MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters

ChengAo Shen; Wenchao Yu; Fangyu Wu; Dongjin Song; Hanghang Tong; Dongsheng Luo; Wei Cheng; Haifeng Chen

cs.LGAug 24, 2026

MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters

Authors: ChengAo Shen, Wenchao Yu, Fangyu Wu, Dongjin Song, Hanghang Tong, Dongsheng Luo, Wei Cheng, Haifeng Chen, +1 more

Organizations: University of Houston · NEC Labs · University of Waterloo · University of Connecticut · University of Illinois at Urbana-Champaign · Singapore Management University

Abstract

Time series forecasting (TSF) is evolving toward multimodal and agentic settings, yet using foundation models remains uneconomical in resource-constrained scenarios, where compact, specialized forecasters are more desirable. However, lightweight forecasters typically require substantial training data, limiting their use in domains with scarce, slowly accumulated, or privacy-sensitive time series. To address this dilemma, we investigate the challenging problem of few-shot learning for lightweight forecasters. We propose MetaCaster, a meta-harness-optimized multi-agent framework that uses agentic data generation to automatically train specialized lightweight forecasters from only a few examples and textual contexts. Our work highlights a new TSF paradigm in which agents act not as forecasters but as intermediary engineers that prepare efficient, task-specific forecasters for deployment. Experiments on 18 datasets, 23 state-of-the-art lightweight forecasters, and 14 baselines demonstrate that MetaCaster achieves both data efficiency and computational efficiency while maintaining high-quality TSF performance.

Figures & tables

Figure 1: Comparison of different paradigms of using LLMs for TSF. MetaCaster is the proposed model.

Figure 2: An illustration of the harness optimization framework of the proposed MetaCaster system.

	Dataset	Ours	Generation Models					Augmentation Methods				References
	Dataset	MetaCaster	TimeDP	VerbalTS	T2S	DiffTS	TimeVAE	Repeat	Bootstrap	Jitter	MagWarp	$\mathcal{D}_{\text{sup}}^{m}$	$\mathcal{D}_{\text{tr}}^{m}$
IND ( $K=10$ )	ETTm1	0.376	1.049	1.004	0.922	1.154	0.749	0.781	0.748	0.746	1.919	0.687	0.316
	Electricity	0.300	0.288	1.061	1.150	0.365	0.328	0.380	0.379	0.375	7.057	0.721	0.121
	Seattle	1.522	1.420	1.800	1.792	1.643	1.759	1.955	1.849	1.925	43.24	2.341	1.024
	SZTaxi	0.104	0.124	0.116	0.118	0.111	0.114	0.134	0.118	0.123	1.122	0.187	0.096
	Sales	2.616	2.587	2.419	2.758	4.323	2.795	4.175	2.949	2.946	4.321	3.004	2.927

Table 1: TSF performance comparison on IND and OOD corpora for

K\in\{10,30,50\}

in terms of MSE. Lower MSE is better. Red ( blue ) values indicate the best (second-best) MSE per row.

\dagger

M4 uses instance-normalized MSE to address large distribution shifts. MAE results are available in Appendix E .

Figure 3: Comparing the selected (trained) forecasters of agent pipelines MetaCaster and TimeScientist .

Figure 4: Performance distribution. Each dot represents the performance of a forecaster $\bar{f}_{l}^{m}$ on a test set $\mathcal{D}_{\text{te}}^{m}$ , covering all forecasters in LT-Lib and all datasets in both of IND and OOD corpora.

Figure 5: Comparing MetaCaster with TSFMs on Solar dataset. Full results are in Appendix E .

Dataset ( $\rightarrow$ )	IND Corpus							OOD Corpus
Method ( $\downarrow$ )	ETTm1	Electricity	Seattle	SZTaxi	Sales	Bitbrains	Solar	Saugeen	USbirths	M4	Overall
MetaCaster	0.345	0.226	1.177	0.114	2.362	0.125	0.152	1.464	0.533	2.112	0.267
(a) Loss $\rightarrow$ MMD	0.708	0.157	1.204	0.150	3.581	0.575	0.254	2.339	0.403	2.556	0.764
(a) Loss $\rightarrow$ Wasserstein	0.702	0.284	1.734	0.152	3.553	0.483	0.254	2.406	0.535	3.105	0.940
(b) Remove context $\mathsf{C}_{m}$	0.386	0.293	1.425	0.144	2.595	0.158	0.281	2.004	0.534	2.312	0.521
(c) LLM $\rightarrow$ Gemini-3.1-Pro	0.430	0.226	1.177	0.114	2.362	0.120	0.151	1.396	0.533	2.129	0.288

Table 2: Ablation analysis in terms of MSE. “Overall” assesses the normalized MSE (Eq. ( 4 )) across datasets.

