In this paper, we introduce a novel sufficient dimension reduction method, namely, weighted inverse regression ensemble (WIRE). WIRE is developed based on the conditional characteristic function of the response given the predictors, and it is slicing free and is readily applicable to multivariate response data. In the next, we provide a new perspective for model-free variable selection based on conditional characteristic function, which reveals many parallels between sufficient dimension reduction and variable selection. Based on such findings, we build up a new framework for testing the predictors contributions. And we further propose a forward stepwise algorithm incorporating with WIRE for ultrahigh dimensional model-free variable screening and selection. We show that, the WIRE method is a root-n consistent sufficient dimension reduction method, and the forward WIRE algorithm enjoys the variable screening consistency when the predictor dimensionality p diverges at an exponential rate of the sample size n. Finally, we compare our proposal with existing methods by simulation and on a real data set.