Measuring systemic risk of banking sector in Uzbekistan

This paper examines systemic risk in Uzbekistan's banking sector using the Conditional Value-at-Risk (CoVaR) framework proposed by Adrian and Brunnermeier (2016), one of the most widely used approaches for measuring systemic risk. Specifically, we analyze bank-level risk spillovers under extreme market conditions and empirically investigate the determinants of systemic risk.

The global financial crisis highlighted that the social costs of financial distress arise primarily from system-wide spillovers rather than from the failure of individual institutions. In response, a substantial body of literature has developed measures designed to capture the externalities imposed by financial institutions on the broader financial system. Among these, the CoVaR framework has emerged as a leading market-based measure of systemic risk.

CoVaR is defined as the value-at-risk (VaR) of the financial system conditional on a specific institution being in distress, while ΔCoVaR captures the marginal contribution of that institution to system-wide tail risk by comparing distress and median states (Adrian and Brunnermeier, 2016). The framework is typically implemented using quantile regression, which accommodates asymmetric and nonlinear dependence structures and avoids restrictive distributional assumptions. These features make CoVaR particularly well suited to financial data characterized by fat tails and state-dependent behavior.

Growing literature applies CoVaR-based measures to emerging and developing economies, where banking systems are often dominant, highly exposed to external shocks and subject to structural transformations. Across regions, studies consistently find that systemic risk is strongly time-varying and increases sharply during periods of financial stress (Girardi and Ergün, 2013).

Empirical evidence from emerging Asia demonstrates that CoVaR and ΔCoVaR are effective in identifying systemically important banks and tracking systemic risk dynamics over financial cycles. Studies on China and other Asian economies show that large, state-owned or highly leveraged banks contribute disproportionately to system-wide tail risk, particularly during crisis episodes (Jiang et al., 2020; Wen et al., 2020; Xu et al., 2021). Similar findings are reported for India and ASEAN economies, where quantile-based measures reveal strong interbank spillovers and heightened sensitivity to macro-financial conditions and external shocks (Verma et al., 2019; Khan et al., 2021; Pham et al., 2021). Evidence from other emerging economies further confirms that systemic risk is closely linked to bank characteristics such as size, leverage, capitalization and liquidity, as well as to global financial conditions (Rizwan and Ahmad, 2019; De Jonghe et al., 2015; Mendonça et al., 2019).

Overall, the emerging-economy literature highlights three stylized facts. First, systemic risk is strongly time-varying and rises nonlinearly during stress periods. Second, global financial conditions often play a dominant role in shaping domestic systemic risk, sometimes exceeding the explanatory power of bank-specific factors. Third, rankings of systemically important banks can vary across risk measures, underscoring the need for robustness analysis and complementary indicators.

Uzbekistan represents a distinctive case among emerging and transition economies. Its banking system has undergone rapid transformation in recent years, including financial liberalization, monetary policy reform and a strategic shift toward commercialization and privatization of state-owned banks. In this context, international institutions emphasize the need to strengthen system-wide risk monitoring and macroprudential surveillance (IMF, 2025). Consistent with this policy focus, the Central Bank of the Republic of Uzbekistan has highlighted systemic risk indicators conceptually aligned with the ΔCoVaR framework in its financial stability assessments (Central Bank of the Republic of Uzbekistan, 2022). However, despite extensive evidence for other emerging economies, bank-level empirical analysis of systemic risk in Uzbekistan remains scarce.

This paper fills this gap by providing the first comprehensive CoVaR-based assessment of systemic risk in Uzbekistan's banking sector. Using data for 21 commercial banks, we combine quantile regression, panel regressions and a PVARX framework to analyze risk spillovers and their determinants. The results show that banks with higher individual risk contribute disproportionately to systemic risk, that systemic risk surged during the COVID-19 period before subsiding and that macroeconomic and global financial factors dominate bank-specific characteristics. These findings have important implications for bank-level risk management and macroprudential policy in transition economies.

The remainder of the paper is organized as follows. Section 2 presents the static and dynamic CoVaR models and estimation results. Section 3 analyzes the determinants of systemic risk using panel regressions and PVARX models. Section 4 concludes with policy implications.

VaR refers to the minimum loss of a financial asset or portfolio at a given confidence level $α$⁠, where $X$ denotes an investment with stochastic returns, as defined in Equation (1).

