Leveraging Time Sequence Deep Learning Models: Impact of Hidden Layers on AI Model Performance in Credit Scoring

The need for a reliable and accurate credit scoring system is more evident than ever, driven by the imperative to enhance customer experience through real-time processing and execution of requests and approvals. Based on the conducted research, there is insufficient utilization of powerful deep learning (DL) algorithms in the domain of credit scoring. Numerous significant limiting factors hinder and decelerate DL implementation in credit scoring: small-sized datasets, scarcity, and unreliability of data, excessive complexity, slow training, and poor explainability of predictive models together with rigorous data protection and financial regulations, etc. This manuscript examines the effect of hidden layer depth in deep learning network architectures for models used in credit risk assessment. Does increasing the depth of the architecture help when working with static, non-sequential credit scoring data? Although rarely used in credit scoring prediction, Long Short-Term Memory (LSTM)-a highly advanced deep learning algorithm-has the potential to be well-suited for applications in the financial sector. The results indicate that, particularly when trained on small datasets like those used for credit scoring, increasing the complexity of neural networks with deepening them, often does not yield significant performance improvements. Findings were validated and analyzed by applying publicly available credit scoring datasets.