基於集成學習框架之信用違約預測－以信用卡客戶為例

Abstract

藉由堆疊 (Stacking) 與勻合 (Blending) 學習器 (base estimators) 所產生的異質 性集成學習(Heterogeneous Ensemble Learning)框架，本文建構消費金融信用卡客 戶之違約風險預警模型。採用 Yeh and Lien (2009) 的資料集，我們的實證結果顯 示，堆疊與勻合集成皆能有效降低誤判信用違約客戶為正常的型二誤差。尤其當 輔以適當的學習器挑選策略，堆疊集成的綜合辨別能力呈現一定的泛化優越成效， 明顯勝出任何非集成的單一學習器。更加地，輔以學習器挑選策略的堆疊集成能 夠提高模型識別違約客戶的準確率 (F1 值)，且在增進識別違約客戶能力的同時 有效降低誤判正常客戶為違約的分類誤差 (AUC 值)。

Based on Heterogeneous Ensemble Learning that allows for the Stacking and Blending of base learners of distinct types, in this study we construct an ensemble-learning assisted credit-risk prediction model in an attempt to prewarn consumer banks of their credit card holders’ possibility of default. Using the dataset as in Yeh and Lien (2009), our empirical results show that ensemble learning models that exploit either Stacking or Blending can effectively reduce the Type II error in mis-judging defaulted entities as normal. In particular, when equipped with a learner-selection strategy, heterogeneous ensemble learners that exploit Stacking tend to exhibit superior predictive power over all single base learners. Furthermore, ensemble learners with Stacking are found to be capable of improving the rate of accuracy in nailing down defaulted entities (F1-score); they demonstrate the ability to identify credit-critical customers while at the same time reduce the possibility of misjudging normal customers as defaulted ones (AUC-value)