Time Series Forecasting: Unlocking New Frontiers with Multimodal AI and Interpretability

Latest 50 papers on time series forecasting: Sep. 14, 2025

Time series forecasting, the art and science of predicting future data points based on historical observations, remains a cornerstone of decision-making across industries—from finance and energy to supply chain and healthcare. The inherent complexity of temporal data, often marred by non-stationarity, missing values, and the need for explainability, presents continuous challenges for AI and ML practitioners. Recent breakthroughs, as highlighted by a collection of innovative research papers, are pushing the boundaries of what’s possible, integrating advanced deep learning techniques, multimodal fusion, and a renewed focus on interpretability.### The Big Idea(s) & Core Innovationsoverarching theme in recent time series research is a multifaceted approach to harness richer data and more sophisticated models while maintaining efficiency and transparency. A significant trend is the integration of multimodal data and Large Language Models (LLMs). For instance, FinMultiTime: A Four-Modal Bilingual Dataset for Financial Time-Series Analysis by researchers from Central University of Finance and Economics and National University of Singapore introduces a dataset that aligns financial news, tables, K-line charts, and stock prices, demonstrating that data scale and quality, alongside multimodal fusion, significantly boost financial prediction accuracy. Building on this, Tsinghua University and Pengcheng Laboratory’s FinZero: Launching Multi-modal Financial Time Series Forecast with Large Reasoning Model proposes FinZero, a multimodal pre-trained model with an Uncertainty-adjusted Group Relative Policy Optimization (UARPO) method, outperforming GPT-4o by enhancing reasoning and providing confidence scores and reasoning traces. This quest for multimodal synergy extends to general time series, with SynLP Research Group, Tsinghua University’s Text Reinforcement for Multimodal Time Series Forecasting presenting TeR-TSF, an RL-driven data augmentation framework that uses LLMs to generate high-quality, aligned text, drastically improving forecasting performance when textual data is sparse or misaligned.critical innovation lies in enhancing LLMs for temporal reasoning. Several papers explore how to bridge the “modality gap” between continuous time series and discrete language. For example, Sun Yat-Sen University and National University of Singapore’s Integrating Time Series into LLMs via Multi-layer Steerable Embedding Fusion for Enhanced Forecasting introduces MSEF, a framework that integrates time series patterns across multiple LLM layers for improved few-shot forecasting. Similarly, Tianjin University and Nanyang Technological University’s Adapting LLMs to Time Series Forecasting via Temporal Heterogeneity Modeling and Semantic Alignment proposes TALON, which uses a Heterogeneous Temporal Encoder and Semantic Alignment Module, achieving up to 11% MSE reduction. Further pushing the boundaries, Semantic-Enhanced Time-Series Forecasting via Large Language Models by University of Science and Technology Beijing presents SE-LLM, which uses Temporal-Semantic Cross-Correlation and a Time-Adapter architecture to effectively model both long and short-term dependencies.multimodal advancements, efficiency, robustness, and interpretability remain paramount. South China University of Technology’s Temporal Query Network for Efficient Multivariate Time Series Forecasting introduces TQNet, an efficient model leveraging a single-layer attention mechanism with periodically shifted learnable queries for state-of-the-art accuracy and high computational efficiency. For scenarios with missing data, University of Science and Technology of China’s IBN: An Interpretable Bidirectional-Modeling Network for Multivariate Time Series Forecasting with Variable Missing introduces IBN, which uses Uncertainty-Aware Interpolation and Gaussian kernel-based Graph Convolution for reliable reconstruction and explicit spatial correlation modeling. Interpretability is also central to PAX-TS: Model-agnostic multi-granular explanations for time series forecasting via localized perturbations, a framework that explains feature importance at various temporal scales. Furthermore, University of California, Berkeley’s Quantum-Optimized Selective State Space Model for Efficient Time Series Prediction explores quantum-inspired components for efficient long-term prediction with minimal computational overhead.### Under the Hood: Models, Datasets, & Benchmarksresearch has not only introduced novel models but also enriched the ecosystem with critical datasets and robust benchmarking tools, essential for driving progress:FinMultiTime: A pioneering, large-scale, bilingual, four-modal dataset for financial time-series analysis (text, tables, images, time series) covering S&P 500 and HS 300. Code available at Hugging Face.FVLDB (Financial Vision-Language Database): Introduced with FinZero, this dataset provides diverse financial image-text pairs for multimodal model training and evaluation.Real-E: The largest electricity dataset to date, covering 74+ power stations across 30+ European countries, enabling robust energy forecasting. Code and benchmark links are anticipated from Karlsruhe Institute of Technology.CHRONOGRAPH: The first graph-based