Model-agnostic noise reduction for high-dimensional time series data

RESUMO

[PARTIAL/PARAPHRASED] We develop a model-agnostic framework for noise reduction in high-dimensional time series that explicitly targets optimal recovery of a low-dimensional latent dynamic component contaminated by observational white noise. The researchers characterize optimal linear projections onto dynamic subspaces and describe residual error geometrically. Their method employs lagged covariance matrices, bootstrap dimension selection, and a low-rank representation of the structured noise. The denoised series converges at parametric rates. Simulations demonstrate improvements in subspace estimation, reconstruction error, and forecast accuracy compared to orthogonal projection methods. Applications include high-dimensional stock returns and a 20-variable macroeconomic time series.