Time Series Forecasting: Unpacking the Latest Trends in Multi-modal, Causal, and Efficient Models

Latest 50 papers on time series forecasting: Sep. 21, 2025

Time series forecasting is a cornerstone of modern AI/ML, driving critical decisions across finance, energy, healthcare, and beyond. As data grows in volume and complexity, so do the challenges of accurately predicting future trends, handling non-stationarity, incorporating diverse data modalities, and maintaining efficiency. Recent research has been pushing the boundaries, offering exciting breakthroughs that promise more robust, interpretable, and scalable forecasting solutions. This blog post dives into some of these cutting-edge advancements, synthesizing insights from a collection of recently summarized papers.### The Big Idea(s) & Core Innovationsoverarching theme in recent time series forecasting (TSF) research revolves around multimodality, interpretability, and efficiency. Researchers are increasingly recognizing that numerical time series alone often lack the complete picture. The paper, FinZero: Launching Multi-modal Financial Time Series Forecast with Large Reasoning Model, by Yanlong Wang et al. from Tsinghua University, introduces FinZero, a multimodal pre-trained model fine-tuned with Uncertainty-adjusted Group Relative Policy Optimization (UARPO). This approach integrates various financial signals (like image-text pairs from their FVLDB dataset) to provide not just predictions, but also confidence scores and reasoning traces, significantly enhancing interpretability and reliability over models like GPT-4o.on the multimodal trend, BALM-TSF: Balanced Multimodal Alignment for LLM-Based Time Series Forecasting by Shiqiao Zhou et al. from the University of Birmingham and Siemens AG, tackles the critical issue of modality imbalance when integrating Large Language Models (LLMs) into TSF. Their BALM-TSF framework uses learnable prompts and balanced alignment to semantically and distributionally balance textual and temporal inputs, achieving state-of-the-art results with minimal parameters. Similarly, Text Reinforcement for Multimodal Time Series Forecasting by Zhang, Wei and Li, Yifei from SynLP Research Group, Tsinghua University, proposes TeR-TSF, an RL-driven data augmentation framework that generates high-quality text using LLMs to address missing or misaligned textual data, leading to substantial performance gains.integration of LLMs with numerical data is further explored in Integrating Time Series into LLMs via Multi-layer Steerable Embedding Fusion for Enhanced Forecasting by Zhuomin Chen et al. from Sun Yat-Sen University and the National University of Singapore. Their MSEF framework allows LLMs to directly access time series patterns across all architectural depths, improving information retention and achieving a 31.8% MSE reduction. This deep integration is echoed by Xiaoyu Tao et al.’s From Values to Tokens: An LLM-Driven Framework for Context-aware Time Series Forecasting via Symbolic Discretization, where TokenCast converts continuous time series into “temporal tokens” for unified, context-aware modeling with LLMs, especially useful for complex contextual data like clinical notes.multimodality, causality and efficiency are key. Xiangfei Qiu et al. from East China Normal University introduce DAG: A Dual Causal Network for Time Series Forecasting with Exogenous Variables. This groundbreaking framework uses dual causal networks to discover and inject causal relationships between endogenous and exogenous variables across temporal and channel dimensions, outperforming methods that underutilize future exogenous information. This move towards understanding ‘why’ forecasts occur, not just ‘what’ will occur, is critical for real-world applications.and robust modeling are not forgotten. Liran Nochumsohn et al. from Ben-Gurion University present Super-Linear: A Lightweight Pretrained Mixture of Linear Experts for Time Series Forecasting, an efficient model that uses frequency-specialized linear experts and a spectral gating mechanism to dynamically select experts, achieving state-of-the-art performance with high interpretability and speed. For long-term multivariate forecasting, Chenheng Xu et al. from UCLA propose the Fourier Neural Filter (FNF) with a Dual Branch Design, which integrates temporal-specific inductive biases and optimizes spatio-temporal modeling for superior performance without auxiliary techniques. Similarly, the ARMA Block: A CNN-Based Autoregressive and Moving Average Module for Long-Term Time Series Forecasting by Myung Jin Kim et al. offers a lightweight CNN-based alternative that inherently encodes positional information, achieving competitive accuracy against complex Transformer models. In another quest for efficiency, GateTS: Versatile and Efficient Forecasting via Attention-Inspired routed Mixture-of-Experts by Kyrylo Yemetsa et al. simplifies MoE training through an attention-inspired gating mechanism, reducing parameters and improving accuracy.