Early Warning of Bond Defaults in Wholesale and Retail Enterprises: An Integrated Random ForestBP Neural Network Approach

Aiming at the problem that traditional default early warning models have insufficient discriminative power and robustness under the scenario of strong correlation among high-dimensional indicators and unbalanced samples, this study proposes a serial fusion model integrating Random Forest (RF) and Back Propagation (BP) Neural Network. First, RF is used for feature selection: the minimum sufficient feature set is screened out based on the criterion of cumulative contribution 90%, which corresponds to the input layer nodes of the BP Neural Network. Subsequently, the BP Neural Network is applied to perform nonlinear fitting on the selected features. During the training phase, ADASYN (Adaptive Synthetic Sampling Approach) oversampling is only implemented on the training set to balance samples while preventing information leakage. The model is trained and evaluated using samples from the wholesale and retail industry, and its early warning effect is verified by predicting the default probability of Gome Electrical Appliances. The results show that the RF-BP model has strong early warning capability for small to medium-sized samples and can issue risk early warning signals in advance.