1: \begin{abstract}

2:   We consider the problem of supervised dimension reduction with a

3:   particular focus on extreme values of the target $Y\in\rset$ to be

4:   explained by a covariate vector $X \in \rset^p$. 

5:     The general

6:     purpose is to define and estimate a projection on a lower

7:     dimensional subspace of the covariate space which is sufficient

8:     for predicting exceedances of the target above high thresholds.

9:     We propose an original definition of Tail Conditional Independence

10:     which matches this purpose.  Inspired by  Sliced Inverse

11:     Regression (SIR) methods, we develop a novel framework (TIREX,

12:     Tail Inverse Regression for EXtreme response) in order to estimate

13:     an extreme sufficient dimension reduction (SDR) space of

14:     potentially smaller dimension than that of a classical SDR space.%

15:    We prove the weak convergence of tail empirical processes involved

16:   in the estimation procedure and we illustrate the relevance of the

17:   proposed approach on simulated and real world data.

18: \end{abstract}

19: