On the Determinants of NPLs: Lessons from Greece

Abstract

We investigate the relationship between non-performing loans (NPLs ) and their fundamentals, mainly bank and macroeconomic variables . This is done based on aggregate portfolio loans from the Greek economy. This economy constitutes an interesting case to study the factors determining NPLs , given the pervasive recessionary conditions held in it after the global financial crisis and the outbreak of its sovereign debt crisis occurred in the year 2010. We suggest a new econometric framework to study the above relationship which extends the SUR (seemingly unrelated regressions) framework to allow for a common break in its slope coefficient of unknown time. We show that the deterioration in the macroeconomic conditions (captured by very high rates of unemployment ) and political uncertainty constitute key factors in explaining the sharp rise of NPLs of the Greek banking sector after the first quarter of the year 2012. With the exception of bank profitability, we find that bank-specific variables associated with bank capitalization and liquidity risk seem to determine NPLs only under normal economic conditions.

1 Introduction

Since the onset of the financial crisis academics and practitioners have shown renewed interest in the credit quality of loan portfolios. Average bank asset quality has deteriorated, sharply, due to the global financial crisis that began at the end of 2008. The rapid increase in non-performing loans (hereafter, NPLs ) has increased banks’ vulnerability to further shocks and, at the same time, has limited their lending operations with major consequences for economic activity. The deterioration of the ratio of NPLs to total bank loans can be attributed to macroeconomic and bank-specific factors [see, e.g. Berger and DeYoung (1997); Louzis et al. (2012)]. Empirical evidence suggests that NPLs exhibit anti-cyclical behaviour. A deterioration in the macroeconomic conditions, with a fall in GDP and high unemployment rates, have negative effects on NPLs , as it reduces the ability of borrowers to service their debt . Among the bank-specific factors that have been found in the literature to affect NPLs are size, cost efficiency and management performance, credit conditions, market power and banks’ risk profile.

Based on the aggregate data from the Greek banking system, in this study, we focus on the factors that affect NPLs during recessions. Answering this question has important implications for banking policies trying to mitigate the effects of recession on NPLs . The Greek economy constitutes an interesting case to study the factors determining NPLs , given the pervasive recessionary conditions that have characterized the economy since 2008. In 2009, the economy entered into an economic recession phase leading to a fall in GDP of around 3% in the year 2009 and an increase in the NPL ratio by 3.5% points. In 2010, financial markets start to lose faith in Greece ’s ability to service its public debt and, after some months of negotiations between the country and the EU leaders, Greece received its first bailout from the European union and the IMF to ensure debt servicing and prevent a default. Greece committed to adopt a sharp fiscal consolidation which led to further recessionary conditions of the economy and rapidly raised NPLs . The undervaluation of the assets in the banking sector along with a loss of deposits and a high ratio of NPLs to total bank loans caused liquidity problems to the Greek banks. Therefore, the need for substantial recapitalizations was inevitable. The increase of NPLs also opened a vicious cycle between them and unemployment (or other macroeconomic variables reflecting recessionary conditions).

The data used in our analysis consists of three different categories of loan portfolios: mortgages, business and consumer loans. The relationship between NPLs of these three categories of loans and their determinants (bank-specific or macroeconomic variables ) was estimated based on the seemingly unrelated regressions (SUR) framework. Using SUR estimation method, we properly address endogeneity by allowing for cross-correlation across the error terms of the equations of the system of NPLs and possible sources of heterogeneity in the slope coefficients of the estimated regressions. Also, estimation and inference can be drawn based on the time-dimension of our data, which is reasonable and much larger than its cross-section one. One innovation of our econometric analysis is that the SUR framework is extended to allow for a common break in the relationship between NPLs and their determinants. The existence of such a break may capture the influence of exogenous events (i.e. deterioration of the economic conditions, sovereign debt crisis , political events, etc.) on the relationship between NPLs and their determinants, and if this applies to the bank-specific or the macroeconomic conditions.

