In this issue of The Journal of Credit Risk we present three research papers.
Our first paper is "Valuation differences between credit default swap and corporate bond markets" by Oliver Entrop, Richard Schiemert and Marco Wilkens. The paper studies both the size and the cross-sectional and time series behavior of the "basis" between European corporate bonds and credit default swaps (CDSs). While this basis has been studied before, Entrop et al's paper contributes in several ways. Firstly, the basis is constructed differently, by calibrating to the CDS curve, yielding a bond-specific rather than issuer-specific basis and taking into account contractual differences between bonds and CDSs. Secondly, the authors show that the basis is smaller than has been reported in previous studies. Finally, the authors try to explain the basis in terms of firm specific and market factors.
The issue's second paper, "The effect of training set selection when predicting defaulting small and medium-sized enterprises with unbalanced data" by Giovanna Menardi and Nicola Torelli, deals with the accuracy of a default scoring model when the training set has characteristics that are different from those of the population that it is finally applied to. This could be the case when applying a scoring function to small firms while the model's parameters are estimated using data from larger firms; this could happen because of limited data availability. Another characteristic of default data is the so-called data imbalance problem: that is, the fact that defaults remain rare events, leading to parameters estimated on a small number of data points. The authors build their analysis mainly on a link between the impact of the choice of training set and the data imbalance problem. More precisely, the data imbalance might affect the sample selection bias. To be clear, the sample selection bias becomes visible when the data imbalance is small: that is, when the samples are more balanced. This suggests that the class imbalance problem might be of first-order importance and should be solved before handling the sample selection bias. In addition, the authors evaluate the sample selection bias by applying various sampling strategies to real data. In evaluating the accuracy of differing sample selection methods on the biases induced by the choice of training sample, the authors consider successively two methods compared to the benchmark.
Our final paper, "The art of probability-of-default curve calibration" by Dirk Tasche, deals with the problem of calibrating probability-of-default (PD) curves to rating systems when the unconditional PD is given. Calibrating PD curves is a key issue for anyone concerned with rating systems, estimation of PDs or stress testing. In hispaper, Tasche creates a theoretical framework that is then used to compare different calibration techniques; some of the techniques are quite well-known ("scaling the PD curve", for example) while others are new ones proposed by the author. The framework allows for a comparison not only in terms of performance but also in terms of the underlying ("invariance") assumptions.
JPMorgan Chase, New York
The effect of training set selection when predicting defaulting small and medium-sized enterprises with unbalanced data