What is a More Powerful Model Worth?
by Roger M. Stein of Moody's KMV, and
November 13, 2003
Abstract: In this paper, we provide empirical evidence of the economic impact of differences in power between various default models. In evaluating credit risk models, it is common to use metrics such as power curves and their associated statistics. However, power curves are not necessarily easily linked intuitively to common lending practices and in use, many users request a specific rule for defining a cutoff above which credit will be granted and below which it will be denied. A companion paper, Stein (2003), presented a framework for defining optimal lending cutoffs. It turns out that the framework for defining cutoffs can be extended to a more complete pricing approach that is more flexible as was shown in that paper. In both cases, the analysis permits the evaluation of models in terms of economic benefits and demonstrates that more powerful models are generally more cost effective than weaker ones. However, due in part to the non-parametric nature of power curves as well as the heterogeneity of the middle market banking environment, it can be difficult to size the magnitude of this economic value analytically. In this paper, using the CRD database of middle-market financial statement and loan performance data, we simulate lending using default prediction models of various qualities. We examine both the cutoff- and pricing-based lending approaches. In both cases we find that the use of a more powerful model, on average, leads to substantial economic benefit over that of a weaker model. We also provide a case study giving some insight into why the more powerful model was preferred. We discuss examples of dollar values of typical expected differentials in profit for prototypical small, medium and large institutions at various power levels. The methodology we present is a general one and can be used to size the financial value of lending policies based on any many varying sets of models.