Dynamic relational network embedding-based anomaly detection for predicting financial risks

Financial bubbles are major causes of economic crises and substantial market losses. Existing bubble detection methods mainly rely on financial indicators while overlooking the complex relationships among stock prices and indicators. To address this limitation, this study proposes a dynamic relational network embedding framework for financial bubble detection. First, a multi-layer dynamic network is constructed to model the inconsistent relationships among stock prices and financial indicators. Then, an entropy-transformer-based network embedding model is developed to learn low-dimensional representations of dynamic financial interactions. Finally, an autoencoder-based anomaly detection module is employed to identify financial bubbles. Experiments on Chinese and U.S. stock markets demonstrate that the proposed method consistently outperforms existing baselines in terms of precision, recall, and F1-score, achieving average F1-score improvements of 7.92% and 8.88%, respectively. The results verify the effectiveness of dynamic relational modeling for financial bubble detection.