### Under the Hood: Models, Datasets, & Benchmarksadvancements are often underpinned by novel architectures, new datasets, and robust benchmarking efforts. Here are some key contributions:Super-Linear: A lightweight Mixture-of-Experts (MoE) model with frequency-specialized linear experts and a spectral gating mechanism, demonstrating significant improvements in inference speed and model size. Code available.DAG: A dual causal network framework for time series forecasting with exogenous variables, accompanied by newly collected TSF-X datasets. Code available.TimeAlign: A dual-branch framework leveraging global and local alignment mechanisms for distribution-aware alignment, with theoretical justification on mutual information. Code available.GTS Forecaster: An open-source Python toolbox for geodetic time series forecasting, integrating deep learning models like KAN, GNNGRU, and TimeGNN for complex spatiotemporal patterns. Code available.ARMA Block: An efficient CNN-based module inspired by ARIMA for long-term time series forecasting, inherently encoding absolute positional information.Fourier Neural Filter (FNF): A novel neural architecture with Dual Branch Design (DBD) for multivariate long-term time series forecasting, achieving SOTA across 11 benchmarks.FinMultiTime: The first large-scale, bilingual, cross-market, four-modality dataset for financial time-series analysis (text, tables, images, time series). Code available.Temporal Query Network (TQNet): A simple yet efficient model employing a single-layer attention mechanism with periodically shifted learnable queries for multivariate time series forecasting. Code available.GLinear: A novel architecture balancing simplicity and sophistication for enhanced time series prediction, outperforming existing models on benchmark datasets. Code available.FinZero: A multimodal pre-trained model for financial time series forecasting, fine-tuned with UARPO and evaluated on the FVLDB dataset (Financial Vision-Language Database). Code via UARPO fine-tuning is provided.IBN: Interpretable Bidirectional-modeling Network for MTSF with variable missingness, featuring Uncertainty-Aware Interpolation (UAI) and Gaussian kernel-based Graph Convolution (GGCN). Code available.Real-E: The largest electricity dataset to date, covering 74+ power stations across 30+ European countries, serving as a foundation benchmark for robust energy forecasting. [Code available at [Real-E Link] and [Benchmark Link] (placeholders)].ARIES: A framework for assessing relationships between time series data properties and modeling strategies, implementing an automated model recommendation system. Code available.ADAPT-Z: A novel online adaptation method leveraging current features and historical gradients to address delayed feedback and distribution shifts in multi-step forecasting. Code available.RDIT: Residual-based Diffusion Implicit Models for Probabilistic Time Series Forecasting, integrating diffusion processes with bidirectional Mamba networks. Code available.TeR-TSF: An RL-driven data augmentation framework for multimodal TSF, using dual-objective reward mechanisms for text generation. Code available.BALM-TSF: A dual-branch framework using learnable prompts and balanced alignment for LLM-based TSF. Code available.Quantum-Optimized Selective State Space Model: Integrates quantum optimization for efficient long-term prediction. Code available.Compositionality in Time Series: Investigates symbolic dynamics and compositional data augmentation for clinical time series. Code available.FLAIRR-TS: A test-time prompting optimization framework for time series forecasting with LLMs, using iterative refinement and retrieval.BinConv: A convolutional neural architecture designed to work with Cumulative Binary Encoding (CBE) for ordinal time series data. PAX-TS: A model-agnostic framework for multi-granular explanation in time series forecasting via localized perturbations. Code available.GateTS: A sparse MoE architecture tailored for univariate time-series forecasting with an attention-inspired gating mechanism. [Code available at [Code link for GateTS] (placeholder)].DLTransformer: Distributed Lag Transformer for explainable multivariate time series forecasting, integrating time-variable-aware learning. Code available.APRNet: An Amplitude-Phase Reconstruct Network for multivariate time series forecasting, leveraging amplitude-phase relationships in the frequency domain. WaveTS-B and WaveTS-M: Lightweight models combining wavelet transforms with MLPs and MoE for efficient time series forecasting.TFB: An automated and comprehensive benchmarking framework for time series forecasting methods, addressing biases and covering diverse domains. Code available.DeepEDM: A framework integrating Empirical Dynamic Modeling (EDM) with deep learning for robust and scalable time series forecasting. Code available.ChronoGraph: The first graph-based multivariate time series dataset from real production microservices, with expert-annotated incident windows. [Code available at PLACEHOLDER (publicly available)].### Impact & The Road Aheadadvancements herald a new era for time series forecasting. The emphasis on multimodal inputs, particularly the synergy with LLMs, promises to unlock deeper contextual understanding and more human-like reasoning in predictions. Frameworks like FinZero and BALM-TSF demonstrate how combining diverse data types can lead to significant accuracy and interpretability gains, crucial for high-stakes domains like finance. The focus on causal networks, as seen in DAG, is moving the field beyond correlation to true understanding, enabling better intervention strategies.remains a critical concern, with lightweight models like Super-Linear, ARMA Block, and GateTS proving that performance doesn’t always necessitate massive parameter counts. The development of robust benchmarks like Real-E and TFB is indispensable for fair comparison and accelerating research in challenging, real-world scenarios. Moreover, the increasing attention to privacy risks, highlighted by Privacy Risks in Time Series Forecasting: User- and Record-Level Membership Inference, underscores the need for responsible AI development.future of time series forecasting appears bright, characterized by increasingly intelligent, context-aware, and efficient models. We can expect further integration of LLMs with numerical data, a stronger emphasis on causal inference, and more sophisticated approaches to uncertainty quantification and explainability. The goal is clear: to deliver forecasts that are not only accurate but also reliable, interpretable, and adaptable to the dynamic, often unpredictable, rhythm of the real world.