University credit risk assessment via XGBoost and PFI with OCR-processed financial data

Currently, artificial intelligence has been widely applied in fields such as finance and fiscal management. However, existing intelligent assessment methods still face challenges when processing dynamic and complex economic data, particularly due to limited flexibility and insufficient model interpretability. To this end, the study proposes a credit risk assessment method that integrates XGBoost and replacement feature importance and introduces optical character recognition technology to structure unstructured texts such as budget reports and financial announcements issued by universities and construct a multi-dimensional, computable risk feature dataset. The method quantitatively assesses local government debt risk and identifies key influencing variables. Experimental results show that the model achieves an accuracy of 92.08%, outperforming logistic regression and random forest by 9.13% and 3.31%, respectively. Compared with the original unoptimized extreme gradient boosting model, the proposed method increases the area under the feature importance curve by 1.72% when using the top two features. The analysis highlights that debt service coverage ratio, fiscal self-sufficiency, and estimated implicit debt are the most critical indicators affecting local government credit risk. The proposed assessment framework has strong predictive performance and good transparency, which provides effective technical support for the assessment of educational financial risk in the context of smart education.