3.1. Data Source
The current study has included various independent variables to study the impact of financial development on NPLs among the selected emerging countries. Data is extracted from the global financial indicator database of the World Bank, covering the period from 1995 to 2020. This study includes the data of 12 emerging economy countries selected based on their growth rate, per capita income, and the classification done by the world bank. Only 12 countries are included out of the total emerging countries due to the lack of data availability of some of the variables and the convenience of sampling. In context to the bank-specific variable affecting NPLs, this study has included bank cost-to-income ratio (B.I.) to measure banking efficiency [
47], loan to deposit (L.D.) for measuring banking liquidity [
48], bank non-interest income to total Income (N.I.) for considering bank dependencies on interest income of loans [
49], and regulatory capital (R.A.) for measuring banking stability [
50].
This study has used the following variables to measure financial development. Bank deposit to GDP (BDGDP) for measuring the size of the bank [
11], private credit to GDP(PCGDP) for measuring financial intermediation [
51], and measure the influence of financial liberalization, foreign bank asset to total bank asset ratio (FABA) is considered [
40]. Apart from the above to measure banking competitiveness and stability, this study has also used Lerner (L.R.) index for competitiveness and Z-score for banking stability besides incorporating macroeconomic variables growth rate (G.R.), unemployment (U.N.), inflation (IN), and interest rate (I.R.).
Table 1 shows the expected relationship between the explanatory and the outcome variables. In addition, it also highlights the reasonable justifications for the above relationship supported by previous empirical literature.
Table 1. shows that we expect that the following variables, bank cost to income ratio, bank non-interest income to total income, regulatory capital, bank deposit to GDP ratio, foreign bank asset to total bank asset ratio, Z-Score, and growth rate exert a negative influence on the NPLs, whereas, loan to deposit ratio, private credit to GDP ratio, unemployment rate, inflation rate, and interest rate exerts a positive impact on the NPLs.
3.2. Econometric Methodology
This study has used the Autoregressive Distribution Lag (ARDL) method given by Pesaran et al. [
52] and Nonlinear Autoregressive Distribution Lag (NARDL) method as suggested by Shin et al. [
1], as these methods provide a robust result when the data set is small, and variables are mixed integrated, i.e., I(0) and I(1). NARDL approach is utilized because this method provides the best results after jointly checking the cointegration and asymmetry of the data. Moreover, it also considers hidden cointegration [
53]. The only precondition before applying ARDL or NARDL is that none of the variables is of the second order of integration, and to check this Augmented dicky fuller test and CIPS unit root test is used.
An extension of Auto-Regressive Distribution Lag is the Non-linear Auto-Regressive Distribution Lag; thus, we provide the basic form of unrestricted error correction linear ARDL method:
where
vector of regressors,
dependent variable;
Intercept;
shows variables are in difference;
,
represents short-run coefficient;
p and
q represent restricted lags and
is the error term.
In the ARDL methodology, we determine cointegration among variables based on the upper and lower bound values. The null hypothesis states no cointegration, whereas the alternative hypothesis entails the presence of cointegration. The Upper bound values take variables as I(1) order of integration; the lower bound value considers that variables are of I(0). In our study, some variables are I(1), and some are I(0), so we can consider both the lower and upper bound values. Subsequently, based on the F-statistic value for the upper and lower bound mentioned in the below tables, we can accept or reject the null and alternative hypothesis. The ARDL technique assumes a symmetric response between the dependent variable and all exogenous variables. Therefore, we have used the following equation:
In Equation (2), Ln shows that all the explanatory and outcome variables are in the natural log form. Here, NPL denotes non-performing loans, B.I. denotes bank cost to income ratio, L.D. denotes loan to deposit ratio, N.I. Indicates non-interest income to total income ratio, R.A. depicts regulatory capital, BDGDP denotes bank deposit to GDP ratio, PCGDP denotes private credit to GDP ratio, FABA denotes foreign bank asset to total bank asset, C.O. represents Lerner index, S.T. denotes Z-score, GDP shows growth rate, U.N. denotes unemployment, IN denotes inflation rate, I.R. shows interest rate, and finally is the error term.