The results of the paper lead to a number of interesting conclusions. They show that responsible for the sharp rise in NPLS of the Greek banking system, that began with the aftermath of the global economic and financial crisis in the year 2008, is the severe deterioration of the recession conditions of the economy and the political uncertainty occurred in the first quarter of the year 2012, i.e. 2012:Q1. These conditions changed, structurally, the relationship between NPLs and their determinants after that period. In particular, we find that unemployment and inflation determine the NPLs of the Greek banking system, over the whole sample, but their effects become stronger after 2012:Q1, due to the above conditions. From the bank-specific variables examined, we find that only changes in the return on assets can explain NPLs changes after 2012:Q1. Bank-specific variables, like changes in equity and the loans-to-deposits, are found to determine, significantly, the NPLs of the Greek banking system only during the period before the year 2012. Summing up, our results support the view that the abrupt shifts to the NPLs of the Greek banking system can be mainly attributed to macroeconomic deterioration and political uncertainty.

The paper is organized as follows. Section 13.2 presents the model that we will employ to estimate the relationship between NPLs and their determinants, and it discusses hypotheses of interest that can be tested regarding the bank-specific variables employed in our analysis. Section 13.3 describes the data and econometric analysis. Section 13.4 concludes.

2 The Model

Our empirical analysis is based on the following reduced form model for non-performing loans (denoted as ${\text{NPL}}_{it}$):

$$ \begin{aligned} \Delta {\text{NPL}}_{it} & = (c_{i} + b_{1} \Delta {\text{ROA}}_{t - 1} + b_{2} \% {\text{EQTY}}_{t - 1} + b_{3} \Delta {\text{LTD}}_{t - 1} + \gamma_{1} \Delta {\text{UNPL}}_{t - 1} + \gamma_{2} {\text{INFL}}_{t - 1} ) * {\text{DUM}}_{t - 1} \\ & \quad + \;(c_{i}^{ * } + b_{1}^{ * } \Delta {\text{ROA}}_{t - 1} + b_{2}^{ * } \% {\text{EQTY}}_{t - 1} + b_{3}^{ * } \Delta {\text{LTD}}_{t - 1} + \gamma_{1}^{ * } \Delta {\text{UNPL}}_{t - 1} + \gamma_{2}^{ * } {\text{INFL}}_{t - 1} ) * {\text{DUM}}_{t - 1}^{ * } , \\ & \quad + \,\rho \Delta {\text{NPL}}_{It - 1} + u_{it} \\ \end{aligned} $$

(1)

where $\Delta$ denotes first-difference, % denotes percentage change of a variable, i = 1, 2 and 3 denote the three aggregate categories of loans (i.e. business, mortgages and consumer, respectively), $t = 1,\,2, \ldots ,T$ denotes the time series observations of our sample, and ${\text{DUM}}_{t - 1}$ is a dummy variable which takes the value of 1 when $t - 1\; \le \;T_{ 0}$, where a structural change in model (1) occurs, and unity otherwise. ${\text{DUM}}_{t - 1}^{ * }$ is the complementary variable to ${\text{DUM}}_{t - 1}$, which takes the value of 1 when $t - 1\; > \;T_{ 0}$, and zero otherwise. The definitions of the bank-specific and macroeconomic variables included in the RHS of (1) are as follows.

Bank-specific:

$\Delta {\text{ROA}}_{t}$ is the first-difference of ROA, defined as earnings before interest and taxes divided by total assets. ROA is a measure of bank profitability. We use this variable as a proxy for quality of management to investigate the bad management hypothesis. In particular, a less profitable bank is more likely to exhibit poor performance in credit scoring, appraisal of pledged collaterals and monitoring borrowers which in turn leads to higher ${\text{NPL}}_{it}$ ratios. Therefore, we expect a negative effect of profitability on NPLs ; see, for example, Berger and DeYoung (1997), Podpiera and Weil (2008), Louzis et al. (2012).

$\% {\text{EQTY}}$ is the percentage change (%) in equity (denoted ${\text{EQTY}}$). This variable can capture the effects of bank capitalization on NPLs . According to the moral hazard hypothesis, low capitalization of banks increases NPLs , as bank managers tend to increase the riskiness of the bank’s loan portfolio when the bank is weakly capitalized and, as a result, NPLs will increase; see, for example, Berger and DeYoung (1997) and Salas and Saurina (2002). We thus expect a negative relationship between $\% {\text{EQTY}}_{it - 1}$ and $\Delta {\text{NPL}}_{it}$. Apart from the empirical literature, the moral hazard problem in the banking sector has received increasing attention in recent theoretical DSGE models; see, for example, Gertler, Kiyotaki and Queralto (2012) and Borio (2014). Note that we do not employ the ratio of Equity-to-Assets (${\text{ETA}}$) in our analysis to capture the effects of capitalization on NPLs , due to the sharp devaluations of the bank assets occurred during our sample.

