財務報表舞弊欺騙了財務報表的使用者，也會造成投資人重大 的損失，及時且有效的偵測財務報表舞弊是相當重要的。本研究以 2000 年至2019 發生財務報表舞弊之上市櫃公司為主要研究對象，先 使用隨機森林(random forest)來進行重要變數的篩選，再以類神經網 路(artificial neural network)、長短期記憶模型(long short-term memory model)與卷積神經網路(convolutional neural network)分別來建立偵測 模型。實證結果顯示，由隨機森林篩選重要變數加上卷積神經網路所 建立之模型(RF-CNN)有最佳的偵測能力，其整體準確率達95.66%； F1 分數達91.67%；接收者操作特徵曲線之曲線下面積達99.75%。
Fraud in financial statements deceives the users of financial statements and can also cause great losses to investors. It is very important to detect fraud in financial statements timely and effectively. This study focuses on Taiwan listed companies with fraudulent financial statements from 2000 to 2019 as the research object. First, the random forest is used to screen important variables, and then the artificial neural network, long short-term memory model and convolutional neural network are used to establish the detection models separately. The empirical results show that the RF-CNN model has the best detection ability, with an overall accuracy rate of 95.66%; F1 score of 91.67%; area under the receiver operating characteristic curve of 99.75%.
