What Factors Drive Credit Risk

Credit risk remains the most significant risk facing financial institutions. An understanding of the drivers of credit risk (i.e. the possibility that a borrower will fail to repay a loan) is therefore a key analytical and practical challenge for lenders. Furthermore, to the extent that on a number of occasions in the past, banking system problems have been largely associated with simultaneous loan default across borrowers, an appreciation of the potential sources of heightened credit risk is also highly relevant for policymakers. This paper investigates this issue by applying statistical factor analysis to an indicator of credit risk derived from equity prices, namely KMV’s expected default frequency measure, which itself is derived from the Merton option-based model of default. Specifically, based on a large panel data set of firms, we apply principal component techniques to extract the different (systematic) factors influencing individual default tendencies. Recent authors (e.g. Stock and Watson (1998), Forni, and Reichlin (1998)) have shown that panel data sets with a large number of cross-sections are particularly suitable for extracting common factors. As the number of time series becomes large (large N), q “appropriately chosen” linear combinations of the variable under consideration become increasingly collinear with the common factor. This implies that q averages of the observations converge on the set of q common factors. Exploiting these insights, we investigate the common factors underlying individual default tendencies where the principal components are constructed using different aggregations across our panel of firm defaults. More precisely, we assume that default tendencies follow a factor structure they can be decomposed into two parts: a common part driven by a few factors, common to firms in the sample, and an idiosyncratic part, which is firm specific. This decomposition is carried out in a sequential fashion. The first step is to isolate the European-wide sources of default risk and evaluate their importance for the firms in the sample. This is achieved by extracting principal components of the full sample of firms, for the entire sample period and across rolling time periods. Importantly, firms belonging to the same country or sector are not constrained to respond equally to these factors; they can “load” the factor in different ways. The next step is to isolate the country and sector factors, which are implicitly left in the estimated idiosyncratic component estimated in the first stage. Again principal component analysis is employed on this idiosyncratic component to establish co-movement at the country or sectoral level. Finally, the relative contributions of the three types of factors in explaining default risk are evaluated by projecting each firm’s default tendency on the three components. Without an underlying structural model for (joint) default, it is difficult to attach particular economic meaning to the estimated statistical common factors. And indeed formally, more information would need to be imposed on the model to separately identify the factors and their corresponding loading coefficients. Nonetheless, the statistical model is helpful in assessing the relative importance of EU, country, sector and firm-level effects on firm default and how this may have changed over time. This general theme has been explored before in the literature, but the focus has almost exclusively been on equity returns rather than measures of credit risk per se. Moreover, from a portfolio perspective, the model can throw light on the potential diversificationbenefits from different lending strategies across countries, sectors and firms. In the spirit ofrecent developments in portfolio credit risk models, we simulate variants of the model toconstruct joint default probability distributions, which allow for different systematic andunsystematic (i.e. idiosyncratic) components of measured credit risk. We then investigate theoptimal return-risk frontiers for the different variants of the model. Importantly, since defaultcorrelation between firms is typically low, the distribution of default is unlikely to be normallydistributed and may not even be symmetric. Consequently variance is not a good indicator ofportfolio risk. Instead, we investigate efficient portfolio frontiers using measures of downsiderisk such as the semi-variance. ReferencesForni, M and Reichlin, L, (1998), “Let’s Get Real: A Dynamic Factor Analytic Approach toDisaggregated Business Cycle”, Review of Economic Studies, 65, pp. 453-473. Stock, J and Watson, M, (1998), “Diffusion Indexes”, NBER Working paper No. W6702.