Shin et al. (2014) used the following asymmetric long-run model for checking the asymmetric relationship among the variables:
where
and
represents the long-run asymmetric coefficients;
and
shows error term and vector regressors, respectively.
is further divided into negative and positive terms, shown in Equation (4).
The partial sum of negative and positive shocks can be further represented as follows:
Moving further with the model formulation, we combine Equations (1) and (2) to derive the asymmetric ECM, which is mentioned below:
where
and
are short-run adjustments for positive and negative shocks. Finally, the NARDL equation for short and long-run asymmetries in our case can be expressed as follows:
where, L
n denotes Natural Log, and + and – sign denotes the positive and negative variation of explanatory and outcome variables.
Prior to utilizing the Non-linear ARDL methodology, it is essential to check variables’ stationarity. We have ensured the stationarity of the variables via ADF and CIPS unit root tests, and no I(2) integrated variable was found. The conventional unit root test does not have the power to detect breaks in the series. To this end, we utilized Zivot and Andrews unit root test, which not only captures breaks in the data but also further validates the integration order. The study further reported to Brock, Dechert, and Scheinkman (BDS) test in order to confirm nonlinearity. The outcomes of the BDS test indicate that the data is not identical and independently distributed through the rejection of the null hypothesis. The outcomes of the BDS test allow us to proceed with the Asymmetric/non-linear ARDL approach. The Non-linear autoregressive distribution lag method is used similarly to the linear autoregressive distribution lag is used. Initially, the error correction model is evaluated through ordinary least square then the asymmetric long-run relationship is investigated by employing the bound test along with examining the null hypothesis of no-cointegration by considering lower and upper bounds. Finally, the asymmetric and the symmetric effect is inspected on exogenous variables both for the short-run period and long-run period by using the Wald test.
3.3. Analysis and Discussion
Before proceeding with the symmetric and asymmetric analysis, descriptive properties of the variables are discussed.
Table 2 reports the details of the descriptive statistics of the explanatory and outcome variables. The mean of NPLs in the selected emerging countries is comparatively higher, which is 11.57%. Likewise, the proportion of non-interest income (32.42) and the mean of economic growth (4.02) is low. In addition, the descriptive statistic also reveals a low level of financial sector development in the panel of selected countries. For instance, the mean score of private credit to GDP ratio (42.70%), bank asset to GDP ratio (39.01), and foreign bank presence (14.05) is low compared to other developed economies. Moreover, competitive and stability levels are also low in the selected emerging countries. Furthermore, the descriptive statistic also shows that the data is not normal, which motivates us to proceed with the non-linear approach.
Assessing the level of data integration is one of the preconditions of applying the ARDL and NARDL models. Therefore, the augmented dicky fuller technique and CIPS unit root test is employed to confirm the stationarity of data. A broad strand of the literature suggests that NARDL and ARDL approach is best suited when the data is of the mixed level of integration, and none of the variables is of the second order of integration. The result of the ADF test and CIPS test (
Table 3) further substantiate mixed integration and reveal that variables like private credit and bank deposit to GDP, non-interest income, growth rate and unemployment are integrated at levels, whereas all the other variables are integrated at first difference.
Sometimes traditional unit root test provides spurious results due to ignoring the concept of a structural break. Thus, Zivot and Andrew’s (Z.A.) 1992 structural break test is used to make this study robust and error-free.
Table 4 shows the outcome of the Z.A. test, which substantiates a structural break in the series. The result of the Z.A. test also confirms that none of the variables is of second-order integration.
As the data suffers from the issue of time break, it is required to check the nonlinearity of the series. The study further uses Brock, Dechert, and Scheinkman (BDS) test to confirm nonlinearity. The findings of the BDS test conclude that the data is not identical and independently distributed by rejecting the null hypothesis.
Table 5 shows the result of the BDS test. As all the results of the unit root project that none of the variables is of the second order of integration, the study further proceeds with the analysis of variables using the ARDL and NARDL approaches.
To analyze the short-run and long-run relationship between the variables, ARDL and NARDL methods are employed, using a maximum of four lags. Using Akaike’s Information Criteria (AIC), various models are estimated, and the appropriate model is selected based on the lower value of AIC.
