Financial fragility of euro area households

Accepted Manuscript Title: Financial Fragility of Euro Area Households Author: Miguel Ampudia Has van Vlokhoven Dawid ˙ Zochowski PII: DOI: Reference:

S1572-3089(16)00024-3 http://dx.doi.org/doi:10.1016/j.jfs.2016.02.003 JFS 421

To appear in:

Journal of Financial Stability

Received date: Revised date: Accepted date:

10-11-2014 19-7-2015 18-2-2016

˙ Please cite this article as: Ampudia, M., van Vlokhoven, H., Zochowski, D.,Financial Fragility of Euro Area Households, Journal of Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.02.003 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Financial Fragility of Euro Area Households1 Miguel Ampudia2, Has van Vlokhoven3 and Dawid Żochowski4

ip t

January 2016

cr

Abstract

Ac ce p

JEL-codes: D10, D14, G21

te

d

M

an

us

Until recently, the lack of appropriate harmonised micro data covering both income and wealth has been the major obstacle in analysing financial vulnerability of the household sector in the euro area. This data problem has been partially circumvented by the dissemination of the Household Finance and Consumption Survey (HFCS). Based on this unique data set, we put forward a stress testing method of household balance sheets in a consistent manner across euro area countries. To this end, we put forward a metric of distress which takes into account both the solvency and liquidity position of the household and demonstrate that this metric outperforms the most common metrics used in the literature, which do not take into account the households’ asset holdings. We calibrate this metric using the country level data on non-performing loan ratios and estimate stress-test elasticities in response to an interest rate shock, an income shock and a house price shock. We find that, albeit euro-area households are relatively resilient as a whole, there are large discrepancies in the impact of macroeconomic shocks across countries. Finally, we demonstrate that our framework could be used to assess some measures mitigating losses to the banks, such as engaging in the restructurings of loans that are at risk of defaulting.

Keywords: household indebtedness, stress testing, household finance, financial stability.

1

This paper uses data from the Eurosystem Household Finance and Consumption Survey. It presents the authors’ personal opinions and does not necessarily reflect the views of the European Central Bank or the Eurosystem Household Finance and Consumption Network. We are grateful to participants at seminars at the ECB, the Household Finance and Consumption Network, Loughborough University, Eurofound and Oesterreichische Nationalbank for useful comments. We also thank Iftekhar Hasan(editor), Alistair Mirne (special editor) and two anonymous referees for comments which have greatly improved the paper. All remaining errors are our own. 2 European Central bank; Sonnemannstrasse 20, 60314 Frankfurt am Main; Germany; phone: +49 (0) 69 1344 8860; [email protected]. Corresponding author. 3 Stockholm University; Universitetsvägen 10 A, 106 91 Stockholm; Sweden; phone: +46 (0) 816 4245; [email protected]. 4 European Central bank; Sonnemannstrasse 20, 60314 Frankfurt am Main; Germany; phone: +49 (0) 69 1344 5896; [email protected].

1 Page 1 of 28

1. Introduction and Related Literature

cr

ip t

The recent financial crisis has underscored the importance and need for in-depth surveillance and the analysis of risks faced by financial institutions in a consistent and uniform manner. For this purpose, four rounds of EU-wide stress tests were conducted in Europe, and their results were published by the European Banking Authority (EBA) and its predecessor, the Committee of European Banking Supervisors (CEBS), between 2009 and 2014.5 Those tests examined the way in which the capital ratios of the banks would be affected in the event of an adverse macroeconomic scenario and sovereign risk shock. Additionally, in the run-up to the assumption of the new supervisory and macro-prudential powers by the ECB, a top-down macro stress testing framework was developed to conduct regular forward-looking bank solvency assessments (Henry and Kok, 2013).

an

us

While credit risk is the most important factor in determining bank solvency, macro-stress tests, by their nature, do not observe the build-up of vulnerabilities and imbalances in the household or corporate sector but rather link macro-variables to aggregate probabilities of defaults. However, stress testing corporate or household balance sheets directly could provide useful insights into risks arising from the real banking sector. In this way, they may further enhance the accuracy of the macro stress tests, for instance, by providing micro-based estimates of the elasticity of the real sector to macro-shocks and accounting for the distributional aspects of households’ ability to pay.

te

d

M

In particular, studying the vulnerabilities of the household sector is important for at least two reasons. First, the household sector holds the largest share of wealth in developed economies (the rest being held by non-profit organizations, foreigners or the state). As wealth is one of the most important factors in determining household consumption through its lifecycle, household consumption decisions are influenced by its solvency position, thereby impacting the economic activity. Second, vulnerable households pose a threat to financial stability due to their ties to financial institutions.

Ac ce p

Until recently, the lack of appropriate harmonized micro-data covering both income and wealth has been the major obstacle in conducting the vulnerability exercise for the household sector in the euro area. This data problem has been partially circumvented by the dissemination of the Household Finance and Consumption Survey (HFCS), a novel dataset that collects information on sociodemographic variables, assets, liabilities, income and consumption for a sample of households that is representative of both the national and euro area level. This paper uses the HFCS to conduct stress tests of household balance sheets by quantifying the impact of hypothetical adverse shocks on their ability to continue servicing their debt. Moreover, thanks to micro information on the distribution of wealth and income, it detects groups of households or countries that are particularly vulnerable to adverse shocks. Therefore, it gives the policy makers a tool to adequately impose macro-prudential policy measures. In particular, from a central bank perspective, the impact of monetary policy decisions on credit risk stemming from the household sector across euro area countries can be assessed and quantified. This, in turn, could inform the macro-prudential policy makers’ decisions on optimal preventive measures and their application across countries to mitigate the risks to the extent possible.

5

EU-wide stress tests, European Banking Authority, https://www.eba.europa.eu/risk-analysis-and-data/eu-wide-stress-testing

2 Page 2 of 28

There is one obvious limitation to our reported results. The household level micro-data set we used is only available so far as a cross-section, and the time series is not available yet. This means that we cannot fully validate our simulations of household stress at the micro-level over different stages of the economic cycle. However, we are able to provide a more limited robustness test at the macrolevel in Section 5 by constructing a time series of the data that combines the cross-sectional information in the data set with the evolution of variables at the macro level.

cr

ip t

The remainder of this section provides a review of the literature that uses micro-level survey data in assessing the vulnerabilities in the household sector. This literature strand links most closely with our paper.

an

us

Johansson and Persson (2006), Vatne (2006) and Żochowski and Zajączkowski (2007) were the first researchers to use micro-level data on households to assess the vulnerability of households in Sweden, Norway and Poland, respectively. Subsequently, similar studies were also conducted for Finland (Herrala and Kauko, 2007), Hungary (Holló and Papp, 2007), Canada (Dey, Djoudad and Terajima, 2008), Czech-Republic (Bičáková, Prelcová and Pašaličová, 2010), Croatia (Sugawara and Zalduendo, 2011.) and Italy (Michelangeli and Pietrunti, 2014). Tiongson et al. (2009) conduct a cross-country study of household indebtedness using survey data for a large group of European countries.

M

In most of these papers, the authors use two types of measures of household vulnerability or, put differently, ability to repay debt: (1) the debt-service-to-income ratio and (2) financial margin.

Ac ce p

te

d

Concerning the debt-service-to-income ratio, defined as the ratio of monthly loan installments (both interest and principal) to total disposable income, the key is to find an adequate level of threshold for the identification of vulnerable households. This is done using various criteria. For instance, Tiongson et al. (2009) argue that a 30-percent threshold is adequate for recalling studies on mortgage delinquencies (May and Tudela, 2005.). The same threshold is used by Michelangeli and Pietrunti, 2014, while Dey, Djoudad and Terajima (2008), using data on mortgage-debt delinquency rates in Canada, find that for households with a ratio beyond 35 percent, there is a significant increase in the probability of delinquency. As the Italian dataset is partially composed of a panel, Michelangeli and Pietrunti (2014) check the forecasting performance of a 30-percent threshold in the debt-service-to-income ratio by conducting a back-test on previous waves of the survey. They are able to fit the percentage of vulnerable households to the realizations in the following two surveys relatively well. Financial margin is the most popular measure of household vulnerability in the literature that exploits micro-level data on households. It is defined as disposable income after deducting basic living costs and loan installment payments. Households with negative financial margins are considered vulnerable (see: Johansson and Persson (2006), Żochowski and Zajączkowski (2007) and Holló and Papp (2007)). The basic idea behind this measure is that households with negative current cash flows are much more likely to default on loans than those with positive financial margins. Hence, financial margin may be a good proxy for default risk. Vatne (2006) compares various definitions of financial margin and finds that there is a positive correlation between the share of debt held by households with negative margins and default rates. Moreover, the turning points of the exposed debt seem to precede the turning points of default rates, which suggest that the financial margin measured at the individual household level may be an early warning indicator for 3 Page 3 of 28

non-performing loans. An algebraic transformation of the financial margin is used by Bičáková, Prelcová and Pašaličová (2010). Their measure is defined as the ratio of loan repayments to net income minus basic living costs, and they show that it outperforms the standard debt-service-toincome ratio as a predictor of default risk for their particular sample of Czech households.

