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

1 Introduction

• Two deployment challenges motivate feedback and selection: plausible pipeline changes can yield different accuracy, and a pipeline that performs well on average may not suit changing conditions.Historical forecast errors help evaluate and refine choices such as whether calendar information enters through model inputs or a separate predictor.Source: §1 · p. 1
• TIMERSI applies recursive self-improvement to deployment pipelines around frozen TSFMs, using task knowledge to guide what to try and forecasting feedback to guide revision and use.The agent edits input transformations, features, and forecast corrections while keeping model weights fixed.Source: §1 · p. 1, Figure 1 · p. 2
• TIMERSI iteratively evaluates code revisions on historical forecasting windows, returns errors to the agent, and selects among resulting pipelines for each new forecast using performance under similar conditions.Task knowledge guides revisions, while completed forecasts under similar conditions inform deployment choices.Source: §1 · p. 1
• Across 198 tasks on TIME and FEV-bench, TIMERSI achieves the best scores among existing submissions across six point and probabilistic forecasting metrics.Ablations favor feedback-guided improvement over non-iterative pipeline search with the same number of candidates; further experiments report gains with different coding agents and TSFMs.Source: §1 · p. 1
• TSFM deployment work has studied direct deployment and agentic deployment for realistic applications via recursive self-improvement.The cited direct-deployment works include Kamarthi and Prakash (2024), Das et al. (2024), Woo et al. (2024), and Ansari et al. (2024); the RSI works include Lee et al. (2026) and Srikanth et al. (2026).Source: §1 · p. 2

2 TIMERSI: Self-Improved Deployment for Time-Series Foundation Models

• TIMERSI automates pipeline improvement and selection for new forecasts while keeping the TSFM’s weights fixed.A code agent develops pipelines through repeated execution, evaluation, and revision; an adaptive selector uses past outcomes under similar forecasting conditions.Source: §2 · p. 2
• The agent writes code across preprocessing, feature construction, context selection, and post-processing to change model inputs and outputs.It can transform observations, add task-knowledge variables, select histories or variable groups, combine forecasts, fit auxiliary predictors, and correct systematic errors.Source: §2 · p. 2
• Earlier historical windows form an adaptation set for code revision, while later windows form a selection set for choosing deployment candidates.The system runs candidates from each window’s forecast origin and scores predictions for the following H steps.Source: §2 · p. 2
• Starting from direct TSFM inference, the agent proposes code revisions and uses forecast scores, window-level errors, and execution logs to decide what to refine, combine, or replace.Earlier code and evaluation records remain accessible so later revisions can build on successful ideas and address observed errors.Source: §2 · p. 2
• After the search budget is exhausted, TIMERSI evaluates candidates on later selection windows and retains the highest-ranked pipelines together with the reference pipeline.The deployment set provides several task-specific strategies for subsequent forecasts.Source: §2 · p. 2
• For each target and forecast origin, the selector describes current conditions using recent-history and reference-forecast features, then compares retained pipelines on similar completed windows.The selector weights retrieved windows by similarity and chooses the pipeline with the lowest weighted error relative to the reference.Source: §2 · p. 3

3 Experiments

• Across 111,177 TIME and 235,039 FEV-bench forecast instances, evaluation reports normalized MASE and CRPS for TIME and Skill Scores for MASE, SQL, WQL, and WAPE for FEV-bench.Comparisons use published results from all 32 TIME submissions and 31 FEV-bench submissions.Source: §3 · p. 3
• TIMERSI ranks first among submissions across six metrics on TIME and FEV-bench; against direct frozen TimesFM 3.0, it reduces TIME MASE by 1.52% and CRPS by 1.61%.On FEV-bench, it raises MASE Skill by 2.79 percentage points and SQL Skill by 2.26 points.Source: §3 · p. 4
• Non-iterative search with the same number of candidate pipelines underperforms direct TimesFM 3.0 inference, whereas TIMERSI improves performance on both benchmarks.The comparison supports feedback-guided pipeline revision rather than only proposing and selecting candidates.Source: §3 · p. 4, Table 1 · p. 4
• With Antigravity (Gemini 3.8 Flash) fixed, normalized MASE/CRPS are 0.637/0.532, outperforming compared external submissions on TIME.With Claude Code fixed, TIMERSI reduces MASE/CRPS from 0.662/0.556 to 0.647/0.545 for Chronos-2 and from 0.669/0.572 to 0.641/0.549 for LPTM; LPTM’s FEV-bench MASE/SQL Skill rises from 30.70/43.20 to 34.63/46.33.Source: §3 · p. 4
• Case-study pipelines reduce task-level MASE by 2.86–6.70% over direct TimesFM 3.0 by translating domain knowledge into task-specific changes.Examples include calendar features for parking, tidal decomposition for ocean currents, history-supported holiday corrections for demand, and daily-reset shape corrections for solar counters; feedback also prompted a strategy change in the parking case.Source: §3 · p. 4, Figure 3 · p. 4

