Adaptive Multi-Scale Anomaly Detection for Cross-Sector Financial Time Series via End-to-End Wavelet-Attention Fusion

Traditional two-stage pipelines for financial anomaly detection suffer from a critical yet often overlooked flaw: the wavelet decomposition and the downstream detector are optimized independently. When a fixed Haar filter encounters the fat-tail shocks typical in A-share sectors, gradient flow breaks at the decomposition boundary, and the detector receives suboptimal frequency segments. We introduce EWAF-AE, an autoencoder that embeds discrete wavelet transform as a learnable layer, allowing the spectral filters to co-train with the encoder. Instead of a homogeneous backbone, we split the encoding: a shallow Transformer handles the low-frequency trend branch, while a masked locality-sensitive attention module processes high-frequency details. A gating network then calibrates the fusion ratio per sector—nonferrous metals receive heavier trend weighting (0.888), whereas banking shifts toward detail sensitivity (0.252)—based on rolling volatility and a learnable sector embedding. The anomaly score is not a single MSE, but a weighted composite of pixel-level, trend-consistency, detail-sparsity, and spectral losses, augmented by a triplet margin loss on pseudo-anomalous samples. On six A-share sector indices from 2015 to 2025, EWAF-AE hits 83.3% of benchmark crisis events, compared with 25.0% for Isolation Forest and 41.7% for LSTM-AE.