Confidence regions for entries of a large precision matrix

TL;DR

This paper develops a data-driven method to construct confidence regions for entries of high-dimensional precision matrices in time-dependent data, enabling structure testing and component recovery.

Contribution

It introduces a novel procedure combining penalized regressions, Gaussian approximation, and bootstrap for inference on precision matrices without requiring stationarity.

Findings

Method performs well in simulations

Effective in real stock return data analysis

Does not rely on second order stationarity

Abstract

Precision matrices play important roles in many practical applications. Motivated by temporally dependent multivariate data in modern social and scientific studies, we consider the statistical inference of precision matrices for high-dimensional time dependent observations. Specifically, we propose a data-driven procedure to construct a class of simultaneous confidence regions for the precision coefficients within an index set of interest. The confidence regions can be applied to test for specific structures of a precision matrix and to recover its nonzero components. We first construct an estimator of the underlying precision matrix via penalized node-wise regressions, and then develope the Gaussian approximation results on the maximal difference between the estimated and true precision matrices. A computationally feasible parametric bootstrap algorithm is developed to implement the proposed procedure. Theoretical results indicate that the proposed procedure works well without the second order cross-time stationary assumption on the data and sparse structure conditions on the long-run covariance of the estimates. Simulation studies and a real example on S&P 500 stock return data confirm the performance of the proposed approach.