TIME
收藏资源简介:
TIME是由南洋理工大学等机构联合构建的新一代时间序列预测基准,包含50个原创数据集和98个任务,严格避免数据泄露风险。数据集通过政府公开数据、产业合作等多渠道获取,并采用人机协同的自动化筛查流程确保数据质量,涵盖多元领域和频率的时序特征。其创新性地提出模式级评估视角,通过结构分解和时序特征编码实现跨数据集的通用能力分析,旨在为零样本时序基础模型提供更贴近真实场景的评估框架。
TIME is a next-generation time series forecasting benchmark jointly developed by Nanyang Technological University and other institutions. It includes 50 original datasets and 98 tasks, strictly avoiding data leakage risks. The datasets are collected from multiple channels such as government open data and industrial collaborations, and an automated human-machine collaborative screening process is adopted to ensure data quality, covering time series features across diverse domains and sampling frequencies. It innovatively proposes a pattern-level evaluation perspective, enabling cross-dataset general capability analysis through structural decomposition and time series feature encoding, aiming to provide a more real-scenario-aligned evaluation framework for zero-shot time series foundation models.

- 1It's TIME: Towards the Next Generation of Time Series Forecasting Benchmarks南洋理工大学; 格里菲斯大学; DataDog; 清华大学; Squirrel Ai Learning · 2026年



