Input space optimization for neural network-based forecasting of non-stationary time series

Abstract

Forecasting time series under instability remains a challenging task due to structural discontinuities, nonstationarity, and the influence of external factors. This paper proposes a forecasting approach focused on optimizing the model's input space rather than architectural changes to the forecasting algorithm itself. This study examines the impact of exogenous variables on neural network-based forecasting using data on infectious disease incidence in Ukraine. A baseline neural network forecasting model based solely on lagged target values is compared with models incorporating exogenous information. A systematic procedure for constructing and optimizing the input space is proposed, including evaluating various subsets of exogenous variables for fixed model architecture, training parameters, and forecast horizon. Experimental results demonstrate that the proposed input space optimization strategy leads to a significant reduction in forecast error, and these improvements are achieved without increasing model complexity or changing the training procedure. Copyright © 2026 for this paper by its authors.