<p>Existing credit risk assessment models exhibit strong predictive power but often remain limited in capturing complex interactions among risk factors and lack interpretability demanded by regulatory frameworks. These limitations create a persistent gap between accurate prediction and transparent decision-making. To address these challenges, this study proposes a credit risk assessment model based on feature engineering approach and interpretable sub-models, denoted as the FEIS model, which explicitly links methodological innovations to distinct problem dimensions. Specifically, to overcome the inability of traditional models to identify non-linear moderating effects, the FEIS model constructs interpretable interaction variables through feature engineering. To counter sample heterogeneity and imbalance that often distort risk estimation, cohesive clustering and propensity score matching are introduced to form homogeneous and balanced subsets. To reconcile the trade-off between predictive accuracy and interpretability, random forest and Lasso sub-models are employed as self-interpretable learners for probability of default (PD) and loss given default (LGD) prediction, respectively. Finally, a credit grant rule integrating PD, LGD, and expected loss rate (ELR) is optimized via the firefly algorithm to ensure risk-adjusted profitability under transparent and auditable logic. Empirical results on two datasets show that the FEIS model outperforms 12 benchmark PD predicting models and 6 LGD predicting models, while maintaining high stability and interpretability. These findings demonstrate that the design of the FEIS model, including linking feature engineering, interpretable sub-modelling, and intelligent optimization, naturally mitigates key challenges of opacity, heterogeneity, and inconsistency in contemporary credit risk assessment.</p>