<![CDATA[Credit risk prediction is important for supporting accurate lending decisions and reducing the potential risk of loan default. However, deep learning performance on tabular credit data can vary according to model architecture and TPOT hyperparameter configuration. Therefore, this study aims to comparatively evaluate Multilayer Perceptron, Convolutional Neural Network, TabNet, and FT-Transformer for credit risk prediction and examine the effect of TPOT hyperparameter optimization. The study uses the Lending Club dataset from 2007–2020, with 2,000 records from 2019–2020 sampled for analysis. The data were preprocessed through feature selection, missing-value handling, outlier treatment, categorical encoding, and feature scaling, while Synthetic Minority Over-sampling Technique was applied to address class imbalance. A stratified 70:30 train-test split was used, and the models were evaluated using accuracy, precision, recall, and F1-score before and after optimization. The results show that all four models improved after optimization, although the magnitude varied. Convolutional Neural Network achieved the highest optimized accuracy of 95.50%, while TabNet achieved the highest recall of 86.05% and F1-score of 86.29%. Multilayer Perceptron obtained the highest precision of 91.02%, whereas TabNet achieved the largest F1-score improvement of 6.84 percentage points. Overall, TPOT hyperparameter optimization improved model performance, with TabNet providing the most favorable overall balance for credit risk prediction.]]>
