The Next Wave: Breakthroughs in Time Series Forecasting with LLMs and Beyond

Latest 73 papers on time series forecasting: Aug. 25, 2025

Time series forecasting is the heartbeat of countless industries, from predicting stock market trends to managing global supply chains and even anticipating extreme weather events. The ability to accurately predict future values from historical data is a cornerstone of modern decision-making. However, the inherent complexity of temporal data—non-stationarity, intricate patterns, and the need to integrate diverse data types—has always posed significant challenges. Recently, the AI/ML community has seen an explosion of innovative research, pushing the boundaries of what’s possible, particularly through the clever integration of Large Language Models (LLMs) and novel architectural designs.

The Big Idea(s) & Core Innovations

One of the most exciting themes emerging from recent research is the synergistic integration of Large Language Models (LLMs) with time series forecasting. Several papers tackle the challenge of bridging the modality gap between numerical time series and textual data, leveraging the powerful contextual understanding of LLMs to enhance predictive accuracy. For instance, TALON: Adapting LLMs to Time Series Forecasting via Temporal Heterogeneity Modeling and Semantic Alignment by **Yanru Sun et al. from Tianjin University and A*STAR introduces a framework that combines heterogeneous temporal encoding with semantic alignment to effectively adapt LLMs for time series tasks. Similarly, Hao Liu et al. from University of Science and Technology Beijing in their paper, Semantic-Enhanced Time-Series Forecasting via Large Language Models, propose SE-LLM, a framework that leverages a Temporal-Semantic Cross-Correlation (TSCC) module and a Time-Adapter architecture to effectively model long-term dependencies and short-term anomalies. Furthering this, DP-GPT4MTS: Dual-Prompt Large Language Model for Textual-Numerical Time Series Forecasting by Chanjuan Liu et al. from Dalian University of Technology and Guangzhou University uses a dual-prompt mechanism to better capture both explicit task instructions and context-aware embeddings from timestamped text, showing superior performance in complex scenarios. The common thread here is the understanding that rich semantic context can drastically improve forecasting, even for inherently numerical data. This is further reinforced by TokenCast: From Values to Tokens: An LLM-Driven Framework for Context-aware Time Series Forecasting via Symbolic Discretization by Xiaoyu Tao et al. from University of Science and Technology of China**, which transforms continuous time series into ‘temporal tokens’ to enable unified modeling with contextual features using LLMs.

Another significant innovation lies in enhancing model robustness, interpretability, and efficiency through architectural advancements and novel regularization techniques. iTFKAN: Interpretable Time Series Forecasting with Kolmogorov-Arnold Network by Ziran Liang et al. from Hong Kong Polytechnic University introduces an interpretable framework using Kolmogorov-Arnold Networks (KAN) to provide symbolic representations, bridging the gap between deep learning’s power and explainability. In a similar vein, DeepKoopFormer: A Koopman Enhanced Transformer Based Architecture for Time Series Forecasting by Ali Forootani integrates the Koopman operator with Transformers to better capture nonlinear and oscillatory behaviors, while Synaptic Pruning: A Biological Inspiration for Deep Learning Regularization by Gideon Vos et al. from James Cook University proposes a biologically inspired pruning method that dynamically eliminates low-importance neural connections, reducing predictive error rates by up to 52%.

Beyond these, advancements in handling specific data challenges are noteworthy. Sagar G. Patel and Ankit K. Mehta from University of Technology, India, and Research Institute for Data Science, USA, introduce a 2D Time Series Approach for Cohort-Based Data (the URL needs to be inferred from the ID, but the input doesn’t provide one, so a placeholder is used for now), significantly improving forecast reliability in small data environments. Meanwhile, SPADE-S: A Sparsity-Robust Foundational Forecaster by Malcolm Wolff et al. from Amazon SCOT Forecasting offers a specialized architecture for sparse and low-magnitude time series, addressing a common pain point in demand forecasting.

Under the Hood: Models, Datasets, & Benchmarks

Recent advancements are often underpinned by new models, robust datasets, and comprehensive benchmarking frameworks that allow for rigorous evaluation and comparison. Here are some key resources and methodologies highlighted in these papers:

