Time Series Paper Digest (Aug 23–Sep 5 2025)

Uncertainty Awareness on Unsupervised Domain Adaptation for Time Series Data

Unsupervised domain adaptation methods aim to generalize effectively to unlabeled test data, especially facing distribution shift between training and test sets common in time‑series data. This paper proposes combining multi‑scale feature extraction with uncertainty estimation to improve model generalization and robustness across domains. The approach first employs a multi‑scale mixed‑input architecture that captures features at different scales, increasing training diversity and reducing feature discrepancy between source and target domains. Built on this architecture, the authors introduce an evidence‑based uncertainty‑aware mechanism that applies a Dirichlet prior to labels to facilitate target prediction and uncertainty estimation. The uncertainty‑aware mechanism aligns features with the same label across domains, enhancing adaptation and yielding significant performance gains on the target domain. Moreover, the model achieves lower Expected Calibration Error (ECE), indicating better confidence calibration. Experiments on several benchmark datasets show state‑of‑the‑art results, demonstrating the method’s effectiveness for unsupervised domain adaptation of time‑series data.

GateTS: Versatile and Efficient Forecasting via Attention‑Inspired Routed Mixture‑of‑Experts

Accurate univariate forecasting remains a pressing need in real‑world systems such as energy markets, hydrology, retail demand, and IoT monitoring, where signals are often intermittent and prediction horizons span short‑ and long‑term. Although Transformers and Mixture‑of‑Experts (MoE) architectures have gained popularity for time‑series forecasting, a key gap persists: MoE models typically require complex training involving a primary prediction loss, an auxiliary load‑balancing loss, and careful routing/temperature tuning, which hinders practical adoption. This paper proposes a simplified architecture for univariate time‑series forecasting that effectively handles both short‑ and long‑term horizons, including intermittent patterns. The method combines sparse MoE computation with a novel attention‑inspired gating mechanism that replaces the conventional single‑layer softmax router. Extensive empirical evaluation shows that the gating design naturally promotes balanced expert utilization and achieves superior forecasting accuracy without the auxiliary load‑balancing loss required by classic MoE implementations. The model attains better performance while using far fewer parameters; for example, it outperforms state‑of‑the‑art Transformers such as PatchTST that need many parameters. Moreover, experiments across multiple datasets confirm that, compared with standard MoE, the proposed gating yields higher computational efficiency than LSTM for both short‑ and long‑term forecasting, enabling cost‑effective inference.

An Empirical Evaluation of Factors Affecting SHAP Explanation of Time Series Classification

Explainable AI (XAI) has become increasingly important for understanding and attributing predictions of complex time‑series classification (TSC) models. Among attribution methods, SHapley Additive exPlanations (SHAP) is widely regarded as effective, but its computational cost grows exponentially with the number of features, limiting its use on long time series. Recent work suggests that segmenting features and aggregating a single attribution value for a group of consecutive time points can dramatically reduce SHAP runtime. However, selecting the optimal segmentation strategy remains open. This study evaluates eight different time‑series segmentation algorithms to examine how segment composition affects explanation quality. Two established XAI evaluation metrics, InterpretTime and AUC Difference, are used for assessment. Experiments on multivariate (MTS) and univariate (UTS) time series reveal that the number of segments influences explanation quality more than the specific segmentation method. Notably, equal‑length segmentation consistently outperforms most custom segmentation algorithms in most cases.

Text Reinforcement for Multimodal Time Series Forecasting

Recent time‑series forecasting (TSF) research employs multimodal inputs, such as text and historical time‑series data, to predict future values. These works focus on developing advanced techniques to fuse textual information with time‑series data, achieving promising results. However, they rely on high‑quality text and series inputs, and in some cases the text fails to accurately or fully capture information carried by the historical series, leading to unstable multimodal TSF performance. Enhancing the textual modality to improve multimodal TSF is therefore necessary. This paper proposes a Text Reinforcement model (TeR) that generates strengthened text to address weaknesses in the original text, and then uses this reinforced text to aid the multimodal TSF model’s understanding of the series, improving forecasting performance. To guide TeR toward higher‑quality reinforced text, the authors design a reinforcement‑learning scheme that assigns rewards based on each reinforced text’s impact on multimodal TSF performance and its relevance to the forecasting task. TeR is optimized accordingly to improve the quality of generated text and boost TSF performance. Extensive experiments on real‑world benchmark datasets across various domains demonstrate the method’s effectiveness.

Online Time Series Prediction Using Feature Adjustment

Time‑series forecasting is important across many domains, but distribution shift poses a major challenge. In online deployment scenarios, the issue is amplified because data arrive sequentially, requiring the model to continuously adapt to evolving patterns. Existing online learning methods for time series focus on two aspects: selecting parameters to update (e.g., the final layer weights or adapter modules) and designing appropriate update strategies (e.g., using recent batches, replay buffers, or averaged gradients). The authors challenge the conventional parameter‑selection approach, arguing that distribution shift originates from changes in underlying latent factors that affect the data. Consequently, updating the feature representations of these latent factors may be more effective. To address the critical problem of delayed feedback in multi‑step prediction—where true values arrive much later than predictions—the authors introduce ADAPT‑Z, which continuously tracks automatic adjustment increments in a Z space. ADAPT‑Z employs an adapter module that combines the current feature representation with historical gradient information, enabling parameter updates that are unaffected by delay. Extensive experiments show that the method consistently outperforms non‑adaptive baseline models and surpasses state‑of‑the‑art online learning methods on multiple datasets.

One‑Embedding‑Fits‑All: Efficient Zero‑Shot Time Series Forecasting by a Model Zoo

The rise of time‑series foundation models (TSFMs) has dramatically advanced zero‑shot forecasting, enabling prediction on unseen series without task‑specific fine‑tuning. Extensive research shows that no single TSFM universally outperforms others, as different models exhibit preferences for different temporal patterns. This diversity suggests an opportunity to leverage the complementary strengths of TSFMs. To this end, the authors propose ZooCast, which characterizes each model’s unique forecasting advantage. ZooCast can intelligently assemble the current TSFMs into a model zoo and dynamically select the best model for each forecasting task. The key innovation is the One‑Embedding‑Fits‑All paradigm, which builds a unified representation space where each model in the zoo is represented by an embedding, enabling efficient similarity matching across all tasks. Experiments demonstrate that ZooCast achieves strong performance on the GIFT‑Eval zero‑shot forecasting benchmark while retaining the efficiency of a single TSFM. In real‑world scenarios, as models are released sequentially, the framework can seamlessly incorporate new models to obtain incremental accuracy gains with negligible overhead.

CCE: Confidence‑Consistency Evaluation for Time Series Anomaly Detection

Time‑series anomaly‑detection metrics are essential tools for model evaluation, yet existing metrics suffer from several limitations: insufficient discriminative power, strong dependence on hyperparameters, sensitivity to perturbations, and high computational cost. This paper introduces a new evaluation metric—Confidence‑Consistency Evaluation (CCE)—which jointly measures prediction confidence and uncertainty consistency. By employing Bayesian estimation to quantify uncertainty of anomaly scores, the authors construct global and event‑level confidence and consistency scores for model predictions, yielding a concise CCE metric. Theoretically and experimentally, the authors demonstrate that CCE possesses strict boundedness, Lipschitz robustness to score perturbations, and linear time complexity. Additionally, they establish RankEval, a benchmark for comparing the ranking ability of different metrics.