• TIMERSI automates task-specific deployment through iterative pipeline improvement and adaptive selection around frozen TSFMs.Source: §4 · p. 5
• The system’s deployment approach centers on improving pipelines and selecting among them while keeping the forecasting model’s weights fixed.Source: §4 · p. 5
• On TIME and FEV-bench, TIMERSI ranks first across six forecasting metrics compared with competitive external submissions.Source: §4 · p. 5

Limitations

• TIMERSI relies on clear task and variable descriptions to connect numerical patterns with domain knowledge; historical-window gains do not guarantee future gains under distribution shift.Source: Limitations · p. 5

A Additional Related Work

• TIMERSI complements numerical-backbone and semantic-context research by adapting executable pipelines around frozen forecasters and coupling chronological feedback with strategy selection.Source: A Additional Related Work · p. 7

B Pipeline Interface

• The pipeline interface separates preprocessing, feature engineering, context selection, and postprocessing; prediction inputs exclude target values beyond each window’s origin.Source: B Pipeline Interface · p. 7

C Historical Selection Details

• The selector represents targets with up to 96 recent observations and a reference median forecast, then retrieves up to 12 similar completed windows using standardized features and distance-based weights.Source: C Historical Selection Details · pp. 7–8
• Candidate ranking uses selection-window MASE with a quantile-score tiebreaker, retaining three ranked candidates plus the reference when needed.Source: C Historical Selection Details · p. 8

D Discovered Deployment Strategies

• The illustrated deployment programs span task-specific features, domain-informed representations, business rules, measurement structure, and auxiliary predictors; these selected examples are not aggregate benchmark results.Source: D Discovered Deployment Strategies · p. 8

• On July 4, the first customer’s observed orders were zero; direct inference forecast 195.23, while TIMERSI forecast zero.Source: §D.1 · p. 8
• The holiday rule required at least four prior observations for a target, with at least 80% zero, and then set all nine forecast quantiles to zero.Source: §D.1 · p. 8
• Historical feedback narrowed the day-after-holiday adjustment to July 4 rather than dates such as Black Friday; the first holiday-aware correction delivered the main adaptation gain.Source: §D.1 · p. 8
• 6.70%: the complete pipeline reduced MASE across 36 targets and 12 windows, with zero predictions on four holidays where all 144 target values were zero.Source: §D.1 · p. 8

• After a calendar-predictor blend increased adaptation error, the agent supplied 22 calendar features to the frozen model, including periodic terms, weekend indicators, and weekend–time-of-day interactions.Source: §D.2 · p. 8
• On all 354 car parks, MASE fell from 0.678730 to 0.641658 and CRPS from 0.046609 to 0.044105.Source: §D.2 · p. 8, Table 2 · p. 9
• Parking MASE comparisons use separate adaptation and selection windows, and revision 4—not the last revision or adaptation winner—was deployed.Source: Table 4 · p. 9

D.3 Ocean Currents: Using Tidal Knowledge in Model Inputs

• 4.18% lower MASE and 0.64% lower CRPS: the tidal decomposition improves TimesFM forecasts across ten series, six variables, and five 72-hour windows.The agent selects distinguishable tidal constituents, fits their amplitudes and phases, and backtests the fit against a constant-mean predictor before subtracting and restoring the tidal component.Source: §D.3 · p. 9

