How Micro Data Improve the Estimation of Household Credit Risk Within the Macro Stress Testing Framework

Abstract

Macro stress testing has become an increasingly important part of central banks’ and macroprudential authorities’ toolkits in the wake of the Global Financial Crisis. One of the most important parts of the stress testing framework is the estimation of credit risk losses under adverse circumstances. However, standard satellite models based on econometrics of time series may not be well suited to estimate credit risk losses for countries with short time series, incomplete credit cycle or structural breaks, as in the case of Slovakia. Incorporating micro data into the stress testing framework becomes important for such countries. In this paper, we show that using micro data leads to more plausible results compared to time series in form of higher estimated share of nonperforming loans. For the estimations, we use a unique set of individual retail loan data collected from banks operating in Slovakia. As the Slovak banking sector did not face a rapid worsening of the credit quality of retail loans during the Global Financial Crisis, the recent Covid-19 pandemic in 2020 is the only reference period of increased economic and financial tensions. This period also confirms our approach, as using micro data leads to nonperforming loan ratios that are closer to the ratio of indebted households asking for loan payment deferral during the pandemic.

Appendices

Appendix 1: Data Used for the Logit Regression

Variable	Possible states
Change in economic status	1: employed in quarter t-1 and t
	0: employed in quarter t-1, unemployed in quarter t
Sex	1: male
	2: female
Education	0: at most primary
	1: at most secondary
	2: tertiary
Marital status	1: single
	2: married
Type of activity	1: employed
	2: self-employed
Age: dummy	Up to 25 years
	Up to 35 years
	Up to 45 years
	Up to 55 years
	Up to 65 years
	More than 65 years

Appendix 2: Estimation of Households’ Credit Risk Using Time Series

Under the previous framework, quarterly time series of the NPL ratio for four types of loans—housing loans, consumer loans, current account overdrafts, and credit cards/other loans—were estimated using data since the beginning of 2006. Due to the short time series and the large number of possible explanatory variables, the Bayesian Model Averaging (BMA) method was used similarly as in Dees et al. (2017).

A standard ADL model structure is estimated:

$$\begin{aligned} \begin{aligned} \textit{Y}_t = \alpha + \rho _1\text {Y}_{t-1} + {...} + \rho _p\text {Y}_{t-p} + \sum _{i=1}^{K}\left( \beta _1^i X_{t-1}^i + {...} + \beta _q^i X_{t-q}^i \right) + \epsilon _t \end{aligned} \end{aligned}$$

(3)

In the below table we give a rough overview of the explanatory variables used for the regressions. At most 4 of the below variables entered the model. Variables closely related to each other were divided into several groups (e.g. interbank interest rates, GDP, GDP growth). At most 1 variable from each group entered the respective regressions.

Variables
Real GDP	1W EURIBOR
Nominal GDP	1M EURIBOR
Real GDP, annual growth	2M EURIBOR
Nominal GDP, annual growth	3M EURIBOR
HICP	6M EURIBOR
Unemployment rate	9M EURIBOR
Unemployment rate, annual growth	12M EURIBOR
Residentail real estate prices	2Y Slovak government bond yield
Flat prices	5Y Slovak government bond yield
RRE price, annual growth	10Y Slovak government bond yield
Flat price, annual growth	Slope, 10 vs 2 years Slovak government bonds
Average real wages	Slope, 10 vs 5 years SK government bonds
Average nominal wages	Slope, 5 vs 2 years SK government bonds
Average real wages, annual growth	Slope, 10Y SK government bonds vs 1Y EURIBOR
Average nominal wages, annual growth	Slope, 5Y SK government bonds vs 1Y EURIBOR
Loan interest rate	Credit-to-GDP gap
2 years average loan growth	Credit-to-GDP gap, alternative definition

Appendix 3: Loan Losses Under Different Assumptions as a % of the Volume of Loans