Zeus: Towards Tuning-Free Foundation Model for Time Series Analysis

We present Zeus, a unified tuning-free Time Series Foundation Model (TSFM) that delivers superior performance across diverse analysis tasks without any task-specific fine-tuning. Unlike prior studies that primarily focus on zero-shot forecasting but require task-specific tuning for other tasks, Zeus bridges this gap by addressing two fundamental challenges in multi-task generalization. First, to reconcile point-level granularity with long-sequence scalability, Zeus incorporates a multi-scale Transformer featuring point-wise tokenization and a U-shaped hierarchy, effectively balancing fine-grained fidelity with computational efficiency. Second, to accommodate varying inductive biases across different tasks, Zeus introduces Multi-Objective Temporal Masking (MOTM), a unified strategy that supports heterogeneous tasks (e.g., extrapolation, interpolation, and global abstraction) within a single framework. Extensive experiments across five representative tasks demonstrate that Zeus consistently achieves competitive results in tuning-free settings, underscoring its potential as a general-purpose TSFM. The code is available at https://github.com/GestaltCogTeam/Zeus.

Modern AI systems for time series data are usually designed for only one task at a time, such as forecasting future values or filling in missing observations. As a result, researchers often need to retrain or fine-tune separate models for different applications, which is costly and limits real-world usability. We developed Zeus, a general-purpose AI model that can handle multiple time series tasks without any task-specific retraining. Zeus addresses two key challenges. First, time series data often require both fine local details and long-range context, which are difficult to model efficiently at the same time. Zeus solves this by combining detailed point-level representations with a multi-scale architecture that can process long sequences efficiently. Second, different tasks require different ways of understanding temporal patterns, so we designed a unified training strategy that teaches the model to learn forecasting, interpolation, and global sequence understanding simultaneously. Across five representative tasks, Zeus consistently achieves strong performance in tuning-free settings. Our results suggest that a single foundation model can serve as a versatile tool for many time series applications, reducing the need for expensive task-specific model development.