Cross-Timescale Transformer with One-Dimensional Convolution for Integrated Financial Risk Anomaly Detection and Discrimination

Abstract

This paper addresses the characteristics of risk anomalies in financial time series, such as concealed signals, multi-scale coupling, and non-stationary distributions. It proposes a risk anomaly detection and discrimination model based on one-dimensional convolution and a cross-timescale Transformer. The method takes multivariate sequences as input, employs a one-dimensional convolutional encoder to extract local fluctuation patterns and short-range mutation features, and constructs sequence representations with different time resolutions through multi-scale downsampling. Long-range dependencies and cross-cycle correlations are learned in a cross-timescale attention module. Subsequently, the multi-scale context is upsampled and aligned to a unified time axis, then fused with the local representations to form a joint representation that combines fine-grained localization capabilities with global semantic consistency. The model simultaneously outputs time-by-time anomaly scores and sequence-level risk category probabilities within the same framework, achieving integrated modeling for anomaly localization and risk discrimination. Comparative experiments show that this method achieves superior performance in classification, ranking ability, and reliability metrics, validating the effectiveness and stability of local pattern modeling and cross-scale context fusion for identifying financial risk anomalies.