A Transfer Learning-Based Model for Predicting Internal Control Deficiencies

Abstract

This paper addresses the inefficiencies and heavy reliance on human judgment in traditional internal control auditing, as well as the challenges posed by sparse sample data. In this context, we propose a transfer learning–based predictive model for internal control deficiencies. The core idea uses U.S. SOX 404 data as the source domain and Chinese A-share listed companies as the target domain. The model is designed using a “pre-training—transfer—adaptation—calibration” framework, and is developed along four dimensions: the data and labeling system, model pre-training, model fine-tuning, and evaluation metrics. The research provides explainable risk indicators and practical guidelines for regulators, audit practitioners, and corporate governance.