Long-Term Time Series Forecasting: The Good, the Bad, and the Ugly

Long-term time series forecasting (LTF) aims to predict time series over extended time horizons, offering significant cross-domain advantages and competitive insights. However, there is no universal solution to this task, and different methods may prove to be effective depending on the specific context. This survey provides a critical investigation of the state-of-the-art and the current research landscape in LTF. To structure this analysis, the paper is built around three key research questions: (1) Why are long-term forecasts more intriguing and valuable than short-term ones? (2) Why are specialised models essential for accurate long-term forecasting? (3) What are the typical challenges, pitfalls and critical points that must be addressed in this field of research? By addressing these questions, this work critically examines existing approaches, highlighting weak and strong points of current research trends. This discussion is further enriched by an overview of the benchmark datasets, the evaluation metrics, a taxonomy for model classification, as well as an analysis of recent methodological advances.