$\Delta {\text{LTD}}$ is the first-difference of the loan-to-deposit ratio, which is considered as a proxy for liquidity risk. One would expect that an increase in $\Delta {\text{LTD}}$ will increase NPLs , as it increases the banks’ probabilities of default; see, for example, Louzis et al. (2012), Makri et al. (2014) and Anastasiou et al. (2016).

Macroeconomic:

$\Delta {\text{UNPL}}_{t - 1}$ is the change in the unemployment rate of the economy. This variable captures the business and macroeconomic conditions in the economy, at any point of time. Instead of this variable, we could have used the real GDP growth rate. As in Monokroussos and Thomakos (2016), we find that choosing one of these two macro variables is sufficient to capture the macroeconomic conditions in the economy. Changes in unemployment may be thought as a better indicator of how deep and persistent the recession in an economy is. As expected a priori, an increase in $\Delta {\text{UNPL}}_{t - 1}$ leads to an increase in NPLs , for all categories of loans. The positive effect of the unemployment has also been documented in Quagliarello (2007), Louzis et al. (2012) Anastasiou et al. (2016) and Monokroussos et al. (2016).

${\text{INFL}}_{t - 1}$ is the quarter inflation rate. The effect of inflation on NPLs should be positive, since an increase in inflation leads to a fall in the real income of borrowers. This is in line with prior evidence; see, among others, Beck et al. (2013), Klein (2013).

In addition to the above variables, note that in the RHS of the model, we have also included variable $\Delta {\text{NPL}}_{it - 1}$ to capture the own dynamic (trend) effects of NPLs on $\Delta {\text{NPL}}_{it}$, over time.

Model (1) can be employed to test a number of hypotheses about NPLs . It can test for a regime change in the relationship between NPLs and their determinants associated with a structural change in the financial, banking, and economic conditions of the economy, after break point $T_{0}$. These changes could be associated with exogenous events, which can be identified by the data through model (1). Given the existence of such a change, the model can reveal if the effects of bank-specific or macroeconomic variables on NPLs are asymmetric across the different regimes identified by the data. Although one may argue that bank-specific variables, like $\Delta {\text{ROA}}$ and changes in equity or credit, constitute valid explanatory variables of NPLs , these effects may considerably change across the different economic conditions after break point $T_{0}$. Similar arguments can be applied to the macroeconomic variables of the model.

In our analysis, $T_{0}$ will be treated as an unknown quantity and it will be estimated, endogenously, from the data. This can shed light on the particular conditions of the economy (or the banking sector) triggered a structural change in the relationship between NPLs and their determinants, after this point. To identify $T_{0}$, we rely on a search procedure [see, e.g., Andrews (1993); Dendramis et al. (2014; 2017)] solving the following optimization problem:

$$ T_{0} = { \arg }\,\mathop { \sup }\limits_{{T_{0} \in Q}} \,{ \log }\,{\text{L(}}\theta |T_{0} ) , $$

where $Q$ is the set of possible structural break points of the sample such that $Q \subseteq \left\{ {1,\;2, \ldots ,T} \right\}$, and $L (\theta |T_{0} )$ is the likelihood function of model (1) conditional on $T_{0}$, where $\theta$ denotes the vector of parameters. In words, the above procedure will select the break point $T_{0}$ which maximizes the log-likelihood function of the model, over all possible break points in the sample.

Before proceeding to estimation of the model, a number of final remarks are necessary in order to justify its econometric specification. First, both dependent and independent variables of the model are expressed in first differences (or percentage rates) to become stationary series. This is done in order estimation procedure and inference to rely on standard asymptotic results, holding over the time ($T$)-dimension of our data. Second, a number of bank-specific or macroeconomic variables , like the size of banks and loan interest rates, are not present in analysis. These variables were found to be insignificant, for our sample, either when allowing for a common break in the model or not. Third, the lag specification of the model is chosen based on the Akaike information criterion. The inclusion of lagged values of the regressors in the model also helps to avoid inference and estimations problems that could arise from the contemporaneous correlation between the explanatory variables and the error terms of the model.

