Deciphering Complexity. Modeling the Future.

We propose an adaptive penalization framework that couples sliced inverse regression with a data-driven bandwidth selector, yielding consistent recovery of the central subspace even when the number of candidate predictors vastly exceeds the sample size. The estimator introduces a slice-adaptive weighting scheme that stabilizes the eigenstructure of the conditional covariance operator, and a refinement step that discards spurious directions before the final projection is constructed. On the theoretical side, we establish selection consistency under mild moment conditions and derive a rate of convergence that matches the sharpest results currently available in the sparse regression literature. Extensive simulations across synthetic and biomedical benchmarks — including gene-expression panels with more than forty thousand predictors — show meaningful gains in both selection accuracy and estimation stability relative to leading sliced inverse and sparse principal component baselines. A companion R package and reproducibility notebook accompany the manuscript so that practitioners can adopt the method on their own high-dimensional datasets without reimplementing the estimator from scratch.