Time Series Paper Digest: Extreme Event Prediction, Multimodal Fusion & Anomaly Detection

xTime: Extreme Event Prediction with Hierarchical Knowledge Distillation and Expert Fusion

Paper link: http://arxiv.org/pdf/2510.20651v1

Authors: Quan Li, Wenchao Yu, Suhang Wang, Minhua Lin, Lingwei Chen, Wei Cheng, Haifeng Chen

Abstract: Extreme events such as floods, heatwaves, or acute medical incidents occur infrequently but have severe consequences. Existing time‑series models optimise overall performance and struggle with these rare events because of data imbalance and the neglect of informative intermediate patterns. xTime addresses this by first using hierarchical knowledge distillation to transfer knowledge from models trained on less‑rare events to improve prediction of rarer events. It then adds a mixture‑of‑experts (MoE) module that dynamically selects and fuses outputs from experts specialised for different rarity levels. Experiments on multiple datasets show consistent gains, with extreme‑event accuracy improving between 3 % and 78 % over baselines.

MGTS‑Net: Exploring Graph‑Enhanced Multimodal Fusion for Augmented Time Series Forecasting

Paper link: http://arxiv.org/pdf/2510.16350v1

Authors: Shule Hao, Junpeng Bao, Wenli Li

Abstract: Recent work integrates multimodal features (temporal, visual, textual) into time‑series models but faces three challenges: insufficient fine‑grained temporal pattern extraction, sub‑optimal multimodal fusion, and limited adaptability to dynamic multi‑scale features. MGTS‑Net proposes three components: (1) a multimodal feature extraction (MFE) layer that tailors encoders for each modality to capture fine‑grained temporal patterns; (2) a multimodal feature fusion (MFF) layer that builds a heterogeneous graph modelling intra‑modal temporal dependencies and cross‑modal alignments, then dynamically aggregates multimodal knowledge; (3) a multi‑scale prediction (MSP) layer that adaptively weights and fuses short‑, medium‑ and long‑term predictors. Extensive experiments demonstrate lightweight design and high efficiency with strong performance.

Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection

Paper link: http://arxiv.org/pdf/2510.16511v1

Authors: Dongchan Cho, Jiho Han, Keumyeong Kang, Minsang Kim, Honggyu Ryu, Namsoon Jung

Abstract: Real‑world multivariate time‑series anomalies are rare and often unlabeled. Existing methods rely on increasingly complex benchmark‑adjusted architectures that detect only parts of anomalous segments and over‑state performance. The authors introduce OracleAD, a simple yet interpretable unsupervised framework. OracleAD encodes each variable’s past into a causal embedding, jointly predicts the current point, and reconstructs the input window, thereby modelling temporal dynamics. The embeddings pass through a self‑attention module that projects them into a shared latent space, capturing spatial relationships that evolve with time. These projected embeddings are aligned with a stable latent structure (SLS) representing normal relationships. Anomalies are scored using both prediction error and deviation from SLS, enabling fine‑grained diagnosis per time point and across variables. Because any SLS deviation stems from violated temporal causality, OracleAD can pinpoint root‑cause variables at the embedding level. Experiments on multiple real datasets and evaluation protocols achieve state‑of‑the‑art results while preserving interpretability via SLS.

Abstain Mask Retain Core: Time Series Prediction by Adaptive Masking Loss with Representation Consistency

Paper link: http://arxiv.org/pdf/2510.19980v1

Authors: Renzhao Liang, Sizhe Xu, Chenggang Xie, Jingru Chen, Feiyang Ren, Shu Yang, Takahiro Yabe

Abstract: Time‑series forecasting is critical in energy management and financial markets. Although deep models (MLP, RNN, Transformer) have advanced performance, the common “long‑sequence information gain” assumption is flawed. Systematic experiments reveal a counter‑intuitive phenomenon: truncating history appropriately can improve accuracy, indicating that models learn redundant features (noise or irrelevant fluctuations) that hurt signal extraction. Guided by the information‑bottleneck theory, the authors propose Adaptive Masking Loss with Representation Consistency (AMRC). AMRC comprises (1) a dynamic masking loss that adaptively identifies highly discriminative time fragments to guide gradient descent, and (2) a representation‑consistency constraint that stabilises the mapping among inputs, targets, and predictions. Experiments show AMRC suppresses redundant feature learning and significantly boosts model performance, challenging traditional assumptions about temporal modelling.