cr

ip t

To the best of our knowledge, only one study combines data for both income and assets to identify distressed households. Brunetti et al. (2015) consider a household to be fragile if its income is higher than the expected expenses and its liquid assets are lower than its unexpected expenses. They, however, focus on a non-optimal portfolio allocation analysis rather than measuring the credit risk of a household.

te

d

M

an

us

Once vulnerable households are identified, some studies try to determine the causes behind this situation. Using a logit model and the financial margin as a dependent variable, Herrala and Kauko (2007) find that a household's risk of being financially distressed depends on the net income after tax and loan servicing costs. Holló and Papp (2007) argue that the financial margin as a dependent variable in a logit model has shortcomings, given that, per definition, it is highly correlated with consumption and income, i.e. the potential independent variable. Instead, the authors estimate a logistic model defining a binary dependent variable as households whose arrears exceed one month. They find that the main individual factors affecting household credit risk are disposable income, the income share of monthly debt servicing costs, the number of dependents and the employment status of the head of household. Tiongson et al. (2009) construct a probit model of unemployment in which the unemployment status is a function of an individual's educational level, age, gender, and place of residence. While Bičáková, Prelcová and Pašaličová (2010) use a non-parametric measure of financial vulnerability, they construct a probit model for the probability that a household takes a loan in the next year and a tobit model to estimate the size of new loans. Finally, Michelangeli and Pietrunti (2014) construct a micro-simulation model to monitor the financial vulnerability of Italian households by modelling income and growth dynamics.

Ac ce p

Many of the studies mentioned above, as well as ad-hoc work inside central banks, perform stress testing exercises to assess the degree to which households are vulnerable to adverse financial shocks and the effect of these shocks on the stability of the financial system. The structure of most of these studies is common. The first step is to choose a measure with which to classify a household as vulnerable. Most studies use the financial margin or the debt-service-to-income ratio. The shocks are then defined, both their typology and quantification. For instance, Tiongson et al. (2009) simulate the impact of shocks for various countries based on actual changes in recent years as well as on uniform, hypothesized magnitudes. The third step is to show the impact of these shocks on the household vulnerability measures. Last, the impact on the banks is analyzed by looking at measures such as the Exposure at Default (EAD) and the Losses Given Default (LGD). The former measure represents the debt held by vulnerable households as a percentage of total debt, while the latter represents potential losses faced by the banking sector as a percentage of total debt. We contribute to this literature strand twofold. First, we propose a micro indicator of household financial vulnerability, which takes into account household liquidity position as well as solvency. Second, based on the unique dataset encompassing euro area countries, we put forward a stress testing method which, in the absence of a micro level stress indicator, is calibrated based on country-level data on non-performing loan ratios and therefore is not entirely ad hoc. Using the 4 Page 4 of 28

ip t

calibrated version of our indicator, we compare the financial vulnerability of households and the banking sector across the euro area against the same hypothetical adverse scenarios using micro data on the household balance sheet structure. The hypothetical adverse shock scenario can be composed of any combination of interest rates, housing prices and income shocks. To the best of our knowledge, this is the first paper that uses a harmonized micro-level data set to conduct stress tests on household balance sheets in a consistent manner across euro area countries. Such a method is potentially valuable also because, due to the common monetary policy, the euro area-wide stress testing gives consistent insights into the vulnerability of the euro area household sector.

an

us

cr

The paper is structured as follows. Section 2 presents the dataset used in the paper and more generally discusses the data challenges related to analyzing the financial vulnerability of households. In section 3, we define a measure of household distress that accounts for both the liquidity and solvency situation of a household. Section 4 presents the results of a stress test exercise by analyzing the effects of adverse shocks to the interest rate, income and housing prices on the household distress measure and quantifying the potential losses of the financial industry that could materialize as a result. In section 5, we provide a set of robustness checks. Section 6 demonstrates a possible use of our measure of distress in loan restructuring. Section 7 presents the paper’s conclusions.

M

2. Data Availability and Data Challenges

Ac ce p

te

d

Due to the lack of micro-level information on households, debt burden indicators, such as the debtto-income ratio or the debt-to-asset ratio, have been calculated at a macro-level using aggregated information from national accounts statistics. However, this aggregate view masks information about the distribution of debt and debt burden among individual households. Overindebtedness can also emerge at relatively low levels of aggregate debt burden indicators, for instance, if debt is concentrated among a relatively low percentage of households. Hence, the debt-to to-income ratio or the debt-to-asset ratio calculated at the individual household level and their distribution across the population provide useful insights into household debt burden. However, debt burden measures are simply indicators of household vulnerability. To better predict the household probability of default, all household characteristics determining their credit risk combined with its loan repayment history, such as those embedded in credit register data, are required. The ideal dataset, thus, would include complete balance sheet (household assets and liabilities) and income information for every household in the country combined with information about the collateral and loan repayment history or at the least whether the household defaulted on its debt. Unfortunately, at the euro area level, such information is not yet available and will likely not be available in the near future. With these caveats in mind, the Eurosystem’s Household Finance and Consumption Survey (HFCS) provides a useful component of this ideal dataset. To date, it is the best available data source to study household financial fragility and perform household stress test exercises in a consistent manner for euro area countries. The HFCS contains information regarding socio-demographic variables, assets, liabilities, income and consumption for a sample of households that is representative both at the national and euro area level. A set of population weights is provided to ensure the representativeness of the sample. It provides ex-ante comparable data for more than 62,000 households in 15 euro area countries (all 5 Page 5 of 28

but Ireland, Estonia, Latvia and Lithuania)6. For the purposes of this paper, we exclude Finland from the analysis because the data on debt payments were not collected in this country.

ip t

Another important feature of the HFCS is that missing observations for all variables necessary to construct wealth and income aggregates (i.e. questions that were not answered by the respondent households) are imputed five times. The survey was conducted around 2010, but the reference periods have not been fully harmonized. In particular, the reference period for the Spanish data is 2008/2009; for Greece and the Netherlands, 2009; and for the other countries, 2010.

us

cr

There are two main features that make this dataset unique with respect to all other data sources. First, it contains complete information on both wealth and income for each household. This allows us to properly asses the buffers that a household has to cover its debt payments (which may come from income streams or accumulated wealth) and to quantify the losses for the financial system in cases of default. This can only be done when complete information on the household asset portfolio and its composition is available, which makes it possible to evaluate the amount a creditor could recover in the event of a borrower’s default.

te

d

M

an

Nevertheless, the data also have some shortcomings. First, there is no explicit information on whether a household has defaulted on its debt. Therefore, we define the measure of default ourselves. Second, there is no information about the region of residence leading to limitations in the estimation of basic living costs and introducing heterogeneity to collateral valuation (possibly different housing price elasticities in different regions). In addition, so far, only one wave of the HFCS has been conducted, and therefore we lack a time series component that would be useful for performing a validation of our measure through back testing. Finally, the HFCS oversamples wealthy households, whereas, from a financial stability point of view, it is recommended to oversample indebted households.

Ac ce p

The way we address these challenges constitutes the essence of section 3 of the paper, but we outline the main ideas here. To identify the households that are likely to default on their debts, we define a metric that takes into account both the liquidity and solvency position of the household. When constructing such measures, it is necessary to calibrate some parameters. To this end, we use macro data on Non-Performing Loans (NPL) at the aggregated country level from the EBA EU-wide transparency exercise results. To perform an evaluation of this metric of default, in section 5, we first “update” the HFCS by using the evolution of asset prices and income observed in the macro data. We then compare the changes in the Loss Given Default (LGD) that we calculate using our measure with the changes in the Loan Loss Provisions (LLPs) observed in the aggregated banking data.

3. Financially Vulnerable Households 3.1 Debt burden indicators

6

For more details on the survey, see http://www.ecb.europa.eu/home/html/researcher_hfcn.en.html. The results from the first wave are described in detail in the Household Finance and Consumption Network (2013a).

6 Page 6 of 28

The most common measures of household debt burden are the debt-to-income ratio, debt-to-asset ratio and debt-service-to-income ratio.

Table 1a Summary statistics

ip t

To better understand how the distribution of debt can impact household vulnerability, we study how these measures (more precisely, the percentage of households exceeding a certain thresholds of these measures) vary across countries and the income distribution.