D.4 Solar: Respecting a Daily-Reset Pattern

• 2.86% lower MASE and 2.66% lower CRPS: solar shape projection is selected across three targets and 35 72-hour windows.The program estimates a normalized daily profile from the previous 28 valid days, then fits a level to model output and reconstructs each forecast path.Source: §D.4 · p. 9

D.5 Auto Production: Building a Calendar-Aware Specialist

• Across five 12-month test windows, TIMERSI achieves MASE 0.233980 versus 0.756537 for direct inference, a 69.07% reduction; six further final-search revisions reduce test MASE by 25.25% from 0.313005.Source: §D.5 · p. 10

D.6 COVID Reports: Revising an Overextended Trend

• After a logarithmic transformation raises adaptation MASE from 3.9173 to 4.1816, the agent inspects errors and revises the overextended upward trend.Later revisions blend the model with a damped-trend predictor, address reporting spikes and repeated values, and adjust damping based on recent growth duration.Source: §D.6 · p. 10
• Across nine countries and five 30-day windows, the selected COVID pipeline reduces task-level MASE by 7.91% and CRPS by 15.62%, not necessarily in every country or window.Selection favors revision 6 with milder damping over revision 5, despite revision 5 having the lowest adaptation error.Source: §D.6 · p. 10, Table 5 · p. 10

D.7 Atmospheric Variables: Separating the Daily Cycle

• 2.22% lower MASE across four variables and 30 official 24-hour windows: the selected reversible transformation separates daily cycles from the forecast signal.The agent fits daily harmonics in overlapping historical segments, forecasts the residual with TimesFM, then restores the future cycle.Source: §D.7 · p. 10

Evaluation scope

• TIME covers 98 tasks from 50 datasets and 111,177 forecast instances, while FEV-bench covers 100 tasks, 1,308 specified windows, and 235,039 instances.The evaluation follows released task definitions; available past-only covariates stop at the origin, while known-future covariates may span the horizon.Source: E Benchmark Protocol and Full Results · p. 10

Pipeline configuration and search budget

• The default setup uses frozen TimesFM 3.0 weights and Claude Code, with a candidate archive that also includes programs developed with Claude Opus 5.Search allows up to six iterations, one evaluated candidate per iteration, with a 256-turn maximum and a 90-minute timeout per session.Source: E Benchmark Protocol and Full Results · pp. 10–11

Chronological feedback

• Historical adaptation windows guide code revisions, later windows select deployment candidates, and only completed outcomes can inform forecasts at deployment.The forecast under evaluation cannot supply its own target values to the pipeline or selector; the setup is task-adapted deployment with a frozen backbone, not zero-shot operation of the whole system.Source: E Benchmark Protocol and Full Results · p. 11

Comparison scope

• Public baseline scores use the October 10, 2026 benchmark snapshot and published configurations, not a common compute budget.The comparison includes pretrained forecasters and, where available, task-trained and statistical methods; the FEV-bench TIMERSI entry was pending merger into the public files.Source: E Benchmark Protocol and Full Results · p. 12

Generalization across agents and TSFMs

• On all 100 FEV-bench tasks, TIMERSI with LPTM scores SQL, MASE, WQL, and WAPE Skill of 46.33%, 34.63%, 50.96%, and 38.58%, versus 43.20%, 30.70%, 47.60%, and 34.18% for direct LPTM inference.The comparison tests a different forecasting backbone within the deployment framework.Source: E Benchmark Protocol and Full Results · p. 12

E.1 Additional FEV-bench Metrics

• TIMERSI ranks first across all four FEV-bench metrics, including WQL Skill of 55.759% and WAPE Skill of 44.788%.The comparison covers all 100 tasks and 31 publicly visible baselines; higher Skill Scores are better.Source: §E.1 · p. 13, Table 9 · p. 14

F Use of AI Assistants

• AI assistants supported language refinement, manuscript organization, coding, and experimental workflows; the authors guided the study and take responsibility for its design and reported results.Source: F Use of AI Assistants · p. 13