multivariate time series dataset derived from real production microservices, complete with expert-annotated incident windows for anomaly detection and structure-aware forecasting. Resources are available at https://arxiv.org/pdf/2509.04449.TFB (Time Series Forecasting Benchmark): An automated, comprehensive, and fair benchmarking framework from East China Normal University that addresses limitations of previous benchmarks by covering 10 diverse domains and supporting various evaluation strategies. Code available at https://github.com/decisionintelligence/TFB.DeepEDM: A framework by University of Wisconsin-Madison integrating Empirical Dynamic Modeling and deep learning, with code at https://abrarmajeedi.github.io/deep_edm.TQNet: A lightweight, efficient model from South China University of Technology for multivariate time series forecasting, with code available at https://github.com/ACAT-SCUT/TQNet.IBN: An interpretable network from University of Science and Technology of China for multivariate time series forecasting with variable missingness. Code available at https://github.com/zhangth1211/NICLab-IBN.ADAPT-Z: An online adaptation method from Tongji University for time series forecasting that addresses delayed feedback by adjusting latent factor representations. Code available at https://github.com/xiannanhuang/ADAPT-Z.RDIT: A framework from MIT for probabilistic time series forecasting, combining point estimation and residual modeling, with code at https://anonymous.4open.science/r/RDIT-16BB/.BALM-TSF: A lightweight framework from University of Birmingham addressing modality imbalance in LLM-based time series forecasting. Code available at https://github.com/ShiqiaoZhou/BALM-TSF.MSEF: A framework from Sun Yat-Sen University for integrating time series into LLMs, with code at https://github.com/One1sAll/MSEF.TALON: A framework from Tianjin University that adapts LLMs to time series forecasting, with code at https://github.com/syrGitHub/TALON.TokenCast: An LLM-driven framework from University of Science and Technology of China for context-aware time series forecasting via symbolic discretization, with code at https://github.com/Xiaoyu-Tao/TokenCast.DLTransformer: A distributed lag transformer from Korea government (MSIT) for explainable multivariate time series forecasting. Code at https://github.com/kYounghwi/DLFormer_official.Synaptic Pruning: A biologically inspired regularization method from James Cook University for deep learning, with code at https://github.com/xalentis/SynapticPruning.TSF-Prompt: An LLM-based framework from Central South University for time series forecasting. Code at https://github.com/SanMuGuo/Time-Prompt.cohort-based-2D-time-series: Code for enhancing forecasting with a 2D approach for cohort data from University of Technology, India, available at https://github.com/Lightricks/cohort-based-2D-time-series.PAX-TS: Model-agnostic multi-granular explanations for time series forecasting via localized perturbations. Code at https://anonymous.4open.science/r/pax-ts-6410.N-BEATS-MOE: N-BEATS with a Mixture-of-Experts Layer for Heterogeneous Time Series Forecasting. Code at https://github.com/zaai-ai/mixture_of_experts_time_series.### Impact & The Road Aheadadvancements have profound implications. The focus on multimodal integration, particularly with LLMs, promises more intelligent and context-aware forecasting systems. Imagine financial models that not only analyze numerical trends but also interpret global news and earnings reports, or energy grids that factor in weather patterns and social media sentiment. The development of specialized datasets like FinMultiTime, Real-E, and ChronoGraph, along with comprehensive benchmarks like TFB, is crucial for fostering fair comparisons and accelerating research.emphasis on interpretability (seen in IBN, PAX-TS, DLTransformer) is equally critical, moving us beyond black-box predictions to systems that can explain why a forecast was made, building trust and enabling better human-AI collaboration. The exploration of efficiency and adaptability (TQNet, ADAPT-Z, Quantum-Optimized SSSM, GateTS, WaveTS-B/M) ensures that these powerful models can be deployed in real-world, dynamic environments with limited resources and non-stationary data. Innovations like Synaptic Pruning even hint at biologically-inspired efficiency gains., the road ahead also presents challenges. While LLMs show great promise, their inherent text-centric biases and the need for robust cross-modal alignment are being actively addressed by frameworks like BALM-TSF, TSF-Prompt, and TokenCast. Moreover, the increasing sophistication of models brings privacy concerns to the forefront, as highlighted by Privacy Risks in Time Series Forecasting: User- and Record-Level Membership Inference. The development of targeted adversarial attacks like Fre-CW further underscores the need for robust and secure forecasting systems. Future research will likely focus on even more seamless multimodal fusion, real-time adaptive learning, stronger privacy-preserving mechanisms, and continued efforts to make complex models both powerful and transparent. The synergy between classical statistical methods and cutting-edge deep learning, exemplified by VARMAformer, promises a future of time series forecasting that is not only accurate but also robust, interpretable, and adaptable to the ever-evolving complexities of our world.