This study investigates the effect of training period length on the detection accuracy of accounting fraud detection models, focusing on Japanese companies. Using data from companies listed on the Tokyo Stock Exchange from 2008 to 2018, models were constructed using three classification algorithms: logistic regression, random forest, and eXtreme Gradient Boosting (XGBoost). The training periods were set from 3 to 7 years, and the detection accuracy was evaluated using the area under the receiver operating characteristic curve (ROC-AUC), area under the precision-recall curve (PR-AUC), and Recall@k.

The results showed that a 4-year training period yielded the best ROC-AUC across all models. However, no clear trend was observed between the training period length and accuracy for the other durations. For PR-AUC and Recall@k, all models showed low performance with a 3-year training period. Extending the training period to 4-5 years improved performance, whereas further extension to 6-7 years led to a decrease in accuracy. The optimal training period varied from model to model: 5 years for logistic regression and XGBoost, and 4 years for random forest.

A key finding of this study is that training periods that are too short provide insufficient data for effective learning, whereas training periods that are too long can cause noise from older data, which reduces the accuracy. These results underscore the importance of selecting an appropriate training period to improve model performance and provide practical insights into the development of accounting fraud detection models.