How resilient is the Italian banking system? : a VAR approach

How Resilient is The Italian Banking System?

A VAR approach

Beatrice Armanini

15

th

_July

Supervisor: Dr. Alex Clymo

E-mail: beatrice.armanini@gmail.com

Student number: 11084043

ABSTRACT

The aim of this study is to analyse the resilience of the Italian banking system through a stress-testing exercise. Stress-tests have been broadly used by financial authorities and regulators as a tool to monitor credit and market risk. The more recent field of macroeconomic stress-testing incorporates macroeconomic variables when studying the vulnerability of the financial system as a whole. In this thesis, I perform a macroeconomic stress-test on the Italian banking sector and study the interactions between the macro-economy and the banking sector through a vector autoregressive (VAR) model. To model the banking sector, I use seven financial soundness indicators using aggregated bank balance sheet data. The findings show that shocks to GDP have a strong worsening impact on banking variables involving bad debts. Shock to inflation and interest rate do not lead to a significant response. Overall, most feedback effects I find flow from the real economy to the financial soundness indicator.

C

ONTENTS 1.INTRODUCTION ... 1 2.LITERATURE ... 4 3.METHODOLOGY ... 9 3.1 MODEL AND METHODOLOGY ... 9 3.1.1 Vector autoregressive model ... 9 3.1.2 Cholesky decomposition of the variance-covariance ... 11 3.2 MODEL SPECIFICATION ... 12 3.2.1 Macroeconomic sector ... 12 3.2.2 Banking sector ... 14 3.2.3 Financial indicators ... 15 4. DATA SAMPLE ... 17 5. RESULTS ... 20 5.1 IMPULSE RESPONSE FUNCTIONS ... 20 5.1.1 Bad Debts to Capital Ratio ... 20 5.1.2 Bad Debts to Gross Loans (with and without cassa deposit and prestiti) ... 21 5.1.3 Capital to assets ... 23 5.1.4 Liquid Assets to Total Assets ... 24 5.1.5 Deposits to Total Loans ... 24 5.1.6 Default Rate ... 25 5.2 ROBUSTNESS CHECKS ... 26

5.2.1 Different Ordering of the Variables ... 26 5.2.2 VAR model with FSIs in Levels ... 27 5.2.3 VAR model with One Lag ... 27 5.2.4 VAR model with Four Lags ... 28 6.CONCLUSION ... 29 REFERENCES ... 31 APPENDIX I – DICKY-FULLER TEST RESULTS ... 34 APPENDIX II – GRAPHS ... 37 APPENDIX III – IMPULSE RESPONSE FUNCTIONS OF ROBUST CHECK ... 41

1

1.INTRODUCTION

The Global Financial Crisis and subsequent Sovereign Debt crisis have underlined the far-reaching role of the financial sector in the economies of the Eurozone’s single currency area. The experience of recent years has underlined the need for a better understanding and assessment of the riskiness of the financial system. Stress-testing financial systems requires a stronger emphasis on the link between the real economy and the financial system, as spill-overs between the two have shown to be crucial to financial stability (Goméz-Puig et al., 2016). In the past, stress-testing has been mainly employed by financial authorities as an instrument to measure financial risks and analyse the soundness of the banking system and its individual banks. Stress-testing at the individual institution level has been used since the beginning of the 1990s and bank regulators around the world have since generally imposed the use of this tool to monitor credit and market risk (Sorge, 2004). A version of this tool which has moved to the forefront of financial regulation is macroeconomic stress testing, which investigates the weaknesses of the financial system as a whole by incorporating macroeconomic variables. This type of stress test moves beyond the analysis of individual banks and broadens its scope to include non-financial variables. Since crisis in the financial system cannot solely be explained as the result of individual bank failures, macroeconomic stress testing has become an essential part of the policymaker’s toolkit to assess financial stability issues (Kapinos, 2015). The most extensive appliance of this kind of macroeconomic stress testing to date was done by the IMF and the World Bank as part of its Financial Sector Assessment Programs in 1999 (FSAPs). Macroeconomic stress-testing is a relatively new academic field but has become an important instrument for financial intermediaries, national authorities and international financial institutions. This instrument represents an indispensable tool for macro-prudential policy. Furthermore, it is relevant from a public policy perspective as the exercise has to be undertaken according to a “coherent macroeconomic and financial set up and to incorporate a macro-prudential perspective” (Vítor Constâncio, 2015 ).

While the global crisis of 2008-2009 heavily and immediately hit the credit institutions of many developed economies, Italian banks emerged relatively unscathed and, unlike other states,

2

the Italian government did not intervene in its financial sector. Italy is the fourth largest economy of Europe but at the moment also one of its weakest: public debt stands at 135% of GDP and the unemployment rate has reached 11.5%; with a historical peak of 37.9% of unemployment amongst the young (Italian statistic institution ISTAT). Weak economic recovery, combined with little government intervention following the crisis has led to a curb in new lending and an accumulation of bad loans which has already pushed small and medium lenders, such as Banca delle Marche and Banca d’Etruria into insolvency. The government has recently aimed to address this issue, but is restricted by the tighter regulation and supervision imposed by the EU (Reuters, 2016). The main difficulty in this context is the bail-in procedure that according to European law will be enforced following a bank failure. According to these rules investors will carry some of the burden in case the bank collapses. In Italy, a majority of households hold bonds from their own bank, which means that in case of a bank failure the households will be “bailed-in” and see their bond savings disappear.

As such, worries about the Italian financial sector have grown stronger as the sector has increased in vulnerability. Bank balance sheets have deteriorated sharply due to the enormous amount of non-performing loans (NPL) present on the balance sheets of Italian banks (reaching €350bn), the highest level within the Eurozone (Sanderson, 2016). Shares in Unicredit Banca, Italy’s biggest bank and the only one of globally systemic proportions, have lost half their value since April 2016. The third bank of Italy, Monte dei Paschi di Siena, is also experiencing financial troubles. Shares in the bank have fallen sharply on the 4th and 5th of July after news emerged that the European Central Bank issued a call to reduce non-performing loans on the bank’s balance sheet from €46.9 billion to €32.6 billion within two years (The Economist, 2016). The FTSE Italian All-Share Bank index has fallen losing overall 56% of its value in the last year (Sanderson, 2016). Both the national and European authorities are now concerned that the collapse of the Italian banking system could trigger a new financial crisis (The Guardian, 2016). In light of these developments, this study reproduces an application of macroeconomic stress-testing to the aggregate Italian banking system in order to highlight its weaknesses and strengths. A broad part of the literature on stress-testing analyses the resilience of singular institutions without considering the system as a whole. Moreover, these stress-test exercises

3

are applied to a limited number of banking variables and do not include macroeconomic variables. In this study I will estimate seven VAR models, each model involving a different financial system indicator (FSI) as a proxy for the aggregate banking system. The financial system indicators are computed using aggregated banking balance sheet data according to guidelines provided by the IMF. I construct two capital based and four assets based indicators, respectively: capital on assets ratio, nonperforming loans to capital ratio, return on equity ratio, liquid assets on total assets ratio, deposit on total loans ratio, nonperforming loans on total gross loans ratio and the loan default rate. The subsequent identification of impulse response functions is done by using a Cholesky decomposition. The finding of this study reveals that shocks to GDP have have a strong worsening effect on banking variables involving bad debts. This shows that weak economic recovery in Italy can rightfully be identified as a cause for the large amount of bad debt in its banking system. Furthermore, I find that the capital to assets ratio has a strong link with the real economy as shocks to this ratio positively influences GDP and inflation but negatively affects the interest rate.

