Macroprudential policy, bank risk and efficiency: some evidence from China

Abstract

This paper examines the impact of macroprudential policy on bank risk by utilizing the panel data of 360 commercial banks in China during the period of 2007–2018. The results demonstrate that overall bank risk decreases in response to tightened macroprudential policy. Furthermore, listed banks and banks with larger size benefit more from macroprudential tightening, and macroprudential policy works more effectively when monetary policy is tightened. Additionally, our findings uncover the mediating role of bank efficiency in the relationship between macroprudential policy and bank risk. We observe that bank efficiency increases when macroprudential policy is tightened, which subsequently contributes to a reduction in bank risk. This impact on bank risk is primarily achieved by mitigating banks’ leverage and enhancing their profitability.

Keywords:

• Macroprudential policy
• bank risk
• bank efficiency
• mediation analyses
• china
• listed banks

JEL classification:

• E58
• G18
• G21

Acknowledgement

Minghua Chen and Ji Wu thank the support of the Guanghua Talent Project of Southwestern University of Finance and Economics.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 Our sample period is selected due to the following reasons. First, the data of the majority of Chinese commercial banks prior to 2007 are not available in the Wind Financial Database and the CSMAR Database. Second, new accounting rules were introduced in China in 2007, which may cause the data prior to 2007 not comparable to those in the aftermath. Third, we truncate the sample period up to 2018 to avoid possible influences of COVID-19 to our data.

2 The Macroprudential Policy Database provided by Alam et al. (2019) integrates information from some major existing databases, central bank reports, and IMF’s annual Macroprudential Policy Survey. It excels some other earlier databases by covering a richer set of macroprudential instruments for a longer period.

3 The data for MaPP_C are available from 2000 to 2018, while those for MaPP_KS are only available from 2000 to 2015.

4 Stochastic frontier approach is commonly employed in many researches related to bank efficiency due to its advantage in distinguishing inefficiency from random disturbances and measurement errors (e.g. Berger, Hasan, and Zhou 2009; Sun, Harimaya, and Yamori 2013; Pessarossi and Weill 2015). Different from some prior SFA approaches, Belotti and Ilardi (2018) further rule bank-specific fixed effects out from the measurement of bank efficiency.

5 We appreciate an anonymous reviewer for pointing out that most of our control variables in baseline estimations are not statistically significant. We selected our control variables based on existing literature, which typically includes various bank characteristics and macroeconomic conditions (e.g. Laeven and Levine 2009; Demirgüç-Kunt and Huizinga 2010; Delis and Kouretas 2011; Chen et al. 2017). On one hand, incorporating a relatively rich set of covariates in our estimation helps address the issue of “omitted variables” and reduces the likelihood of biased estimation. This is particularly important since our study primarily aims to identify the impact of macroprudential regulations, which can potentially be related to various bank characteristics. On the other hand, including multiple covariates in our estimation may weaken their individual explanatory power, as these bank characteristics and macroeconomic conditions could be correlated to some extent. Although we have examined the correlation coefficients among these control variables and found only mild correlation, their mutual correlations could still diminish the significance of our estimates, especially given the limited observations in our empirical tests.

6 We have explored the heterogeneous effects of macroprudential instruments with different targets. We find that their impacts significantly vary in terms of magnitude and direction. The results pertaining to these examinations are not reported in this paper due to the limited space, but they are available upon request.