3 Empirical Analysis

In this section, we estimate model (1) and we discuss the estimation results. In our analysis, we also compare the estimates of the model to those of a version of it which does not allow for a structural break. The estimation of both these models is carried out using maximum likelihood (which is asymptotically equivalent to three stage least squares based on the SUR framework of the model, for i = 1, 2 and 3 equations (categories of loans). This estimation method allows for the disturbance terms $u_{it}$ to be cross-sectionally correlated, across $i$, as is assumed in SUR equations. To formally test, if there is a structural break in the model, we will carry out a likelihood ratio test (denoted as LR-stat), with the null hypothesis:

$$ H_{0} :\;c_{i} = c_{i}^{ * } ,\;b_{1} = b_{1}^{ * } ,\;b_{2} = b_{2}^{ * } ,\;b_{3} = b_{3}^{ * } ,\;\gamma_{1} = \gamma_{1}^{ * } ,\;\gamma_{2} = \gamma_{2}^{ * } $$

against its alternative

$$ H_{a}:\;c_{i} \ne c_{i}^{ * } ,\;{\text{or}}\;b_{1} \ne b_{1}^{ * } ,\;{\text{or}}\;b_{2} \ne b_{2}^{ * } ,\;{\text{or}}\;b_{3} \ne b_{3}^{ * } ,\;{\text{or}}\;\gamma_{1} \ne \gamma_{1}^{ * } ,\;{\text{or}}\;\gamma_{2} \ne \gamma_{2}^{ * } $$

Testing the above null hypothesis is a crucial step to examining if there is a break in model (1) and, hence, the model constitutes a consistent specification with the data. The test statistic LR-stat is defined as $LR{\text{-stat}} = 2({ \log \,L({\theta|T_{0}}) - \log \,L({\theta_{0}})})$, where $L (\theta_{0} )$ is the likelihood function of the model under the null hypothesis $H_{0}$ (i.e. without a break; $\theta_{0}$ is the vector of parameters of this version of the model, without a break).¹ Since $T_{0}$ (and, hence, the slope coefficients of the model) is not identified under the null hypothesis, the significance levels (probability values) of LR-stat will be obtained based on the bootstrap statistical technique. The steps of this procedure are described below.

First, we estimate model (1) without a structural break and obtain estimates of its vector of slope coefficients $\theta_{0}$ and its residuals, denoted as $\hat{u}_{it}$. Based on these estimates and the values of our explanatory variables, next we generate bootstrap values of $\Delta {\text{NPL}}_{it}$ by replacement from the residuals $\hat{u}_{it}$. We generate $B$ bootstrap samples of size $3 \times T$. For each bootstrap sample, we estimate the model with and without a break at $T_{0}$ and calculate statistic LR-stat, defined above. The above procedure is repeated $B = 1000$ times. Based on these repetitions, we then compute the 5% (or 1%) quantile value of the empirical distribution of LR-stat, which constitutes its 5% (or 1%) critical value. The null hypothesis is rejected for values of LR-stat bigger than the above 5% (or 1%) level.

3.1 The Data

Our data set consists of quarterly observations of the macroeconomic and bank-specific variables of the model covering the period from 2005:Q1 to 2015:Q4, implying $T = 44$ observations. They are obtained from the Bank of Greece . Regarding NPLs data, these consist of three different type of loans: business, mortgage and consumer and they also include restructured loans. Pagratis et al. (2017) provide a more detailed analysis on the new loan restructuring framework. The inclusion of restructured loans is important. It measures more accurately the size of NPLs . In particular, the NPL ratio excluding the restructured loans at the last quarter of 2015 was 35.6% whereas the NPL ratio including the restructured loans was 43.5% for the same period. The sample period of the study captures different phases of the business cycle in the Greek economy. It refers to the pre-sovereign debt crisis period, i.e. 2005–2010 and to its aftermath, i.e. 2010–2015. Thus, it can provide useful insights into the determinants of the NPLs before, or after, the crisis.

