Anomalous Payment Behavior Detection and Risk Prediction for SMEs Based on LSTM-Attention Mechanism

Abstract

This paper proposes a novel approach for detecting anomalous payment behaviors and predicting financial risks in Small and Medium-sized Enterprises (SMEs) using an enhanced LSTM-Attention mechanism. The model integrates bi-directional LSTM networks with a multi-head attention mechanism to capture complex temporal dependencies in payment patterns while focusing on significant transaction features. The approach addresses the challenges of imbalanced datasets and evolving payment behaviors through a comprehensive risk assessment framework and dynamic threshold adjustment mechanism. Experimental results on a dataset containing 2.85 million transactions from 7,500 SMEs demonstrate the model's superior performance, achieving 98.5% accuracy and 94.2% precision in anomaly detection. The proposed model significantly outperforms traditional approaches and contemporary deep learning methods, showing a 15-20% improvement in detection accuracy while maintaining low false positive rates. The integration of behavioral risk indicators with operational metrics enables early risk prediction with an AUC-ROC score of 0.982. The model's effectiveness is validated through extensive case studies across various industry sectors, demonstrating robust generalization capabilities and practical applicability in real-world scenarios. The research contributes to the field by introducing an adaptive risk assessment framework that combines temporal pattern analysis with contextual business information for enhanced payment risk detection.