A distribution-free approach to estimating best response values with application to mutual fund performance modeling

Frontier regression models seek to model and estimate best rather than average values of a response variable. Our proposed frontier model has similar intent, but also allows for an additional error term. The composed error approach uses the sum of two error terms, one an inefficiency error and the other as white noise. Previous research proposed assumptions on the distributions of the error components so that the distribution of this total error can be specified. Here we propose a distribution free approach to specifying these errors. In addition, our approach is completely data driven, rendering model specification an unnecessary step. We also outline, step-by-step, an approach to implementing this procedure. Our entire approach is illustrated with a mutual fund data set from the Morning Star database.