近年許多公司破產、被下市的案例頻傳，皆導致投資人莫大的損失，也造成市場的恐慌。繼續經營的評估具複雜性，因此，如何出具正確繼續經營之意見，對於審計人員的專業判斷就相當重要。相較於以往傳統的迴歸模型，本研究則是使用可處理更具複雜且龐大資料的深度學習方法，建構一個準確率高的繼續經營預測模型。研究資料取自於臺灣經濟新報資料庫(TEJ)，研究對象為2009年至2018年有繼續經營疑慮及無繼續經營疑慮之上市櫃公司，並以1:3進行配對，變數方面包含財務變數及非財務變數。本研究以決策樹CHAID篩選出具有重要性之變數，再以循環神經網路建立預測模型，並跟未經變數篩選之循環神經網路模型進行比較。實證結果顯示，以決策樹CHAID搭配循環神經網路為最佳預測模型，其準確率達91.65%。
In recent years, it has become frequent to see cases of bankruptcy and delisting of entities, which have caused huge losses to investors and caused panic in the capital market. The assessment of going concern is beyond complicated. Therefore, it is more than important for auditors’ exercising their professional judgements to bring a rather appropriate opinion forward. In contrary to traditional regression model, this study focuses on using deep learning to deal with complicated and massively enormous data base, to construct a going concern predict model with accuracy. The research data was taken from Taiwan Economic Journal (TEJ), the research objects are listed companies with and without going concern doubts for the financial statements issued for the period from 2009 to 2018, with a matching ratio of 1:3. The variables identified for assessment include financial and non-financial ones. The study scrutinizes variables of importance by applying decision tree CHAID. Then the recurrent neural network is adopted to establish a model for predicting if a going concern issue does exist. In addition, it is also compared with the recurrent neural network model without performing the scrutiny of variables. The empirical results show that decision tree CHAID combined with Recurrent Neural Network is the best prediction model with accuracy rate at 91.65%.
