面向类别不平衡的家庭非房产债务压力风险预测——基于代价敏感XGBoost的研究

Predicting Household Non-Housing Debt Pressure under Class Imbalance—A Cost-Sensitive XGBoost Approach

摘要: 家庭非房产债务往往对应消费信贷、经营周转及亲友借款等短期资金需求，其债务负担上升可能加剧家庭金融脆弱性。本文利用中国家庭追踪调查(CFPS) 2010~2022年数据构建家庭–年份非平衡面板，以下一调查期非房产债务收入比超过50%识别债务压力风险，并以当期家庭特征实施时间外推预测。针对高风险样本占比较低、漏识别成本较高的特点，本文在XGBoost框架中引入类别代价、风险强度动态权重和面向F2准则的分类阈值选择，同时以风险评分分组考察模型的筛查效能。结果表明：风险强度动态加权模型在测试集上的AUC为0.7426、KS为0.3650，召回率和F2值分别为0.6683和0.4512；风险评分最高10%家庭的实际风险发生率为33.76%，约为总体风险率的3.33倍。更换风险定义、滚动时间窗口以及剔除当期直接债务变量后，模型仍保持稳定的样本外区分能力。特征组消融结果进一步显示，资产缓冲和债务惯性是提高预测能力的主要信息来源。本文的贡献在于将家庭债务研究由当期相关性描述扩展为面向未来状态的预测问题，并给出适用于类别不平衡家庭调查数据的风险排序与筛查框架。

Abstract: Non-housing debt typically stems from short-term funding needs—such as consumer credit, business working capital, and loans from friends or family members—and a rising debt burden may exacerbate household financial vulnerability. This paper constructs an unbalanced household-year panel using data from the China Family Panel Studies (CFPS) spanning 2010 to 2022. It identifies debt pressure risk based on a non-housing debt-to-income ratio exceeding 50% in the subsequent survey wave and employs current household characteristics for out-of-time predictive modeling. Addressing the challenges posed by a low proportion of high-risk samples and the high cost of false negatives, the study incorporates class-specific costs, dynamic weighting based on risk intensity, and F2-oriented classification threshold selection into the XGBoost framework; it further evaluates the model’s screening performance by segmenting households according to risk scores. Results indicate that the dynamic risk-intensity weighted model achieves an AUC of 0.7426 and a KS statistic of 0.3650 on the test set, with recall and F2 values of 0.6683 and 0.4512, respectively. Among the top 10% of households by risk score, the actual risk incidence rate is 33.76%, approximately 3.33 times the overall risk rate. The model maintains stable out-of-sample discriminative power even when the risk definition is altered, rolling time windows are applied, or current direct debt variables are excluded. Feature group ablation analysis further reveals that asset buffers and debt inertia are the primary sources of information for enhancing predictive capability. This study contributes to the literature by shifting the focus of household debt research from describing contemporaneous correlations to predicting future states, and by providing a risk ranking and screening framework tailored to class imbalance household survey data.