Rethinking Agentic Model Orchestration for Time Series Forecasting

Tong Guan; Tianxiang Zhan; Shirui Pan; Ming Jin

cs.LGOct 21, 2026

Rethinking Agentic Model Orchestration for Time Series Forecasting

Authors: Tong Guan, Tianxiang Zhan, Shirui Pan, Ming Jin

Organizations: Griffith University · Zhejiang University

Abstract

Time series foundation models (TSFMs) perform unevenly across forecasting tasks. Recent work uses agentic model orchestration (AMO) to select or combine TSFMs, often outperforming individual forecasters. Yet these gains may reflect stronger or more complementary candidates rather than better orchestration. To disentangle these effects, we introduce FairAMO, a controlled evaluation protocol with frozen candidate forecasts. Experiments on GIFT-Eval show that AMO improves accuracy, but its gains are bounded by the candidate TSFMs. The best orchestrators reduce MASE by less than 6% relative to the best single TSFM across all tasks. These gains depend on complementarity (different candidates excel on different tasks) rather than the number of candidates. Adding members yields diminishing returns, and a redundant member can erase these gains. The value of historical evidence also saturates quickly. A few recent windows capture all gains, and further history adds none. These findings separate system-level forecasting gains from the contribution of orchestration itself.

Figures & tables

Figure 1: Effects of candidate strength and complementarity on orchestration performance.

Figure 2: Overview of FairAMO. (A) Each candidate TSFM is run once and its quantile forecasts are cached as training backtests, validation backtests and test forecasts, with the test labels hidden. (B) The orchestrator reads only the permitted evidence and commits to one model or to a set of weights over $K$ of the $M$ candidates. (C) The frozen decision is applied to the cached test forecasts and scored by the official GIFT-Eval protocol. (D) The 11 deployable orchestrators on the main pool under val+B10 (MASE, lower is better); \scriptsizeR⃝ , \scriptsizeE⃝ , and \scriptsizeO⃝ mark reference, external, and our re-implemented methods. Detailed results are in Appendix Table B1 .

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Reference
Method	MASE	$G$	HR (%)
Oracle (window-level, not deployable)	0.580	+0.117	100.0
Argmin Router	0.662	+0.035	29.8
Dev-Selected Best	0.702	$-$ 0.004	$-$ 3.7
Equal Weight	0.710	$-$ 0.013	$-$ 10.9
Random	0.731	$-$ 0.034	$-$ 29.2

Appendix

Table B1: Main-pool results under val+B10 , grouped as in Figure 2 (D) into reference, external and re-implemented methods.

G

is the gain over the best single TSFM and HR is the fraction of the 0.117 headroom realized, so HR

=G/0.117

. Bold marks the best deployable result. The oracle row may switch member per window and is a non-deployable ceiling.

Size	Best single TSFM	Ceiling	$H$	Argmin MASE	Argmin HR (%)
2	0.699	0.647	0.052	0.689	18.4
3	0.698	0.617	0.081	0.678	23.5
4	0.697	0.596	0.101	0.670	26.7
5	0.697	0.580	0.117	0.662	29.8

Appendix

Table B2: Effect of pool size, averaged over all subsets of each size, at the val+B10 budget. The ceiling is the window-level oracle and

H

is the pool headroom, the gap between the best single TSFM and that ceiling. HR is the fraction of

H

that the Argmin Router realizes. Bold marks the largest headroom and the best Argmin result.

Method	Val only	+B3	+B10	+B100	+B300
Random	0.731	0.731	0.731	0.731	0.731
Equal Weight	0.710	0.710	0.710	0.710	0.710
TimeCopilot Median [ TimeCopilot, 2026 ]	0.690	0.690	0.690	0.690	0.690
Dev-Selected Best	0.703	0.698	0.702	0.702	0.702
RAES-Conductance-Ensemble-V [ AriGoldbach, 2026 ]	0.672	0.672	0.672	0.672	0.672
Argmin Router	0.672	0.661	0.662	0.664	0.663

Appendix

Table B3: Effect of the evidence budget on the main pool, the sweep plotted in Figure 4 . Entries are MASE (lower is better) and bold marks the best budget for each method whose decision reads the training backtests. The remaining rows are constant because their decisions do not.

