Exploration of the MLP Neural Network Evaluation Model for Risk Control in Digital Inclusive Finance

To address the challenges of data complexity and high-dimensional feature selection in risk control for digital inclusive finance systems, this study introduces a digital inclusive finance risk control evaluation (DIFRCE) model. DIFRCE integrates a multi-layer perceptron neural network, the tree-structured Parzen estimator Bayesian optimization algorithm, and the layer-wise adaptive moments optimizer for batch training algorithm. It is validated using five financial risk datasets. Shapley additive explanations highlight the important roles of customer churn rate (value = 0.939) and market growth rate (contribution score = 0.806), underscoring their significant impact on customer behavior and market performance evaluation. System robustness is demonstrated through stability tests: the model's noise-resistance scores of 0.92 under data noise and 0.87 under simulated cyber-attack scenarios, with a maximum recovery time of less than four seconds and minimal deviation from original data. Prediction performance is evaluated using time-series analysis, indicating stable long-term errors (ranging from –0.50 to 0.49), low mean squared error (0.05 to 0.20), and acceptable autocorrelation and partial autocorrelation values (–0.2 and 0.20). Value at risk (VaR) model, which estimates the maximum potential loss at a given confidence level, confirms the DIFRCE model's effectiveness: the 95% VaR for credit risk and operational risk are 0.12 and 0.10, respectively, while the 99% VaR values are 0.25 and 0.20. These results confirm the model's effectiveness in delivering precise and comprehensive risk management tools, optimizing resource allocation, reducing potential losses, and enhancing system stability and security for financial institutions.