TS-PET: A Novel Framework for Fine-Tuning Pretrained Time-Series Models

Time series foundation models offer powerful zero-shot capabilities but face significant adaptation challenges: (1) inefficient fine-tuning due to computational constraints and unreliable sensitivity scores in existing variants to the low-rank adaptation (LoRA) method, and (2) parameter-heavy prediction heads causing overfitting. We propose time series parameter-efficient transformer (TS-PET), a novel fine-tuning framework featuring: (1) A lightweight prediction module that reduces parameters by >80%, mitigating overfitting while maintaining performance; (2) Specialized pruned LoRA, which introduces robust rank allocation via identifying synergistic parameter interactions, enabling stochastic approximation for efficiency. Extensive experiments on eight benchmarks (Electricity Transforming Temperature, Exchange, Weather, etc.) show TS-PET achieves state-of-the-art accuracy-outperforming MOMENT (linear probing), PatchTST, and adaptive LoRA variants-while demonstrating superior parameter efficiency. Our solution enables scalable adaptation of time series foundation models without compromising performance.