Subset	Best single TSFM	$H$	MASE	HR (%)
Chronos2+TiRex2	0.697	0.060	0.675	36.4
FlowState+TiRex2	0.697	0.058	0.679	31.3
TiRex2+Toto2-313m	0.697	0.051	0.684	26.1
FlowState+Toto2-313m	0.702	0.055	0.691	18.7
Moirai+TiRex2	0.697	0.042	0.690	16.1
Moirai+Toto2-313m	0.703	0.047	0.696	14.9

Appendix

Table B4: The Argmin Router on every size-2 and size-3 subset at the val+B10 budget. Best single TSFM is the strongest single member of that subset,

H

its headroom over the window-level oracle, and HR the fraction of

H

realized, so HR

=

(Best single TSFM

-

MASE)

/H

. Rows are sorted by HR and bold marks the highest within each subset size.

Agentic Time Series Forecasting

See all 14 papers

TempusBench: An Evaluation Framework for Time-Series Forecasting
Apr 13, 2026Denizalp Goktas, Gerardo Riaño-Briceño, Omkar Tekawade +14Time Series Forecasting Forecasting Benchmarks
Simulacrum New York City, NY, USA
TimeRouter: Efficient and Adaptive Routing of Time-Series Foundation Models
Jun 10, 2026Kanghui Ning, Yushan Jiang, Kashif Rasul +3Adaptive Model Routing Time Series Forecasting
School of Computing, University of Connecticut, Storrs, USA · Department of Machine Learning Research, Morgan Stanley, New York, USA.
Latent Inference-Time Guidance of Time Series Foundation Models
Sep 29, 2026Chloé Hashimoto-Cullen, Amaury Durand, Laurent Bozzi +3Time Series Forecasting Inference-Time Guidance
LPSM, Sorbonne Université · EDF R&D · UCL

Apr 13, 2026J/K move · Enter open · S save

TempusBench: An Evaluation Framework for Time-Series Forecasting

Denizalp Goktas, Gerardo Riaño-Briceño, Omkar Tekawade, Alif Abdullah+13

Simulacrum New York City, NY, USA

Time Series Forecasting Forecasting Benchmarks Time-Series Foundation Models Time Series Modeling

Open paper AskPDF arXiv

Foundation models have transformed natural language processing and computer vision, and a rapidly growing literature on time-series foundation models (TSFMs) seeks to replicate this success in forecasting. While recent open-source models demonstrate the promise of TSFMs, the field lacks a comprehensive and community-accepted model evaluation framework. We see at least four major issues impeding progress on the development of such a framework. First, existing evaluation frameworks comprise benchmark forecasting tasks derived from often outdated datasets (e.g., M3), many of which lack clear metadata and overlap with the corpora used to pre-train TSFMs. Second, these frameworks evaluate models along a narrowly defined set of benchmark forecasting tasks, such as forecast horizon length or domain, but overlook core statistical properties such as non-stationarity and seasonality. Third, domain-specific models (e.g., XGBoost) are often compared unfairly, as existing frameworks do not enforce a systematic and consistent hyperparameter tuning convention for all models. Fourth, visualization tools for interpreting comparative performance are lacking. To address these issues, we introduce TempusBench, an open-source evaluation framework for TSFMs. TempusBench consists of 1) new datasets which are not included in existing TSFM pretraining corpora, 2) a set of novel benchmark tasks that go beyond existing ones, 3) a model evaluation pipeline with a standardized hyperparameter tuning protocol, and 4) a tensorboard-based visualization interface. We provide access to our code on GitHub: https://github.com/Smlcrm/TempusBench and maintain a live leaderboard at https://smlcrm.com/tempusbench.

Abstract

Agentic Time Series Forecasting

Explore similar work

TempusBench: An Evaluation Framework for Time-Series Forecasting

TimeRouter: Efficient and Adaptive Routing of Time-Series Foundation Models

Latent Inference-Time Guidance of Time Series Foundation Models