VaR provides a concise and intuitive framework for assessing risk in complex investment portfolios composed of multiple financial instruments, and its theoretical foundations are relatively straightforward. However, extensive empirical research has identified several fundamental shortcomings of VaR. First, VaR may violate the principle of diversification as the risk of a portfolio can exceed the sum of risks calculated separately for individual assets. Second, VaR fails to capture information about extreme tail losses, which are often of greatest interest for risk management. Third, VaR typically relies on the assumption of normally distributed returns, which is unrealistic for financial time series that commonly exhibit volatility clustering and fat tails. Finally, VaR measures only institution-specific risk and does not account for spillover effects across financial institutions.

These limitations of VaR became particularly evident following the Global Financial Crisis (GFC) of 2008. To address these shortcomings, Adrian and Brunnermeier (2016) proposed the CoVaR, a risk measure that captures the tail risk of the financial system conditional on an institution being under distress. CoVaR is derived from the conditional distribution of system-wide losses beyond the VaR threshold. Compared with VaR, CoVaR provides a more informative framework for assessing systemic risk and portfolio optimization. If the loss of a financial asset exceeds its VaR at confidence level $α$⁠, the conditional expectation is expressed in Equation (2).

When the stochastic loss of the portfolio is denoted by −X (−X<0), $VaRα$ at the $1−α$ quantile represents the loss level that is exceeded with probability $1−α$⁠, as shown in Equation (3).

CoVaR optimization yields results consistent with VaR optimization under normal and elliptical distributions. However, unlike VaR, which may be discontinuous, CoVaR is a continuous and convex function with respect to portfolio positions, making it more suitable for systemic risk analysis. To examine spillover effects in a banking system consisting of multiple institutions, we specify the quantile regression model in Equation (4).

where $Xˆqsystem|Xi$ denotes the estimated losses at q quantile when the loss of institution i is X under a certain confidence level. Based on the definition of VaR, CoVaR is obtained as the predicted system loss conditional on institution $i$’s loss reaching its VaR level, as shown in Equations (5) and (6). Once $VaRqi$ is estimated from the return series of institution $i$⁠, the risk spillover effect, measured by $ΔCoVaR$⁠, can be computed as shown in Equation (7).

Given that financial time series typically follow leptokurtic rather than normal distributions, conventional linear regression methods are often inappropriate for risk estimation. Moreover, ordinary least squares (OLS) regression focuses on conditional means and thus provides limited insights into tail behavior. Quantile regression addresses these limitations by modeling the relationship between explanatory variables and specific quantiles of the dependent variable. As a result, quantile regression is particularly well suited for analyzing tail risk and extreme events. It is also more robust to outliers and does not impose restrictive assumptions such as homoscedasticity. These features make quantile regression widely used in financial economics and systemic risk analysis.

According to internal data from the Central Bank of Uzbekistan (CBU), 33 commercial banks are recorded. Among them, 12 banks (banks 4, 6, 10, 12, 15, 17, 19, 25, 30, 31, 32 and 33) are excluded due to data limitations and extreme outliers. Based on data availability and research feasibility, the final sample includes 21 banks over the period from March 2019 to March 2022, yielding 37 monthly observations.

Table 1 reports summary statistics for individual bank returns. All banks exhibit positive average returns. Although skewness varies substantially across banks, all return series display leptokurtic behavior. Except for Bank 20 and Bank 29, the Jarque–Bera test rejects the null hypothesis of normality at the 10% significance level for all banks, consistent with the typical characteristics of financial return series, such as fat tails and asymmetry. Due to the lack of stock price data, earnings before interest and taxes (EBIT) are used to compute individual bank returns. The return of the banking system is calculated as the weighted sum of individual bank returns, where weights are based on each bank's share of total equity. The book value of equity (BVE) is employed as a proxy for market value, and end-of-period data are used to maintain consistency.

To quantify risk spillovers, we first estimate the quantile regression model specified in Equation (8). The estimated coefficients are then substituted into Equation (9) to compute CoVaR for each bank. Using these estimates, $ΔCoVaR$ is obtained from Equation (10), which measures the marginal contribution of an institution to systemic risk by comparing distress and median states.

where $VaR0.05Banki$ denotes Bank i's return value at 5% quantile, $ΔCoVaR0.05SYS|Banki$ represents an institution's contribution to systemic risk (spillover effects) and $CoVaR0.05SYS|Banki$ and $CoVaR0.5SYS|Banki$ are CoVaR under distress and median (normal) circumstances, respectively.

