Environmetrics Ecological Modeling

Bootstrap‐Based Forecasting for Locally Stationary Long‐Memory Processes: An Application to Daily PM2.5 Concentrations in Coyhaique, Chile

Guillermo Ferreira, Oscar Rubilar‐Parra, Jorge Mateu
Wiley · 2026-10-01 · DOI 10.1002/env.70157

Abstract

ABSTRACT Forecasting is a central objective in time series analysis, with increasing interest in obtaining accurate predictions and reliable prediction intervals. State‐space models provide a natural framework for this purpose through the Kalman filter, although conventional predictive inference often relies on Gaussian innovations and time‐invariant system matrices. We propose a state‐space bootstrap (SSB) procedure for constructing prediction intervals that accommodates time‐varying system matrices, incorporates parameter estimation uncertainty through bootstrap reestimation, and relaxes the Gaussian assumption by resampling standardized innovations. The methodology is applied to daily concentrations in Coyhaique, Chile, using a regression– model formulated within the SS framework. The systematic component accounts for trend and annual seasonality through Fourier harmonics, while the stochastic component captures short‐ and long‐range dependence, allowing the fractional memory parameter and innovation scale to evolve smoothly over time.

Read at publisher →