Countries Austria

Obs. 2380

% indebted households 35.6%

% adjustable rate lending HMR 74.4%

% DSI > 0.4 6.8%

Belgium

2327

44.8%

35.8%

20.0%

6.0%

21.1%

Cyprus

1237

65.4%

68.7%

30.4%

4.4%

25.4%

France

15006

46.9%

13.7%

13.1%

8.3%

12.2%

3565

47.4%

19.5%

12.1%

16.2%

11.4%

2971

36.6%

49.4%

7951

25.2%

57.9%

Luxembourg

950

58.3%

82.6%

Malta

843

34.1%

87.3%

Netherlands

1301

65.7%

Portugal

4404

37.7% 44.5% 26.8%

Spain

6197

50.0%

Total

51532

43.4%

2 3 4 5

cr 7.2%

9.7%

6.1%

11.3%

14.1%

6.5%

20.6%

4.2%

2.4%

9.1%

83.9%

15.4%

17.8%

32.8%

87.4%

20.1%

6.9%

26.2%

93.5%

26.9%

5.0%

9.6%

44.4%

13.7%

4.4%

9.0%

87.4%

26.7%

7.2%

23.2%

49.7%

16.0%

11.1%

15.7%

% DSI > 0.4 14.4%

% DI > 4 12.1%

14.8%

16.1%

2

16.4%

18.0%

21.2%

3

12.1%

14.4%

13.4%

4

7.9%

17.8%

15.5%

5

4.0%

Ac ce p

Net wealth quintile 1

te

Table 1b Summary statistics

14.1%

M

343 2057

d

Slovenia Slovakia

% DI > 4 9.5%

15.6%

an

Greece Italy

us

Germany

% DA > 1 14.9%

Income quintile 1

% DA > 1 27.7%

Sources: HFCS & own calculations. Note: HMR stands for household main residence, DSI, DA and DI stand for debt-service-to-income ratio, debtto-asset ratio and debt-to-income ratio, respectively.

Table 1a, column 4 shows the proportion of households with a debt-service-to-income ratio greater than 0.4. This ratio reflects the capacity of the household to repay its debt without resorting to selling assets. Because the majority of household assets are illiquid7, this indicator reflects the ability of households to repay their debt on time and thus focuses on the short-term horizon. The debtservice-to-income ratio is constructed as total monthly debt payments to monthly net income. There is substantial heterogeneity across countries. In some countries, such as Cyprus, Slovenia or Spain, 7

For a complete picture of the composition of euro area household balance sheet, see the Household Finance and Consumption Network (2013a).

7 Page 7 of 28

the ratio exceeds 25%, while in other countries, such as Austria or Malta, the ratios are well under 10%. Table 1b, column 1 shows the distribution of the ratio for the euro area across wealth quintiles and shows a rather even picture.

an

us

cr

ip t

The debt-to-asset ratio is constructed as total debt over total assets. From a financial stability perspective, it is important to monitor households with a debt-to-asset ratio greater than one (table 1a, column 5). These households are “under water”; should their debt be liquidated, banks would have to face losses. This ratio is greater than one for 11.1% of indebted households in the euro area. However, cross country variation is meaningful — as much as 17.8% of Dutch and only 2.4% of Maltese indebted households are “under water.” While in most jurisdictions there is no legal cap on the LTV ratio, a threshold can be put in place for capital and provisioning requirements, leading to differences in typical LTV ratios set by the banks across countries (ECB, 2009) and partially explaining the variation in debt-to-asset ratios. Furthermore, the last column in Table 1b shows that at the euro area level, there are important differences across the income distribution. While less than 5% of indebted households in the highest income quintile exhibit a debt-to-asset ratio greater than 1, more than 25% of indebted households in the first income quintile do so.

Ac ce p

te

d

M

Table 1a, column 6 and Table 1b, column 2 show the numbers for heavily indebted households according to their debt-to-income ratio (those with a debt-to-income ratio greater than 4). This ratio indicates the number of years a household needs to generate income to repay its entire debt. Although this ratio has some drawbacks as it is comprised of a stock and a flow variable, it does provide some useful insight into the financial risk a household faces. For instance, households with high debt-to-income ratios are more sensitive to shocks, particularly interest rate shocks, and they are therefore more likely to default should they materialize. We again identify substantial crosscountry variation, from 32.8% of indebted households in the Netherlands to 9.0% in Slovakia. In total, 15.7% of indebted households in the euro area have a debt-to-income ratio of greater than four. Furthermore, looking at the net wealth distribution, the proportion of heavily indebted households is quite stable across the net wealth quintiles. These financial burden indicators are useful for forming a general impression of households under financial stress. Nevertheless, they could hide important aspects of the problem. For example, while a household with a very high debt-service-to-income ratio may face difficulties covering its debt installment from the current income stream, if it owns liquid assets, it can sell them to continue servicing the debt without ever being at risk of missing payments. In a similar vein, a household may be “under water,” but if its income is sufficient to cover monthly installments, it may never default in the absence of any negative income shock.

3.2 Moving toward default — a measure of distress In view of the drawbacks of financial burden indicators, which we discussed above, we propose a comprehensive measure of financial distress that would proxy the household’s probability of default. To this end, we propose a measure of distress that accounts for both the liquidity and solvency conditions of a household, as the household is forced to default only if these two conditions are met.

8 Page 8 of 28

To express this concept formally, we first define the financial margin FM iq of household i in country q as

FM iq  I iq  Ti q  DPi q  BLC q 8 ,

(eq. 1)

household i in country q, respectively. Finally, we define basic living costs as

~ BLC q   I q ,

ip t

where I iq is the i-th household income and Ti q and DPi q are the taxes and debt payments paid by

(eq. 2)

cr

~

where I q is the median income in country q and  is a specific fixed percentage.9

an

us

This construction of the financial margin assumes that a household uses its income to pay taxes, repay its debt and cover basic living costs. We consider that households that are not able to cover all their spending from income, i.e. those with a negative financial margin, are in financial distress. We do not take into account any possible changes in future income and hence do not consider restructuring.

d

M

We calculate the financial margin using information on income and debt payments from the HFCS. Regarding taxes, we only consider income taxes, which we estimate using information on tax brackets and tax credits from the OECD10 and individual household income. To estimate basic living costs, we follow the literature on poverty lines.11 The European Commission uses a relative poverty line in their poverty studies for European countries, in which the poverty line is a fixed percentage of median income (European Commission, 2011). Following this approach, we set the basic living costs as a fixed percentage,  , of median income — specifically, to 40% of median income.12

te

q

Ac ce p

A household with a negative financial margin can still service its debt in case it has sufficient assets to sell to cover the payments. Therefore, we introduce a second condition related to the ability to cover the negative financial margin with liquid assets. It states that a household is considered to be in distress if the household’s negative financial margin for a determined number of months, M q , is greater than the household’s liquid assets. In other words, this condition says that a household is not in distress even if it has a negative financial margin if it can cover a given number of months of the flow of negative financial margin from its liquid assets.

8

In addition, the rent paid is subtracted from the financial margin for tenants because housing costs are already covered by the debt payments for those with mortgages, and hence, the basic living costs exclude housing costs. 9 In addition, the basic living costs are adjusted by the number of members for each household, in line with the OECDmodified scale, which assigns a value of 1 to the household head, 0.5 to each additional adult and 0.3 to each child. 10 We explored the possibility of including real estate taxes because of its relative importance in some countries, but the fact that these taxes are often determined at the local level and on the basis of many different factors made any estimation unreliable. 11 In this literature, a household is considered impoverished if its income is below a set poverty line, in which being impoverished can be understood as being socially excluded (Laderchi, Saith and Stewart, 2003). In other words, we set the basic living costs as the minimum amount of money needed for a household necessary to avoid social exclusion — i.e., we set the basic living costs as equal to the poverty line. 12 The percentage used by the European Commission (2011) ranges from 40% to 70%.

9 Page 9 of 28

To allow for some uncertainty in our measure of default, we attach a probability of default to each distressed household based on the relationship between its financial margin and liquid assets (conditional on being in distress). That is, not all of our households in distress will default. We will explain how this probability distribution is determined in section 3.3.

 FM iq  0  | FM iq | *M q  LIQiq  FM iq  0  | FM iq | *M q  LIQiq

(eq. 3)

cr

1 qi   0

ip t

Formally, we define the measure of distress, qi 13, of household i in country q as the following:14

where LIQiq are the i-th household liquid assets in country q. Liquid assets are defined as the sum of

us

deposits, money invested in mutual funds, bonds, shares and managed accounts, the value of nonself-employment private business and other financial assets, such as derivative products. Finally,

M q stands for the number of months in which the negative margin is covered by liquid assets and is

an

allowed to vary from country to country.

A household is considered to be in distress ( qi  1 ) if its financial margin is negative and the sum of the household’s negative flow of financial margin for a determined number of months is greater than the household’s liquid assets. If any of these conditions do not hold, the household is not in

M

distress ( qi  0 ).

d

We need to determine the exact number of months to be used as threshold for comparing the flow of negative financial margins with liquid assets. This parameter will be calibrated in section 3.3.

13 14

Ac ce p

te

It is important to note that we only consider households as being in distress if they are unable to pay their debts, i.e., do not have enough income to cover their spending and do not have sufficient liquid assets. Given the available data, we do not and cannot consider households that are able but unwilling to service their debts. Issues such as strategic defaults are beyond the scope of this paper.15

 stands for δυστυχία, which means distress, misfortune or adversity in Greek. Note that these two conditions for distress can be rewritten as one sufficient condition as follows. A household is in distress if the following two conditions hold:  FM iq  I iq  Ti q  DPi q  BLC q  0   LIQiq q q q q FM M LIQ FM | | *     i i i  Mq  where the second equation can be rewritten because that condition is only relevant if the financial margin is negative. Combining the two gives the following:

I iq  Ti q  DPi q  BLC q   15

LIQiq Mq

, or equivalently,

DPi q  I iq  Ti q  BLC q 

LIQiq Mq

.

