Time Series Forecasting Methods: A Systematic Review from Classical Models to Deep Learning

Time series forecasting (TSF) is a foundational task in computational modeling, with broad applications across finance, health sciences, energy, transportation, manufacturing, and environmental modeling. Over the decades, statistical and classical machine learning (ML) methods have provided effective forecasting solutions, while recent advances in deep learning have substantially expanded the modeling capabilities for complex, high-dimensional, and nonlinear time series data. This survey provides a comprehensive and structured overview of TSF methods, covering statistical models, classical machine learning approaches, deep learning architectures, and hybrid methods. We introduce a taxonomy of TSF models based on six key perspectives and discuss major methodological paradigms, including statistical models, ML approaches, and nine core deep learning architectures such as RNNs, CNNs, Transformers, Graph Neural Networks, generative models, and emerging methods including Neural ODEs, foundation models, and Mamba-based networks. In addition, we examine forecasting strategies, benchmark datasets, evaluation scores, software availability, and major application domains. Finally, we provide both qualitative and quantitative comparisons across methodological approaches and model architectures to support practical model selection and highlight current challenges and future research directions in TSF.