Corporate governance, fraud learning cycles, and financial fraud detection: Evidence from Chinese listed firms

Corporate governance indicators play an important role in detecting financial fraud. Compared to the multilayer perceptron neural network (MLP NN) model, the extreme gradient boosting (XGBoost) model detects financial fraud more reliably, but suffers from a parameter search problem. An ant colony optimization algorithm can effectively optimize the model and increase its accuracy. Using data from 1,660 Chinese listed firms between 2015 and 2021, adding corporate governance indicators considerably increased the XGBoost model's accuracy. Model optimization and empirical evidence show that fraud detection accuracy is higher in the early fraud learning cycle than in the most recent cycle. Moreover, the accuracy of detecting fraud is higher in the short fraud learning cycle than in the long cycle, while two years is the optimal fraud learning cycle. This study also analyzes the mechanisms through which corporate governance indicators affect financial fraud detection. Notes：Model 1 = ORTA + FER + GMI + TPR + TFA + ET + GRTA + ROA + ONCS + ONCE = Fin Model 2 = Frate + Board + Indep + Dual + Top1 + Top5 + Balance + SOE + Inst + FinBack + OverseaBack + Fin Please verify if the modification meets the publishing requirements, and if there are any other issues, please feel free to contact me at any time. • This study compares the ability of XGBoost and MLP NN models to detect fraud • Models that include corporate governance indicators more accurately detect fraud • Fraud prediction is more accurate when ACO algorithms are added to XGBoost models • Accuracy of detecting fraud is higher in early learning and short learning cycle • Two-year is the optimal fraud learning cycle in order to conceal fraud