Figure 6: An illustration of Harness optimization by MetaCaster : (a) hinge loss (Eq. ( 4 )) during optimization iterations. (b)-(d) key traces of MGagent and HPagent from Epoch 4 to 5 on alibaba_cluster_2018 dataset.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Freq.	$D$	$N$	$T$	Domain	Source
australian_electricity_demand	30min	1	5	231k	energy	Godahewa et al. (2021)
solar_power	4s	1	1	7.4M	energy	Godahewa et al. (2021)
PEMS_BAY	5min	1	325	52k	traffic	Li et al. (2018)
traffic_hourly	1h	1	862	17.5k	traffic	Godahewa et al. (2021)
weather	1d	1	3,010	var	weather	Godahewa et al. (2021)
cdc_fluview_ilinet	1w	5	5	1.2k	health	Centers for Disease Control and Prevention ()

Appendix

Table 3: Summary of Training Corpus

\mathcal{C}_{\text{har}}

	Dataset	Freq.	$D$	$N$	$T$	Domain	Source
IND	ETTm1	15min	7	1	50k	energy	Zhou et al. (2021)
	Electricity	1h	1	370	35.1k	energy	Trindade (2015)
	Seattle	1h	1	323	8.8k	traffic	University of Washington STAR Lab ()
	SZTaxi	15min	1	156	3.0k	traffic	Zhao et al. (2020)
	Sales	1d	1	118	1.8k	retail	Godahewa et al. (2021)
	Bitbrains	5min	2	1,250	8.6k	cloud	Shen et al. (2015)

Appendix

Table 4: Summary of Test Corpora.

Family	Forecaster	Params	MACs (M)	Latency (ms)	Peak VRAM (MB)	Reference
Linear	Vanilla Linear	65K	0.45	0.03	8.4	Zeng et al. (2023)
	DLinear	129K	0.90	0.10	8.7	Zeng et al. (2023)
	NLinear	65K	0.45	0.04	8.4	Zeng et al. (2023)
	RLinear	65K	0.45	0.17	8.4	Li et al. (2023)
	CrossLinear	2.5M	28.45	0.72	23.1	Zhou et al. (2025)
	MixLinear	243	0.08	0.43	8.3	Ma et al. (2024)

Appendix

Table 5: Summary of the 23 lightweight forecasters in LT-Lib .

	Pre-training	Deployment	Inference
Active agents	All	MGagent + FTagent	None
LLM-driven	Yes	Yes	No
GPU	Required	GPU or CPU	GPU or CPU
Time cost	5-7 h	30-40 min	a few ms
LLM tokens	${\sim}46$ M	${\sim}150$ K	0

Appendix

Table 6: Computational cost of different stages under MetaCaster framework.

	Dataset	Agent	Generation					Augmentation				Ref
	Dataset	MetaCaster	TimeDP	VerbalTS	T2S	DiffTS	TimeVAE	Repeat	Bootstrap	Jitter	MagWarp	Sample	Full
IND ( $K=10$ )	ETTm1	0.457	0.568	0.567	0.531	0.582	0.620	0.650	0.637	0.636	0.930	0.625	0.400
	Electricity	0.413	0.415	0.860	0.866	0.474	0.436	0.468	0.466	0.464	1.652	0.652	0.243
	Seattle	0.849	0.844	0.989	0.989	0.971	0.948	0.995	0.962	0.986	4.395	1.127	0.663
	SZTaxi	0.255	0.277	0.274	0.276	0.260	0.261	0.285	0.266	0.272	0.814	0.339	0.246
	Sales	0.661	0.795	0.771	1.027	1.510	0.830	0.945	0.766	0.763	1.169	0.866	0.868

Appendix

Table 7: TSF performance comparison on IND and OOD corpora for

K\in\{10,30,50\}

in terms of MAE. Lower MAE is better. Red ( blue ) values indicate the best (second-best) MAE per row.

\dagger

M4 uses instance-normalized MAE to address large distribution shifts.

Figure 7: Comparing MetaCaster with TSFMs on the datasets in the test corpora.

