Title: TimeBlocks: Foundational and Continual Time-Series Blockbase -- Extended Version

Abstract:

The continuous wave of digitization has resulted in an explosion of time-series data streams monitoring diverse processes, offering significant potential for insight extraction. Simultaneously, the success of foundational language models raises a critical question: Can we develop time-series models that possess foundational capabilities—such as handling multiple tasks—while remaining lightweight enough for real-time stream processing? Current foundational time-series models tend to be bulky and are primarily optimized for offline environments where strict computational and temporal limits do not apply, and where frequent model recalibration is unnecessary. However, these large architectures struggle in data stream contexts; their size and absence of continual calibration support hinder their accuracy, durability, and viability in hardware-constrained settings.

To address these challenges, we introduce TimeBlocks, a framework designed to facilitate the efficient creation of lightweight, multi-task models adaptable to varying conditions. TimeBlocks operates by maintaining a repository of modular, interchangeable blocks that can be combined to form new time-series models. Upon receiving specific time-series data, a routing mechanism iteratively identifies and selects the most appropriate blocks to assemble a model that is both accurate and resource-efficient. Furthermore, we integrate a technique named StreamCore, which constructs a representative, compact subset of the incoming data stream. This subset ensures a guaranteed approximation of the stream over time, thereby enabling continuous model calibration. Our experimental evaluation across various datasets and tasks demonstrates that TimeBlocks can generate models that surpass existing baseline methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC