We compare linear regression, local polynomial regression and selected machine learning methods for modeling credit spread changes. Using partial dependence plots (PDPs) and H-statistic, we find that the outperformance of machine learning models compared to regression ones is mostly attributable to complex non-linearities and not to interactions. The PDPs are additionally used to perform a factor hedging. For the first time, credit spread changes are decomposed by applying SHapley Additive exPlanation (SHAP) values. The proposed framework is applied to US and Euro Area corporate and covered bond credit spread changes of different maturities to quantify the influence of several macroeconomic and financial variables. Despite several commonalities between the decompositions of US and Euro Area credit spread changes, we also observe some differences — particularly related to the impact of certain explanatory variables during crisis periods.

中文翻译：

---

使用可解释的人工智能分析信用利差变化

我们比较了线性回归、局部多项式回归和选定的机器学习方法来对信用利差变化进行建模。使用部分依赖图 (PDP) 和 H 统计量，我们发现机器学习模型的性能优于回归模型主要归因于复杂的非线性而不是交互。 PDP 还用于执行因子对冲。首次通过应用 SHapley Additive exPlanation (SHAP) 值来分解信用利差变化。拟议的框架适用于美国和欧元区不同期限的企业和担保债券信用利差变化，以量化多个宏观经济和金融变量的影响。尽管美国和欧元区信用利差变化的分解存在一些共同点，但我们也观察到一些差异，特别是与危机期间某些解释变量的影响有关。