Stress-testing is considered an important crisis management tool (Kapinos et al., 2015). Therefore, the results are relevant for policy makers and country authorities as it provides a clear and strong signal of the weaknesses of the Italian banking system.

The rest of the paper is organised as follows. The next section describes the literature review. Section 3 delineates the empirical methodology used to perform the stress-testing. It provides an extensive explanation of VAR models and the Cholesky decomposition of the variance-covariance matrix of the residuals. Section 4 describes the data sample and plots the trends of the variables. Section 5, presents and interprets the empirical findings as well as providing

4

2.LITERATURE

The literature about stress-testing presents a high level of heterogeneity. The approaches used differ in empirical methodology, the type of risk considered, the choice of relevant institutions and the type of scenario studied. Empirical methodologies range from time series analysis and panel regressions to structural macro models, as for example Dovern et al. (2009). The various types of risk that stress-test focus on include credit, market and liquidity risk. The institutions and scenarios incorporated in stress-tests vary greatly; from small scale, focused event studies to complete financial market stability studies.

Stress-tests focused on credit and market risk are common used and strictly enforced instruments used by regulators. A more recently adopted type of stress-test incorporates variables from the macro economy to study the vulnerability of the financial system. This type of stress testing, also known as macro-economic stress-testing, is the type of test I will follow in this thesis. The broadest application of macro-economic stress-testing to date has been carried out by the IMF and World Bank (1999) in the Financial Sector Assessment Program (FSAP). This test involved developing, emerging and advanced economies, for a total of three-quarters of the World Bank’s member countries. The objective was to identify for different financial systems the weaknesses and vulnerabilities that could have serious macroeconomic repercussions. This was done by analysing the resilience of the individual country’s financial system to macroeconomic shocks. In this study the most common approach used was a single factor sensitivity tests which look at the impact of a change in one variable.

Blaschke et al. (2001) provides an overview of the macroeconomic stress-tests computed by the IMF and World bank in the FSAP. The paper evaluates the main analytical tools and approaches used in the FSAPs. Specifically, the authors distinguish between stress-testing at the individual institution or portfolio level and the aggregate financial system level. In the first case, risk is measured by estimating the losses that could arise on a portfolio as a consequence of abnormal markets. Methodological challenges with this type of stress-testing stem from the choice of the type of risk, scenario and shock examined. The second type of stress-testing,

5

testing the aggregate financial system, is described as a measurement of the risk exposure of a group of institutions to the same shock. The aim of this type of stress-testing is to identify risk exposures and structural vulnerabilities that could harm the stability of the financial system. Blaschke et al. identify the main difficulties of this approach as the selection of firms involved in the analysis and the process of aggregation across institutions.

The paper by Blaschke et al. also provides a basic toolkit of stress-testing methods and describes how several risks have been investigated in the FSAP. Credit risk is examined by looking at how changes in the extension of provisions effect bank solvency, by analysing liquidity risk and by simulating responses of non-performing loans to macroeconomic shocks. All the different techniques have in common that they rely on a single factor sensitivity test. A disadvantage of this type of stress-testing is that it uses a single equation methodology by which it remains difficult to analyse how the macroeconomic variables interact between each other.

Cross-correlation effects between financial and macroeconomic variables, also known as financial spill-over effects, are defined as the extent to which financial shocks feedback into the real economy. These effects have been a major source of financial instability and macroeconomic uncertainty during the crisis and post-crisis years (Claessens et al., 2013). The new crisis emerging in the Italian banking sector is another example of this, as the accumulation of bad loans cannot be seen as an issue separate from macroeconomic developments.

The above highlights the crucial importance of incorporating the interaction between macroeconomic and financial variables when stress-testing. A type of model that allows for the inclusion of these correlated effects is the vector autoregressive (VAR) model. This model allows for interaction between its factors and manages to create shock-scenarios concerning shocks to more than one macroeconomic indicator. In this way it explicitly models the two-way interaction between the financial sector and macro-economy. I will proceed by presenting several stress-test studies which have used VAR models as the core of their methodology, the first study focusses on soundness indicators at singular

6

financial institution level, whereas the third and fourth analyse the soundness of the aggregate banking sector. In my analysis I will focus on the Italian banking system as a whole. First, Hoggarth et al. (2005) conduct a VAR model stress-test for the UK banking system. The paper provides an analysis of banks’ fragility by taking into consideration the write-off to loan ratios as a financial soundness indicator. Sectoral VARs for the UK are estimated that focus on corporate write-offs and household defaults using write-off data on a quarterly basis between 1993 and 2004. In order to design a small scale macroeconomic model, the authors use endogenous variables including the output gap, the nominal short-term interest rate, inflation and the real exchange rate. To obtain the impulse response functions, the authors use a Cholesky decomposition. The main result shows that an increase in the output gap leads to a decrease in the write-off ratio at both aggregate and corporate sector level. Moreover, the household write-off ratio is found to be sensitive to shocks in income gearing: the percentage of post-tax profits that is spend on debt interest payments. Overall, the paper finds that the UK banking sector is robust to shocks they apply. A study which has a strong focus on macroeconomic stress-testing is Dovern et al. (2009). This paper aims to study the resilience of the German banking system to macroeconomic shocks. The empirical work relies on annual data spanning 39 years (four business cycles) between 1968 and 2007. The financial stress indicators are modelled using bank balance sheet data at the aggregate level. A particular focus is on the write-off ratios: identified as the quotient of the sum of total write-offs and value-adjustments relative to the total amount of out-standing loans and the return on equity ratio, defined as profits after tax divided by the amount of total equity. The macro-economy is characterised by a standard monetary VAR model that includes real GDP, consumer prices, the 3-months interest rate and US GDP as exogenous variable. The impulse response functions are derived using two methods: short-run sign-restriction on the vector and applying a Cholesky decomposition approach to check the robustness of the findings.

