TSFLab: A Self-Improving Agent-Agnostic Workbench for Time Series Forecasting Research

TSFLab. A Self-Improving Agent-Agnostic Workbench for Time Series Forecasting Research. By ChengAo Shen and nine others, from the University of Houston, the University of Illinois Urbana-Champaign, and five other institutions. Published on the twenty-ninth of October.

Publications and products on time series forecasting, TSF, keep multiplying, and researchers carry a growing burden just analysing existing work. Not every publication releases code: in the surveyed period, only about twenty percent of arXiv TSF papers had a code link. Code that is released still needs manual maintenance. Researchers end up constantly adapting or reimplementing earlier work, and independent implementations can differ in experimental settings and are hard to keep up to date.

One response is to hand this to AI agents, for benchmarking and even for autoresearch, such as developing new TSF models. Applying existing coding agents like Claude Code and Codex directly falls short on three counts: the costs are prohibitive, they scale poorly to large research projects such as benchmarking, and they lack any self-maintenance.

TSFLab is a workbench placed between the agent and the research. It is agent-agnostic, so it works with any compatible coding agent, and it equips that agent with shared resources and tools. The aim is research at low cost, together with continuous maintenance of the research resources themselves, which is the sense in which the workbench is self-improving.

Rather than each agent starting from scratch on every task, compatible agents draw on what the workbench already holds, and the workbench keeps those resources maintained over time. The experiments exercise it on three kinds of work: reproduction, benchmarking, and the development of new methods.

For reproduction and benchmarking, TSFLab improves task completion by between twenty and one hundred and sixty-three percent across baselines. It also cuts fresh tokens per completed task by between eleven and sixty-six percent.

A pilot study of agent-driven method development points the same way. With TSFLab, forecasting accuracy improves, while using between thirty-two point eight and seventy-three point six percent fewer fresh tokens. Development time is between nine point two and sixty-eight point eight percent shorter than in the other settings.

Instead of a smarter agent, a better-stocked bench. The proposal is that the cost and upkeep problems of agent-run research sit in the shared environment as much as in the agent, and that fixing the environment once lets whichever agent comes next do more of the work for fewer tokens.