock data The last column indicates in which panels a stock is included. "o" indicates that the stock was in Panel 1, "oo" indicates that the stock was in Panel 1 and 2, and "ooo" indicates that the stock was in all the panels. Source: S&P Global ESG ratings and authors' calculations.  Notes: The last column indicates in which panels a stock is included. "o" indicates that the stock was in Panel 1, "oo" indicates that the stock was in Panel 1 and 2, and "ooo" indicates that the stock was in all the panels. Source: S&P Global ESG ratings and authors' calculations. Notes: The last column indicates in which panels a stock is included. "o" indicates that the stock was in Panel 1, "oo" indicates that the stock was in Panel 1 and 2, and "ooo" indicates that the stock was in all the panels. Source: S&P Global ESG ratings and authors' calculations.