The extent to which the banking system is affected by an individual bank is a reliable indicator of systemic risk. The impact is inversely correlated with the stability of the bank's own returns, which will affect the fluctuations in the aggregate yields of the banking sector. As seen in Table 2, Bank 2 is the largest, followed by Bank 11 and Bank 1, in terms of the absolute magnitude of risk spillovers from the specific bank to the banking system. There is little change in the ranking of unconditional VaR, which suggests that banks with higher risk will exert greater spillover effects on the banking sector, and vice versa. The positive association demonstrates that by enhancing and perfecting internal risk management, the whole banking sector can effectively withstand threats.

For ease of interpretation, all values are multiplied by −1 and converted to positive numbers. Figure 1 illustrates that VaR and $ΔCoVaR$ exhibit similar dynamics, indicating that both measures capture key aspects of systemic risk. Comparing these indicators allows for the identification of systemically important banks and provides valuable information for risk monitoring and early-warning systems.

In practice, the risk faced by a financial institution is influenced not only by its own characteristics but also by shocks originating from other financial sectors. The transmission of such shocks is commonly referred to as the risk spillover effect. Given its ability to capture tail dependence, CoVaR is particularly suitable for analyzing dynamic systemic risk. To model time-varying risk spillovers, we estimate a quantile regression framework in which bank returns are regressed on lagged state variables, as shown in Equations (11) and (12).

where $Xti$ represents the return rate of financial institution i at time t, q is the quantile and $Mt−1$ denotes the state variable with a lag of one period. The parameters of the model can be estimated using quantile regression. The inclusion of lagged variables accounts for delayed transmission effects in financial markets. Based on the estimated coefficients, time-varying VaR and CoVaR series are computed using Equations (13) and (14). Subsequently, dynamic ΔCoVaR is calculated as shown in Equation (15).

Following the literature, extreme risk scenarios are captured using the 1 and 5% quantiles, while the 50% quantile represents normal market conditions. The same sample of 21 banks is used, and several banking-sector risk factors are included as state variables, as summarized in Table 3.

All state variables can measure the bank's systemic risk from different perspectives. We consider the liquidity spread (LS) as a key indicator of both the bank's credit policy and the market participants' capability to finance. The LS is the difference between the weighted average short-term interest rate and the 1-year treasury bill rate, which is known as the risk-free interest rate. The interest rate spread (IRS) is generally measured by the difference between long-term and short-term treasury bill rates, which are proxied by the weighted average long-term deposit rate and the weighted average short-term deposit rate, respectively. The credit spread (CS) is the difference in yields between government bills with the same maturity and bonds issued by commercial banks. In this study, the one-year deposit rate is used in place of corporate bond yields. IRS and CS reflect interest rate risk and credit risk in the market and exert a significant influence on banks' asset allocation and leverage ratio. During economic expansion (contraction), owing to decreasing (increasing) credit risk and credit spreads, overall risks are alleviated (intensified). To achieve a more efficient asset allocation, banks are expected to ramp up leverage and raise liabilities.

Uzbekistan's sole stock exchange, the Tashkent Stock Exchange (TSE), provides the basis for measuring domestic market volatility. Since an official volatility index is unavailable, we estimate the conditional variance of the Uzbekistan Composite Index (UCI) using a GARCH (1,1) model. As reported in Table 4, all state variables are stationary, with LS and IRS approximately normally distributed, while CS and VUCI exhibit leptokurtosis and asymmetry.

Figure 2 shows that systemic risk measured at the 1% quantile consistently exceeds that at the 5% quantile. Systemic risk rose sharply during 2020 and peaked in March 2021, reflecting the impact of the COVID-19 shock, before declining rapidly in subsequent months. More recently, systemic risk has shown an upward trend, approaching pre-crisis levels. Figure 3 indicates that while the magnitude of spillovers varies across banks, their dynamic patterns are largely synchronized, with some degree of heterogeneity.

To examine the determinants of systemic risk, we estimate the panel regression model specified in Equation (16). The lagged dependent variable is included to account for persistence and potential endogeneity in systemic risk measures.

