LLMs Improve Time-Series Anomaly Detection with Frequency Data

Existing large language model (LLM) approaches for time-series anomaly detection (TSAD) primarily rely on time-domain data, such as indexed values or de-seasonalized representations. However, anomalies can manifest as subtle changes in recurring patterns or oscillatory fluctuations, which are better captured in the frequency domain. This research introduces a novel framework that augments LLM-based TSAD with explicit frequency-domain evidence. The proposed method uses the Fast Fourier Transform (FFT) to generate both global (sequence-level) and local (time-localized) frequency-domain insights. These spectral features are then provided to LLMs alongside the conventional time-domain inputs. Evaluations with various LLMs on a benchmark dataset demonstrate that this addition of structured frequency-domain evidence leads to clear improvements in anomaly detection accuracy, suggesting a more comprehensive approach to identifying complex time-series deviations.