Explainable Shapley Additive Explanations-Enhanced Stacked Ensemble for Digital Financial Risk Prediction

Nowadays, predicting Digital Financial s prediction is crucial for financial institutions to evaluate the likelihood of borrower default. Machine Learning (ML) models have improved prediction accuracy; however, existing models lack interpretability and struggle with high-dimensional data, where the complexity of the models makes it difficult for stakeholders to trust automated decisions. Hence, this study proposes an explainable Shapley additive explanation (SHAP) Stacked Ensemble (X-SHAP-Stack) for Digital Financial prediction. Initially, the input data were collected from the German Credit dataset, which contains borrower attributes and credit outcomes. Preprocessing is then performed by K-Nearest Neighbor (KNN) imputation, one-hot encoding, z-score normalization for handling missing values, converting categorical variables and scale the numerical attributes respectively. Then, an Information Gain (IG) technique is employed to select only the most relevant predictors. Furthermore, a stacked ensemble model is constructed using base learners with a logistic regression meta-learner to generate the final Digital Financial predictions. Finally, SHAP was employed to provide both global and individual-level interpretability of model decisions. From the results, the proposed X-SHAP-Stack established better results in terms of the root mean square error (9.943) and mean absolute error (7.075) than the existing Singular Spectrum Analysis with Long Short-Term Memory (SSA-LSTM) model.