Figures 13.1, 13.2, 13.3 and 13.4 present graphs of the dependent and explanatory variables of model (1). In particular, Figs. 13.1 and 13.2 present graphs of the three different ${\text{NPL}}$ series, ${\text{NPL}}_{it}$, and their first differences $\Delta {\text{NPL}}_{it}$, used in the estimation of the model, respectively. Figure 13.3 presents the bank specific variables $\Delta {\text{ROA}}_{t}$, $\% {\text{EQTY}}_{t}$ and $\Delta {\text{LTD}}_{t}$, while Fig. 13.4 the macroeconomic variables ${\text{UNPL}}_{t}$ and ${\text{INFL}}_{t}$, in levels. In Table 13.1, we present correlation coefficients across the above variables, as defined in the model, i.e. the independent variables are lagged one period. A number of comments can be drawn from an inspection of the above figures and table. First, the ratio of the non-performing to total loans rocketed to its highest level in 2015:Q4 from its low in 2005:Q1. From 2010 to 2015, there was a 45% increase in the NPLs on consumer loans. The NPLs ratio of business and mortgage loans increased by 33 and 31.8%, respectively. Figure 13.2 shows that the biggest quarter on quarter increase in NPLs for consumer and mortgage loans occurred from 2011:Q4 to 2012:Q1. For NPLs on business loans, the highest increase in this ratio was from 2012:Q4 to 2013:Q1.

Fig. 13.1

Source Bank of Greece

Evolution of NPLs by type of loans.

Fig. 13.2

Source Bank of Greece

Changes in NPLs by type of loans.

Fig. 13.3

Source Bank of Greece

Change in ROA and LTD, and equity growth rate.

Fig. 13.4

Source Bank of Greece

Unemployment rate and CPI seasonally adjusted.

Table 13.1 Correlation coefficients

An inspection of the unemployment rate in Fig. 13.4, indicates that the dramatic increase in NPLs within the above period can be attributed to the need to eliminate the unsustainable fiscal and current account imbalances in the Greek economy appeared in the year 2009. The elimination of fiscal and current account deficits came at the expense of growth and unemployment . Greece lost more than a quarter of its GDP within 2009–2015 period. Figure 13.4 indicates that the unemployment rate has been increasing since year 2008, with a sharp increase of this rate occurred, immediately, after the implementation of the first fiscal stabilization program in the year 2010. This trend in unemployment was stabilized in the year 2013 and it was reverted after the year 2014, where the real economy exhibited a slight positive growth rate. Note that the high levels of unemployment after the year 2013, were associated with deflation of the economy. Greece lost more than a quarter of its GDP within 2009–2015 period.

Turning our discussion on the bank-specific variables (see Fig. 13.3), we observe that the sharpest drop in profitability in the banking sector occurred in the second quarter of 2012. This was the outcome of a heavy loss in the banking system incurred, mainly, by the implementation PSI (Private Sector Involvement) program of debt restructuring . In particular, in the period between January and September of the year 2012, the Greek banking groups listed in the Athens stock exchange recorded after tax losses of 5.1 billion euros, which, on the one hand, reflect additional write-downs on their Greek government bonds as a result of the PSI, and, on the other hand, impairment charges on loans to the private sector. For more details, please see the Annual Report of Bank of Greece (2012).The change in the loan to deposit ratio (LTD) is highly volatile during our sample period. The ultimate increase in LTD ratio occurred in the second quarter of 2012 can be attributed to the massive bank deposits withdrawal, which in turn can be due to the political uncertainty (double elections) and the fears of exit of Greece from the Eurozone (known as GREXIT ). From 2012:Q3 to 2014:Q4, there was a drop in the LTD ratio, which can be attributed to reduced new lending. At the last quarter of 2014, the LTD ratio raised again owing to deposit outflows triggered by the hightened political risk, the failure of the Parliament to elect a new President of the Republic and the need, thus, for elections in January 2015. Note that, due to the fears of GREXIT , the change in the LTD ratio remained positive until the imposition of capital controls at the end of June of the year 2015. Finally, looking at the change of the equity growth, Fig. 13.3 indicates that, from 2012:Q1 to 2013:Q3, there was an impairment in the capital base of the Greek banking system mainly due to the restructuring of public debt occurred in March 2012, due to the PSI program, and the continuous deposit outflows due to the high economic and political uncertainty of Greece since the start of the sovereign debt crisis .

