Time-PEFT: Temporal and Multichannel Complexity-Based Fine-Tuning for Time-Series Foundation Models

Recent studies have attempted to fine-tune time-series foundation models to enhance a target dataset's forecasting performance. However, these approaches proceed without a clear criterion for identifying complex datasets that require fine-tuning due to performance degradation in zero-shot forecasting. To distinguish datasets that are more challenging than standard benchmarks, we introduce data-driven temporal complexity and multichannel complexity. *Temporal complexity* captures the difficulty of identifying distinct patterns by quantifying spectral entropy in the frequency domain, while *multichannel complexity* captures cross-channel information flow that can impact predictive uncertainty. These metrics serve as *effective proxies for performance gains* achievable through fine-tuning. Based on the two metrics, we develop *Time-PEFT*, a parameter-efficient fine-tuning framework that incorporates a frequency adapter for top-$k$ filtering and a channel adapter for multichannel modeling. With the base variant of MOMENT as a backbone, *Time-PEFT* improves performance by up to 38\% over LoRA on complex datasets.

Many real-world problems require forecasting how measurements change over time, such as signals from medical sensors. Large models trained on time-series data can often make useful forecasts, but they may struggle with new complex time-series datasets. Thus, it would be practical to identify when datasets are sufficiently complex that pre-trained time-series models require additional adaptation. We introduce two simple ways to measure this complexity. The former measure captures how hard it is to find useful patterns over time, while the latter captures how much knowing one variable helps explain another. Based on these measurements, we design Time-PEFT, a lightweight method for adapting large time-series models to complex datasets. Our experiments show that this approach can improve forecast accuracy on complex datasets.