Getting more from your regression model: A free lunch?

Abstract: We consider a simple approach for approximating detailed information about the conditional distribution of a real-valued response variable, given values for its covariates, using only the outputs from a standard regression model. We validate this approach by assessing its performance in the context of quantile regression; when applied to the outputs of linear, gradient boosted tree ensemble and random forest models. We find that it compares favourably to the standard approach for estimating quantile regression functions, especially for commonly selected tail probabilities, and is highly competitive with the quantile regression forest model, across a large collection of benchmark data sets.