Given that a full recourse system is in place in all euro area countries, i.e. banks have full recourse on all household assets, the likelihood of strategic default is much lower than in the U.S,, where a household has the right to cancel a mortgage by returning the collateral if the value of the collateral is lower than the current amount of debt.

10 Page 10 of 28

To conclude, we consider a household to be financially vulnerable if it has a negative financial margin and the negative monthly cash flow for a specific time period in the future cannot be covered by liquid assets.

ip t

3.3 Calibration

M

an

us

cr

If our measure of vulnerability fairly reflects the household’s probability of default (PD) on its debt, it should reflect the data on defaults at the macro level. Because data on household sector PDs are not available across all euro area countries, we use the ratio of non-performing loans to total loans (NPL ratio) as a proxy16. In particular, we are interested in defaults on loans to households. Hence, we use the NPL ratios in this particular loan market segment. We use publicly available data from the 2015 EU-wide transparency exercise of the EBA.17 It is the only data source of harmonized NPLs for mortgage loans across European countries. Harmonized NPL ratios were first published in the 2015 EBA transparency exercise and are only available for December 2014 and June 2015. Because we need a harmonized NPL ratio for mortgage loans for the HFCS reference year, i.e. 2009 or 2010, we reconstruct those ratios. To this end, we assume that the growth rate of the NPLs between the HFCS reference year and June 2015 using the national, rather than harmonized, definition is equal to the growth rate of the harmonized NPLs. To establish a link between the micro data on defaults as we define them ( qi ) and macro data (

), we follow a two-step approach. First, we define two additional measures that aim to

te

d

capture the gross and net amount of debt of distressed households in the HFCS data, namely the Exposure at Default (EAD) for the entire debt of distressed households and Loss Given Default (LGD) for the entire net debt, i.e., after deducting the value of the collateral. In our context, both these measures should be understood as expected values as each distressed household defaults at a specific probability, which we implicitly determine by calibrating the measure. In the second step,

Ac ce p

we calibrate the remaining parameter of the distress statistic, namely, the number of months, M q , of paying the negative financial margin out of the stock of the liquid assets (see section 3.2); thus, .18 By EAD expressed as a fraction of the debt in distress over total debt for country q mimics determining M q , we implicitly determine a probability distribution of default. Households with no liquid assets will default at a probability of one, while households with liquid assets greater than

M q multiplied by its negative financial margin are not in distress and, therefore, their probability of default equals zero. All households in between these two polar cases are assigned a probability of default ( ) based on a linear function determined by those two cases. We define EAD as follows: 16

A priori, other measures can also be considered for this purpose. For example, the rates of personal bankruptcy can provide a benchmark that contains information about households that are already in default. However, due to the vastly different bankruptcy laws prevailing in each country (some countries in the euro area do not even have such a law in place) and, consequently, a lack of data on personal bankruptcies, it is not possible to conduct a harmonized analysis based on this measure. 17 2015 EU-wide transparency exercise results, European Banking Authority, http://www.eba.europa.eu/risk-analysis-anddata/eu-wide-transparency-exercise/2015/results 18 We could also compare LGD with actual loan losses that banks face, i.e. after deducting collateral. However, on account of the lack of data on loan loss provisions on household debt by country, we leave this exercise for the future.

11 Page 11 of 28

(eq. 4) where

is the total debt of household i in country q and

is the probability that the distressed

household defaults. EAD reflects the “expected” amount of debt held by financially vulnerable households as a percentage of the debt held by all households in the country.

ip t

We define Loss Given Default (LGD) as follows:

where

cr

(eq. 5)

stands for the assets that the bank can recover in case of a default of household i.

is a

us

binary indicator that takes a value of 1 if the debt of household i is higher than the assets that can be recovered and 0 otherwise. LGD measures the potential losses to the banks coming from the household sector.

an

We calibrate the parameter M q by minimizing the absolute difference between EAD and NPL for each country in our sample.19

d

te

Months 25.6 1.3 0 14.6 13.0 7.5 3.9 2.4 8.6 0

Ac ce p

Countries Austria Belgium Cyprus France Germany Greece Italy Portugal Slovakia Spain

M

Table 2 Number of months of negative financial margin covered by liquid assets (M)

Sources: HFCS & own calculations.

Table 2 shows the results of this calibration. There is substantial heterogeneity across countries, from 0 months in Spain and Cyprus to more than 25 months in Austria. For most countries, this number is closer to 10 months. This may reflect the differences between the countries in terms of the willingness to utilize savings for the repayment of debt, which in turn may be linked to the extent to which banks are able, in practice, to collect debts from defaulted households. In addition, family transfers to cover the negative financial margin gap could also play a role in these differences. Table 3 shows the NPL ratio in column 1 and the EAD in column 2 for each country, where the EAD follows from the definition of our metric as obtained in the calibration. We are able to match most 19

We exclude Slovenia, Malta, Luxembourg and the Netherlands from the calibration. Those countries have relatively low sample sizes for the purposes of this calibration. For example, although Luxembourg has 580 indebted households in the HFCS, it has an NPL of only 0.3%. This would mean there are about 2 vulnerable households in the HFCS in Luxembourg, which would make the calibration too sensitive to sampling errors.

12 Page 12 of 28

of the countries perfectly; only Cyprus and Spain show a divergence between the two numbers (this is due to the fact that M q cannot be negative). This may suggest that for Cyprus and Spain, the vulnerabilities in the household sector were already present at the time the HFCS was conducted but were not yet reflected in NPL ratios.20

ip t

4. Stress Testing Euro Area Households

us

cr

Having established a measure of household distress, in this section, we present the results of stress tests, which aim at assessing cross-country sensitivity to various types of adverse shocks. We begin with analyzing the impact of an interest rate shock, an asset price shock and an income shock on the percentage of distressed households. We then turn to quantifying the risk that these households pose to the financial system by calculating the changes in EAD and LGD following the shocks. We also consider some combinations of these shocks.

te

d

M

an

For the interest rate shock, we consider hypothetical increases in the interest rate ranging from 100 to 300 basis points. We chose the level of the shocks such that the most adverse scenario is equivalent to the reduction of interest rates carried out by the ECB between October 2008 and mid2010; hence, the shock reflects the transition back to the pre-crisis interest rate level. The change in the interest rate affects our metric of distress via two channels — first, through the increase of debt payments, and second, through the increase of financial income received from interest-paying accounts. We test various levels of an interest rate shock to unveil possible non-linearities in response to an interest rate shock. Furthermore, we chose the shocks in an arbitrary manner rather than based on the standard deviations due to the very small variation in the interest rates in recent years. Nevertheless, the magnitude of the interest rate shock is in line with that used in the literature. 21

Ac ce p

Considering the first channel, the pass-through of official rates to the lending rates needs to be assessed. This, however, is a challenging task, especially in a cross-sectional dimension of countries with different financial products and different banking practices. In particular, the pass-through depends on the conditions of the debt contract, namely, whether the loan is subject to an adjustable or fixed rate. Furthermore, the bank practices regarding the pass-through of interest rates on

20 21

NPL ratios are lagging indicators of default. The level is also in the range used in other stress testing exercises — for instance, IMF (2012), Albacete and Lindner (2013) or BdE economic bulletin (2011).

13 Page 13 of 28

ip t cr

Table 3 Exposure at default for different metrics of distress

DSI > 70% (5) 9.8%

us

DSI > 35% (4) 19.8%

NPL (1) 5.1%

Belgium

4.2%

4.2%

62.7%

51.2%

15.0%

15.6%

19.4%

23.0%

4.8%

Cyprus

7.2%

9.1%

74.3%

65.6%

28.5%

25.5%

29.3%

32.9%

29.9%

France

3.5%

3.5%

55.3%

42.2%

8.2%

6.6%

9.3%

12.4%

Germany

3.9%

3.9%

50.3%

39.7%

13.0%

9.3%

11.3%

16.9%

8.6%

Greece

9.3%

9.3%

40.9%

32.8%

6.7%

7.5%

9.1%

14.7%

12.6%

Italy

7.2%

7.2%

42.2%

33.7%

11.4%

12.0%

16.1%

18.0%

Portugal

3.8%

3.8%

43.9%

34.8%

8.8%

12.1%

16.0%

Slovakia

5.0%

5.0%

48.4%

39.3%

6.2%

6.0%

9.0%

14.4%

6.3%

Spain

2.9%

4.3%

69.5%

61.3%

22.5%

18.8%

23.0%

30.4%

18.6%

4.4%

52.5%

42.8%

14.1%

12.3%

15.4%

19.9%

ed

M an

Countries Austria

ce pt

9.8%

Inability to meet expenses (9) 11.5%

7.2%

Behind with debt payment last 12 months (10)

Behind with debt payments now (11)

11.9%

5.1%

21.7%

Sources: HFCS, EBA, various Central Banks, IMF& own calculations. Notes: the NPLs are the non-performing loan ratios of the household sector constructed using the harmonized NPL ratios for June 2015 as published by the EBA and assuming that the growth rate of the NPLs between the HFCS reference year and June 2015 using the national, not harmonized, definition is equal to the growth rate of harmonized NPLs. The EBA did not publish harmonized data for Greece and Slovakia, therefore the NPL for those comes directly from their respective Central Banks, DSI stands for debt-service-to-income ratio, BLC stands for basic living costs as a percentage of median income. The last 3 columns follow from specific questions asked in the HFCS. The column inability to meet expenses considers households with expenses above income, and who finance this by means of a loan, asking help from friends or relatives or leaving some bills unpaid, as in distress.

