TimeRSI: Towards Self-Improved Agentic Deployment for Time-Series Foundation Models

TimeRSI. Towards Self-Improved Agentic Deployment for Time-Series Foundation Models. By Haoxin Liu and six others, from the Georgia Institute of Technology, the University of Southern California, and one other institution. Published on the fifteenth of October. Its code repository is worked out: about sixty source files, with a readme and a licence.

Time-series foundation models forecast well zero-shot, yet putting them to work on real applications still calls for task-specific designs around their inputs and outputs. Running the model directly can leave domain knowledge unused and recurring forecast errors uncorrected. Building a tailored pipeline fixes that, but costs substantial human expertise and effort. So the choice has been between a cheap deployment that ignores what is known about the task and a good one that someone has to build by hand.

TimeRSI hands the pipeline work to an agent. It is a self-improving deployment system built around forecasting models that stay frozen. A coding agent turns a task description into executable changes to the pipeline. The forecaster is never retrained. What gets revised is everything around it: the code that prepares what goes in and handles what comes out.

Each revision is tested on historical forecasting windows, and the errors it makes there steer the next revision. Once deployed, an adaptive selector chooses among the pipelines this process produced, guided by completed forecasts made under similar conditions. There are two loops, then. One runs before deployment and improves the pipelines by learning from historical errors. The other runs during deployment and decides which pipeline to trust, based on how each has done in comparable situations.

Across one hundred and ninety-eight tasks on TIME and fev-bench, TimeRSI outperforms all submissions on six point and probabilistic forecasting metrics. Controlled comparisons favour revision guided by feedback over a pipeline search that does not iterate. Further experiments show the approach generalizes to different coding agents and different forecasting models.

The model itself never changes here. Everything that improves sits around it, written by an agent and corrected by the system's own past mistakes. The comparison with pipeline search that does not iterate is the telling one: the gain is credited to the feedback, not merely to trying many pipelines.