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

Abstract

Time-series foundation models provide strong zero-shot forecast, but deployment for real- world applications still requires task-specific designs for model inputs and outputs. Direct inference can leave domain knowledge and re- curring forecast errors unaddressed, while tai- lored pipelines require substantial human ex- pertise and effort. We introduce TIMERSI, a self-improving agentic deployment system around frozen forecasting models. A cod- ing agent translates task descriptions into ex- ecutable pipeline changes. It tests each re- vision on historical forecasting windows and uses observed errors to guide subsequent re- visions. During deployment, an adaptive se- lector chooses among the resulting pipelines using completed forecasts under similar con- ditions. Across 198 tasks on TIME and FEV- bench, TIMERSI outperforms all submissions on six point and probabilistic forecasting met- rics. Controlled comparisons support the value of feedback-guided revision over non-iterative pipeline search. Further experiments verify the generalization with different coding agents and forecasting models.