Figure 8: Comparing the selected (trained) forecasters of MetaCaster and TimeScientist using LT-Lib .

Dataset	Ours	Generation Models					Augmentation Methods				References
	MetaCaster	TimeDP	VerbalTS	T2S	DiffTS	TimeVAE	Repeat	Bootstrap	Jitter	MagWarp	$\mathcal{D}_{\text{sup}}^{m}$	$\mathcal{D}_{\text{tr}}^{m}$
Top-1
ETTm1	0.243	0.967	0.982	0.886	0.921	0.437	0.382	0.382	0.385	0.358	0.489	0.300
Electricity	0.148	0.209	0.947	1.017	0.237	0.128	0.125	0.125	0.125	0.125	0.274	0.107
Seattle	0.949	1.271	1.658	1.674	1.288	1.678	1.097	1.097	1.114	1.647	1.295	0.953
SZTaxi	0.099	0.091	0.103	0.105	0.091	0.095	0.092	0.092	0.091	0.088	0.090	0.075

Appendix

Table 8: TSF performance on test corpora for top selected forecasters by the compared methods in terms of MSE. Lower MSE is better. Red ( blue ) values indicate the best (second-best) MSE per row.

\dagger

M4 uses instance-normalized MSE to address large distribution shifts.

Figure 10: Performance distribution. Each dot represents the performance of a forecaster $\bar{f}_{l}^{m}$ on a test set $\mathcal{D}_{\text{te}}^{m}$ , covering all forecasters in LT-Lib and the datasets in the IND corpus (left) and OOD corpus (right).

Figure 11: Performance of the individual forecasters (rows) trained by the compared methods (columns) on the IND datasets (top) and OOD datasets (bottom).

Figure 12: Comparison between the generated time series by different generation methods and the ground truth time series on different datasets.

Figure 13: Visualization of data distributions generated by MetaCaster on the IND datasets and OOD datasets.

Explore similar work

Model-agnostic Retrieval-Augmented Extended Forecasting for time series
Aug 14, 2026Juan Pablo Villa Serna, Rohan Asthana, Vasileios BelagiannisTime Series Forecasting Retrieval-Augmented Time Series Forecasting
Friedrich-Alexander-Universität Erlangen–Nürnberg
CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting
Apr 30, 2026Bokai Pan, Mingyue Cheng, Zhiding Liu +6Time Series Forecasting LLM-Based Time Series Forecasting
State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China, Hefei 230026, China
KairosAgent: Agentic Time Series Forecasting with Fused Semantic Reasoning
May 28, 2026Kun Feng, Ziwei Shan, Yuchen Fang +6Multivariate Time Series Forecasting Zero-Shot Time Series Forecasting
ShanghaiTech University · Ant Group

Aug 14, 2026, cs.LGJ/K move · Enter open · S save

Model-agnostic Retrieval-Augmented Extended Forecasting for time series

Juan Pablo Villa Serna, Rohan Asthana, Vasileios Belagiannis

Friedrich-Alexander-Universität Erlangen–Nürnberg

Time Series Forecasting Retrieval-Augmented Time Series Forecasting Time-Series Foundation Models

Open paper AskPDF arXiv

Time series forecasting with pretrained foundation models has demonstrated strong zero-shot capabilities. However, achieving optimal performance on time series with short or negligible historical data in domain-specific applications typically requires adaptation via either fine-tuning or RAG. While fine-tuning is effective, it incurs substantial computational costs. This work explores RAG within univariate time series (Retrieval Augmented Generation) as a more efficient alternative, in particular RAF (Retrieval Augmented Forecasting), and introduces RAEF (Retrieval-Augmented Extended Forecasting), a model-agnostic method built upon RAF. RAEF incorporates key refinements to the retrieval and aggregation mechanisms: (1) direct retrieval in input-space rather than embedding-space, reducing inference overhead, and (2) concatenation-based aggregation that preserves temporal structure instead of averaging. Empirical evaluation across multiple benchmark datasets demonstrates that RAEF outperforms RAF in both accuracy and inference overhead. Furthermore, comprehensive comparisons with zero-shot and fine-tuned foundation models show that RAEF achieves competitive or superior performance to fine-tuning while avoiding its computational burden, establishing it as a practical and scalable approach for domain adaptation in time series forecasting.

Abstract

Explore similar work

Model-agnostic Retrieval-Augmented Extended Forecasting for time series

CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting

KairosAgent: Agentic Time Series Forecasting with Fused Semantic Reasoning