Impermanent
收藏资源简介:
Impermanent是由TimeCopilot团队联合多所顶尖学术机构构建的动态时序预测基准数据集,聚焦GitHub开源生态中400个高星仓库的四种开发活动(议题创建、拉取请求、推送事件和新增关注者)。该数据集以GH Archive事件流为原始数据源,通过滚动窗口机制实现每日更新,包含小时级至月级的非平稳时间序列,具有显著的概念漂移特性。数据集构建采用预序评估框架,严格隔离训练与测试时段以避免数据泄漏,旨在评估模型在开放环境下的时序泛化能力。其核心应用场景是验证时间序列基础模型在真实动态系统中的持续预测性能,解决传统静态评估导致的过拟合和性能虚高问题。
Impermanent is a dynamic time series prediction benchmark dataset jointly constructed by the TimeCopilot team and several leading academic institutions. It focuses on four development activities (issue creation, pull requests, push events, and new follower additions) of 400 high-star repositories within the GitHub open-source ecosystem. The dataset uses GH Archive event streams as its raw data source and is updated daily via a rolling window mechanism. It contains non-stationary time series with granularities spanning from hourly to monthly, exhibiting significant concept drift. The dataset is built using a temporal-split evaluation framework that strictly isolates training and test time periods to prevent data leakage, with the goal of evaluating the temporal generalization ability of models in open environments. Its core application scenario is to validate the sustained prediction performance of time series foundation models in real-world dynamic systems, addressing the issues of overfitting and inflated performance induced by traditional static evaluations.

- 1Impermanent: A Live Benchmark for Temporal Generalization in Time Series Forecasting牛津大学; ELLIS研究所·蒂宾根; Mila·魁北克人工智能研究所·蒙特利尔大学; 亚马逊网络服务 · 2026年