The study’s conclusion assert that monetary policy plays a central role for the stability and resilience of the banking sectors. Indeed, the empirical results indicate that monetary policy shock have the biggest impact on the financial stress indicators. This cannot be stated for

7 aggregate demand shocks, where the results show only a barely significant reaction. In the end, no significant responses are found for aggregate supply shocks. In my study I follow the methodology employed by Dovern et al. and apply it to the Italian banking sector. An empirical study focusing explicitly on Italian banks is Filosa (2007). To stress-test the Italian banking sector, Filosa adopts a VAR model. This work, unlike the previous one, incorporates as endogenous variables: the output gap, the inflation rate, the spread between the loan and inflation rate, one measure of the amount of free capital held by banks and an indicator of the banks’ soundness. The model is enriched by adding three exogenous variables: the Euro effective exchange rate, the short rate policy rate of the ECB and a linear time trend. The banks’ risk indicators that are selected are two default rates and the interest rate margin. Moreover, the author uses quarterly data from 1990 to 2006. The empirical results show that the financial soundness indicators are barely pro-cyclical. Asset price volatility does not have a significant impact on Italian banks’ stress variables. In addition, Filosa does find a statistically significant relation between the real economy and the soundness indicator of the banking sector. Finally, they find that the banks’ soundness hardly reacts to a hypothetical tightening of the monetary condition. I expand this analysis using a standard monetary VAR approach with a different time spam, from 1997 until 2015. The data therefore also includes the last crisis. Moreover, I enrich the analysis by involving additional financial soundness indicators. The findings of Filosa diverge from the results of my study. Although both studies estimate a link between the real economy and the banking sector with a central role played by GDP, Filosa’s conclusion underlines a weak pro-cyclicality of the financial soundness indicators to shock to the real economy. In contrast, I find that shocks to GDP have a worsening effect on financial variables.

In summary, there has been an extensive amount of research done in the field of stress-testing. Stress-tests incorporating macroeconomic variables are less common but highly relevant when considering the significant cross-correlation effects that have emerged in recent years between the real and financial economy. Furthermore, previous stress-tests that have focused on the Italian banking system have not included macroeconomic variables. As such, my research contributes to the existing literature by incorporating both macroeconomic

8

9

3.METHODOLOGY

In light of the attention drawn by Italian banks in the last months I will stress-test the Italian banking system using macroeconomic stress-testing tools. I am proceeding with a replication of the methodology in Dovern at al. (2009). The study focuses on the interaction between the Italian banking sector and macro-economy. First, I will describe in general the VAR model and the Cholesky decomposition of the variance-covariance matrix that I will use in my analysis. After, I will describe my specific model set-up and the dataset I use.

3.1

M

ODEL AND METHODOLOGY

3.1.1

V

ECTOR AUTOREGRESSIVE MODEL

The core of the vector autoregressive model is the autoregressive model. The autoregressive (AR) model is a model in which the output variable is represented as a linear function of its own lagged values. Therefore, a 𝑝"#_{-order AR model (AR(p)model) defines its output as a}

linear function of a number p of its own past values (Stock and Watson, 2011). VAR models generalize the AR models by allowing for more than one evolving variable. It is generally used to capture the interdependencies among several time series and has proved to be particularly helpful for analysing, describing and forecasting the behaviour of financial and economic time series (Zivot and Wand, 2006). This is due to the fact that financial, economic, and business variables are not only correlated, but often also cross-correlated for several time periods. Indeed, by analysing the temporal interdependencies between several variables, the model’s ability to explain and forecast responses increases. The VAR model has three versions (the reduced form, recursive and structural VAR). For the aim of this paper I focus on the reduced and structural form, which are distinguished between each other for the way they allow correlations to be interpreted casually

In the reduced form, each variable is explained as a linear function of its own past values and the past values of the remaining variables. If the terms are correlated with each other than the error terms are typically correlated across time. These correlations are the object of interest. Therefore, structural VAR (SVAR) models make explicit “identifying assumptions” to

10

allow for a casual interpretation of the correlations between variables (Ibid.).

The stress-test I perform in this thesis uses a structural form VAR. The subsequent identification of the structural effects and of the impulse response functions will be done by using a Cholesky decomposition of the covariance matrix. The model I use thus expresses each variable as a linear function of its own lagged values, the lagged values of the remaining variable and exogenous variables. If the system is influenced by other variables which are determined outside the specified process it can be helpful to use additional exogenous variables. As this is likely to be the case for the Italian banking system, I include an exogenous variable which I will detail below. Each equation is estimated by an ordinary least squared procedure (OLS) where the error term contains the movements in the variables after taking past values into account. A VAR with p lags is usually denoted a VAR(p). The system is therefore defined as 𝑦_"= Α_' + Α₎𝑦_"*)+ Α₊𝑦_"*+… + Α_-𝑦_"*-+ a_" (1) Where 𝑦"= 𝑦)", … , 𝑦0" is a (n x 1) vector, Α_' is a (n x 1) vector of parameters, Α₁ = 𝑗 = 1, … , 𝑝 are (n x n) matrix of parameters and

a_" = (𝑦_)", … , 𝑦_0") ∼ 𝑊𝑁 0, Σ is assumed to be an unobservable zero mean white noise vector. Therefore, a hypothetical VAR (1) with two variables (K=2) can be written as follow, 𝑦_)" 𝑦_+" = 𝛼_)' 𝛼_+' + 𝛼₎₎𝛼₎₊ 𝛼₊₎𝛼₊₊ 𝑦_),"*) 𝑦_+,"*) + a_)" a_+" Which is equivalent to the compact form 𝑦_"= Α_' + Α₎𝑦_"*)+ a_" (2) Where a" ∼ 𝑊𝑁 0, Σ with Σ covariance matrix Σ = 𝜎₎₎ 𝜎₎₊ 𝜎₊₎ 𝜎₊₊

11 and, 𝑦_)" = 𝛼_)'+ 𝛼₎₎𝑦_),"*)+ 𝛼₎₊𝑦_+,"*)+ a_)" (3) 𝑦+" = 𝛼+'+ 𝛼+)𝑦),"*)+ 𝛼++𝑦+,"*)+ a+" (4) Where 𝑐𝑜𝑣 a)", a+" = 𝜎)+ for t=s; zero otherwise. The dependence between the two variables is specified by the coefficients of the matrix Α) = MNNMNO MONMOO and by the matrix of covariance Σ = 𝜎₎₎ 𝜎₎₊ 𝜎₊₎ 𝜎₊₊ . More specifically, the coefficients 𝛼₎₊ and 𝛼+) show the dynamic effects between 𝑦)and 𝑦+. However, the cross-equation error

variance–covariance matrix Σ contains all the information about contemporaneous correlations in a VAR (Stock and Watson, 2001). Without any restriction it is not possible to interpret the result in a correct way and overcome the issue of the correlated errors. As a consequence, I will need to apply a method that allows to distinguish between the dynamic and single effects of the matrix, in order to show how the endogenous variables react to a one-time shock of the disturbances. In technical terms, I must orthogonalize the disturbances to capture structurally interpretable impulse response functions. In order to do so, I will apply a Cholesky factorization of the structural VAR covariance matrix.

