Time Series Forecasting: Unpacking the Latest AI/ML Innovations

Latest 50 papers on time series forecasting: Nov. 16, 2025

Time series forecasting remains a cornerstone of decision-making across industries, from finance and healthcare to energy and manufacturing. However, the dynamic, often non-stationary, and inherently complex nature of temporal data presents a persistent challenge for AI/ML models. Researchers are constantly pushing the boundaries, developing novel architectures, loss functions, and learning paradigms to enhance accuracy, robustness, and interpretability. This post delves into recent breakthroughs, exploring how the latest research is tackling these challenges and shaping the future of time series prediction.

The Big Idea(s) & Core Innovations

The recent surge in time series forecasting research highlights a collective effort to move beyond traditional methods and integrate advanced AI techniques. A prominent theme is the quest for more robust and nuanced uncertainty quantification, as seen in the work from the Institute of Big Data Science and Industry, Shanxi University. Their paper, “Beyond MSE: Ordinal Cross-Entropy for Probabilistic Time Series Forecasting”, introduces OCE-TS, replacing Mean Squared Error (MSE) with Ordinal Cross-Entropy (OCE) to provide superior stability and outlier robustness, crucial for real-world applications like financial risk control. Building on the limitations of MSE, the same institution also presents “RI-Loss: A Learnable Residual-Informed Loss for Time Series Forecasting”, which uses the Hilbert-Schmidt Independence Criterion (HSIC) to explicitly model noise structures and capture temporal dependencies more effectively, leading to significant performance gains.

Another significant innovation centers around improving model architecture and data processing. Researchers from Changsha University and Central South University, in their paper “MDMLP-EIA: Multi-domain Dynamic MLPs with Energy Invariant Attention for Time Series Forecasting”, address challenges like weak seasonal signal loss and insufficient channel fusion with an adaptive dual-domain seasonal MLP and an energy invariant attention mechanism, achieving state-of-the-art results. Similarly, the “EMAformer: Enhancing Transformer through Embedding Armor for Time Series Forecasting” from Beijing Jiaotong University and Beijing Normal University revisits Transformer limitations, introducing three inductive biases—global stability, phase sensitivity, and cross-axis specificity—to stabilize inter-channel correlations in multivariate time series forecasting (MTSF), significantly outperforming existing methods.

The interpretability and adaptability of models are also gaining traction. “CaReTS: A Multi-Task Framework Unifying Classification and Regression for Time Series Forecasting” by researchers from Cardiff, Newcastle, and Leeds Universities proposes a dual-stream architecture that unifies classification for macro-level trend prediction with regression for micro-level deviation estimation, improving both accuracy and interpretability. Addressing the complex challenges of non-stationary data, East China Normal University’s “Towards Non-Stationary Time Series Forecasting with Temporal Stabilization and Frequency Differencing” introduces DTAF, a dual-branch framework that combines temporal stabilization with frequency differencing for robust long-term predictions. Meanwhile, “CometNet: Contextual Motif-guided Long-term Time Series Forecasting” from Tianjin University leverages contextual motifs to capture long-range dependencies, overcoming the receptive field bottleneck common in traditional models.

An exciting new frontier involves integrating human expertise and large language models (LLMs). The University of Science and Technology of China’s “AlphaCast: A Human Wisdom-LLM Intelligence Co-Reasoning Framework for Interactive Time Series Forecasting” redefines forecasting as an interactive, step-by-step collaboration between human experts and LLMs, offering greater adaptability and interpretability. Furthermore, the advent of foundation models is revolutionizing the field. Datadog AI Research’s “This Time is Different: An Observability Perspective on Time Series Foundation Models” introduces TOTO, a 151-million-parameter zero-shot forecasting model, and BOOM, a large-scale observability benchmark, setting new state-of-the-art performance. The University of Freiburg, ELLIS Institute Tübingen, and Prior Labs contribute “TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting”, a foundation model pre-trained solely on synthetic data that achieves competitive zero-shot performance using linear RNNs. Building on the potential of pre-trained models, “TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning” from NXAI GmbH and JKU Linz introduces an xLSTM-based model that excels in zero-shot forecasting for both short and long horizons, thanks to Contiguous Patch Masking and novel data augmentation. For specific challenges like missing values, “Revisiting Multivariate Time Series Forecasting with Missing Values” by the University of Illinois at Chicago proposes CRIB, a direct-prediction framework based on the Information Bottleneck principle, outperforming imputation-then-prediction methods, especially at high missing rates.