Finally, the correlation coefficients between the dependent and independent variables of the model indicate that there is a positive and very high correlation among the three different categories of NPLs ratio changes. This is not surprising, given that ${\text{NPL}}_{it}$ or $\Delta {\text{NPL}}_{it}$ seem to move very closely, over the whole sample (see, Figs. 13.1 and 13.2). As expected, we find a negative correlation between $\Delta {\text{NPL}}_{it}$ and $\Delta {\text{ROA}}_{t - 1}$, and $\Delta {\text{NPL}}_{it}$ and $\% {\text{EQYTY}}_{t - 1}$, for all $i$, but their degree is not big enough. The only explanatory variable which exhibits the highest degree of correlation with $\Delta {\text{NPL}}_{it}$, for all $i$, is the change in unemployment rate. As it was expected this is positively associated with $\Delta {\text{NPL}}_{it}$. Another interesting finding of the table is that there is a low degree of correlation between the bank-specific and macroeconomic sets of variables used in the estimation of the model. Thus, these two groups of variables can be taken to reflect different sources of information. This also holds within the variables of each of these groups. It may attributed to the fact that the variables of both of these groups are appropriately transformed (e.g. differenced) to remove any common trend driving them.

3.2 Estimates

Maximum likelihood (ML) estimates of the model (1) and its alternative versions, without a break and/or the macroeconomic variables , are reported in Tables 13.2, 13.3 and 13.4 respectively. Table 13.2 presents results for the model without a break based on single equation ML estimates, for each category of loans (i.e., business, mortgages and consumer). These estimates can reveal if there is high degree of heterogeneity in the slope coefficients estimates of the model, across $i$. This table also reports the adjusted coefficient of determination $\bar{R}^{2}$ and the maximum likelihood value of the model (denoted ${\text{loglik}}$), at its optimal estimates. These can be used for model comparison and to show how well the model fits into the data.

Table 13.2 Single equation estimates of the model without a break

Table 13.3 System (SUR) estimates of the model without a break

Table 13.4 System (SUR) estimates of the model with a break

Tables 13.3 and 13.4 present ML estimates of the model without and with the break, respectively. This is done based on the SUR framework, assuming homogeneity in the slope coefficients of the model, across the different categories of loans $i$. This assumption can improve upon the efficiency of the estimates of the model, given the small number of degrees of freedom available, for all $i$. It can be justified, empirically, by the single equation estimates of the model without a break, reported in Table 13.2, which indicate that there is not a high degree of heterogeneity in the slope coefficient estimates, across $i$. Note that, where there is some degree of heterogeneity, the estimates of the slope coefficients of the model tend to be insignificant, at the 5% level. To see if there is evidence of cross-correlation of error terms $u_{it}$, across $i$, both Tables 13.3 and 13.4 present estimates where $u_{it}$ are assumed to be correlated across $i$. The correlation matrix across $u_{it}$ is denoted as $\varSigma$.

The values of $\bar{R}^{2}$, reported in Table 13.2, indicate that the full specification of the model, with the set macroeconomic variables , fits better into the data, compared to that without. The relationship between $\Delta {\text{NPL}}_{it}$ and $\Delta {\text{UNPL}}_{t - 1}$ is positive as was expected by the theory. This is true for all different sets of estimates reported in Tables 13.2, 13.3 and 13.4. $\Delta {\text{NPL}}_{it}$ is also positively related to ${\text{INFL}}_{t - 1}$, but this relationship is less strong, compared to that between $\Delta {\text{NPL}}_{it}$ and $\Delta {\text{UNPL}}_{t - 1}$. For the SUR-based estimates, reported in Table 13.3, the slope coefficient of $\Delta {\text{UNPL}}_{t - 1}$ becomes significant at 10% level. The positive relationship between $\Delta {\text{NPL}}_{it}$ and ${\text{INFL}}_{t - 1}$ can be attributed to the fact that an increase in inflation reduces the real income of borrowers. Regarding the relationship between NPLs and the bank-specific variables, the single equation results of Table 13.2 indicate that, although the sign of the slope coefficients of these variables is consistent with the theory, they are not always significant, at 5 or 10% level, across the three equations. Note that the estimates of the slope coefficient of $\Delta {\text{LTD}}_{t - 1}$ are not found to be significant, for all $i$, at the 10, or 5%, level. This is true even for the SUR based estimates of the model, reported in Table 13.3. The SUR-based estimates of the model clearly indicate that the relationship of $\Delta {\text{ROA}}_{t - 1}$ and $\% {\text{EQTY}}_{t - 1}$ with $\Delta {\text{NPL}}_{it}$ is negative and significant, as predicted by the bad management and moral hazard hypotheses, respectively.