3.1.2

C

HOLESKY DECOMPOSITION OF THE VARIANCE

-

COVARIANCE

VAR models do not allow to make statements about causal relationships. The literature suggests several techniques to overcome this issue, identify the structural shocks and obtain the impulse response functions (IRFs). The IRF indicates how the endogenous variable react over time to a one-time shock to one of the disturbances.

To obtain consistent and interpretable IRFs I will use the Cholesky decomposition of the variance-covariance matrix of residuals. It is represented by a matrix which is the squared root of the covariance matrix. Given a (n x n) positive definite variance–covariance positive matrix Σ, it can be factorised into a product Σ = LLʹ where L is a lower triangular matrix with real positive diagonal elements

12 L = 𝛾₎₎ 0 0 𝛾₊₎ 𝛾₊₊ 0 𝛾_R) 𝛾_R+ 𝛾_RR Essentially the original covariance matrix is reduced to a lower triangular matrix. This methods “uncorrelates” the analysed variables and thereby explicitly identifies the interdependence between the endogenous variables through a casual ordering. The ordering of the variables depends on the type of macroeconomic model used. The terms have to be placed following an increasing order of endogeneity.

3.2

M

ODEL SPECIFICATION

Having described the VAR model and the Cholesky decomposition method, I will now proceed by detailing my model specification. First, I will describe how I model the macro-economy, followed by the banking sector.

3.2.1

M

ACROECONOMIC SECTOR

In this study I use a structural VAR model which permits the interaction between the variables involved and manages to create shock-scenarios. This type of approach overcomes the drawback of a singular equation methodology which does not allow to analyse how the variables respond to the interaction between themselves.

To model the macro-economy, I use a standard monetary VAR model. This framework is suitable to summarize the dynamics of macroeconomic data, as they provide empirical evidence on the response of variables to specific exogenous impulses. In this way, VAR models offer a method that incorporates dynamics across several time series. A standard monetary VAR model involves three endogenous elements: real GDP, consumer prices (CPI), and the three-month interest rate on deposits (i) (Christiano et al. 1999). This standard approach leaves room for further additions. I use EU GDP as exogenous variable, thereby taking into consideration the fact that Italy is the 8th _{largest economy in the world and}

13

trades mainly within the old continent. The model is further enriched by an endogenous variable that represents the situation of the aggregate Italian banking system: a financial soundness indicator (FSI). I use seven financial soundness indicators and as such obtain seven “varieties” of the model. By expanding the VAR model it is easy to run into issues regarding the degree-of-freedom, which could lead to significant of inefficient estimators. Therefore, I choose not to add more than one exogenous variable. The model as described above results in a reduced VAR with four equations and five variables (four exogenous, one endogenous): current Italian GDP as a function of the past values of Italian GDP, inflation, interest rate, the FSI and the exogenous European GDP; inflation as a function of the past values of inflation, Italian GDP, interest rate, a FSI and the exogenous Europe GDP; and similarly equations for the interest rate and FSI. In functional form, the equations are, in specific order, represented as 𝐺𝐷𝑃" = 𝛼)+ 𝛼))𝐹𝑆𝐼"*)+ 𝛼)+𝐹𝑆𝐼"*++ 𝛽)) 𝐼𝑇[\]^_N+ 𝛽)+ 𝐼𝑇[\]^_O+ 𝛾))𝐶𝑃𝐼"*)+𝛾)+𝐶𝑃𝐼"*++ 𝛿))𝑖"*)+ 𝛿)+𝑖"*++ 𝜌))𝐸𝑈_𝐺𝐷𝑃"*)+ 𝜌)+𝐸𝑈_𝐺𝐷𝑃"*++ 𝜀)" 𝐶𝐼𝑃_" = 𝛼₊+ 𝛼₊₎𝐹𝑆𝐼_"*)+ 𝛼₊₊𝐹𝑆𝐼_"*++ 𝛽₊₎ 𝐼𝑇_{[\]^_N}+ 𝛽₊₊ 𝐼𝑇_{[\]^_O}+ 𝛾₊₎𝐶𝑃𝐼_"*)+𝛾₊₊𝐶𝑃𝐼_"*++ 𝛿+)𝑖"*)+ 𝛿++𝑖"*++ 𝜌+)𝐸𝑈_𝐺𝐷𝑃"*)+ 𝜌++𝐸𝑈_𝐺𝐷𝑃"*++ 𝜀+" 𝑖" = 𝛼R+ 𝛼R)𝐹𝑆𝐼"*)+ 𝛼R+𝐹𝑆𝐼"*++ 𝛽R) 𝐼𝑇[\]^_N+ 𝛽R+ 𝐼𝑇[\]^_O+ 𝛾R)𝐶𝑃𝐼"*)+𝛾R+𝐶𝑃𝐼"*++ 𝛿R)𝑖"*)+ 𝛿R+𝑖"*++ 𝜌R)𝐸𝑈_𝐺𝐷𝑃"*)+ 𝜌R+𝐸𝑈_𝐺𝐷𝑃"*++ 𝜀R" 𝐹𝑆𝐼_" = 𝛼_h+ 𝛼_h)𝐹𝑆𝐼_"*)+ 𝛼_h+𝐹𝑆𝐼_"*++ 𝛽_h) 𝐼𝑇_{[\]^_N}+ 𝛽_h+ 𝐼𝑇_{[\]^_O}+ 𝛾_h)𝐶𝑃𝐼_"*)+𝛾_h+𝐶𝑃𝐼_"*++ 𝛿_h)𝑖_"*)+ 𝛿_h+𝑖_"*++ 𝜌_h)𝐸𝑈_𝐺𝐷𝑃_"*)+ 𝜌_h+𝐸𝑈_𝐺𝐷𝑃_"*++ 𝜀_h" Each equation is estimated by an OLS regression. In a VAR model it is important to ensure that the variables used are plausibly related to each other and a small number of coefficients is

14 used. Indeed, having to estimate a large number of coefficients increases the estimation error which can lead to a worsening of the analysis (Stock and Watson, 2011). I therefore have set the number of lagged values to two in these tests. To ensure stationary of the time series, all the variables enter the reduced form VAR as log-differences, in this way I obtain the growth rates of the variables which are stationary.

For the Cholesky decomposition of the covariance matrix, the ordering of the variables is fundamental. The terms have to be placed following an increasing order of endogeneity. I therefore order GDP first, followed by the inflation rate, the interest rate and, finally, the FSI. In this way, it is assumed that the macroeconomic shocks have a delayed effect on the real economy, but all hit the banking sector at the same time (Dovern at all., 2009).

