A Comprehensive Survey of Time Series Forecasting: Concepts, Challenges, and Future Directions
Abstract
Time series forecasting (TSF) is a fundamental task in diverse decision-making scenarios, such as business intelligence, energy management, healthcare, and scientific discovery. Over the past decades, TSF has been extensively studied, with methods evolving from traditional statistical models to recent data-driven deep learning approaches. However, despite this rapid progress, the TSF literature still lacks a unified problem formulation and a systematic analysis of its core entities, data characteristics, and fundamental challenges. In particular, a systematic understanding is still needed regarding how existing TSF models address key issues such as data sparsity, long-term dependency, non-stationarity, missing observations, irregular sampling, and covariate utilization. This survey provides a comprehensive and structured overview of TSF from both problem and methodology perspectives. We first identify the key entities involved in TSF and analyze their essential characteristics, such as frequency, length, missing values, and distribution shifts. We then present a general problem formulation and summarize the major challenges that shape TSF research. Based on this foundation, we propose a taxonomy that organizes representative methods from the perspectives of preprocessing and forecasting, covering traditional statistical methods, machine learning models, deep neural architectures, and emerging data-driven paradigms such as foundation models and LLM-based forecasting. Furthermore, we discuss important research trends, including transfer learning, trustworthy forecasting, standardized evaluation, and other emerging directions. Finally, we outline promising future opportunities and deployment challenges that may further advance TSF in complex real-world scenarios.