MetaCaster

Time series forecasting 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.

Unlike existing generative models that focus on data realism, MetaCaster generates data to ensure that forecasters trained on it perform comparably to those trained on real data of the same size in the target domain. 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.

The table compares MSE across generation models, augmentation methods, and MetaCaster. We randomly hold out 3 forecasters from LT-Lib and report the average performance over the remaining 20; the held-out forecasters are used to assess generalization to unseen forecasters. The table also reports forecasters trained with the K-shot support set 𝒟_sup and with the full set 𝒟_tr, serving as lower and upper performance references.

Several observations: MetaCaster outperforms the generation and augmentation baselines in most cases, demonstrating the benefit of optimizing data generation for forecasting; its performance improves with larger K, showing effective few-shot utilization; when K ≥ 30 it approaches or even surpasses 𝒟_tr, suggesting that raw data may be noisy and optimized data can improve training; at K = 10 it remains competitive; and despite the increased difficulty it generalizes well to OOD datasets, where it generally outperforms the baselines.

The ablation uses averaged MSE with K = 30, with MetaCaster as the original model. First, we study the effect of forecasting-oriented optimization by replacing the objective with MMD and Wasserstein distances, which directly align the generated set with the authentic set. Second, we remove the contextual cues 𝖢_m to evaluate their contribution. Third, we examine the impact of different LLMs.

Minimizing data distribution discrepancy generally degrades performance, as it is not directly aligned with the forecasting objective. Contextual cues are crucial, as they guide MGAgent in selecting domain-relevant knowledge for generating time series. Interestingly, different LLMs yield comparable results in many cases, suggesting the Harness — rather than the LLM backbone — is the key factor in MetaCaster, consistent with prior findings. Although GPT-5.3-Codex wins in some cases, it is unstable on datasets such as ETTm1 and USbirths, leading to worse overall results; thus we adopt GPT-5.4 as the default LLM for its consistent performance.

The hinge loss evolves over 8 harness optimization epochs, where MetaCaster converges quickly and selects the final Harness from epoch 5. The traces below follow MGAgent and HPAgent from epoch 4 to 5 on the alibaba_cluster_2018 dataset. MGAgent analyzes the few-shot examples 𝒟_sup, inspects statistics, retrieves domain knowledge about the two variates CPU and memory, and constructs a TS-Generator.

HPAgent detects the performance degradation at epoch 4 via memory, then performs Analysis and Diagnosis. The Analysis uses analytical tools pertinent to the issue identified in the reasoning traces of MGAgent and the training logs of FTAgent. The Diagnosis identifies broken inter-variate correlation as the root cause. It then updates MGAgent's skills, leading to an improved TS-Generator that strengthens the CPU–memory correlation and adds consistency checks. Consequently, epoch 5 produces correlated time series that are more compliant with the few-shot examples than epoch 4.

This example illustrates a simple Harness update. In fact, MetaCaster's optimization is more complex and deals with B = 8 datasets in the harness training corpus concurrently. The final output is a trained lightweight forecaster rather than the generated time series.

We collect 23 state-of-the-art lightweight forecasters proposed from 2022 to 2026, including linear-layer models such as MixLinear, MLP-based forecasters such as TSMixer, and frequency-domain models such as FITS. Their sizes are much smaller than state-of-the-art Transformer-based forecasters (PatchTST, ≈3.2M) and TSFMs (Chronos, ≈700M). We compile them in LT-Lib with a unified interface to facilitate training calls in FTAgent.