Book value of equity (BVE), total assets (AST) and debt-to-equity ratio (DTE) are bank-specific financial variables represented by $Banki,t−1$⁠. $Fint$ denotes financial variables including the CBOE Volatility Index (VIX), Federal Funds Rate (FED) and Uzbekistan Stock Market Volatility Index (VUCI). Besides, the BVE and AST are log-differenced, the VIX and FED are first differenced series and the DTE is original data that was obtained from the CBU.

The specific explanations on independent variables can be summarized as follows. First, the BVE is vital because it provides a snapshot of a bank's financial state at a specific point in time. It can be used to compare a bank's current equity position to its equity position in previous periods. Additionally, it can be used to evaluate the financial leverage of a bank. Second, the AST is defined as the assets owned by a bank that has an economic value whose benefits can be derived in the future. To improve market value and sustainability for the future, the bank should look healthy and a bank's health will be decided on various parameters, among which assets are the most crucial ones as it will help in predicting the range of profit a bank can earn on its current investment over the period. Third, the DTE shows the level of financial leverage being used, reflecting the ability of owners' equity to protect creditors' equity and the bank's long-term solvency. The three financial indicators listed above and the degree of risk are closely related. Fourth, VIX is a popular measure of the stock market's expectation of volatility based on S&P 500 index options, which is often referred to as the fear index. The influence of the US financial market on the global market, especially in emerging markets economies (EMEs), cannot be ignored. Fifth, FED is the target interest rate set by the Federal Open Market Committee (FOMC). FED is the representative overnight call rate, which sensitively reflects the surplus and shortage of excess reserves in the interbank market. It affects monetary and financial conditions, which in turn have a bearing on critical aspects of the broader economy including employment, growth and inflation. And finally, VUCI is an indicator capturing the uncertainty of domestic equity market returns.

Table 5 shows summary statistics for the variables. Applying the Levin et al. (2002) and Im et al. (2003) tests to check if panel datasets are stationary demonstrates that all series except for DTE reject the null hypothesis at the 1% level. Series VIX, FED and VUCI are also stationary.

Table 6 shows the lag terms of ΔCoVaR are significantly positive at a statistical level of 5% or 10%, which accords with the autocorrelation of most financial time series. In other words, the current level of risk spillovers is positively correlated with the level of risk spillovers in the past period.

For the banks with steady earnings and cash flows, a greater DTE is preferred since it uses leverage to boost return on equity (ROE) and lower weighted average cost of capital. However, if the DTE is excessively high, the bank may not be able to repay its debts or even go bankrupt, which drives up risk spillovers. The estimated coefficients consistently demonstrate that a larger DTE is linked to a higher systemic risk, despite their differing degrees of significance.

Based on baseline models (1) and (2), we add two global financial variables to models (3) and (4). Global risks are bound to be transmitted to the domestic financial market and commodity market, which is why the coefficient of VIX is positive and significant. The change in interest rates will have an impact on the bank's lending and investment activities, as well as risk spillovers. For instance, a rise (fall) in interest rate will boost (curb) the cost of borrowing for businesses, and from the standpoint of banks, it will increase (decrease) the yield, which could improve (deteriorate) risk-taking capacity. The significance and sign of the FED coefficients are as expected. Domestic equities market volatility has a positive impact on ΔCoVaR in models (5) and (6) as VIX does.

Regarding the short panel (known as the large-N and small-T case), it is difficult to determine whether the error term exhibits autocorrelation in that the information of each unit is insufficient. However, the error term is assumed to be independent and identically distributed (IID). For a long panel (large-T and N > T case), the assumption can be loosened. Meanwhile, given the possible in-group autocorrelation, between-group heteroscedasticity and contemporaneous correlation, we employ the panel generalized least squares (GLS) as a robustness check. Table 7 substantiates that the results are robust since most variables are slightly different in significance level and magnitude. While the sign of FED is as expected, it is no longer significant.

Additionally, we estimate a three-variable panel vector autoregression (Panel VAR) model to examine the dynamic relationship among several key variables and pave the way for providing policy implications as ΔCoVaR is significantly affected by the VIX and FED, both of which are exogeneous variables within the system. Therefore, we can estimate an extended model by adding lagged endogenous and exogenous simultaneously, Panel VAR with exogenous variables, i.e. Panel VARX as in Equation (17).