Under the Hood: Models, Datasets, & Benchmarks

The innovations highlighted above are built upon significant advancements in models, datasets, and benchmarks. Here’s a breakdown of key resources:

• Loss Functions & Learning Paradigms:
  • OCE-TS (Ordinal Cross-Entropy for Probabilistic Time Series Forecasting, https://arxiv.org/pdf/2511.10200): A new loss function for uncertainty quantification.
  • RI-Loss (Residual-Informed Loss for Time Series Forecasting, https://arxiv.org/pdf/2511.10130): Leverages Hilbert-Schmidt Independence Criterion (HSIC) to handle noise. Code inferred to be available at the paper URL.
  • DBLoss (Decomposition-based Loss Function, https://arxiv.org/pdf/2510.23672): A decomposition-based loss for trend and seasonality, with code available at https://github.com/decisionintelligence/DBLoss.
  • Selective Learning (for Deep Time Series Forecasting, https://arxiv.org/pdf/2510.25207): Utilizes a dual-mask mechanism for filtering non-generalizable timesteps, with code at https://github.com/GestaltCogTeam/selective-learning.
• Architectures & Frameworks:
  • MDMLP-EIA (Multi-domain Dynamic MLPs with Energy Invariant Attention, https://arxiv.org/pdf/2511.09924): A novel MLP-based model for seasonal signals and channel fusion, code at https://github.com/zh1985csuccsu/MDMLP-EIA.
  • CaReTS (Multi-Task Framework Unifying Classification and Regression, https://arxiv.org/pdf/2511.09789): Dual-stream architecture for improved interpretability, code at https://anonymous.4open.science/r/CaReTS-6A8F/README.md.
  • EMAformer (Enhancing Transformer through Embedding Armor, https://arxiv.org/pdf/2511.08396): Transformer enhancement with inductive biases, code at https://github.com/PlanckChang/EMAformer.
  • DTAF (Dual-branch Temporal Stabilization and Frequency Differencing, https://arxiv.org/pdf/2511.08229): Handles non-stationarity in temporal and frequency domains, code at https://github.com/PandaJunk/DTAF.
  • CometNet (Contextual Motif-guided Long-term Time Series Forecasting, https://arxiv.org/pdf/2511.08049): Uses contextual motifs for long-range dependencies.
  • IMA (Imputation-Based Mixup Augmentation, https://arxiv.org/pdf/2511.07930): Combines imputation and Mixup for data augmentation, code at https://github.com/dangnha/IMA.
  • PFRP (Predicting the Future by Retrieving the Past, https://arxiv.org/pdf/2511.05859): Retrieval-based forecasting using a Global Memory Bank, code at https://github.com/ddz16/PFRP.
  • Synapse (Adaptive Arbitration of Complementary Expertise, https://arxiv.org/pdf/2511.05460): Dynamic arbitration framework for TSFMs.
  • ZOO-PCA (Embedding-Space Data Augmentation for Privacy, https://arxiv.org/pdf/2511.05289): Privacy-preserving data augmentation for clinical time series, code at https://github.com/MariusFracarolli/ML4H_2025.
  • AWEMixer (Adaptive Wavelet-Enhanced Mixer Network, https://arxiv.org/pdf/2511.04722): Combines wavelets with mixer architecture, code at https://github.com/hit636/AWEMixer.
  • ForecastGAN (Decomposition-Based Adversarial Framework, https://arxiv.org/pdf/2511.04445): Adversarial training with decomposition for multi-horizon forecasting.
  • StochDiff (Stochastic Diffusion Probabilistic Model, https://arxiv.org/pdf/2406.02827): Diffusion-based model for stochastic time series.
  • IMTS-Mixer (Irregular Multivariate Time Series Forecasting, https://arxiv.org/pdf/2502.11816): MLP-based for irregularly sampled data, with code for component works available.
  • CRISP (Crisis-Resilient Portfolio Management, https://arxiv.org/pdf/2510.20868): Graph-based spatio-temporal learning for finance.
  • ViTime (Foundation Model Powered by Vision Intelligence, https://arxiv.org/pdf/2407.07311): Leverages LVMs for time series, code at https://github.com/IkeYang/ViTime.
  • TEM (Token-level Topological Structures in Transformer-based TSF, https://arxiv.org/pdf/2404.10337): Preserves topological structures in Transformers, code at https://github.com/jlu-phyComputer/TEM.
  • InvDec (Inverted Decoder for Multivariate Time Series Forecasting, https://arxiv.org/pdf/2510.20302): Hybrid architecture for separated temporal and variate modeling (code to be released).
  • QKCV Attention (Enhancing Time Series Forecasting with Static Categorical Embeddings, https://arxiv.org/pdf/2510.20222): Integrates categorical embeddings into attention layers.
  • SST (Multi-Scale Hybrid Mamba-Transformer Experts, https://arxiv.org/pdf/2404.14757): Hybrid Mamba-Transformer for long- and short-range patterns, code at https://github.com/XiongxiaoXu/SST.
• Domain-Specific Models & Tools:
  • MLF (Multi-period Learning for Financial Time Series Forecasting, https://github.com/Meteor-Stars/MLF): Integrates multi-period inputs for financial TSF, code at https://github.com/Meteor-Stars/MLF.
  • LiteCast (Lightweight Forecaster for Carbon Optimizations, https://arxiv.org/pdf/2511.06187): Carbon intensity forecasting for energy efficiency, with code inferred at https://github.com/AbelSouza/LiteCast.
  • DeltaLag (Learning Dynamic Lead-Lag Patterns in Financial Markets, https://arxiv.org/pdf/2511.00390): Deep learning for dynamic lead-lag relationships in finance, code at https://github.com/hkust-gz/DeltaLag.
  • ARIMA_PLUS (In-Database Time Series Forecasting and Anomaly Detection in Google BigQuery, https://arxiv.org/pdf/2510.24452): Scalable, automatic, and interpretable forecasting within BigQuery.
  • Unsupervised Anomaly Prediction (N-BEATS and Graph Neural Network in Semiconductor Process Time Series, https://arxiv.org/pdf/2510.20718): GNNs for anomaly detection in semiconductor manufacturing, with N-BEATS code reference at https://github.com/philipperemy/n-beats.
• Benchmarks & Evaluation:
  • BOOM (Benchmark for Observability Metrics, https://huggingface.co/datasets/Datadog/BOOM): First large-scale benchmark for observability data.
  • SynTSBench (Rethinking Temporal Pattern Learning in Deep Learning Models, https://arxiv.org/pdf/2510.20273): Synthetic data-driven evaluation framework, code at https://github.com/TanQitai/SynTSBench.
  • Gift-Eval benchmark: Heavily utilized by several papers for evaluating zero-shot forecasting.