7 We estimated Equations (3)(3) $$\begin{array}{l}\mathit{\text{lnT}}{C}_{\mathit{\text{it}}}=c+\sum _{h=1}^3 {\alpha }_h{\left(\mathrm{\text{ln}}\hspace{0.17em}{y}_h\right)}_{\mathit{\text{it}}}+\frac{1}{2}\sum _{h=1}^3 \sum _{k=1}^3 {\alpha }_{\mathit{\text{hk}}}\mathit{\text{ln}}{\left(y_h\right)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(y_k\right)}_{\mathit{\text{it}}}+\sum _{m=1}^2 {\beta }_m\mathit{\text{ln}}{\left(p_m\right)}_{\mathit{\text{it}}}\\ +\frac{1}{2}\sum _{m=1}^2 \sum _{n=1}^2 {\beta }_{\mathit{\text{mn}}}\mathit{\text{ln}}{\left(p_m\right)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(p_n\right)}_{\mathit{\text{it}}}+\frac{1}{2}\sum _{h=1}^3 \sum _{m=1}^2 {\varphi }_{\mathit{\text{hm}}}\mathit{\text{ln}}{\left(y_h\right)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(p_m\right)}_{\mathit{\text{it}}}+\delta \mathit{\text{ln}}(\mathit{\text{NP}}{)}_{\mathit{\text{it}}}\\ +\sum _{h=1}^3 {\kappa }_h\mathit{\text{ln}}(\mathit{\text{NP}}{)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(y_h\right)}_{\mathit{\text{it}}}+\sum _{m=1}^2 {\rho }_m\mathit{\text{ln}}(\mathit{\text{NP}}{)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(p_m\right)}_{\mathit{\text{it}}}+{\mathit{\text{Year}}}_t+f_i+{\epsilon }_{\mathit{\text{it}}}\end{array}$$(3) and (4)(4) $${\epsilon }_{\mathit{\text{it}}}=v_{\mathit{\text{it}}}+u_{\mathit{\text{it}}}$$(4) and Equation (7)(7) $${\mathit{\text{Efficiency}}}_{\mathit{\text{it}}}={\mathrm{ }c}_2+{\beta }_2{\mathit{\text{MaPP}}}_t+{\gamma }_2{\mathit{\text{BankChar}}}_{\mathit{\text{it}}-1}+{\delta }_2{\mathit{\text{Macro}}}_{t-1}+f_i+{\epsilon }_{\mathit{\text{it}}}$$(7) separately rather than simultaneously. Some works (e.g. Battese and Coelli (1995), Greene (2005), and Wang and Ho (2010)) enable researchers to estimate the values of the (in)efficiency item u_it, along with the estimates for the effects of some determinants for the mean of u_it. For this purpose, these approaches usually assume that the (in)efficiency item u_it follows a certain distribution form with a specific function for u_it. For example, Battese and Coelli (1995) assumes that u_it is a truncation (at zero) of a normal distribution with a mean of z_itδ, where z_it is a vector of explanatory variables and δ represents coefficients to be estimated. It is important to note that the coefficients of δ are the coefficients on the determinants of the mean of u_it, rather than the coefficients on the determinants of u_it. In other words, the coefficients δ cannot be interpreted as the estimated effects of the determinants on the (in)efficiency term u_it itself, but rather the effects of the determinants on the expected value of u_it given the explanatory variables z_it. Additionally, the true distribution of u_it can be difficult to identify, and the assumed distribution may not accurately reflect u_it, leading to potentially inaccurate estimated results. For these reasons, we first adopted the stochastic frontier approach recommended by Belotti and Ilardi (2018) to estimate Equations (3)(3) $$\begin{array}{l}\mathit{\text{lnT}}{C}_{\mathit{\text{it}}}=c+\sum _{h=1}^3 {\alpha }_h{\left(\mathrm{\text{ln}}\hspace{0.17em}{y}_h\right)}_{\mathit{\text{it}}}+\frac{1}{2}\sum _{h=1}^3 \sum _{k=1}^3 {\alpha }_{\mathit{\text{hk}}}\mathit{\text{ln}}{\left(y_h\right)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(y_k\right)}_{\mathit{\text{it}}}+\sum _{m=1}^2 {\beta }_m\mathit{\text{ln}}{\left(p_m\right)}_{\mathit{\text{it}}}\\ +\frac{1}{2}\sum _{m=1}^2 \sum _{n=1}^2 {\beta }_{\mathit{\text{mn}}}\mathit{\text{ln}}{\left(p_m\right)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(p_n\right)}_{\mathit{\text{it}}}+\frac{1}{2}\sum _{h=1}^3 \sum _{m=1}^2 {\varphi }_{\mathit{\text{hm}}}\mathit{\text{ln}}{\left(y_h\right)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(p_m\right)}_{\mathit{\text{it}}}+\delta \mathit{\text{ln}}(\mathit{\text{NP}}{)}_{\mathit{\text{it}}}\\ +\sum _{h=1}^3 {\kappa }_h\mathit{\text{ln}}(\mathit{\text{NP}}{)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(y_h\right)}_{\mathit{\text{it}}}+\sum _{m=1}^2 {\rho }_m\mathit{\text{ln}}(\mathit{\text{NP}}{)}_{\mathit{\text{it}}}\mathit{\text{ln}}{\left(p_m\right)}_{\mathit{\text{it}}}+{\mathit{\text{Year}}}_t+f_i+{\epsilon }_{\mathit{\text{it}}}\end{array}$$(3) and (4)(4) $${\epsilon }_{\mathit{\text{it}}}=v_{\mathit{\text{it}}}+u_{\mathit{\text{it}}}$$(4) for the values of bank (in)efficiency item u_it, and subsequently used them as the dependent variable in our estimation for Equation (7)(7) $${\mathit{\text{Efficiency}}}_{\mathit{\text{it}}}={\mathrm{ }c}_2+{\beta }_2{\mathit{\text{MaPP}}}_t+{\gamma }_2{\mathit{\text{BankChar}}}_{\mathit{\text{it}}-1}+{\delta }_2{\mathit{\text{Macro}}}_{t-1}+f_i+{\epsilon }_{\mathit{\text{it}}}$$(7) . Unlike previous studies such as Battese and Coelli (1995), Belotti and Ilardi (2018) allows for an arbitrary distribution of the inefficiency term, making their estimates of u_it more robust against possible wrong assumptions about the distribution form of u_it.

