The paper introduces CoSPOT, an online time series forecasting framework that uses a frozen pre‑trained large language model (LLM) as the core forecaster. CoSPOT adapts to evolving data by applying compositional spectral prompts—frequency‑domain basis prompts weighted by their amplitudes—allowing the model to represent unseen patterns as new combinations of learned bases while updating few parameters. Experiments on real‑world datasets show CoSPOT’s effectiveness in extended online phases and cross‑dataset scenarios with significant distribution shifts.
By Seungyoon Choi, Hyunchul Kim, Jae-Gil Lee, Chanyoung Park
arXiv:2606. 19412v1 Announce Type: new Abstract: Time series forecasting leverages historical patterns to predict future values, but traditional methods face challenges when dealing with complex, non-stationary patterns that are difficult to memorize during training.
By Huu Hiep Nguyen, Minh Hoang Nguyen, Dung Nguyen, Hung Le
arXiv:2603. 15506v2 Announce Type: replace-cross Abstract: We argue that the current practice of evaluating AI/ML time-series forecasting models, predominantly on benchmarks characterized by strong, persistent periodicities and seasonalities, obscures real progress by overlooking the performance of efficient classical methods.
By Raeid Saqur, Christoph Bergmeir, Blanka Horvath, Daniel Schmidt, Frank Rudzicz, Terry Lyons
arXiv:2602. 01588v3 Announce Type: replace-cross Abstract: Multimodal time series forecasting is crucial in real-world applications, where decisions depend on both numerical data and contextual signals.
By Huu Hiep Nguyen, Minh Hoang Nguyen, Dung Nguyen, Hung Le
The paper introduces the Continuous Evolution Pool (CEP), a replay‑free framework for online time series forecasting that tackles recurring concept drift. CEP maintains a dynamic pool of specialized forecasters, using lightweight statistical genes to identify concepts, spawn new models when distribution shifts occur, and prune obsolete ones under memory limits. Experiments on real‑world datasets show CEP reduces forecasting error by up to 24% compared to state‑of‑the‑art baselines, especially in scenarios with pronounced recurring drift.
By Tianxiang Zhan, Ming Jin, Yuanpeng He, Yuxuan Liang, Shirui Pan
SMart is a new time series representation learning framework that combines a multi-phase recurrence plot recovery task with a source dataset selector. The recovery task uses three alternative modes to guide the encoder in capturing time series dynamics, while the selector chooses multiple suitable source datasets to augment the target dataset during pre‑training. Experiments demonstrate that SMart surpasses state‑of‑the‑art models, reducing mean absolute error by up to 19.5% in regression and increasing classification accuracy by up to 1.34%.
By Fang He, Wang-chien Lee