Time Series Forecasting: Unlocking New Frontiers with LLMs, Hybrid Models, and Enhanced Interpretability

Latest 50 papers on time series forecasting: Sep. 1, 2025

Time series forecasting, a cornerstone of predictive analytics, continues to evolve at a rapid pace. From financial markets to climate science and demand planning, accurately predicting future trends is paramount. Recent breakthroughs, as highlighted by a wave of innovative research papers, are pushing the boundaries of what’s possible. These advancements leverage the power of large language models (LLMs), pioneer hybrid architectures, and champion greater interpretability, offering a glimpse into the future of robust and versatile forecasting.

The Big Idea(s) & Core Innovations

One of the most exciting trends is the integration of LLMs with time series data. Papers like Integrating Time Series into LLMs via Multi-layer Steerable Embedding Fusion for Enhanced Forecasting by Zhuomin Chen et al. (Sun Yat-Sen University, National University of Singapore) introduce frameworks like MSEF, allowing LLMs to directly access and retain time series patterns across all architectural depths. This is crucial for bridging the inherent modality gap between continuous numerical data and discrete linguistic representations. Similarly, Adapting LLMs to Time Series Forecasting via Temporal Heterogeneity Modeling and Semantic Alignment by Yanru Sun et al. (Tianjin University, A*STAR, Nanyang Technological University) presents TALON, which combines a heterogeneous temporal encoder with a semantic alignment module, achieving up to 11% improvement in MSE. Other LLM-focused innovations include FLAIRR-TS – Forecasting LLM-Agents with Iterative Refinement and Retrieval for Time Series from Google and Carnegie Mellon University, which uses agentic systems for iterative prompt refinement, and DP-GPT4MTS: Dual-Prompt Large Language Model for Textual-Numerical Time Series Forecasting by Chanjuan Liu et al. (Dalian University of Technology, Guangzhou University), which leverages explicit task instructions and context-aware embeddings from timestamped text to boost accuracy. TokenCast: From Values to Tokens: An LLM-Driven Framework for Context-aware Time Series Forecasting via Symbolic Discretization from University of Science and Technology of China and iFLYTEK Research also emphasizes this symbolic discretization for unified modeling.

Beyond LLMs, novel model architectures and data encoding techniques are making significant strides. GateTS: Versatile and Efficient Forecasting via Attention-Inspired routed Mixture-of-Experts by Kyrylo Yemetsa et al. (Lviv Polytechnic National University, Eindhoven University of Technology, University College London) introduces an attention-inspired gating mechanism for sparse Mixture-of-Experts (MoE) models, simplifying training and achieving superior accuracy with fewer parameters. In a similar vein, N-BEATS-MOE: N-BEATS with a Mixture-of-Experts Layer for Heterogeneous Time Series Forecasting enhances N-BEATS with an MoE layer for better handling of diverse time series, improving interpretability through dynamic routing. For more robust data representation, BinConv: A Neural Architecture for Ordinal Encoding in Time-Series Forecasting by Andrei Chernov et al. proposes Cumulative Binary Encoding (CBE) and a tailored convolutional architecture, significantly improving performance by preserving ordinal relationships.

Interpretability and robustness are also key themes. iTFKAN: Interpretable Time Series Forecasting with Kolmogorov-Arnold Network by Ziran Liang et al. (Hong Kong Polytechnic University) utilizes Kolmogorov-Arnold Networks (KAN) to offer transparent, symbolically represented explanations for forecasts, bridging the gap between deep learning and explainable AI. The importance of stability over mere accuracy for real-world applications is underscored by Measuring Time Series Forecast Stability for Demand Planning from Amazon Web Services, which shows ensemble models like AutoGluon offer greater consistency.

Under the Hood: Models, Datasets, & Benchmarks

Recent research is not only introducing new methodologies but also refining existing ones and establishing better evaluation standards:

