SIFT: Enhancing Time Series Foundation Models via Semantic Invariance and Structural Fidelity Fine-Tuning

2026-09-26

AI Summary

Fine-tuning time series foundation models is prone to overfitting and the mean prediction trap. To address these challenges, this work proposes SIFT, a novel method that introduces semantic spectral decomposition perturbation and non-core subspace augmentation to enable semantics-preserving adversarial training. Furthermore, SIFT incorporates a component-level hybrid reconstruction objective to ensure structural fidelity. Extensive experiments across ten real-world datasets demonstrate that the proposed approach significantly enhances the zero-shot generalization performance of various time series foundation models. By effectively mitigating common fine-tuning pitfalls while preserving semantic and structural integrity, this study establishes a new paradigm for the efficient adaptation of pretrained time series models.

Abstract

Time Series Foundation Models (TSFMs) have achieved remarkable zero-shot performance through extensive pre-training on massive time series datasets. Nevertheless, due to the low-dimensional properties and diverse structural patterns of time series data, performing naive fine-tuning on TSFMs often leads to overfitting and falling into the mean-prediction trap. To address these challenges, we propose SIFT, a robust adaptation method that enhances time series foundation models by preserving Semantic Invariance and structural Fidelity throughout the fine-Tuning process. We employ semantic-invariant adversarial augmentation, which utilizes semantic spectrum decomposition to partition the semantic space and then generates perturbations within the non-core semantic subspace to bolster the model's robustness against these perturbations, mitigating overfitting. We implement a component-based structural fidelity enhancement, which facilitates component-wise mixup and imposes a reconstruction objective to improve the model's ability to preserve structural fidelity, alleviating the mean-prediction trap. Extensive experiments on representative TSFMs covering 10 real-world datasets demonstrate that SIFT can significantly enhance the performance of TSFMs.

Problem

• Time Series Foundation Models
• Fine-tuning
• Overfitting
• Mean-prediction trap

Innovation

• Time Series Foundation Models
• Semantic-Invariant Adversarial Augmentation
• Structural Fidelity Enhancement
• Fine-Tuning
• Mean-Prediction Trap