Impact & The Road Ahead

These advancements herald a new era for time series forecasting, promising more accurate, robust, and interpretable predictions across a multitude of domains. The emphasis on advanced loss functions like OCE-TS and RI-Loss will lead to more reliable uncertainty quantification, critical for high-stakes decision-making in finance and risk management. Innovations in architecture, such as MDMLP-EIA and EMAformer, push the boundaries of model performance by specifically addressing challenges like weak signals and channel interactions.

The integration of human-LLM co-reasoning, as proposed by AlphaCast, opens up exciting avenues for more adaptable and context-aware forecasting systems. The rise of foundation models like TOTO and TempoPFN, with their zero-shot capabilities and synthetic pre-training, signals a shift towards highly generalizable models that can perform across diverse tasks without extensive fine-tuning. The recognition of interpretability, exemplified by CaReTS and counterfactual explanations for multivariate time series forecasting with exogenous variables (“Counterfactual Explanation for Multivariate Time Series Forecasting with Exogenous Variables” by Keita Kinjo from Kyoritsu Women’s University), will foster greater trust and adoption in critical applications, particularly in regulated industries like finance and healthcare. Looking ahead, the focus will likely remain on developing hybrid models, leveraging multi-modal data, enhancing robustness to non-stationarity and missing values, and further integrating human expertise to create truly intelligent forecasting systems. The future of time series forecasting is dynamic, collaborative, and increasingly insightful!