Ac

Total

DSI > 30% (3) 24.7%

Negative margin (BLC = 20%) (6) 11.1%

Exposure at Default Negative Negative margin (BLC margin (BLC = 30%) = 40%) (7) (8) 12.8% 15.2%

Fin. margin + liq. Assets (our metric) (2) 5.1%

14

Page 14 of 28

ip t

existing adjustable-rate loans also differ across countries. However, for indexed loans, the contractual interest rate is constructed using a reference rate, typically the EURIBOR plus a margin. For renewable mortgages, the interest rate can change at the banks’ discretion. Indexed adjustablerate mortgages dominated in Europe after 2000 (Dübel and Rothemund, 2011). Furthermore, in many European countries, the legislation requires lenders to pass decreases in interest rates onto the consumer even in the case of renewable mortgages, turning it essentially into an indexed loan (Dübel and Rothemund, 2011).

us

cr

Taking all this into account and, importantly, to ensure the cross-country consistency in interpreting the results, we assume a 100% pass-through of the official interest rate to the individual loan rate for adjustable-rate loans22. Conversely, fixed interest rate loan contracts are not affected by the interest rate shock. Note that we only have information on the type of loan (fixed vs. adjustablerate) for loans linked to the household’s main residence and other real estate property. Nevertheless, these two types of loans account for more than 80% of the total debt for the whole sample23. We treat all non-collateralized loans as if they were adjustable-rate loans.24

M

an

Regarding the second channel, we believe that the change in the interest rate partly translates into the interest rate paid on sight accounts and savings accounts following the pass-through rates reported by Kleimeier and Sander (2006). We also assume that all these accounts are interest bearing.

22

Ac ce p

te

d

Aside from the interest rate shock, we consider unemployment shock and asset price shock a 1- and 2-standard deviation increase in the unemployment rate and decrease in housing prices, respectively. The standard deviations are based on quarterly data for the unemployment rate and a housing price index constructed by Eurostat. For the unemployment rate, we use data from 2002 until the first quarter of 2015. For the housing price index, we use all data available, which is usually from 2005 until 2014.25 Table 4 shows the distribution of the standard deviations across countries. There are substantial differences across countries: in Belgium, the standard deviation of the unemployment rate is only 0.5 percentage points, while in Greece, it is 7.2 percentage points. For the housing price index, the standard deviation for Spain is more than three times higher than the standard deviation for France. We use the standard deviations as shocks, such that the probability of the occurrence of the adverse scenarios is about the same for all countries. For the interest rate, we

An important factor that may anchor interest rates, even for adjustable-rate mortgages, are the caps on the maximum change in the interest rate. In most European countries, such caps apply to less than 5% of outstanding adjustable-rate loans (ECB, 2009). However, in Belgium, 34% and as much as 50% of the outstanding adjustable-rate loans in France are subject to such caps. For loans subject to the cap in France, the interest rates typically cannot increase by more than 2 p.p. over the initial rate (Dübel and Rothemund, 2011), while in Belgium, the interest rates cannot deviate from the initial rate by more than 2 p.p. in the first three years of the contact (iff/ZEW, 2010). However, given that the HFCS was mainly conducted in 2010, just after a period in which the official interest rates had declined substantially, it is unlikely that the caps would be binding under the interest rate shock scenario. 23 In some cases, the respondent does not know whether the household has a fixed or adjustable rate mortgage. We treat these loans as if the proportion of adjustable rate loans to total loans is the same as the loans for which we have information. 24 As a robustness check, we consider what happens when all non-collateralised loans are treated as fixed-rate loans. We find that the 300-basis point shock would increase the proportion of indebted households in distress, the EAD and the LGD all by 0.1 percentage point less compared with the increase if treating non-collateralised loans as adjustable-rate loans. 25 For France and Slovakia, the data is available from 2006 onward; for Portugal, from 2008 onward; for Italy, from 2010 onward; and for Greece, from 2006 until 2011.

15 Page 15 of 28

an

us

cr

Table 4 Standard deviation unemployment rate and house price indicator Standard deviation Standard deviation unemployment housing price index Countries rate (p.p.) (%) Austria 0.5 9.1% Belgium 0.5 9.0% Cyprus 4.6 9.6% France 0.8 3.8% Germany 2.1 4.8% Greece 7.2 3.8% Italy 2.0 5.0% Portugal 3.2 5.6% Slovakia 2.7 10.7% Spain 6.6 14.3%

ip t

use the same shocks across all countries because they all share the same monetary policy. In contrast to the interest rate and unemployment shock, the asset price shock does not affect the metric of distress but will affect the LGD as it alters the value of the collateral to be seized by the bank.

te

d

M

Source: Eurostat Notes: The standard deviations are based on quarterly unemployment rate data and on a housing price index constructed by Eurostat. For the housing price index data is used from 2005 until 2014, however, for France and Slovakia the data is available from 2006 onwards, for Portugal from 2008 onwards, for Italy from 2010 onwards, and for Greece from 2006 until 2011. For the unemployment rate, data from 2002 until the first quarter of 2015 is used.

Ac ce p

Table 5 reports the changes in the number of distressed households after the shocks as compared to the baseline across countries. In general, the 300-basis point interest rate shock has a more adverse impact than the 2-standard deviation income shock, but not for all countries. For instance, in Greece, the number of distressed households increases by 10.9% after the interest rate shock, while the income shock leads to an increase by 50.9%. To the contrary, in Portugal, the number of households in distress increases by 4.8% after the income shock, but the interest rate shock leads to an increase in the number of households in distress by 30.1%. Greece seems to be the most vulnerable country to adverse shocks: for instance, a combined interest and income shock would lead to an increase in the proportion of households in distress by 63.0% compared to the baseline. These figures are mostly a reflection of the restrictiveness of past bank lending policies and/or household risk aversion, which ultimately determines the effective buffer of income and liquid assets of indebted households. We turn now to the analysis of the impact of the shocks on the EAD, which acts as a yardstick for NPLs. Table 6 shows the relative increase in the EAD after the shocks as compared to the baseline. Here, the impact of the interest rate shock is also higher than the impact of the income shock, except for Greece. In addition, there is substantial heterogeneity across countries with the impact of both shocks. An increase in the interest rate by 300 basis points would lead to an increase in the stock of non-performing loans by around 50% in Austria, France and Portugal, while in Belgium and Germany, it would only increase by about 15%. Although the effect of the income shocks is generally 16 Page 16 of 28

ip t

contained, it is still meaningful for some countries, particularly in Greece, where the EAD increases by 45.0% after the shock of 2 standard deviations. A combined interest and income shock increases the stock of non-performing loans by about 50% or more in Austria, France, Greece, Portugal and Slovakia, with the strongest impact in Greece and Portugal. It is important to note that the debt distribution plays a role in determining the vulnerability to shock. All in all, intuitively, the interest rate shock has the strongest impact in countries where adjustable rate mortgages are prevalent, countries where households hold relatively more debt and countries with a high standard deviation in the unemployment rate and housing price index.

Ac ce p

te

d

M

an

us

cr

To analyze the potential credit losses that banks may incur in the aftermath of household defaults, we calculate the impact of the shocks on the LGD. In particular, it provides useful insight into potential threats to the banks that are posed by a housing price shock. Because our metric of distress does not account for real assets, the shock to housing prices does not affect the percentage of households in distress or the EAD. However, it does affect the LGD as this measure does account for the collateral. Table 7 presents the changes in the LGD assuming that in case of a default, banks can seize liquid assets and the house if the household has a mortgage. Evidently, the housing price shock impacts the level of LGDs substantially. In Belgium, the greatest impact — a 2-standard deviations decline in housing prices — led to a 73% increase in LGD compared to the baseline. This is an indication that for many Belgian borrowers already in the baseline, the value of their house is only slightly above their debt level (or the loan-to-value ratio is slightly below one). As a result, many households fall ‘under water’ in the aftermath of the shock. Turning to the impact of the combined shock — i.e., the interest, income and housing price shock — on losses facing banks, a huge crosscountry divergence is noticeable. Losses for French banks incurred in the household sector would increase by 123.9%, whereas they would only increase by 26.5% for German banks.