3.2.2

B

ANKING SECTOR

After having defined the general set-up for the model and the macroeconomic variables used, I here define the financial soundness indicators which act as proxy for the aggregate Italian banking sector. I use seven different financial soundness indicators; all constructed using aggregated balance sheet data. This choice regarding the data is driven by the understanding that a shock in the macroeconomic environment, which hits a single bank, may lead to a larger crisis which can subsequently be observed through a general worsening of banks’ balance sheets (Jonas Dovern et al. 2009). Moreover, the choice of using aggregate banking data instead of single institutes is in line with my objective of investigating the resilience of the complete sector. The stress-testing exercise will be undertaken seven times, each time involving a different financial soundness indicator (FSI). Moreover, the several VAR specifications, which only differ for the choice of the FSI, will help to check the robustness of the results. The study is mainly focused on the responses of the macro-economy variables to shocks to the FSIs. Below I will outline the FSIs used in this analysis.

15

3.2.3

F

INANCIAL INDICATORS

One of the salient point in the stress-testing exercise is the choice of the financial stress indicators which indicate the strength and vulnerability of the financial system (IMF, 1992). The literature suggests different variables that can be used for this purpose, but the IMF provides a guideline to the most relevant indicators. When the topic of examination is the banking sector, as in this thesis, the resilience of the system can be measured in term of capital and assets (IMF, 2006). Below I will detail the proxies I have used for both categories.

Capital based FSIs include both capital and reserves held by banks. These measures provide a comprehensive idea of the capital resources available and the capacity to absorb losses of the sector. I have used two capital based FSIs, the first of which is the capital on assets ratio. This

ratio offers an indication of capital adequacy and a measure of financial leverage. The secondly proxy I use is the nonperforming loans to capital ratio. This ratio will control for the impact of nonperforming loans (NPLs) on capital and indicate the capacity of the banking system’s capital to react to losses related to NPLs. In case of the Italian banking sector, this ratio plays an essential role due to large amount of NPLs existing within the system. Their impact on the capital of the system is therefore highly relevant. Banks are expected to be able to caver some of the bad loans through various form of risk mitigation, but most likely not all. Asset based FSIs focus primarily on solvency and liquidity. These are both important aspects to consider in analysing the resilience of the banking system. Liquidity is the ability of a bank to satisfy its short-term cash requirement. Solvency refers to the ability to meet long-term obligations. The ability of the sector to face expected and unexpected demands for cash can be measured through the liquid assets on total assets ratio. Another measure of liquidity is the costumer deposit on total loans ratio. Indeed, when deposits are low in comparison to loans, the banks’ portfolio has a higher dependence on volatile funds to cover illiquid assets. This means that in case illiquid assets grow the risk of illiquidity arises. Due to the well-known issue of the excessive amount of NPLs present in the Italian banking sector, I include in my analysis other variables that focus on this aspect. First, I consider the nonperforming loans on total gross loans ratio. This ratio captures borrower’s solvency and shows issues regarding the latest quality in the loan portfolio. This measure has to be

16 interpreted in combination with the NPLs to capital ratio described above. If this last ratio rises, it suggests a signal of deterioration in the quality of the credit portfolio. However, this measure is only useful as far as the NPLs are recognised on time. Indeed, NPLs are usually identified only when issues emerge, which gives to this ratio the feature of being a backward-looking indicator. Finally, I include the default rate as a measure of the sector’s soundness. The default rate is general considered a good indicator of potential distress.

17

4.

DATA

SAMPLE

Here I will discuss the origins of my data and the transformations I apply to construct the financial soundness indicators as described above. The data set used for the analysis contains balance sheet data on the Italian banking system and macroeconomic time series both on a quarterly basis. It covers a time period of eighteen years, from Q1 1997 until Q4 2015. From 1997 onwards, the Italian Central Bank (Banca d’Italia) has published detailed balance sheet data for the aggregate banking sector. The data is aggregated on monthly or quarterly frequency. The FSIs are computed by using two time series to obtain a ratio. Each ratio needs data with the same frequency, therefore I transform data from monthly to quarterly, by taking the average of the three months of each quarter. Moreover, the Italian Central Bank database provides data solely for aggregate banks and aggregate banks and cassa depositi e prestiti (a joint stock bank with 80.1% of its share capital owned by the Italian Ministry of Economy and Finance, 18.4% held by various banking foundations, and the remaining 1.5% in treasury shares). Unfortunately, not all balance-sheet data are offered for both categories. Hence, to construct the ratios, this has to be taken into consideration, making sure that both nominator and denominator belong to the same order of data.

Data regarding capital, deposits, total assets, liquid assets and remaining assets (to calculate liquid assets by subtracting liquid and remaining from total assets), total gross loans, performing loans, non-performing loans (classified as bad debts), have been retrieved from the Statistics Data Centre of the Italian Central Bank. Data on the macro-economy (Italian real GDP, consumer prices, the 3-months interest rate, European real GDP) has been retrieved from Datastream. These data series are graphically represented in Fig.1 found in Appendix II. In a VAR model the components should be stationary to ensure the reliability of the results. A time series is stationary if its probability distribution does not change over time. The graphs in Fig.1 show that the series may be non-stationary and that inflation and interest rate may be correlated. I test the hypothesis of non-stationarity in the levels of the variables using the

18 Dicky-Fuller test. Following this test, I cannot reject the null hypothesis that the data series I use in my analysis have a unit root. For this reason, I log-differentiate the variables to ensure stationarity, which is confirmed by the Dicky-Fuller test (Appendix I). However, the Johansen test for co-integration between inflation and interest rate confirms one co-integrating equation which implies that non-stationarity should not alter the result of the analysis. Figure 2-8 in Appendix II shows the trend of the time series used in the stress-testing exercises, in particular it provides some properties of the singular FSIs in relation with the macro-economy. Starting from Figure 2, it shows that bad debts to capital ratio registers an upward trend starting from the Global Financial Crisis. In particular, from Figure 1 as described above, it is clear that the bad debt time series trends upwards since GDP growth slowed down after the crisis. This observation is consistent with the evidence. In fact, since the start of the crisis, NPLs (bad debts) in Italy have raised from 5% of total loans in 2007 to 17% of total loans in 2014. This rapid growth appears to be due to the prolonged and deep recession faced by the country and the stagnant GDP growth, both of which have affected the creditworthiness of Italian borrowers (Jassaud et al., 2015). Moreover, the lack of incentives to write off Italian loans and the inefficient and complex legal processes, have limited the resolution of NPLs (Ibid.). The same fact can be noticed in Figure 3 and 4 which represent the bad debts over gross loans ratio respectively for the banking system and the banking system included cassa depositi e prestiti. Figure 5 presents an upward trend for the ratio of capital to assets. This ratio, as explained above represents a measure of financial leverage. It is an indicator of capital adequacy: whether the system has enough capital to support its assets (IMF, 2006). After the crisis, the Basel committee and banking regulators raised the level of capital required to protect depositors and reinitiate the stability of the banking system (bank’s Tier 1 capital-to-assets ratio). In light of this, the upward path seen in figure 5 can be explained by an increase in capital. The liquid assets to total assets ratio is shown in Figure 6, and seems to follow the trend of GDP and inflation rate declining at the moment the crisis started in 2008. Figure 7 shows a countercyclical tendency. The deposit to total loans ratio went down till 2008 and increases again after the Global Financial Crises. This is due to the increase in deposits