As a nontheoretical model, the VAR framework does not require any prior constraints on variables. Frequently, we are more interested in the dynamics that are predicted by VAR models than the actual coefficients that are estimated. For this reason, it is most common that VAR studies report impulse response functions (IRFs) and forecast error variance decomposition (FEVD). The IRFs are conducted to trace the dynamic path of variables in the system in response to shocks to other variables, focusing more on the dynamic impact on the system when an error term changes or the model is affected by a certain shock, rather than analyzing the impact of a change in one variable on another variable. A common approach to identifying the shocks of a VAR model is to employ orthogonal impulse responses (OIRs). The basic idea is to decompose the variance–covariance matrix so that Σ = PP’, where P is a lower triangular matrix with positive diagonal elements, which is often obtained by a Cholesky decomposition. Note that the results of an OIR might be sensitive to the order of the variables. We specify Cholesky ordering of endogenous variables as follows: VUCI, DTE and ΔCoVaR, as the volatility of equity market could affect the equity-to-debt ratio. In addition, Granger causality tests show that DTE does help predict VUCI. Therefore, we subsequently adjust the order of endogenous variables as a robustness check. FEVD separates the forecast error variance into proportions attributed to each variable in the system. To be more specific, FEVD further evaluates the importance of different structural shocks by analyzing the contribution of each structural shock to the change of endogenous variables (usually measured by variance).

Figure 4 presents that the positive shock of DTE will be transmitted to ΔCoVaR, which will cause the systemic risk to rise in the initial stage. The magnitude reaches a peak in the first period and remains stable until the second period before it starts to fall. The shock only lasts for three months and then becomes insignificant. When an external shock hits the VUCI, ΔCoVaR will ramp up dramatically in the next month and subsequently moves towards the horizontal line. Albeit statistically significant, the shock does not last for long. The results are robust, regardless of the ordering of endogenous variables as illustrated in Figure 5.

Table 8 shows the contribution of each variable to ΔCoVaR variation. At the beginning, approximately 80% of the variation in the ΔCoVaR is from shocks to ΔCoVaR itself, and most of the remaining 20% is from VUCI. The contribution of DTE to the variation in the ΔCoVaR changes slowly over time and eventually seems to converge at around 5.4%. Table 9 reveals that the adjustment of endogenous variables has little impact on the results. The system becomes stable after one year, with the contribution of ΔCoVaR, DTE and VUCI to variation in ΔCoVaR leveling off at roughly 74%, 5% and 21%, respectively.

This study empirically analyzes risk spillovers among 21 commercial banks in Uzbekistan using the CoVaR framework combined with quantile regression, panel regression and PVARX models. The main findings can be summarized as follows. First, static CoVaR results show that banks with higher individual risk generate larger spillover effects on the banking system, as reflected in the close correspondence between VaR and ΔCoVaR rankings. This implies that strengthening internal risk management at the bank level can effectively enhance system-wide resilience. Dynamic CoVaR estimates further indicate that systemic risk surged sharply during the COVID-19 period, peaking in March 2021, before declining rapidly in subsequent months. Risk spillovers across banks also exhibit broadly synchronized dynamics over the sample period.

Second, panel regression results indicate that macroeconomic and global financial variables exert a stronger influence on systemic risk than bank-specific financial indicators. This finding reflects the heightened vulnerability of Uzbekistan's banking sector to external shocks. Finally, PVARX analysis using impulse response functions and forecast error variance decomposition reveals that domestic equity market volatility contributes more to variations in systemic risk than bank leverage, reinforcing the view that Uzbekistan's banking system is highly exposed to global financial conditions.

Based on these findings, several policy implications emerge. Individual banks should prioritize reducing their risk exposure through stronger internal governance and more comprehensive risk measurement systems. At the system level, risk management and supervisory frameworks should be further strengthened. Many Uzbek commercial banks exhibit volatile earnings and high leverage, while asset quality and profitability face increasing pressure from elevated inflation and slowing economic growth. Establishing dedicated risk management institutions and developing multiple risk indicators, rather than relying on a single metric, would improve resilience to diverse financial risks. In addition, a nationwide system for risk monitoring and financial supervision is advisable, particularly given the systemic importance of large banks.

Overall, the results indicate that Uzbekistan's banking sector remains vulnerable to external shocks. Accordingly, the timely establishment of an effective macroprudential policy framework and the proactive design of policy tools to address external systemic risks are essential for safeguarding financial stability.

Earlier version of this manuscript was prepared as part of the final report for the Bank of Korea’s 2022 Knowledge Partnership Program.