Turning the discussion on the estimates of the model with a break, the results of Table 13.4 leads to a number of very interesting conclusions. First, they provide clear cut evidence that there is a structural change (break) in the relationship between $\Delta {\text{NPL}}_{it}$ and its fundamentals, for the Greek economy. This occurs in the first quarter of the year 2012 (i.e., 2012:Q1). Note that, for the specification of the model without the macroeconomic variables , it occurs two quarters later (i.e., at 2012:Q3). This can be obviously attributed to omitting the unemployment rate variable from the model. The values of ${\text{loglik}}$ and statistic LR-stat, reported in the table, indicate that the full specification model (1), with the bank-specific and macroeconomic variables , as well the break point considered, is more consistent with the data, compared to its version without a break and/or the macroeconomic variables as explanatory variables. The p-value of statistic LR-stat, reported in the table, clearly rejects the null hypothesis $H_{0}$ that there is no structural change in the slope coefficients of the model against its alternative $H_{a}$, which assumes that there exists. We have found the better fit of the model with the break into the data, compared to its version with no break, can be also confirmed by the values coefficient of determination $\bar{R}^{2}$ of all SUR of the model estimated. These results are not reported in the table for reasons of space.

The existence of a structural change in the NPLs and its determinants, at point 2012:Q1, may be associated to the deepening of the recession, the political uncertainty and instability, and the strong fears for GREXIT in this year, as mentioned before. As the results of Table 13.4 indicate, the effects of $\Delta {\text{UNPL}}_{t - 1}$ on $\Delta {\text{NPL}}_{it}$ become stronger and more significant in the subsample after the break point 2012:Q1, rather than that before. The same is true for inflation rate ${\text{INFL}}_{t - 1}$. Figure 13.4 shows that inflation was rising in the year 2012, despite the severe recession of the Greek economy in this year. This had negative effects on the real income of borrowers and, hence, on NPLs , for all loan categories considered. The positive effect of the unemployment rate on NPLs is consistent with prior empirical evidence [see Louzis et al. (2012); Monokroussos et al. (2016)].

The results of Table 13.4 also indicate that, apart from the macroeconomic variables , there is also a structural change in the relationship between NPLs and the bank-specific variables of the model, after break point 2012:Q3. The change in NPLs , $\Delta {\text{NPL}}_{it}$, becomes negatively and significantly related to $\Delta {\text{ROA}}_{t - 1}$ only after this break point . This is the only bank-specific variable which can explain future NPLs changes after the break point . Its effects on $\Delta {\text{NPL}}_{it}$ are consistent with the bad management hypothesis. The change in LTD ratio ($\Delta {\text{LTD}}_{t - 1}$) and the percentage change in equity ($\% {\text{EQTY}}_{t - 1}$) are found to have no and little (less significant) effect on $\Delta {\text{NPL}}_{it}$, respectively, after point 2012:Q1. Comparing the results of Table 13.4 to those of 13.3 (which do not consider a break point ), one can see that the significant effects of $\Delta {\text{LTD}}_{t - 1}$ and $\% {\text{EQTY}}_{t - 1}$ on $\Delta {\text{NPL}}_{it}$ are present only in the period before point 2012:Q1, where the economy was not yet suffering from a severe recession and political uncertainty. The positive relationship between $\Delta {\text{NPL}}_{it}$ and $\Delta {\text{LTD}}_{t - 1}$ found for the period before this point is consistent with the liquidity risk hypothesis, while the negative relationship between $\Delta {\text{NPL}}_{it}$ and $\% {\text{EQTY}}_{t - 1}$ is consistent with the moral hazard hypothesis.

4 Conclusion

In this paper, we investigate whether bank-specific or macroeconomic factors determine NPLs using loan portfolios data from the Greek banking sector. Our econometric analysis is based on a SUR (seeming unrelated regressions) framework which allows for cross-correlation across the error terms of the different categories of loans considered. We have extended this framework to allow for a common structural break in the relationship between NPLs and their determinants. This break can be justified by changes in institutional factors and/or exogenous events, including political uncertainty.

The results of the paper lead to a number of interesting conclusions, with banking or macroeconomic policy implications. They show that political instability and the severe deterioration of the macroeconomic conditions constitute the key factors explaining abrupt shifts of the NPLs of the Greek banking system, over the recent years. Under these conditions, we found that the key factors that can explain movements in NPLs are changes in unemployment and inflation rates. With exception of the earning to assets variable, which reflects bank management conditions, bank-specific variables like changes in equity and loan-to-deposit ratio seems that have no significant effect on NPLs under the above conditions. This is a lesson learned from the recent economic crisis Greece .