19

and the sharp drop in loans granted which followed from the crises and the following stagnation of growth. Consumers in Italy kept saving, scared by the uncertainty regarding the future of the economy of the country. The Italian Central Banks claims that deposits in the banks have increased with almost 13% since the beginning of the crisis. The default rate, represented by Figure 8, shows a downward sloping trend until 2008, after which it sharply increases. Overall, inspection of the FSI series suggests correlation with the real economy, which could indicate that the financial sector is responsive to macroeconomic conditions. In the next section, I test this formally using the SVAR methodology.

20

5.

RESULTS

In this section I will present first the results of my empirical analysis, followed by robustness checks of the results. In the first section I discuss the impulse response functions obtained from the SVAR models separate for each financial soundness indicator. In the second section I detail the methods I use to check the robustness of these results. The graphs of the IRFs and the robustness checks can be found below and in the appendix.

5.1

I

MPULSE RESPONSE FUNCTIONS

The seven SVAR models with quarterly frequency, have been computed each with a different FSI. As detailed above, each models consist of macroeconomic sector variables (real Italian GDP, inflation and the three-month interest rate on deposits), one exogenous variable (real European GDP) and an indicator of the aggregate banking sector soundness (bad debts to capital, bad debts to gross loans, with and without cassa deposit and prestiti, capital to total assets, liquid assets to total assets, deposit to total loans and a default rate). The SVAR models analysed in this section involve two lags, and all the variables enter the model as log first differences. The shocks as discussed below are by definition positive shocks to the variable mentioned. The graphs show the response for twenty quarters (five years). This study mainly focusses on the effects of shocks to the FSI on the macro-economic variables.

5.1.1

B

AD

D

EBTS TO

C

APITAL

R

ATIO

The impulse response functions reported in Figure 9 show the responses of the estimated SVAR for the bad debts to capital ratio. The ratio responds to a positive shock in GDP by decreasing over the first few years. This finding is consistent with the expectation that if the economy experiences an upturn, bad debt relative to capital will decline. In the long run there is no effect from a GDP shock on the ratio. Shock in the interest rate leads to a small increase of the ratio in the medium run, the effect of which seems to persist over a period of ten quarters. Inflation shocks do not affect the ratio. Overall, the effects which are found are not statistically significant, the only strongly significant effect is the response of the bad debts

21

ratio to GDP shocks in the medium term.

5.1.2

B

AD

D

EBTS TO

G

ROSS

L

OANS

(

WITH AND WITHOUT CASSA DEPOSIT AND

PRESTITI

)

Figures 10 and 11 below represent the responses of the bad debts to gross loans ratio, the first figure shows the ratio without including cassa deposit and prestiti, whereas the second one does include this institution. The two ratios respond to shocks in a similar way, therefore I analyse them together. The ratios appear only to be influenced negatively in the short run by a shock in GDP. Shocks to the inflation and interest rate do not produce any relevant change in this financial soundness indicator. Overall the effects found are not statistically significant. Fig. 9. Impulse response function via Cholesky decomposition for Bad Debts to Capital ratio with log-difference.

22

Fig. 10. Impulse response function via Cholesky decomposition for Bad Debts to Gross Loans (with cassa crediti e depositi) ratio with log-difference.

Fig. 11. Impulse response function via Cholesky decomposition for Bad Debts to Gross Loans ratio with log-difference.

23

5.1.3

C

APITAL TO ASSETS

The responses of the capital to assets ratio are represented by Figure 12. The ratio responds positively to shock in GDP, although the effect persists just for a short period of time. However, the effect found is statistically significant. The finding that the capital to assets ratio moves in the same direction as GDP is reasonable: during downturns the ratio drops, while during upturns the ratio rises. The interest rate shock has a negative effect on the ratio but this response is barely significant. The inflation shock does lead to a significant response in the ratio. Even though it is not the main aim of this study, it is worth to notice that out of the FSIs I discuss the capital to assets ratio has the most pronounced effect on the real economic variables. In the short run, shocks to this ratio impact both GDP and the interest rate positively. The inflation rate shows a different response, where a shock to the capital to asset ratio first causes an immediate decrease in inflation, followed by an increase over the subsequent 5 quarters

Fig. 12. Impulse response function via Cholesky decomposition for Capital to Assets ratio with log-difference.

24

5.1.4

L

IQUID

A

SSETS TO

T

OTAL

A

SSETS

Figure 13 shows the impulse response functions for the liquid assets to total assets ratio. A shock to GDP leads to a statistically significant increase in the ratio in the short run. As GDP growth goes down, the ratios follows. This might be due to a decrease in the liquid assets which decreases the ability of a bank to satisfy its short-term cash requirement. The other two macroeconomic structural shocks, inflation and interest rate, do not influence in any relevant way this ratio. No significant effect is found from a shock to this ratio on the real economic variables.

5.1.5

D

EPOSITS TO

T

OTAL

L

OANS

Figure 14 shows the response of the deposit to total loans ratio. The ratio has a strong positive short term effect on inflation which is statistically significant. This implies that in the case deposits as percentage of total loans go up, inflation goes up which seems counterintuitive. The effect of a shock to this ratio on the other real economic variables is not visible. Furthermore, the FSI shows an immediate negative response to a shock in GDP. This is an

Fig. 13. Impulse response function via Cholesky decomposition for Liquid Assets to Total Assets ratio model with log-difference.

25

interesting finding: as GDP growth goes up, the deposit to total loans ratio goes down. This statistically significant negative effect persists for a time period of roughly 6 quarters. A decrease in this ratio can either be caused by a decrease in deposits or an increase in total loans. Both factors are likely to contribute to this decrease in the case of GDP growth. The response to an interest rate shock shows an opposite sign: the ratio responds positively to this shock though this effect only persists in the short sun. The inflation shock does not lead to any relevant result. Again, the results found are not statistically significant.

5.1.6

D

EFAULT

R

ATE

The default rate represented by Figure 15 shows that this banking sector indicator decreases when hit by a shock in GDP for a time period of 20 quarters. Moreover, this result is found to be statistically significant. This finding is consistent with expectations: the default rate and GDP are likely to be negatively correlated. The effects of a shock to inflation or the interest rate do not lead to any relevant reaction in the default rate. Fig. 14. Impulse response function via Cholesky decomposition for Deposit to Total Loans model ratio with log-difference.

