We study the impact of machine learning (ML) models for credit default prediction in the calculation of regulatory capital by financial institutions. We do so by using a unique and anonymized database from a major Spanish bank. We first compare the statistical performance of five models based on supervised learning like Logistic Lasso, Trees (CART), Random Forest, XGBoost and Deep Learning, with a well-known model like Logit. We measure the statistical performance through different metrics, and for different sample sizes and features available. We find that ML models outperform, even when relatively low amount of data is used. We then translate this statistical performance into economic impact by estimating the savings in capital when using an advanced ML model instead of a simpler one to compute the risk-weighted assets following the Internal Ratings Based (IRB) approach. Our benchmark results show that implementing XGBoost instead of Logistic Lasso could yield savings from 12.4% to 17% in terms of regulatory capital requirements.

我们研究了机器学习 (ML) 模型对金融机构监管资本计算中信用违约预测的影响。我们通过使用西班牙一家主要银行的独特匿名数据库来做到这一点。我们首先将 Logistic Lasso、Trees (CART)、Random Forest、XGBoost 和 Deep Learning 等五种基于监督学习的模型与 Logit 等知名模型的统计性能进行比较。我们通过不同的指标以及不同的样本大小和可用特征来衡量统计性能。我们发现，即使使用相对较少的数据量，ML 模型也表现出色。然后，我们通过估计使用先进的机器学习模型而不是更简单的模型来计算基于内部评级 (IRB) 方法的风险加权资产时的资本节省，将这种统计性能转化为经济影响。我们的基准测试结果表明，实施 XGBoost 而不是 Logistic Lasso 可以在监管资本要求方面节省 12.4% 到 17%。