• UniCast (https://arxiv.org/pdf/2508.11954) by Sehyuk Park et al. from Pohang University of Science and Technology, is a parameter-efficient multimodal prompting framework extending Time Series Foundation Models (TSFMs) with vision and text. Its code is available at https://shorturl.at/sw7nX.
• TFB (Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods) (https://arxiv.org/pdf/2403.20150) by Xiangfei Qiu et al. from East China Normal University, introduces an automated, comprehensive benchmarking framework to evaluate diverse TSF methods across 10 domains. The code is publicly available at https://github.com/decisionintelligence/TFB.
• DeepEDM (LETS Forecast: Learning Embedology for Time Series Forecasting) (https://arxiv.org/pdf/2506.06454) by Abrar Majeedi et al. from University of Wisconsin-Madison, integrates Empirical Dynamic Modeling (EDM) with deep learning for robust, scalable forecasting. Code is at https://abrarmajeedi.github.io/deep_edm.
• DLTransformer (Distributed Lag Transformer) (https://arxiv.org/pdf/2408.16896) by Younghwi Kim et al., introduces a time-variable-aware Transformer for explainable multivariate forecasting. Its official code is at https://github.com/kYounghwi/DLFormer_official.
• PriceFM (Foundation Model for Probabilistic Electricity Price Forecasting) (https://arxiv.org/pdf/2508.04875) by Runyao Yu et al. from Delft University of Technology, introduces a spatiotemporal foundation model with a comprehensive dataset covering 24 European countries. Code is at https://github.com/runyao-yu/PriceFM.
• DMSC (Dynamic Multi-Scale Coordination Framework) (https://arxiv.org/pdf/2508.02753) by Haonan Yang et al. from National University of Defense Technology, offers a novel architecture for dynamic multi-scale temporal dependency modeling. The code is available at https://github.com/1327679995/DMSC.
• K2VAE (Koopman-Kalman Enhanced Variational AutoEncoder) (https://arxiv.org/pdf/2505.23017) by Xingjian Wu et al. from East China Normal University, an efficient framework for probabilistic time series forecasting, with code at https://github.com/decisionintelligence/K2VAE.
• CITRAS (Covariate-Informed Transformer) (https://arxiv.org/pdf/2503.24007) by Yosuke Yamaguchi et al. from Hitachi Ltd., is a decoder-only Transformer model that flexibly integrates observed and known covariates. Code is at https://github.com/hitachi-ai/citr-as.
• Waltz (Watermarking Large Language Model-based Time Series Forecasting) (https://arxiv.org/pdf/2507.20762) by Wei Yuan et al. from The University of Queensland, a post-hoc watermarking framework for LLM-based time series forecasting, available at https://github.com/wyuanuq/Waltz.
• QuiZSF (An efficient data-model interaction framework for zero-shot time-series forecasting) (https://arxiv.org/pdf/2508.06915) by Shichao Ma et al. from University of Science and Technology of China, combines retrieval-augmented generation with TSPMs to enhance zero-shot forecasting, particularly in sparse-data scenarios.
• MIRA (Medical Time Series Foundation Model for Real-World Health Data) (https://arxiv.org/pdf/2506.07584) by Hao Li et al. from Microsoft Research, is a specialized foundation model for medical time series, handling irregular intervals and missing values with Continuous-Time Rotary Positional Encoding and Neural ODEs.
• TIME-HD and U-CAST (Are We Overlooking the Dimensions? Learning Latent Hierarchical Channel Structure for High-Dimensional Time Series Forecasting) (https://arxiv.org/pdf/2507.15119) by Juntong Ni et al. from Emory University, introduces the first comprehensive benchmark suite for high-dimensional time series forecasting, with an open-source toolkit at https://github.com/emory-university/time-hd-lib.

Impact & The Road Ahead

The collective impact of this research is profound, pushing time series forecasting into a new era of intelligence and adaptability. The burgeoning synergy between LLMs and traditional time series models promises a future where context-rich, multimodal data can be seamlessly integrated, leading to more nuanced and accurate predictions in complex domains like finance, healthcare, and climate science. The focus on interpretability, as seen with KAN-based models like iTFKAN and KANMixer, is critical for building trust and enabling real-world adoption, especially in high-stakes environments. Furthermore, advancements in robustness against adversarial attacks (like Fre-CW) and proactive security measures (Waltz) underscore a growing maturity in the field, addressing the practical concerns of deploying powerful AI models.

The emphasis on lightweight and adaptable frameworks, such as FlowState and ELF, suggests a future where highly efficient, foundation-model-agnostic solutions can dynamically adjust to real-time data shifts without extensive retraining. This is particularly crucial for edge computing and resource-constrained environments. The development of specialized models for challenging data types (e.g., sparse, cohort-based, or high-dimensional) indicates a move towards more tailored and effective solutions across diverse application landscapes.

The road ahead involves further refinement of multimodal integration, exploring more advanced techniques for semantic alignment, and continuously improving the interpretability and explainability of these powerful models. As the community continues to push the boundaries of LLM-powered and biologically inspired architectures, we can anticipate a new generation of forecasting tools that are not only more accurate but also more robust, transparent, and adaptable to the ever-changing dynamics of our world. The future of time series forecasting is dynamic, data-driven, and increasingly intelligent!