Table 6 (section a) shows a detailed analysis of the ARDL model. The short-run result of the ARDL model reveals that banking efficiency and banking regulatory capital has a negative and significant impact on banking NPLs. It implies that higher efficiency and regulation help in reducing NPLs. Higher regulation and efficiency enhance accountability and promote traceability, hence assisting in lowering the proportion of NPLs. However, in the short run, the loan-to-deposit ratio positively and significantly affects NPLs in emerging countries. In context to macroeconomic variables, growth rate, interest rate, and inflation are prominent factors affecting NPLs. The short-run result depicts that financial development is not so important determinant for emerging countries. The short-run outcome of the macroeconomic variables corroborates the findings of Nkusu [
4], Staehr and Uusküla [
23].
To further substantiate the findings, the long-run results of the ARDL model are also analyzed, which shows that in the long run, emerging countries banking NPLs are significantly affected by financial development. Findings show that banking efficiency (B.I.), banking stability capital (R.A.), banking other sources of Income (N.I.), and banking stability index (S.T.) has a significant and negative impact on NPLs. It implies that higher banking efficiency, adequate regulatory capital, stable banking index, and higher dependency on other sources of income other than interest income can help lower NPLs in emerging countries. On the other hand, financial intermediation (PCGDP), banking outreach (L.D.), and competition (C.O.) have a significant and positive impact on NPLs. Meaning that higher competition, higher financial intermediation, and too much credit lead to risky decision and thus increases NPLs. The resulting outcome supports the conceptual viewpoint, which emphasizes that rapid financial sector development raises the probability of systematic risk, and at a certain point, the benefits of financial sector development outweigh the cost [
54]. The long-run outcome of the ARDL model also strengthens the above conceptual argument. In addition, the findings also reveal that foreign bank presence is not so significant for emerging countries. Although the coefficient shows a negative sign as foreign bank presence is low in emerging countries. In context to traditional macroeconomic factors, unemployment and growth are the significant factors contributing to NPLs in emerging countries. The findings do not support the work of Boyd and De Nicolo [
55] in the context of the competition index, who argued that higher competition increases efficiency and reduces NPLs. However, in the context of financial intermediation, regulatory capital, and stability, this study supports the work of Ozili [
56].
Before applying the ARDL test, diagnostic tests are performed, and the results of which are reported in Section b of
Table 6. To check the level of cointegration among the variables Wald test, as proposed by Pesaran et al. [
52], and the error correction model (ECM) is used. F-statistic is 4.15, which is more than the upper bound value. Thus, it can be concluded that variables have a long-run cointegration and error correction model, which further supports the results of the Wald test as the ECM is negative and significant. The L.M. test and RESET test are employed to test serial correlation and model specification. The result shows that there is no serial correlation as the value of the L.M. test (4.2) is insignificant at a 5% level of significance. Ramsey’s RESET test result confirms that the model is correctly specified, as the value (11.13) is more than the critical value. Lastly, the R-square value, which is 0.62, also substantiates that the model enjoys a good fit.
The ARDL model depicts the short-run and long-run symmetric relationships among the variables. However, we employed the NARDL approach to investigate positive and negative shocks and how these alter among the independent and dependent variables. The NARDL model uses a similar procedure as followed by the ARDL model. The NARDL model reported in Section a of
Table 7 shows that in the short-run, a decrease in bank efficiency in terms of cost-to-income ratio increases NPLs, whereas an increase in bank efficiency reduces NPLs. It implies a short-run asymmetric relationship between banking efficiency and NPLs. In addition, in context to the macroeconomic variables, the result shows that negative shocks in the growth rate increase NPLs, and a positive shock in the interest rate and inflation rate increases NPLs. This substantiates that there is a short-run symmetric relationship between macroeconomic variables and NPLs. Furthermore, the result also highlights that financial development indicators have an insignificant relationship with the NPLs in the short run. Because the changing dynamics of financial development vary in the countries studied, the response of these financial development policies to the level of economic growth and the banking industry is a long-term phenomenon [
46].