17 Page 17 of 28

ip t cr

Belgium 7.5%

Cyprus 5.4%

France Germany 8.7% 9.1%

Greece 7.8%

Italy 8.9%

Portugal 5.7%

Slovakia 8.3%

Spain 3.2%

Total 7.8%

100 bps increase

7.1% [5.7%]

7.6% [1.5%]

5.9% [7.9%]

8.8% [1.3%]

9.2% [1.2%]

8.1% [3.2%]

8.9% [0.7%]

6.2% [8.3%]

8.7% [4.6%]

3.4% [6.8%]

7.9% [1.9%]

200 bps increase

7.2% [6.8%]

8.0% [6.5%]

6.1% [11.9%]

9.1% [4.6%]

9.3% [1.6%]

8.4% [6.9%]

9.1% [2.3%]

6.7% [17.8%]

8.7% [5.3%]

3.8% [18.7%]

8.1% [4.4%]

300 bps increase

7.3% [8.5%]

8.1% [8.6%]

7.0% [28.3%]

9.4% [7.6%]

9.5% [3.8%]

8.7% [10.9%]

9.4% [5.7%]

7.5% [30.1%]

8.8% [6.9%]

3.8% [18.8%]

8.3% [7.2%]

Employment shock 1 s.d. increase unemployment rate 2 s.d. increase unemployment rate Combined shock

6.8% [1.2%] 6.8% [2.0%]

7.6% [1.2%] 7.6% [1.9%]

5.8% [6.0%] 6.1% [13.0%]

8.9% [1.7%] 9.0% [3.5%]

9.3% [2.1%] 9.6% [5.0%]

9.7% [23.2%] 11.8% [50.9%]

9.0% [1.7%] 9.1% [2.9%]

5.8% [2.1%] 6.0% [4.8%]

8.7% [4.8%] 9.1% [9.7%]

3.4% [6.1%] 3.6% [11.8%]

8.0% [2.8%] 8.2% [5.9%]

300 bps increase + 2 s.d. increase unemployment rate

7.4% [10.6%]

9.9% [8.7%]

12.8% [63.0%]

9.7% [8.7%]

7.9% [37.3%]

9.7% [17.3%]

4.2% [30.3%]

8.8% [13.2%]

ed

M an

Austria 6.7%

ce pt

Country Baseline Interest rate shock

8.3% [10.3%]

7.7% [41.2%]

us

Table 5 Effect of shocks on the percentage of indebted households having a positive probability of default (absolute levels and relative changes)

9.7% [11.2%]

Ac

Sources: HFCS & own calculations. Notes: The numbers in brackets are the relative changes compared to the baseline.

18

Page 18 of 28

ip t cr

Austria 5.1%

Belgium 4.2%

Cyprus 9.1%

France 3.5%

Germany 3.9%

6.0% [18.4%] 6.9% [34.7%] 7.6% [48.5%]

4.3% [2.8%] 4.6% [9.5%] 4.9% [16.9%]

9.9% [8.7%] 10.3% [12.8%] 11.7% [28.8%]

3.6% [4.5%] 4.6% [31.0%] 5.1% [46.4%]

4.1% [6.3%] 4.2% [9.6%] 4.4% [14.4%]

1 s.d. increase unemployment rate

5.2% [1.3%]

4.2% [0.4%]

9.5% [4.3%]

3.5% [0.9%]

2 s.d. increase unemployment rate

5.2% [1.5%]

4.2% [0.9%]

10.0% [9.5%]

7.6% [50.1%]

4.9% [17.3%]

12.5% [36.9%]

100 bps increase 200 bps increase 300 bps increase

Slovakia 5.0%

Spain 4.3%

Total 4.4%

8.3% [14.0%] 8.7% [20.3%] 9.0% [23.7%]

4.4% [17.2%] 5.0% [33.3%] 5.9% [56.7%]

5.4% [8.7%] 5.8% [16.3%] 6.1% [23.0%]

4.5% [4.5%] 5.3% [22.4%] 5.3% [22.8%]

4.7% [7.4%] 5.2% [19.2%] 5.5% [26.0%]

3.9% [1.8%]

11.0% [18.2%]

7.3% [0.9%]

3.9% [2.4%]

5.4% [7.1%]

4.4% [2.3%]

4.5% [2.1%]

3.5% [2.0%]

4.0% [4.7%]

13.5% [45.0%]

7.4% [2.4%]

4.0% [5.6%]

5.8% [16.6%]

4.6% [6.1%]

4.6% [5.4%]

5.2% [48.8%]

4.6% [19.1%]

16.5% [77.1%]

9.3% [28.0%]

6.2% [64.1%]

7.2% [44.4%]

5.6% [28.7%]

5.7% [31.8%]

ed

ce pt

300 bps increase + 2 s.d. increase unemployment rate

Italy Portugal 7.2% 3.8%

10.2% [9.6%] 11.1% [19.3%] 12.0% [29.4%]

Employment shock

Combined shock

Greece 9.3%

M an

Country Baseline Interest rate shock

us

Table 6 Effect of shocks on the expected exposure at default (absolute levels and relative changes)

Ac

Sources: HFCS & own calculations. Notes: The numbers in brackets are the relative changes compared to the baseline.

19

Page 19 of 28

ip t cr

Country Baseline

us

Table 7 Effect of shocks on the expected loss given default (absolute levels and relative changes) Austria 2.73%

Belgium 0.36%

Cyprus 1.39%

France 0.87%

Germany 1.34%

Greece 2.46%

Italy 1.93%

Portugal 0.59%

Slovakia 0.73%

Spain 0.51%

Total 1.12%

3.15% [15.4%] 3.51% [28.8%] 3.77% [38.1%]

0.36% [1.1%] 0.39% [8.3%] 0.39% [9.2%]

1.41% [1.0%] 1.41% [1.0%] 1.41% [1.4%]

0.93% [6.6%] 1.71% [95.7%] 1.91% [119.3%]

1.51% [12.0%] 1.54% [14.5%] 1.57% [16.4%]

2.48% [0.6%] 2.50% [1.4%] 2.73% [10.7%]

2.51% [30.3%] 2.55% [32.3%] 2.64% [36.8%]

0.63% [6.1%] 0.70% [17.8%] 0.79% [32.6%]

0.74% [1.4%] 0.75% [2.5%] 0.82% [12.9%]

0.53% [4.1%] 0.65% [28.6%] 0.65% [28.6%]

1.26% [12.7%] 1.50% [33.4%] 1.57% [40.3%]

1 s.d. increase unemployment rate

2.79% [2.3%]

0.36% [0.5%]

1.41% [1.4%]

0.88% [0.9%]

1.36% [1.5%]

2.72% [10.4%]

1.94% [0.5%]

0.60% [1.8%]

0.77% [5.6%]

0.54% [6.1%]

1.14% [1.9%]

2 s.d. increase unemployment rate Housing price shock

2.80% [2.6%]

0.36% [0.8%]

2.85% [4.5%] 2.98% [9.1%]

0.47% [31.5%] 0.62% [73.0%]

3.05% [11.7%]

0.62% [74.2%]

100 bps increase 200 bps increase 300 bps increase

2 s.d. decline

0.89% [1.7%]

1.40% [4.2%]

3.03% [23.1%]

2.00% [4.0%]

0.62% [4.3%]

0.81% [10.7%]

0.56% [11.3%]

1.18% [4.8%]

1.64% [17.2%] 1.95% [40.0%]

0.88% [0.6%] 0.88% [1.2%]

1.38% [2.4%] 1.41% [4.7%]

2.53% [2.6%] 2.60% [5.5%]

1.98% [2.9%] 2.05% [6.5%]

0.62% [5.0%] 0.67% [13.4%]

0.73% [0.0%] 0.76% [4.7%]

0.54% [7.0%] 0.63% [24.3%]

1.16% [3.3%] 1.21% [7.8%]

2.00% [43.5%]

0.90% [2.9%]

1.47% [9.3%]

3.22% [30.6%]

2.13% [10.4%]

0.71% [18.9%]

0.85% [17.0%]

0.72% [41.1%]

1.27% [13.4%]

Ac

Combined shocks 2 s.d. increase unemployment rate + 2 s.d. decline house price

1.43% [2.7%]

ce pt

1 s.d decline

ed

Employment shock

M an

Interest rate shock

300 bps increase + 2 s.d. 4.24% 0.68% 2.10% 1.95% 1.70% 3.55% 2.80% 0.97% 0.96% 1.02% 1.77% increase unemployment [55.3%] [90.8%] [50.4%] [123.9%] [26.5%] [44.2%] [45.4%] [63.0%] [31.4%] [101.2%] [57.5%] rate + 2 s.d. decline house price Sources: HFCS & own calculations. Notes: The numbers in brackets are the relative changes compared to the baseline. The bank is assumed to recover liquid assets plus the value of the house if the household has a mortgage.

20

Page 20 of 28

us

cr

ip t

Figure 1 Effects of the shocks on expected loss given default

te

d

M

an

Sources: HFCS & own calculations. Notes: the interest rate shock is a 300 basis points increase in the interest rate. The income shock is defined as a 2 standard deviations increase in the unemployment rate. Those who lose their job are assumed to receive unemployment benefits. The housing price shock is a 2 standard deviations decline of the value of real estate. The graph shows three different estimates based on different assumptions on which assets the bank can recover in case of a default. The lower end of the line is the loss given default if the bank can recover all assets the household has. The diamond indicates the loss given default if the bank is assumed to recover the liquid assets plus the value of the collateral if the household has a mortgage. The top end of the line is based on these same assumptions plus, here, the value of real estate is downgraded by 20% to account for the tendency that forced sales lead to a price lower than the value.

