Oral Talk 2: Shift-Aware Test Time Adaptation and Benchmarking for Time-Series Forecasting

Test time adaptation (TTA) has shown promise in addressing distribution shifts in different areas, but remains significantly underexplored in time-series forecasting (TSF), where temporal dependencies and the evolving nature of the signals present unique challenges. We present DynaTTA, a dynamic TTA framework for TSF that estimates distribution shifts in real time by tracking prediction errors and embedding drift. This estimate allows us to employ two key mechanisms, a dynamic adaptation rate that is adjusted based on the severity of the shift, and shift-conditioned gating that controls the influence of the learned adaptations as required. These mechanisms enable meaningful and appropriate adaptations in the presence of distribution shifts, while retaining the prior knowledge of the source model. DynaTTA is modular and can be used with any existing pretrained model for TSF, without requiring retraining. We also propose TTFBench, a first-of-its-kind benchmark for evaluating TTA for TSF, comprising thousands of time-series with varying types and intensities of shifts. Through extensive experiments with various backbones and datasets including TTFBench, we show that DynaTTA consistently improves performance. The code and data are available at https://github.com/shivam-grover/DynaTTA.