Proceeding to long-run asymmetric estimation, the results show that all the financial development indicators and traditional macroeconomic variables have a significant impact on the NPLs, except for foreign banks’ presence, which is insignificant for emerging countries, as the concentration of foreign banks is less in emerging countries. The long-run result shows that the positive shocks of financial development indicators (financial intermediation, size of the banks) and banking indicators (bank cost to income ratio, banking competition, loan to deposit ratio) have a significant positive impact on NPLs. In other words, when a positive shock is exerted on financial intermediation, size of the banks, bank cost-to-income ratio, banking competition, and loan-to-deposit ratio, NPLs tend to increase. However, when a negative shock is given to the above variables, NPLs show a decreasing trend. It implies an asymmetric long-run relationship between financial development and NPLs. Such a relationship exists because the initial pace of financial deepening in emerging countries generates financial sector instability. It has been documented that a rapid transformation in the financial structure which is weakly regulated and monitored, can be accompanied by greater risk-taking, competition, and high leverage, thereby increasing the consequences of a banking crisis in emerging countries [
57].
Furthermore, prior research suggests that when a threshold limit of private credit to GDP is reached, financial development slows growth since credit expansion beyond a certain point is a major determinant of banking non-performing loans (NPLs) [
58,
59,
60]. The excessive competition created by new entrants also leads to credit expansion. Excessive competition raises the risks of inadequate credit redressal processes, which leads to an increase in non-performing loans (NPLs) and vice versa. The same arguments, on the other hand, do not hold to countries with well-developed and regulated financial systems. A well-established financial system, especially in developed countries, aids in adapting to financial crises and maintaining a sound banking system [
61,
62]. In addition, structured financial development creates a platform for healthy competition in developed economies. The above arguments corroborate the existence of a non-linear relationship between financial development and NPLs and strengthen the outcome that a high level of financial intermediation and competition increases the NPLs in emerging countries and vice versa.
Our study also concludes that when a positive shock (negative shock) is given to non-interest sources of income, stability index, and regulatory capital, NPLs tend to show a decreasing trend (increasing trend). The plausible reason is that with an increase in banking efficiency, regulation and monitoring enhances, and thus NPLs decrease. Berger and DeYoung [
63] also supported the above outcome and concluded that an increase in operating efficiency increases investments in underwriting and thus reduces NPLs. On the other hand, due to skimping hypothesis, low operating efficiency leads to higher NPLs. Similarly, diversifying the bank’s source of income also limits the excessive reliance on bank interest rate spread which ultimately reduces excessive credit disbursement, and hence NPLs decrease.
In contrast, too much dependency on the bank interest source of income increases the loan disbursement, often resulting in higher NPLs. These explanations also support a non-linear relationship between banking efficiency, non-interest source of income, and NPLs and further align with the studies conducted by (Fernandes et al. [
42]; Berger and DeYoung [
63]. Moreover, in the context of traditional macroeconomic variables, our study shows that in the long run, positive shocks in the unemployment rate, inflation, and interest rate positively and significantly impact NPLs. It implies that with the increase(decrease) in the above-mentioned macroeconomic variables, NPLs increase (decrease). Higher unemployment, inflation, and interest rate restrict the debt servicing capacity of the borrowers and thus creates positive pressure on the NPLs.
Conversely, a low unemployment rate, moderate inflation, and stable interest rate assist in increasing the Debt servicing capability, and hence NPLs show a decreasing trend. Lastly, the results also concluded that positive shocks in the growth rate help reduce NPLs in the long run, whereas negative shocks increase NPLs. A favourable economic growth rate creates an optimistic business environment that assists in generating savings, consumption, and investments. An increase in consumption, saving and investments assist in income generation, and higher income promotes fewer loan defaults and vice-versa. The explanation and outcome are in line with the study carried out by (Nkusu [
4]; Staehr, and Uusküla [
23].
Before applying NARDL, diagnostic checks are also performed to confirm the suitability of the model. The result of the diagnostic test (Section (b) of
Table 7) shows that there is a cointegration among the variables in the long run based on the joint significance F-value, which is 9.15. The ECM value also substantiates the findings of F statistics. L.M. and RESET tests show that the data has no serial correlation and is well specified as the value is lower than the critical value. The adjusted R-square value of the NARDL model is 68%, which supports that the findings of NARDL are more robust than the ARDL model, which has an R-square value of 62%. The diagnostic test also shows the WALD value of long-run and short-run analysis, which confirms the presence of an asymmetric relationship among the variables in the long-run and short-run. Refer to
Appendix B for a list of econometric tests carried out.