The flat-maximum effect and generic linear scoring models: A test

Predicting human behaviour patterns with linear correlation models has absorbed researchers for the past five decades. Although most observers generally concede that humans are inferior to such models in combining information, linear scoring models are, unfortunately, plagued by the flat-maximum effect or the ‘curse of insensitivity’. As Lovie & Lovie observe: ‘The predictive ability of linear models is insensitive to large variations in the size of regression weights and to the number of predictors.’ In essence, seemingly different scoring models tend to produce indistinguishable predictive outcomes. Since its demonstration by Dawes & Corrigan, observers have cast the flat maximum in a decidedly negative light. In contrast, Lovie & Lovie present a provocatively contrarian view of the flat maximum's positive potential. In this same vein, we examine the predictive power of a generic credit-scoring model versus individual empirically derived systems. If, as Wainer noted in regard to the flat maximum, ‘it don't make no nevermind’, generic credit-scoring models could provide cheaper alternatives to individual empirically derived models. During the period 1984–8, a series of linear credit-scoring models were developed for ten Southeastern US credit unions. For each credit union, stepwise multiple regression was employed to select a subset of explanatory variables to be used in a discriminant analysis. A generic credit-scoring equation was developed from the resulting discriminant analyses using weighted average coefficients from five systems. The predictive power of the generic model was compared to the predictive power of holdout sample of the five remaining credit-scoring models. In all cases, the generic model's performance was very close to that of the empirically derived models. Thus, our findings support Lovie & Lovie's challenge to the conventional wisdom that the flat maximum casts a pall on the successful modelling of judgement processes. Indeed, the flat maximum implies a positive role for simpler, and hence cheaper, generic models. Although further research is needed, it should be possible to develop hybrid models with generic cores that perform as well as empirically derived linear models.