Ac ce p

Figure 1 combines the effects of the combined shock on all three measures of distress. Furthermore, because there are substantial differences across jurisdictions in the scope of the assets that can be seized by banks in the case of a default, we show a range of possible losses that the banks could face. To this end, we calculate three levels of LGDs depending on the scope of the assets that can be recovered by banks. We consider (1) all assets, (2) liquid assets plus the value of the real estate for households with a mortgage, and (3) liquid assets plus the value of the real estate after a haircut26. In France and Spain, there is little difference between the three variations of the LGD. In particular, there is hardly any effect of the introduction of a haircut on the value of the collateral. This suggests that the debts of Spanish and the French households are sufficiently covered by assets. To the contrary, the additional haircut on the value of the collateral may significantly increase losses faced by banks in Greece and Cyprus.

5. Robustness Analysis

26

The haircut reflects the usual liquidity premium in the case of a forced sale. We devalue the property by 20%. This is slightly less conservative than the 27% haircut reported by Campbell, Giglio and Pathak (2011) for forced sales in Massachusetts, USA.

21 Page 21 of 28

5.1. HFCS simulation

an

us

cr

ip t

To assess the performance of the proposed measure, we conduct a simulation exercise. Because only one wave of the HFCS is available at present, we update the HFCS data using macro data (following Ampudia et al. (2014)) and compare the resulting LGD with information on Loan Loss Provisions embedded in the aggregate data. We simulate changes in LGD across nine euro area countries such that the shocks to interest rates, unemployment and housing prices mimic the actual changes in these variables that occurred between the reference periods of the survey and the last quarter of 2012. The shocks applied are reported in Table 8. We compare the changes in LGD resulting from these shocks to the changes between the reference period of the HFCS and the last quarter of 2012 in the stock of Loan Loss Provisions (LLPs), which measure the credit losses that banks incur on their loan exposures (net of collateral) after the default occurs. Note that we should already expect differences a priori between these two measures. First, LLPs incur losses — i.e., according to the accounting rules, banks may only book losses if there is clear evidence of impairment. Second, the timing may also be an issue. Supervisors usually define the evidence of impairment as a delay in repayment by 90 days and above, but in some countries, this threshold is different (which can be up to 180 days).

te

d

M

Figure 2, which compares changes in LGDs and LLPs as explained above, shows a strong correlation between both measures. For all countries except Italy, the change in LGD is lower than the change in LLP, which can be explained: during the 2010-2012 period, the provisions continued trending upward, while expected loss already stabilized (ECB, 2014). In the case of Italy, a quite conservative provisioning policy of the national supervisor allowed Italian banks to build up a substantial stock of provisions for mortgage loans before the crisis27, which eliminated the need for a steep increase in the provisioning between 2010 and 2012. Nevertheless, provisions increased further in 2013 and 2014.

Ac ce p

Table 8 Evolution of macro variables from HFCS reference period to Q4 2012 Interest rate Unemployment (change in Housing prices rate (change in Country basis points) (change in %) percentage points) Austria -9.66 13.99 0.28 Belgium 13.04 5.68 -0.79 Cyprus 48.28 -10.54 7.47 France -54.39 9.18 0.82 Germany -100.72 5.12 -0.69 Greece -95.35 -23.84 16.33 Italy 73.54 -5.83 3.14 Portugal 138.66 -2.98 5.08 Slovakia -32.12 -4.36 0.48 Spain -13.89 -25.31 8.84 Sources: National Accounts and Eurostat. Notes: Interest rate refers to the house purchase interest rate (calculated by weighting volumes by borrowing purposes). Housing price changes are calculated based on national house price indices.

27

Among the countries considered in this comparison, Italy had the largest ratio of LLPs to total loans, both in 2009 and 2010.

22 Page 22 of 28

an

us

cr

ip t

Figure 2 LGD vs LLP comparison

d

M

Sources: HFCS, ECB consolidated banking data and own calculations. Notes: Changes in LGD are based on shocks reported in Table 8. Changes in LLP are changes in the stock of loan loss provisions between the reference period of the survey lagged by six months and end-2012. The solid line shows the linear best fit.

te

5.2 Comparison with other metrics of distress

Ac ce p

In this subsection, we benchmark our metric of distress against other possible measures that could be obtained from the HFCS. In the existing literature, the two most common ways of identifying households as being in distress are the debt-service-to-income ratio and the negative financial margin. Table 3 shows the respective EADs resulting from these different metrics of distress as well as the NPL ratio and the EAD related to our proposed metric of distress. Columns 3 and 4, respectively, show the EAD if the critical threshold for the debt-service-to-income ratio is 30%, as in Tiongson et al. (2009) and Michelangeli and Pietrunti (2014) and 35% as in Dey, Djoudad and Terajima (2008). Both measures lead to a much higher EAD compared with the EAD from our metric. The estimated EAD for both is more than 10 times higher than the NPL ratio. Even an increase to the critical threshold of 70% (column 5) leads to a substantially higher EAD. Columns 6 to 8 show how the EAD would look if we instead use a negative financial margin as a metric of distress for different values of the basic living costs. These metrics perform better than the debt-service-to-income ratio in terms of the EAD matching the NPL ratio but are still quite far off compared with the results from our metric. This is a strong indication that the current cash flow and the stock of liquid assets determines household default. Another set of robustness checks we conduct is comparing our metric to the alternative metric of distress that we construct based on the following set of questions from the HFCS. First, a question on whether the household’s regular expenses were higher than the household’s income during the previous 12 months, the results were that the expenses were just about the same or lower than its 23 Page 23 of 28

us

cr

ip t

income. The households that reported spending more than their income were asked what they did to cover those expenses.28 We run a regression of late payments on a series of household characteristics29. The results show that having credit card debt, an overdraft facility or another collateralized loan significantly increases the likelihood of making late payments. Therefore, we assume that someone who finances extra expenses with one of these instruments is in distress. Furthermore, we assume that a household that leaves bills unpaid or asks friends and relatives for support is also likely to be in distress. Therefore, we also included them in the alternative metric. Note that the meaning of these two questions overlaps to a large extent with the two dimensions of our metric, namely, the solvency and liquidity. The first question addresses the issue of the solvency position of the household while the second question considers the liquidity dimension. We use these two questions to construct alternative measures of distress and reported the resulting EAD in Table 2, column 9.30 The EAD that is based on the inability to meet expenses tends to be higher than the EAD calculated using our metric and is higher than the NPL ratio for all countries. For Cyprus, Germany and Spain, the discrepancy between the two metrics is large.

te

d

M

an

For some countries in the HFCS, we can also compare our measure of distress with information on late or missed debt payments. In particular, households in Spain and Portugal were asked if they were late with or missed debt payments over the previous 12 months. Column 10 shows the EADs for these households. Finally, for Portugal, an additional follow-up question was asked — whether the household was, at the time of the survey behind in its debt payments — for which the resulting EAD is shown in column 11. The EAD based on late or skipped payments in the previous 12 months overestimates the NPL ratios because the NPL comprises the debt of households that were more than 3 months delayed with their payments and does not consider those who were behind with installments at a certain point over the last 12 months but are on time with their debt payments at the moment. In this regard, the EAD based on the question asked in Portugal as to whether a household is behind with its payments at the moment is only slightly higher than the NPL ratio in that country.

Ac ce p

All in all, it seems that our measure of distress is a useful micro-based yardstick of aggregate data on NPLs. Using alternative definitions of distress leads to substantially different levels of EADs, which perform worse at this matching exercise.

6. Mitigating Defaults by Loan Restructuring

While modelling behavioral reactions of the households and the banks to the shocks is outside the scope of this paper, this section offers some insight into the effects of loan restructuring on LGDs. In case a household is in distress and therefore the likelihood of default on a loan is high, as a mitigating measure, a bank can decide to restructure the loan. Such a restructuring aims at the 28

The possible answers include selling assets, obtaining a credit card / overdraft coverage, obtaining another loan, using savings, asking for help from relatives or friends, leaving some bills unpaid, or other methods. 29 The results are not reported but are available from the authors upon request. The regression is based on observations from Spain and Portugal. 30 For France and Italy, this measure cannot be calculated because the question comparing income and expenses was not part of the questionnaire in France, while in Italy, the question on the source of money to meet the higher expenses was not part of the questionnaire.

