time-series

Mistral AI 发布 Timestral 时间序列基础模型系列（5M/28M/123M/209M），209M 版本在 GIFT-Eval、TIME 和 fev-bench 上实现零样本预测的最先进精度，并开源代码。

论文提出 HART，一种混合注意力与 RWKV 的时间序列基础模型，通过交替注意力和 RWKV 模块解决非自回归多步预测中 RWKV-7 的长程退化问题（Placeholder Idling Decay）。仅用 36M 参数（约为 Toto-2.0-1B 的 1/28）在 GIFT-Eval 上达到可媲美的长程 MASE（0.752 vs 0.754）。

NVIDIA researchers propose NV-Tesseract-AD, a curriculum-enhanced conditional diffusion framework with a Transformer backbone for multivariate time-series anomaly detection, achieving ~3.7× F1 improvement over non-curriculum baselines and strong zero-shot transfer to EasyTSAD benchmarks.

KITE 提出一种知识注入的时间序列编码器，将工程师的领域先验（语义先验）通过可切换的协变量与注意力偏置机制注入自监督事件预测器，在稀有事件预测上显著优于无先验基线，并支持跨域迁移。

这篇论文将事件序列预测重新表述为条件生成建模问题，提出了一个同时衡量预测准确性和分布真实性的基准，并引入Wasserstein BaryBooster算法，将扩散模型的真实感转化为预测精度。

GreptimeDB 通过将 TopK 运行时边界下推到扫描层（基于 DataFusion 的动态过滤器），把一个 ~5-6B 行 trace 表上的 ORDER BY ... LIMIT 10 查询从 28.9 秒优化到 0.21 秒。

Lapras提出了一种针对时间序列语言模型(TSLM)的后训练框架，通过教师-学生自蒸馏实现连续隐空间推理，在五个时间序列问答基准上平均F1提升最高10.79%，同时生成token减少23.9倍。

该论文发现 ECG 到 PPG 的跨模态蒸馏中，冻结 ECG 教师目标的 40–59% 是与个体身份相关的 per-recording 均值（trait），学生模型往往记住身份而非生理状态；作者提出 fixed-effects distillation，通过减去每条记录的均值使 trait 精确抵消，使 within-person 状态一致性提升一倍以上，并在两个骨干网络和两个额外数据库上验证了效果。

GreptimeDB v1.2.0 正式发布，新增 JSON2 结构化类型（支持 SQL 嵌套路径查询）、支持 Prometheus Remote Write 2.0，并带来可续传的并行 Import/Export V2，面向日志与 AI 代理事件场景。

该论文提出一种预处理感知的基准测试框架，在M4数据集的29,000条时间序列上评估11个预测模型与16种可逆预处理管线，发现为每条序列优化预处理可带来约27%至87%的性能提升，并揭示现有基准存在结构性预处理偏差。

The paper introduces Market-1T, a ~1-trillion observation U.S. equity dataset at 1 Hz resolution, a rigorous evaluation protocol, and a large-scale study of 18 encoder-training strategies for financial representation learning across 2008–2025 market regimes, establishing a foundation for financial world models like DINO-WM, V-JEPA 2, and LeWM.

本文提出一种可解释的机器学习分析，利用多中心纵向临床队列预测参与正念干预后患者在12周和24周的BDI-II抑郁评分，发现基线严重程度仍是最强预测因子，而短期结局更依赖临床与医院背景，长期结局则更依赖行为依从性和人口统计因素。

Parseable is an open-source, object-store-native observability datalake that claims to handle 100M time-series per minute, unifying logs, metrics, and traces in a single binary with SQL and natural-language querying, plus AI tooling like MCP connectors and the OTex observability agent.

A follow-up project showing a tiny 31K-parameter encoder-only transformer trained only on synthetic T1DM simulator data that zero-shot predicts blood glucose from real CGM traces, with LoRA fine-tuning and ExecuTorch deployment on an Android app.

An arXiv paper presenting a validated data onboarding pipeline for AI demand forecasting on U.S. building meter data, reporting a controlled evaluation and a corrected negative result regarding its effectiveness.

TRACE is a reproducible benchmark of 7,300 zone–day instances pairing U.S. electricity prices with official operational text reconstructed at each forecast cutoff to prevent leakage, showing that text yields a median 7.4% reduction in upper-tail pinball loss across time-series foundation models.

The paper introduces SSU-LSF, the first machine-unlearning framework for Mamba-family structured state space models applied to land surface forecasting, using EKFac influence functions and projected gradient ascent to remove non-stationary confounding bias from NDVI, LST, and crop phenology predictions at 8.4x lower cost than full retraining.

A new paper proposes Dual-Stream OSSE-LSTM, an episodic metric-learning framework combining an Omni-Scale CNN with a Bidirectional LSTM for label-efficient time series classification, plus Counterfactual Integrated Gradients (C-IG) for explainable prototype-based decisions, achieving strong and stable accuracy on 19 UCR datasets with only 2-8 labeled examples per class.

TIDES 提出一种选择性 SSM 变体，将输入依赖性从步长移到对角状态矩阵上，从而在保持每 token 表达能力的同时原生处理不规则时间戳，并在 UE 分类与 PhysiomeODE 回归基准上取得最佳平均排名。

A NeurIPS 2026 paper introducing DynaBase, a minimal interpretable architecture for zero-shot reconstruction of dynamical systems built from just a one-parameter piecewise affine map and a context selector. Despite its simplicity, it outperforms major time-series and dynamical-systems foundation models on long-term statistics and short-term predictions, and offers a tractable mathematical basis for analyzing such models.