Frozen time-series models no longer ship as-is: a coding agent rewrites the deployment pipeline and tops six metrics across 198 tasks
Synopsis
Time-series foundation models forecast well zero-shot, yet deploying them still means hand-building input and output pipelines. TimeRSI has a coding agent repeatedly rewrite preprocessing, features, context selection and post-processing around a frozen forecaster, using errors on historical forecast windows to guide each revision, then picks among the resulting pipelines at deployment by similar past conditions. Across 198 tasks on TIME and FEV-bench it beats all published submissions on six point and probabilistic metrics; against direct inference with the same frozen TimesFM 3.0, TIME MASE falls 1.52% and CRPS 1.61%.
Interpretation
The deployment layer itself becomes the thing that improves automatically, while the forecasting model's weights stay fixed. Putting a time-series foundation model to work on a real task previously meant hand-writing task knowledge such as holidays or tidal cycles into input and output designs; direct inference left that knowledge and recurring forecast errors unaddressed. On all 98 TIME tasks and all 100 FEV-bench tasks, TimeRSI scores best on six point and probabilistic metrics among the 32 TIME and 31 FEV-bench submissions compared.
The agent translates a task description into executable pipeline changes and uses historical errors to decide what to change next. Edits span four stages, preprocessing, feature construction, context selection and post-processing, so the agent can change both what the model sees and how outputs are combined and corrected. Earlier historical windows drive iterative code revision and a later block of windows selects the deployment candidates; the default pairs TimesFM 3.0 with Claude Code over at most six search iterations.
At deployment the system chooses a pipeline per target for each new forecast instead of fixing one best-on-average pipeline. The selector describes current conditions with numerical features of recent history and the reference forecast, retrieves similar windows whose outcomes are fully observed, and compares each candidate's error against the reference on them. In a controlled comparison, a search proposing the same number of candidates without feedback-guided revision underperforms direct inference on both benchmarks, while TimeRSI improves on it.
Perspective
The demonstrated scope is offline evaluation on two public benchmarks, 98 TIME tasks and 100 FEV-bench tasks, with a frozen TimesFM 3.0 as the default backbone and Claude Code as the default coding agent. The system relies on clear task and variable descriptions to connect numerical patterns with domain knowledge; case studies cover supply chain, parking, ocean currents, solar counters, an auto-production seasonal factor, COVID reports and atmospheric variables.
Gains on historical windows do not guarantee gains on future forecasts, and distribution shift remains an open problem in time-series forecasting; the case-level task improvements are aggregate and not every country or window improves. Comparisons use each baseline's published configuration rather than a common compute budget, and the FEV-bench entry is pending.