Feature selection in jump models

Jump models switch infrequently between states to fit a sequence of data while taking the ordering into account We propose new framework for joint feature selection, parameter and state-sequence estimation in jump models. Feature selection is necessary high-dimensional settings where number features large compared observations underlying differ only with respect subset features. develop implement coordinate descent algorithm that alternates selecting estimating model parameters state sequence, which scales sets numbers (noisy) demonstrate usefulness proposed by comparing it other methods on both simulated real form financial returns, protein sequences, text. By leveraging information embedded data, resulting sparse outperforms all considered remarkably robust noise. • Joint Determine are responsible differences states. Coordinate noisy Evaluation Sparse