Pre-trained models have demonstrated exceptional generalization capabilities in time-series forecasting; however, adapting them to evolving data distributions remains a significant challenge. A key hurdle lies in accessing the original training data, as fine-tuning solely on new data often leads to catastrophic forgetting. To address this issue, we propose Replay Tuning (R-Tuning), a novel framework designed for the continual adaptation of pre-trained time-series models. R-Tuning constructs a unified latent space that captures both prior and current task knowledge through a frequency-aware replay strategy. Specifically, it augments model-generated samples via wavelet-based decomposition across multiple frequency bands, generating trend-preserving and fusion-enhanced variants to improve representation diversity and replay efficiency. To further reduce reliance on synthetic samples, R-Tuning introduces a latent consistency constraint that aligns new representations with the prior task space. This constraint guides joint optimization within a compact and semantically coherent latent space, ensuring robust knowledge retention and adaptation. Extensive experimental results demonstrate the superiority of R-Tuning, which reduces MAE and MSE by up to 46.9% and 46.8%, respectively, on new tasks, while preserving prior knowledge with gains of up to 5.7% and 6.0% on old tasks. Notably, under few-shot settings, R-Tuning outperforms all state-of-the-art baselines even when synthetic proxy samples account for only 5% of the new task dataset.

Language models have been proven to handle time series data after tokenization and show generalization performance on unseen forecasting tasks. However, existing techniques for tokenizing time series data struggle to eliminate redundant information and noise, which can lead to signal aliasing and cumulative quantization errors, making it difficult to further improve prediction performance. In this paper, we propose an Adaptive Component Extraction and Tokenization (ACET) framework, which includes two key novelties to address these challenges: the Dynamic Component Extraction Module (DCEM) and the Time Series Tokenization Module (TSTM). The DCEM dynamically isolates the principal components from the interference in the original signal, eliminating the need for manual parameter tuning. This not only enhances the accuracy of signal tokenization but also mitigates the adverse effects of high-frequency noise. Then, the TSTM tokenizes continuous time series data while preserving long-term trend features, ensuring that critical information is retained for subsequent forecasting. Extensive cross-domain experiments on various real-world datasets demonstrate that, in zero-shot forecasting scenarios, ACET achieves improvements of 19.27% in WQL and 6.63% in MASE, compared to baseline methods.