<p>Financial time series anomaly detection is crucial for financial risk control, but traditional methods and existing deep learning models often suffer from low detection accuracy, high computational overhead, and poor robustness to noise. To address these issues, this paper proposes FEA-Net, a novel anomaly detection model for financial time series. The model integrates three core modules with distinct functions: the multi-scale frequency-domain encoding module captures hidden abnormal patterns in the frequency domain that are difficult to identify in the time domain, improving detection accuracy; the enhanced sparse attention module reduces redundant computations caused by full attention mechanisms, lowering computational overhead and ensuring efficient inference; the attention-driven attribution module optimizes feature weighting and enhances the model’s interpretability. Experiments are conducted on the S&amp;P 500 and CCFD datasets, with detection accuracy, efficiency, and robustness as evaluation metrics. The results show that FEA-Net outperforms baseline methods such as LSTM, Transformer, and CausalFormer in all evaluation metrics, achieving a good balance between detection performance and computational efficiency. It also exhibits strong robustness to noise and stable performance across different financial datasets, providing an effective solution for real-time anomaly detection in financial time series and having important practical value for financial risk control.</p>