Recent Advances in Time Series Forecasting Methods

1. Introduction

Time series forecasting has become a key decision-support tool, with broad applicability within a series of domains within the economic field, among other fields. Given the observed growth in large-scale, high-frequency, heterogeneous temporal data, a real need for accuracy, robustness, and adaptability in the tools used has been observed, which is not properly covered by traditional statistical methods. As a result, in recent approaches to time series forecasting methods, researchers have tried to place more emphasis on overcoming issues such as nonlinearity, long-term dependence, and uncertainty. Furthermore, the shift towards Machine Learning (ML)- and Artificial Intelligence (AI)-based forecasting was motivated by the inherent limitations to the traditional statistical approaches in the field.

This Special Issue, “Advanced Methods for Time Series Forecasting”, aims to capture the recent evolution in the field by bringing together recent methodological advances and innovative applications that illustrate the evolving state of the field. The contributions published in the Special Issue collectively demonstrate how modern forecasting research increasingly relies on hybrid modeling strategies, attention-based architectures, structured state-space formulations, and time–frequency analyses to address real-world challenges.

2. Advances Reflected in the Special Issue

A central theme in this Special Issue is the growing role of deep learning architecture enhanced with attention mechanisms. In the published works, transformer-based and attention-driven models prove to be effective in capturing complex temporal dependencies across multiple horizons in various situations, ranging from traffic flow prediction, where hierarchical and multi-view attention frameworks have been proposed [1,2], to meteorological forecasting tasks related to tropical cyclone track prediction, where frequency-aware attention mechanisms combined with dual-branch architectures have shown notable robustness [3].

The use of hybrid frameworks that integrate signal decomposition and learning-based forecasting models is another methodological direction. By decomposing complex time series into multi-scale or frequency-specific components prior to prediction, it has been observed that the proposed approaches succeed in addressing issues related to non-stationarity and noise in the case of financial time series [4]. Thus, it can be observed that combining domain-aware preprocessing with advanced neural architecture is a viable approach for improving the accuracy and stability of forecasts.

Long-term forecasting and memory retention are addressed in this Special Issue through recent advances based on state-space reformulations and selective memory mechanisms. These mechanisms offer the possibility of mitigating error accumulation and instability over extended forecasting horizons [5], which is particularly relevant for applications requiring reliable long-term predictions across diverse domains, such as, but not limited to, energy systems, finance, healthcare, and transportation.

The uncertainty quantification represents another critical contribution of this Special Issue. How prediction intervals and uncertainty estimates can be obtained alongside point forecasts is demonstrated through the use of probabilistic deep learning approaches—such as Monte Carlo dropout neural networks [6]. This approach has the advantage of offering more support for risk-sensitive decisions.

In addition to these contributions, this Special Issue also encompasses studies on short and irregular time series forecasting [7], hybrid statistical and learning-based approaches for mid-term energy load prediction in smart buildings [8], application-driven analyses in management and economic contexts [9], and demand forecasting in retail environments using optimized attention-based deep learning models [10].

Beyond the individual modeling contributions, this Special Issue also includes a bibliometric analysis of AI-driven time series forecasting research. The analysis maps research growth and trends, while highlighting key journals and collaboration networks [11].

3. Addressing Knowledge Gaps

Collectively, the papers published in this Special Issue address a series of gaps in the time series forecasting literature. First, contrary to other approaches in the field, which rely on isolated model comparisons, the papers in this Special Issue propose integrated and hybrid solutions that combine complementary modeling approaches. Second, the papers demonstrate how advanced forecasting methods can be shaped to better answer to domain-specific constraints, such as, but not limited to, issues related to short time series, managing irregular sampling, or the fusion of heterogeneous data sources. Third, the papers included in this Special Issue focus on uncertainty-aware forecasting, highlighting its importance in the context of high-impact application areas.

The diversity of approaches noted in this Special Issue further highlight that time series forecasting remains a debatable and evolving research problem due to the complexity of real-world systems, which continue to challenge existing methodologies.

4. Future Research Directions and Outlook Toward a Follow-Up Special Issue

Building upon the advances showcased in this Special Issue, a series of promising avenues for future research emerge, such as the development of general-purpose and pre-trained forecasting models capable of transferring knowledge across domains, reducing data requirements, and improving scalability; creating explainable and trustworthy forecasting systems by embedding explainability directly into forecasting architectures rather than treating it as a post hoc process; and integrating forecasting and decision-making for better support in risk management.

Finally, the growth of streaming and real-time data requires the development of adaptive and online learning approaches, which should be capable of handling concept drift. In this situation, hybrid models can be the most suitable way to address these challenges.

Taken together, these research directions motivate the launch of a new Special Issue on next-generation time series forecasting, in which the focus should be on foundation models, explainability, uncertainty-aware decision support, and adaptive learning in complex environments.