ORCA Enables Black-Box Online Adaptation for Time Series Foundation Models

Time Series Foundation Models (TSFMs) have greatly advanced zero-shot forecasting across various domains. As these models increasingly become commercialized, closed-source API services, traditional online adaptation methods that rely on white-box access for fine-tuning or gradient backpropagation become impractical. A new approach called ORCA (Online Residual Contextual Adaptation) has been proposed to enable black-box online adaptation for TSFMs. The core insight behind ORCA is that the predictive errors of a base model are conditioned on both its input and output, forming a "context of errors" that can be learned. Extensive experiments across 5 state-of-the-art TSFMs and 8 datasets demonstrate ORCA's effectiveness. Ablation studies further quantify the impact of different adapter learning hypotheses on adaptation performance in black-box scenarios, validating the approach's robustness and utility for real-world applications.