• Foundational Models and Ensembles: Chronos, a foundation model, shows strong performance in hydrological forecasting (How Effective are Large Time Series Models in Hydrology?). Enhancing these, Enhancing Transformer-Based Foundation Models for Time Series Forecasting via Bagging, Boosting and Statistical Ensembles from Arizona State University demonstrates that integrating statistical ensembles significantly improves accuracy and uncertainty quantification. Meanwhile, PriceFM: Foundation Model for Probabilistic Electricity Price Forecasting from Delft University of Technology introduces a spatiotemporal foundation model with graph-based inductive biases for electricity markets.
• Novel Architectures: DeepKoopFormer: A Koopman Enhanced Transformer Based Architecture for Time Series Forecasting integrates the Koopman operator into Transformers for better nonlinear dynamics. DMSC: Dynamic Multi-Scale Coordination Framework for Time Series Forecasting from National University of Defense Technology leverages dynamic decomposition and adaptive fusion for state-of-the-art efficiency. T3Time: Tri-Modal Time Series Forecasting via Adaptive Multi-Head Alignment and Residual Fusion unifies temporal, spectral, and prompt-based representations for multivariate forecasting.
• Efficiency & Data Handling: OccamVTS: Distilling Vision Models to 1% Parameters for Time Series Forecasting proposes a knowledge distillation framework that drastically reduces model parameters while maintaining performance. FlowState: Sampling Rate Invariant Time Series Forecasting introduces a TSFM that dynamically adjusts to varying sampling rates without retraining, enhancing adaptability.
• Benchmarking & Robustness: To address evaluation challenges, TFB: Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods provides an automated, comprehensive, and unbiased framework. However, new vulnerabilities arise with complex models, as shown by BadTime: An Effective Backdoor Attack on Multivariate Long-Term Time Series Forecasting, which reveals severe vulnerabilities in MLTSF models.
• Multimodal & Cross-Domain: VisionTS++: Cross-Modal Time Series Foundation Model with Continual Pre-trained Visual Backbones from Zhejiang University, National University of Singapore, and Salesforce Research Asia, converts multivariate time series into RGB images for enhanced modeling, while EventTSF: Event-Aware Non-Stationary Time Series Forecasting from Griffith University and Xidian University introduces an autoregressive diffusion framework to integrate textual events.
• Explainability: PAX-TS: Model-agnostic multi-granular explanations for time series forecasting via localized perturbations offers a versatile method for understanding feature importance at different temporal scales. Meanwhile, On Identifying Why and When Foundation Models Perform Well on Time-Series Forecasting Using Automated Explanations and Rating from University of South Carolina combines XAI with rating-driven explanations to guide model selection.

Several papers provide public code repositories for further exploration:

• Compositionality in Time Series: A Proof of Concept using Symbolic Dynamics and Compositional Data Augmentation
• Enhancing Forecasting with a 2D Time Series Approach for Cohort-Based Data
• Enhancing Transformer-Based Foundation Models for Time Series Forecasting via Bagging, Boosting and Statistical Ensembles
• Integrate Time Series into LLMs via Multi-layer Steerable Embedding Fusion for Enhanced Forecasting
• Benchmarking Pre-Trained Time Series Models for Electricity Price Forecasting
• LETS Forecast: Learning Embedology for Time Series Forecasting
• Measuring Time Series Forecast Stability for Demand Planning
• Synaptic Pruning: A Biological Inspiration for Deep Learning Regularization
• From Values to Tokens: An LLM-Driven Framework for Context-aware Time Series Forecasting via Symbolic Discretization
• Distributed Lag Transformer based on Time-Variable-Aware Learning for Explainable Multivariate Time Series Forecasting
• TFB: Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods
• CITRAS: Covariate-Informed Transformer for Time Series Forecasting
• Time Evidence Fusion Network: Multi-source View in Long-Term Time Series Forecasting
• DMSC: Dynamic Multi-Scale Coordination Framework for Time Series Forecasting
• DeepKoopFormer: A Koopman Enhanced Transformer Based Architecture for Time Series Forecasting
• Adapting LLMs to Time Series Forecasting via Temporal Heterogeneity Modeling and Semantic Alignment
• TLCCSP: A Scalable Framework for Enhancing Time Series Forecasting with Time-Lagged Cross-Correlations
• Time-Prompt: Integrated Heterogeneous Prompts for Unlocking LLMs in Time Series Forecasting
• iTFKAN: Interpretable Time Series Forecasting with Kolmogorov-Arnold Network
• FlowState: Sampling Rate Invariant Time Series Forecasting
• PriceFM: Foundation Model for Probabilistic Electricity Price Forecasting
• How Effective are Large Time Series Models in Hydrology? A Study on Water Level Forecasting in Everglades
• VisionTS++: Cross-Modal Time Series Foundation Model with Continual Pre-trained Visual Backbones
• T3Time: Tri-Modal Time Series Forecasting via Adaptive Multi-Head Alignment and Residual Fusion
• Empowering Time Series Forecasting with LLM-Agents

Impact & The Road Ahead

The collective impact of this research is profound. The seamless integration of LLMs promises to unlock new levels of context-aware and semantic understanding in forecasting, particularly in scenarios rich with textual metadata like clinical notes or financial news. Innovations in model architectures, from sparse MoE to KAN-based systems and quantum circuits (Q-DPTS: Quantum Differentially Private Time Series Forecasting via Variational Quantum Circuits), are leading to more efficient, accurate, and interpretable predictions. Furthermore, a renewed focus on practical concerns like forecast stability, robustness against adversarial attacks, and dynamic adaptation to varying data characteristics underscores a maturing field.

The road ahead for time series forecasting is undoubtedly exciting. We can anticipate even more sophisticated multimodal models that learn from diverse data streams—text, vision, and numerical sequences—to build truly comprehensive predictive systems. The push for greater interpretability will continue to make these complex models more trustworthy and actionable for domain experts. As these technologies mature, they will not only enhance our ability to predict the future but also to understand the underlying dynamics that shape it, paving the way for more informed decision-making across all sectors.