Financial fraud risk prediction using a CNN-Mamba-Transformer model

Despite the frequent occurrence of financial fraud, capital markets often exhibit significant delays in disclosing such activities. Existing literature lacks robust methodologies for timely fraud prediction. Using a comprehensive dataset of Chinese listed firms covering fraud occurrence years from 2010 to 2023, this study develops a prediction model to speed up detection and solve the timeliness issue. To ensure predictive rigor, this study employs a strict temporal design where data from year T−1 are used to predict fraud in year T. The model integrates the convolutional neural network (CNN), Mamba, and Transformer. Our research yields three key conclusions. First, this model predicts best when integrating raw financial statements, ratios, and non-financial indicators as inputs. Second, this model outperforms other financial fraud prediction models and significantly enhances fraud prediction. Third, the model shows strong predictive power in case analyses. This study enriches deep learning applications for fraud prediction, reduces detection time for fraudulent behavior and offers policy implications for regulators.