Neural Time Series Forecasting: A Problem-Driven Survey from Data Challenges to Model Choice

Abstract

Surveys of time series forecasting usually proceed by model family, from recurrent networks to Transformers and foundation models. That chronology is useful, but it gives limited guidance when the practical question is why a forecast fails. This article instead organizes the literature around five recurring difficulties: nonlinear and nonstationary behavior, contamination and structural breaks, uncertainty, long contexts and cross-variable dependence, and limited target-domain data. Studies published between 2014–2025 are compared through the assumptions they make, the settings in which they work, and the failure modes they leave unresolved. The review covers recurrent and probabilistic models, decomposition methods, robust and uncertainty-aware forecasting, Transformer variants, theory-guided methods, and foundation models. Quantitative results are used only when the underlying protocol is sufficiently clear. Across these families, the evidence supports a restrained workflow: begin with a strong simple baseline, identify the dominant source of error, and add complexity only when it addresses that source.