Leading indicators of non-performing loans in Greece: the information content of macro-, micro- and bank-specific variables

Abstract

We examine the information content of a unique set of macroeconomic, bank-specific, market and credit registry variables as regards their ability to forecast non-performing loans using a panel data set of nine Greek banks. We distinguish between business, consumer and mortgage loans and investigate their differences with respect to their optimal predictors. The quasi-AIM approach (Carson et al. in Int J Forecast 27:923–941, 2010) is utilized in order to take into account heterogeneity across banks and minimize estimation uncertainty. In addition, we calculate a number of forecasting measures in order to take into account the policy makers’ preferences. We find that market variables, specifically the supermarket sales, confidence indices for the services and construction sector and the business sentiment index represent good forecasting variables for most categories of NPLs. In addition, industrial production is the optimal predictor for consumer NPLs and imports for business NPLs. Finally, bank-specific variables represent top-performing leading indicators for business NPLs. Our results have significant implications for stress-testing credit risk in a top-down manner and for supervisory and macro-prudential policy design.

Appendix: The model confidence set (MCS)

In this appendix, we briefly describe the model confidence set (MCS) method of Hansen, Lunde and Nason (2003; 2011) used to construct a set of models, $M_{1-a}^*\subseteq M_0$, that present statistically superior predictive ability at a given confidence level.

Assuming an initial set of $M=M_0$ models, the MCS method is based on a specific loss function, $L_{m,t}$ with $m=1,\ldots ,M$, and applies an iterative process of sequential equal predictive ability (EPA) tests of the form:

$$\begin{aligned} {H_{0,M_0}} : E\left( {d_{mk,t} } \right) =0\quad \hbox {for all }m,k\in M \end{aligned}$$

where $d_{mk,t} =L_{m,t} -L_{k,t}$ is the loss differential between models m and kand $L_{{\bullet },t}$ is one of the RMSFE or MAFE at each point in time, t. A rejection of the null hypothesis indicates that a model has inferior predictive ability and should not be included in the MCS at an a significance level. This EPA test is repeated for the remaining $M_{1-a}$ models, with $M_{1-a} \subset M$, and this procedure continues until the null hypothesis cannot be rejected. The final set of surviving models forms the MCS at a $1-a$ confidence level, denoted by $M_{1-a}^*$. The models included in the MCS have equal predictive ability, but they outperform the eliminated models, while the MCS p-values indicate the probability of a model being a member of the MCS.⁹