TimeSage-EV: Live Benchmark for Agentic Time Series Analysis in Evolving Environments
A team of researchers has introduced TimeSage-EV, an innovative benchmark designed to assess time series analysis in dynamic contexts. This new framework tackles a significant limitation in current assessment methods, which often depend on static data snapshots and overlook temporal accuracy. TimeSage-EV encompasses 60 real-world scenarios across six sectors, featuring 1,485 question-answer pairs collected from February 2023 through May 2026, with varying release intervals. During evaluations, large language model agents interpret time series data and reports, while withheld targets represent the true outcomes. The benchmark prioritizes state identification and reasoning, as described in the recent arXiv publication.
Key facts
- TimeSage-EV is a live benchmark for agentic time series analysis in evolving environments.
- It tracks 60 real institutional scenarios across 6 domains.
- It comprises 1,485 scenario-period QA pairs from Feb 2023 to May 2026.
- Release cadences include monthly, weekly, daily, and irregular.
- At each period, LLM agents receive time series data and source reports; withheld target release provides ground truth.
- Evaluates state identification, data summarization, and outlook reasoning.
- Experiments include frontier LLM agents and TimeSage-1.0, a self-evolving agent.
- Paper available at arXiv:2608.14270.
Entities
Institutions
- arXiv