A CNN-LSTM Approach to Credit Scoring: Performance Evaluation with Implications for Explainability and Fairness

Abstract

Credit risk assessment is an important process in financial institutions, since it highly influences the decision in giving loans, determining portfolio quality, and overall financial stability. Traditional approaches such as logistic regression and scorecard model is still often used because they can be easily understood and fit perfectly for categorical data. Yet, these models are not effective enough in handling nonlinear connection and complex interactions among features in structured financial datasets. These observations are consistent with our previous argument that we study a hybrid deep learning framework (CNN and LSTM) for binary credit risk classification. The proposed model was evaluated on the German Credit Dataset (Statlog), which consists of 1,000 entries with 20 distinct features and a binary outcome (Good/Bad credit risk). The preprocessing step consists of ordinal encoding, Min-Max normalization, and an 80:20 stratified split for training and testing. We deal with class imbalance (70:30) by using a weighted binary cross-entropy loss function. The CNN layer collects the local representations of features, while LSTM layers amalgamate characteristics to retain sequential patterns in the transformed input at each layer.