8 A caveat of our mediation analysis is that bank efficiency may be endogenous in Equation (8)(8) $${\mathit{\text{BankRisk}}}_{\mathit{\text{it}}}={\mathrm{ }c}_3+{\beta }_3{\mathit{\text{MaPP}}}_t+{\theta }_3{\mathit{\text{Efficiency}}}_{\mathit{\text{it}}}+{\gamma }_3{\mathit{\text{BankChar}}}_{\mathit{\text{it}}-1}+{\delta }_3{\mathit{\text{Macro}}}_{t-1}+f_i+{\epsilon }_{\mathit{\text{it}}}$$(8) and cause biased estimates. The endogeneity problem within the mediation analysis framework has been extensively discussed (and criticized) in the existing literature. Employing instrumental variables could help alleviate this endogeneity problem. However, finding suitable instrumental variables is particularly challenging in our case, as they need to be correlated with banks’ efficiency while remaining uncorrelated with their risk. It is foreseeable that any bank feature assumed to be related to efficiency could also be argued to be related to risk. Additionally, a valid mediation analysis requires that the remaining effect of macroprudential regulations on bank risk, after accounting for their effects through efficiency, is lower than their overall effect. In our context, this means comparing the estimated coefficient of MaPP in Equation (8) wi(8) $${\mathit{\text{BankRisk}}}_{\mathit{\text{it}}}={\mathrm{ }c}_3+{\beta }_3{\mathit{\text{MaPP}}}_t+{\theta }_3{\mathit{\text{Efficiency}}}_{\mathit{\text{it}}}+{\gamma }_3{\mathit{\text{BankChar}}}_{\mathit{\text{it}}-1}+{\delta }_3{\mathit{\text{Macro}}}_{t-1}+f_i+{\epsilon }_{\mathit{\text{it}}}$$(8) th that in Equation (6) to(6) $${\mathit{\text{BankRisk}}}_{\mathit{\text{it}}}={\mathrm{ }c}_1+{\beta }_1{\mathit{\text{MaPP}}}_t+{\gamma }_1{\mathit{\text{BankChar}}}_{\mathit{\text{it}}-1}+{\delta }_1{\mathit{\text{Macro}}}_{t-1}+f_i+{\epsilon }_{\mathit{\text{it}}}$$(6) validate the existence of a mediation effect. If we were to use instrumental variable approaches to estimate Equation (8)(8) $${\mathit{\text{BankRisk}}}_{\mathit{\text{it}}}={\mathrm{ }c}_3+{\beta }_3{\mathit{\text{MaPP}}}_t+{\theta }_3{\mathit{\text{Efficiency}}}_{\mathit{\text{it}}}+{\gamma }_3{\mathit{\text{BankChar}}}_{\mathit{\text{it}}-1}+{\delta }_3{\mathit{\text{Macro}}}_{t-1}+f_i+{\epsilon }_{\mathit{\text{it}}}$$(8) , comparing these two coefficients would become challenging. Therefore, while our mediation analysis offers valuable insights into the mechanisms by which macroprudential regulations impact bank risk, we approach the specific estimates in Equation (8) wi(8) $${\mathit{\text{BankRisk}}}_{\mathit{\text{it}}}={\mathrm{ }c}_3+{\beta }_3{\mathit{\text{MaPP}}}_t+{\theta }_3{\mathit{\text{Efficiency}}}_{\mathit{\text{it}}}+{\gamma }_3{\mathit{\text{BankChar}}}_{\mathit{\text{it}}-1}+{\delta }_3{\mathit{\text{Macro}}}_{t-1}+f_i+{\epsilon }_{\mathit{\text{it}}}$$(8) th caution. We highly appreciate an anonymous reviewer for bringing up this issue.