24 Page 24 of 28

M

an

us

cr

ip t

reduction of the debt-service burden by the extension of the maturity of the loan and/or the level of interest rate or, rarely, the debt level. In this way, the restructuring of a loan may prevent the household from actual default. Restructurings affect bank income in two opposing ways. While they reduce the income stream from interest, they also increase the likelihood that the whole loan will be repaid. Figure 3 shows the baseline LGD (before the shocks), the LGD after the joint 300 basis point interest rate shock, and the 2-standard deviations income shock for each country. For the latter shock, we show how it is affected by loan restructurings, where restructuring is defined as a percentage decrease in monthly debt service expenses. Such a reduction increases the household financial margin; therefore, some households can avoid distress after the loan restructuring. Hence, the higher the reduction in the monthly installment, the lower the LGD is. For each country, the figure shows where the two lines cross, at which level the reduction in the monthly installment of the LGD after the shocks equalizes with the LGD in the baseline. This shows the amount of restructuring effort that would be required from banks to mitigate adverse shocks. There is substantial heterogeneity across countries, with levels ranging from less than 20% in reduction in the monthly installments in Germany to 40% in Spain. This means that in France, where the LGD doubles in the aftermath of the shocks, a 20% reduction in the installment would be sufficient to eliminate the effects of the shocks. Figure 3 shows that the effects of loan restructurings are non-linear. This underscores the importance of the non-linear effects that arise from the distribution of debt burden across households in assessing the household sector credit risk. Figure 3 Expected LGD dependent on level of restructuring Austria 5%

Belgium 0.8%

d

0.7%

4%

0.6% 3%

te

0.5%

2%

0.4%

0

10 20 30 40 50

0

1.5%

0

Italy

3.5%

2.5%

3%

2%

1.25%

2.5%

1.5%

2%

Slovakia

10 20 30 40 50

Greece

1.5%

1.25%

0.5%

0

1.75%

0

1%

1.6%

3%

0 10 20 30 40 50

2%

1.8%

4%

1%

France 2.5%

2%

10 20 30 40 50

Ac ce p

Germany

2%

Cyprus

2.2%

Portugal 1.25% 1% 0.75%

1%

10 20 30 40 50

10 20 30 40 50

0.5%

0

10 20 30 40 50

0 10 20 30 40 50

Spain

1%

1%

0.75%

0.75%

0.5% 0.25%

0.5%

0 10 20 30 40 50

0 10 20 30 40 50

% restructuring Sources: HFCS & own calculations. Notes: The solid line shows how the LGD after the 2 standard deviations unemployment and 300 basis points interest rate shock is affected by loan restructurings. The x-axis shows the percentage reduction of the monthly debt service expenses. The dashed line shows the baseline LGD (before the shocks) and is not affected by restructuring. In addition, for the calculation of the LGD we assume a haircut on the collateral of 20%.

25 Page 25 of 28

7. Conclusions

us

cr

ip t

In this paper we propose a metric of household distress, which is constructed by combining the data on income, expenditure, assets, debt and collateral from the Household Finance and Consumption Survey, which enables us to conduct a consistent analysis across the different euro area countries. Our metric of household distress is a micro-level yardstick of default and can be aggregated, for instance, at the country level to calculate credit risk indicators such as the Probability of Default (PD), Exposure at Default (EAD) or Loss Given Default (LGD) and their distributions. We demonstrate how these indicators can be calibrated using macro-data and are used for stress testing, for which a scenario can consist of an unemployment shock, interest rate shock, housing price shock or any combination of these.

Ac ce p

te

d

M

an

The use of micro-level data for the purposes of measuring household credit risk could shed more light on the vulnerabilities in this sector. We find that, overall, the risks posed by the household sector to the stability of the financial system in the euro area are generally contained. In the worst case scenario of a combined interest rate, income and housing price shock, the potential losses for the banking system are not higher than 4.3% of the total household debt in any euro area country. However, there is substantial heterogeneity across countries, and the relative impact of the shocks on bank losses is significant in many countries. Moreover, the impact of shocks depends also on the type of shock. In particular, countries where adjustable-rate mortgages predominate are more affected by an interest rate shock, while countries where households hold relatively more debt are generally more vulnerable to any type of shock. Nevertheless, one caveat requires due consideration: low LGDs as calculated using our metric depend on the value of the collateral. Hence, any factors hindering the seizure of the collateral or lowering its value, such as an inefficient legal system, moratoria on foreclosures, or deadlocks in the courts may increase losses to the banking sector. Overall, the effect of the shocks depend on both the household initial distribution of assets, liabilities and income, and the institutional factors prevailing in each country. For example, in the case of the housing price shock, countries with high initial LTV ratios are affected the greatest. We also show that the magnitude of restructuring loans to mitigate the impact of adverse shocks differs substantially across countries, which is a result of a combination of several factors: different propensities to shocks of various countries and varying distributions of financial margin across countries, including some non-linear effects in the tails of the distributions. Our results call for a systematic monitoring of the risks stemming from the household sector by the regulators. We propose a framework that could prove useful for this purpose. This requires the availability of harmonized data across the euro area, for which the HFCS is a big step forward. Nonetheless, there is still room for improvements to be made. For instance, in the HFCS, there is no explicit information on whether a household has defaulted on its debts or on the region of residence (leading to limitations in the estimation of basic living costs). Additionally, as more waves of the HFCS become available with time, a dynamic analysis can be performed, making use of the time series. This would open an important avenue for future research.

26 Page 26 of 28

References Albacete, N., and Lindner, P., 2013. “Household vulnerability in Austria – a microeconomic analysis based on the household finance and consumption survey”, Financial stability report 25 – June 2013, Oesterreichische Nationalbank (Austrian Central Bank), pages 57-73.

cr

ip t

Ampudia, M., Pavlickova, A., Slacalek, J. and Vogel, E., 2014. “Household Heterogeneity in the Euro Area Since the Onset of the Great Recession”, Journal of Policy Modeling (2015), DOI: 10.1016/j.jpolmod.2015.09.001.

us

Bičáková, A., Prelcová, R., Pašaličová, Z., 2010. “Who Borrows and Who May Not Repay”, CNB Working Paper, No. 10/2010. Brunetti, M., E. Giarda, and C. Torricelli, 2015. “Is Financial Fragility a Matter of Illiquidity? An appraisal for Italian Households”, Review of Income and Wealth, DOI: 10.1111/roiw.12189.

an

Campbell, J.Y., Giglio, S., and Pathak, P., 2011. "Forced Sales and House Prices", American Economic Review, American Economic Association, vol. 101(5), pages 2108-31.

M

Dey, S., R. Djoudad, and Y. Terajima, 2008. “A Tool for Assessing Financial Vulnerabilities in the Household Sector.” Bank of Canada Review, (Summer): 45−54.

d

Dübel, H., and Rothemund, M., 2011. “A New Mortgage Credit Regime for Europe – Setting the Right Priorities”, CEPS Special report.

te

ECB, 2009. “Housing Finance in the Euro Area”, Task Force of the Monetary Policy Committee of the European System of Central Banks, European Central Bank Occasional Paper, No. 101.

Ac ce p

ECB, 2014. “Financial Stability Review”, May 2014. European Commission 2011. “The measurement of extreme poverty in the European Union”. Holló, D., Papp, M. (2007): Assessing Household Credit Risk: Evidence from a Household Survey, Magyar Nemzeti Bank Occasional Papers, No. 70. Henry, J., and Kok, C., 2013. “A macro stress testing framework for assessing systemic risks in the banking sector”, ECB Occasional Paper, No. 152. Herrala, R., and Kauko, K., 2007. “Household Loan Loss Risks in Finland – Estimations and Simulations with Micro Data”, Bank of Finland Research Discussion Papers 5/2007. Household Finance and Consumption Network (2013a). “The Eurosystem Household Finance and Consumption Survey – Results from the First Wave”, ECB Statistics Paper No. 2. iff/ZEW, 2010. “Study on interest rate restrictions in the EU”, Final Report for the EU Commission DG Internal Market and Services, Project No. ETD/2009/IM/H3/87, Brussels/Hamburg/Mannheim. IMF, 2012. “Spain: vulnerabilities of private sector balance sheets and risks to the financial sector technical notes”, IMF Country Report, No. 12/140. 27 Page 27 of 28

Johansson, Martin and Mattias Persson, 2006, ―Swedish Households’ Indebtedness and Ability to Pay: A Household Level Study,‖Sveriges Riksbank Economic Review, Vol. 3: pp. 24– 40. Kleimeier, S., and Sander, H., 2006. “Expected versus unexpected monetary policy impulses and interest rate pass-through in euro-zone retail banking markets”, Journal of Banking & Finance, Vol 30, pages 1839-1870.

ip t

Laderchi, C.R., Saith, R., Stewart, F., 2003. "Does it Matter that we do not Agree on the Definition of Poverty? A Comparison of Four Approaches", Oxford Development Studies, Taylor & Francis Journals, vol. 31(3), pages 243-274.

cr

May, O. and Tudela, M., 2005. "When is mortgage indebtedness a financial burden to British households? A dynamic probit approach," Bank of England working papers 277, Bank of England.

us

Michelangeli, V., Pietrunti, M., 2014, A microsimulation model to evaluate Italian households’ financial vulnerability (with), The International Journal of Microsimulation, Volume 7(3) winter

an

Sugawara, N. and Zalduendo, J., 2011. “Stress-Testing Croatian Households with Debt-Implications for Financial Stability”, World Bank Policy Research Working Paper 5906.

M

Tiongson, Erwin, Sugawara, N., Sulla, V., Taylor, A., Gueorguieva, A., Levin, V. and Subbarao, K., 2009. “The Crisis Hits Home: Stress-Testing Households in Europe and Central Asia”, Washington, DC: The World Bank.

Ac ce p

te

d

Zajączkowski, S. and Żochowski, D., 2007. “The distribution and dispersion of debt burden ratios among households in Poland and its implications for financial stability”, a chapter in Proceedings of the IFC Conference on "Measuring the financial position of the household sector", Basel, 30-31 August 2006 - Volume 2, 2007, vol. 26, pp 62-74.

28 Page 28 of 28