26

5.2

R

OBUSTNESS

C

HECKS

To check the robustness of the results obtained through the Cholesky decomposition of the variance-covariance matrix I use three methods. First, I repeat the Cholesky decomposition but change the ordering of the variables. Second, I calculate the IRFs using a VAR model in which the FSIs enter in levels instead of in log-differences as in Dovern et al. And third, I calculate the IRFs with a different number of lags. I reduce the number of lags, as in Marcucci et al (2009), and increase the number of lags, as in Dovern et al. (2010). Below I detail these robustness checks.

5.2.1

D

IFFERENT

O

RDERING OF THE

V

ARIABLES

The ordering of the variable is GDP, inflation rate, followed by the short term interest rate. In each model, the FSI is placed as a fourth variables. This ordering allows to assume that all the shocks hit the banking system at the same time. I repeat this exercises with several different orderings of the endogenous variables but the results do not show significant differences.

Fig. 15. Impulse response function via Cholesky decomposition for Default Rate model ratio with log-difference.

27

5.2.2

VAR

MODEL WITH

FSI

S IN

L

EVELS

I repeat the stress-testing exercised and obtain the IRFs by using the FSIs in levels. The results are reposted in Appendix, Figure 16-22. Although the interpretation of the results in this identification scenario is less clear, the IRFs do confirm the original findings. For the first three ratios (bad debts to capital, bad debts to gross loans, with and without cassa deposit and prestiti) the only significant response is a negative change in each ratio due to a positive shock in GDP. Overall, only the effect of GDP on bad debts to gross loans ratio in statistically significant. The fourth and fifth ratios, bad debts over gross loans and capital to assets, do not present any relevant response. The liquid assets to total assets ratio does show a relevant change but only in the response to its own shock. As a matter of fact, the response is found to be immediately positive. The change persists for ten quarters and the result is statistically significant.

The IRFs for the model with deposits to total loans as FSI confirm the analysis presented above even if in this case the banking sector variable does not immediately react to the positive change in GDP but smoothly decreases for five quarters. The response is statistically significant. Again, the response of this ratio to his own shock is found to be relevant and positive but, different from the previous analysis, in the long run for a time period of twenty quarters. Finally, the last financial soundness indicator studied in this paper, the default rate, presents the same characterises found above with the only difference that the responses persist for a longer range of time. Although, these results are not statistically significant. Overall, the results found show that the main analysis is robust. Entering the variables as levels into the VAR model does not seem to significantly alter the findings presented above.

5.2.3

VAR

MODEL WITH

O

NE

L

AG

As a second robustness check of my results I use the same VAR model described in the methodology but with one lag. This reduction leads to a decrease in the explanatory capability of the variables. The results are similar to the ones found in the previous analysis. In addition,

28 a positive shock of GDP to bad debts to capital, bad debts to gross loans, with and without cassa deposit and prestiti and capital to total assets are all statistically significant. This analysis confirms the results obtained above.

5.2.4

VAR

MODEL WITH

F

OUR

L

AGS

As last robustness check I increase the number of lags in the model from two to four. By adding lags I increase the explanatory power of the variables but it also increases the number of the coefficients entering the model which could lead to an increase in the amount of estimation errors (Stock and Watson, 2001). The results obtained do not differ from the one described

29

6.CONCLUSION

This thesis analyses the effects of macroeconomic shocks on the soundness of the Italian banking sector trough a VAR model.

I estimate seven SVAR models and identify the structural shocks using a Cholesky decomposition of the variance-covariance. The seven models involve five variables, three endogenous macroeconomic variables (current Italian GDP, inflation, interest rate) and one exogenous (European GDP). The models differ with respect to the fourth endogenous variables, the financial soundness indicator. I use two capital based indicators (bad debts to capital and capital to assets) and five asset based (bad debts to gross loans, with and without cassa deposit and prestiti, liquid assets to total assets, deposit to total loans and the default rate). Data for all variables covers a time period of eighteen years, from Q1 1997 until Q4 2015. The time span of this data set is small due to limited availability of banking balance sheet data which represent a drawback of this paper. Several conclusions have been found from this analysis. First, the ratio’s involving bad debt (bad debts to capital and bad debts to gross loans) are found to have no significant impact on the real economic variables: GDP, inflation and interest rate. The other way around, shocks to these real economic variables are found to have an effect on the bad debt ratios. Especially shocks to GDP are shown to have a strong positive effect on the bad debt ratios. This implies that when GDP decreases, the bad debt ratios increase. This provides further evidence that the current state of the Italian banking system, burdened by a large amount of bad debt, is partially the result of weak economic recovery. Out of the asset ratios, the capital to assets ratio is found to have the strongest link with the real economy. In the short run, shocks to this ratio impact both GDP and the interest rate positively. The inflation rate shows a somewhat different response, where a shock to the capital to asset ratio first causes an immediate decrease in inflation, followed by an increase over the subsequent five quarters.

30

Furthermore, as we find both significant effect of financial soundness indicators on real economic variables as we do the other way around, this research confirms the presence of spill-over effects between the real and financial economy. Most feedback effects I find flow from the real economy to the financial soundness indicator. I applied several robustness checks to these results. I use a different ordering of the variables for the Cholesky decomposition of the variance-covariance and enter the financial soundness indicators in levels, both of which have no relevant impact on the previously detailed results. I also use a different lag length (one lag model and four lag model). The results of these robustness checks support the findings of the original analysis. This study faces several limitations which could be addressed in future research. The main issues are regarding the data availability: banking balance sheet data are accessible only for a short time spam, which does not allow to take into consideration a more extensive period of time. Moreover, it is hard to find data at bank level. Further research could also try to involve different FSIs and compute the analysis I present here with a variating approach to investigate the effect of other shocks.

31

R

EFERENCES

Bazley, N. (2016), “The next EU crisis – in the Italian banks – needs to be addressed.” The Guardian Christiano, L., Eichenbaum, M. and Evans, C. (1998). Monetary policy shocks. Cambridge, MA: National Bureau of Economic Research.

Čihák, M. (2007). Introduction to applied stress testing. Washington, D.C.: International Monetary Fund, pp.1-76. Claessens, S., Kose, M., Laeven, L. and Valencia, F. (2014). Understanding Financial Crises: Causes, Consequences, and Policy Responses. SSRN Electronic Journal. Dovern, J., Meier, C. and Vilsmeier, J. (2010). How resilient is the German banking system to macroeconomic shocks?. Journal of Banking & Finance, 34(8), pp.1839-1848. European Central Bank, (2006) Financial stability review. Frankfurt am Main: European Central Bank. European Central Bank, (2015). The role of stress testing in supervision and macroprudential policy. Londra. Filosa, R. (2016). Stress testing of the stability of the Italian banking system: a VAR approach. Heterogeneity and monetary policy, 703(1), pp.1-46.

