A Survey of Automated Time Series Forecasting: From Deep Learning to Foundation Models with Financial Applications

Abstract

Time-series forecasting is a fundamental research topic in artificial intelligence and machine learning, with broad applications in finance, healthcare, energy management, transportation, and industrial systems. As the scale and complexity of temporal data continue to increase, forecasting methodologies have evolved significantly from traditional statistical models to modern intelligent forecasting systems. Understanding this technological evolution is essential for both researchers and practitioners seekin g to develop more accurate, scalable, and automated forecasting solutions. This survey reviews the development of automated time-series forecasting methods from a historical and technological perspective. The evolution of forecasting paradigms is examined across multiple stages, including statistical learning methods, machine learning approaches, deep learning architectures, Transformer-based forecasting models, AutoML frameworks, foundation models, and large-language-model-based intelligent agents. Particular attention is given to how these paradigms differ in terms of automation capability, long-term forecasting performance, interpretability, and application adaptability. The survey highlights the key motivations driving major technological transitions, including limitations in statistical assumptions, challenges in feature engineering, difficulties in long-sequence modeling, and the growing demand for generalization and autonomous decision-making capabilities. In addition, representative forecasting applications in the financial domain are reviewed to illustrate the practical impact of modern forecasting technologies. Finally, current challenges and future research directions are discussed, including multimodal forecasting, foundation model development, intelligent agen t systems, explainable forecasting, and computational efficiency. The findings suggest that time-series forecasting is gradually evolving from task-specific prediction models toward intelligent forecasting ecosystems capable of learning, reasoning, and autonomous decision-making.