9 Two alternative tests, namely the causal steps approach proposed by Baron and Kenny (1986) and the Sobel test suggested by Sobel (1982), have some disadvantages compared to the approach of bias-corrected non-parametric percentile bootstrap, e.g., low testing power and inappropriate assumption for the distribution of the coefficient product.

10 Due to the likelihood that financial regulators may adjust macroprudential policy with the variation in bank efficiency, we checked whether our estimation result in Equation (7) ho(7) $${\mathit{\text{Efficiency}}}_{\mathit{\text{it}}}={\mathrm{ }c}_2+{\beta }_2{\mathit{\text{MaPP}}}_t+{\gamma }_2{\mathit{\text{BankChar}}}_{\mathit{\text{it}}-1}+{\delta }_2{\mathit{\text{Macro}}}_{t-1}+f_i+{\epsilon }_{\mathit{\text{it}}}$$(7) lds by using the 2SLS approach. We still used macroprudential policy indices from other emerging economies (Alam et al. 2019) and calculated the average as an instrument for Chinese macroprudential policy. Although not explicitly reported, the results confirm that our previous finding regarding the impact of macroprudential policy on bank efficiency remains qualitatively unchanged.

11 We conducted a robustness check by performing the Tobit regression. In this analysis, we used bank efficiency as the dependent variable and adjusted it to be within the range of (0, 1). Our findings from the Tobit regression align with our previous results, showing that bank efficiency increases in response to tighter macroprudential activities.

12 This “direct effect” comprises of unexplored “indirect effects.” We focus on the mediating effect of bank efficiency in this research, leaving other mechanisms to be explored in our future research agenda.

13 We appreciate an anonymous reviewer for pointing out that bank efficiency could also have a moderating effect in the relationship between macroprudential regulations and bank risk. We experiment with incorporating the interactive term MaPP × Efficiency into the bank risk equation and find some supportive evidence that efficiency may act as a moderating factor that influences the varying effects of macroprudential regulations across different banks. Specifically, (tightening) macroprudential activities tend to have more pronounced impacts on banks with higher efficiency. However, it is noteworthy that the estimate for MaPP × Efficiency is only marginally significant, indicating a modest moderating influence of bank efficiency on the relationship between macroprudential policy and bank risk.

14 The 95% bias-corrected confidence interval is [−0.0004, 0.00004], hence the hypothesis that the indirect effect here is not different from zero fails to be rejected.