Fong, T. and Wong, C. (n.d.). Stress Testing Banks' Credit Risk Using Mixture Vector Autoregressive Models. SSRN Electronic Journal.

Gómez-Puig, M. and Sosvilla-Rivero, S. (2016). Causes and hazards of the euro area sovereign debt crisis: Pure and fundamentals-based contagion. Economic Modelling, 56, pp.133-147.

Hamilton, J. (1994). Time series analysis. Princeton, N.J.: Princeton University Press.

Hoggarth, G., Sorensen, S. and Zicchino, L. (2005). Stress Tests of UK Banks Using a VAR Approach. SSRN Electronic Journal.

32

International Monetary Fund and World Bank, (1999). Financial sector assessment program. Washington: International Monetary Fund.

International Monetary Fund, (2006) Financial soundness indicators. (2006). Washington, D.C.: International Monetary Fund, pp.Chapter 4,6,8,13.

Jassaud, N. and Kang, K. (2015). A Strategy for Developing a Market for Nonperforming Loans in Italy. IMF Working Papers, 15(24), p.1.

Kapinos, P., Mitnik, O. and Martin, C. (2015). Stress Testing Banks: Whence and Whither?. SSRN Electronic Journal.

Marcucci, J. and Quagliariello, M. (2009). Asymmetric effects of the business cycle on bank credit risk.Journal of Banking & Finance, 33(9), pp.1624-1635.

Mr. Winfrid Blaschke., Mr. Matthew T. Jones., Mr. Giovanni Majnoni., and Mr. Maria Soledad Martinez Peria., (2001). Stress Testing of Financial Systems: An Overview of Issues, Methodologies, and FSAP Experiences. International Monetary Fund, pp.1-88. Romer, D. (2012). Advanced macroeconomics. 4th ed. New York: McGraw-Hill Companies. Sanderson, R. (2016). Bad-debt warning triggers fresh fears for Italian banks. Financial Times. Sims, C. (1986). Are forecasting models usable for policy analysis?. Federal Reserve Bank of Minneapolis 10, pp.2–16. Sims, C. (1992). Interpreting the macroeconomic time series facts. European Economic Review 36, pp.975–1000.

Sorge, M. (2004). Stress-testing Financial Systems: An Overview of Current Methodologies. SSRN Electronic Journal.

Stock, J. and Watson, M. (2014). Introduction to econometrics. Pearson College Division, pp.631-658

33

Za, V. (2016), Bank of Italy calls for revision of EU "bail in" rules, Reuters

Zivot, E. and Wang, J. (2006). Modeling financial time series with S-plus. New York, NY:

34

A

PPENDIX

I

–

D

ICKY

-F

ULLER TEST RESULTS

Bad Debts to Capital

Dickey-Fuller test for unit root Number of obs = 69

--- Interpolated Dickey-Fuller --- Test 1% Critical 5% Critical 10% Critical Statistic Value Value Value

Z(t) -3.554 -3.553 -2.915 -2.592 MacKinnon approximate p-value for Z(t) = 0.0003 Bad Debts to Gross Loans with cassa depositi and prestiti Dickey-Fuller test for unit root Number of obs = 69 --- Interpolated Dickey-Fuller ---

Test 1% Critical 5% Critical 10% Critical Statistic Value Value Value

Z(t) -3.554 -3.553 -2.915 -2.592 MacKinnon approximate p-value for Z(t) = 0.0067 Bad Debts to Gross Loans Dickey-Fuller test for unit root Number of obs = 74 --- Interpolated Dickey-Fuller --- Test 1% Critical 5% Critical 10% Critical Statistic Value Value Value

Z(t) -6.593 -3.546 -2.911 -2.590 MacKinnon approximate p-value for Z(t) = 0.0000

35 Capital to Assets Dickey-Fuller test for unit root Number of obs = 74 --- Interpolated Dickey-Fuller --- Test 1% Critical 5% Critical 10% Critical Statistic Value Value Value

Z(t) -7.563 -3.546 -2.911 -2.590 MacKinnon approximate p-value for Z(t) = 0.0000

Liquid Asset to Total Assets Dickey-Fuller test for unit root Number of obs = 74 --- Interpolated Dickey-Fuller --- Test 1% Critical 5% Critical 10% Critical Statistic Value Value Value

Z(t) -7.048 -3.546 -2.911 -2.590 MacKinnon approximate p-value for Z(t) = 0.0000

Deposit to Tot Loans Dickey-Fuller test for unit root Number of obs = 74 --- Interpolated Dickey-Fuller --- Test 1% Critical 5% Critical 10% Critical Statistic Value Value Value

Z(t) -8.364 -3.546 -2.911 -2.590

MacKinnon approximate p-value for Z(t) = 0.0000

36 Default Rate Dickey-Fuller test for unit root Number of obs = 77 --- Interpolated Dickey-Fuller --- Test 1% Critical 5% Critical 10% Critical Statistic Value Value Value

Z(t) -18.455 -3.542 -2.908 -2.589

MacKinnon approximate p-value for Z(t) = 0.0000

37

A

PPENDIX

II

–

G

RAPHS

Fig. 1. Financial soundness indicators and the macroeconomy Q1 1997-Q4 2015 – the figure presents the trend of the variables used in the VAR models. Fig. 2. Bad Debts to Capital model.

38

Fig. 3. Bad Debts to Gross Loans (with cassa crediti e depositi) model. Fig. 4. Bad Debts to Gross Loans model.

39 Fig. 5. Capital to Assets model. Fig. 6. Liquid Assets to Total Assets model.

40 Fig. 8. Default Rate model. Fig. 7. Deposit to Total Loans model.

41

A

PPENDIX

III

–

I

MPULSE

R

ESPONSE

F

UNCTIONS OF

R

OBUST

C

HECK

Fig. 16. Impulse response function via Cholesky decomposition for Bad Debts to Capital ratio with levels.

Fig. 17. Impulse response function via Cholesky decomposition for Bad Debts to Gross Loans (with cassa crediti e depositi) ratio with levels.

42 Fig. 18. Impulse response function via Cholesky decomposition for Bad Debts to Gross Loans ratio with levels. Fig. 19. Impulse response function via Cholesky decomposition for Capital to Assets ratio with levels.

43 Fig. 20. Impulse response function via Cholesky decomposition for Liquid Assets to Total Assets ratio model with levels. Fig. 21. Impulse response function via Cholesky decomposition for Deposit to Total Loans model ratio with levels.

44 Fig. 22. Impulse response function via Cholesky decomposition for Default Rate model ratio with levels.