Abstract: In recent years, the frequent occurrence of major international economic and political events has intensified the turbulence of the global financial system and increased the risks. How to establish a scientific and effective financial risk early warning system to identify and prevent potential risk crises is of great significance for maintaining financial security and national security.
With the rapid growth of data, significant improvement in computing power and the rapid development of financial technology, the deep learning models represented by deep convolutional neural networks are gradually becoming a research focus in the academic community. Among them, the residual network(ResNet)deep learning model is good at preserving the original features of data through residual connections, which can effectively avoid feature layer by layer disappearance. Therefore, it is widely used in the research field of image processing. Unfortunately, the ResNet deep learning model has not yet been applied to research in the financial field, especially in the field of financial risks warning. Therefore, how to apply the ResNet deep learning models to financial risk early warning research has important practical significance for government financial management departments to scientifically respond to financial risk crises, and maintain financial security and national security.
Based on SSEC, this paper firstly uses the multifractal method to calculate the multifractal volatility, and introduces spectral clustering method to adaptively mine normal state samples and attentional state samples, which measures the financial risk sates. Further, this paper combines Synthetic Minority Oversampling Technique(SMOTE) of the imbalanced sample processing technology, ResNet of deep learning technology and Support Vector Machine (SVM) of the machine learning technology to propose a deep warning model of ResNet-SMOTE-SVM, and conducts an empirical research on financial risk deep warning.
The empirical results show that the financial risk states measured by the multifractal spectral clustering method not only has significant statistical significance, but also is highly consistent with the actual operation of Chinese financial market. Meanwhile, compared with other models, the deep early warning model of ResNet-SMOTE-SVM can effectively extract deep features of the financial market, overcome imbalanced sample problems, and accurately predict financial risks, which has excellent capability of financial risk early warning.
Based on the empirical results, the following conclusions can be drawn: (1)The financial market is a typical multifractal market, and the use of multifractal spectral clustering method can accurately measure the financial risk state. (2)The deep warning model of ResNet-SMOTE-SVM combines the advantages of ResNet network, SMOTE technology and SVM model. It can extract high-level and complex deep-seated features of the financial market, overcome imbalanced sample problems, and accurately predict financial risks. It has strong financial risk warning capabilities. (3)The research on deep waring of financial risks in this paper has innovated research methods in the field of financial risk early warning, provided reference value for risk management of financial systems with complex volatility characteristics, and also offered operable application tools for government financial regulatory and stability.

Key words: multifractal, spectral clustering, ResNet-SMOTE-SVM, deep learning, risk early warning

摘要： 金融安全关系国家安全。如何建立科学有效的金融风险预警体系,对于维护金融安全与国家安全具有重要意义。本文以上证综指长达20余年的5分钟高频数据为研究样本,构建多分形谱聚类方法提取多分形波动率并测度金融风险状态,进而将ResNet,SMOTE与SVM相结合,提出ResNet-SMOTE-SVM深度学习模型,并开展金融风险深度预警的实证研究。实证结果表明,本文构建的多分形谱聚类方法所测度出的金融风险状态不仅具有显著的统计检验意义,而且与我国金融市场实际运行情况高度相符。同时,与其余预警模型相比,ResNet-SMOTE-SVM深度预警模型能够有效提炼金融市场的深层次特征,克服非均衡样本问题以及最为准确地预测金融风险,具有优异的金融风险预警能力,且表现出良好的稳健性。

关键词: 多分形, 谱聚类, ResNet-SMOTE-SVM, 深度学习, 风险预警