MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters
Abstract Overview
This paper studies few-shot learning for lightweight time series forecasters, targeting resource-constrained settings where large foundation models are too costly and only small support sets are available. The proposed MetaCaster framework employs a multi-agent architecture in which a meta-generator produces domain-compliant synthetic training data from a few examples and textual context, a forecaster trainer trains and selects models from a 23-model lightweight library, and a meta-agent optimizes the generator's harness. Rather than producing forecasts directly through an LLM, the system acts as an intermediary engineer that prepares a compact specialist model for standalone deployment. Experiments on 18 datasets across in-domain and out-of-domain benchmarks evaluate MetaCaster against generative models, augmentation techniques, foundation models, and an agentic training pipeline.
Novelty
MetaCaster introduces an "agent-as-engineer" paradigm for time series forecasting, using agents to generate task-aligned training data and configure lightweight models rather than performing direct inference. It also introduces automatic meta-harness optimization that optimizes the generator's prompt and skill infrastructure based on downstream forecasting performance instead of pure data distribution matching.
Results
Across 30 dataset-and-shot evaluation settings, MetaCaster achieves the top MSE result in 19 cases and consistently outperforms established generative and augmentation baselines across shot sizes K in {10, 30, 50}. At K >= 30, forecasters trained on generated data frequently approach or exceed the performance of models trained on full authentic datasets, while generalizing effectively to out-of-domain benchmarks and held-out forecasters. On the Solar dataset, MetaCaster's selected lightweight forecaster attains performance comparable to time series foundation models with up to 10^3x lower latency and 10^5x fewer parameters.
Key Points
- MetaCaster aligns synthetic data generation directly with downstream forecasting error minimization rather than standard distributional realism metrics.
- The framework coordinates a Meta-Generator, a Forecaster Trainer managing the 23-model LT-Lib library, and a Harness Proposer that optimizes generator skills and prompts across training epochs.
- Ablation studies demonstrate that the forecasting-oriented objective and textual domain context are essential, while performance remains relatively stable across different LLM backbones, validating the centrality of harness optimization.