Based on data-driven graph relationship feature selection and support vector data description two-stage financial risk warning model: Evidence from Chinese listed companies

• A new two-stage financial risk early warning framework is proposed. • Based on the correlation between indicators, the graph structure is constructed to screen the features and build the financial risk early warning index system. • Based on the characteristics of unbalanced data sets, the financial risk warning problem is studied from the perspective of outlier detection. • The performance of the proposed model is better than other comparison models. As an emerging research method in economics and finance, machine learning models exhibit strong data processing capabilities and generalisation performance, giving them a significant advantage over traditional statistical learning models in constructing risk warning models. This paper explores the feasibility of applying machine learning for risk warnings in listed companies by starting with focusing on financial risk identification. First, it constructs a financial risk warning indicator by considering various influencing factors from multiple dimensions. We employ a graph structure model of GRFS type based on feature correlations for effective data preprocessing. Second, due to the inherent imbalance in risk data, a machine learning model(SVDD) is selected. Ultimately, a two-stage financial risk warning model, graph relationship feature selection(GRFS-SVDD), is constructed. Results indicate that first, the indicators identified strongly correlate with corporate financial risk, aligning with practical situations and economic theory; second, the GRFS-SVDD model significantly improves the predictive accuracy regarding corporate financial risk and demonstrates robustness across sample sets of different proportions, achieving over 95 % accuracy on training sets of normal samples; and third, the GRFS feature screening method is interpretable, and the results of coefficient estimation indicate cost expense margin, total executive compensation and text sentiment are positively correlated with financial risk, whereas separation degree of two rights, external financing ratio, and resilience of enterprises show negative correlations.