EVENT DOSSIER
MetaCaster is a meta-framework-based proxy system designed to address the contradiction between high costs of basic models and the need for extensive training data for lightweight predictors in resource-constrained scenarios. This system utilizes proxy data generation technology to automatically train specialized lightweight time-series predictors with just a few examples and text context. In this architecture, proxies are positioned as intermediate engineers for efficiently preparing task-specific predictors, rather than direct predictors. Experiments verified the effectiveness of this method on 18 datasets, 23 advanced lightweight predictors, and 14 baseline models, demonstrating that MetaCaster achieves both data efficiency and computational efficiency while maintaining high-quality time-series prediction performance.
MetaCaster 提出一种基于元框架的代理系统，利用代理数据生成技术，仅凭少量示例和文本上下文即可自动训练专用轻量级时间序列预测器。该研究针对资源受限场景下基础模型成本高昂及轻量级预测器需大量训练数据的矛盾，将代理定位为准备高效、任务特定预测器的中间工程师而非直接预测者。实验在 18 个数据集、23 种最先进的轻量级预测器和 14 个基线模型上验证，证明 MetaCaster 实现了数据效率与计算效率的双重提升，同时保持了高质量的时间序列预测性能。