近年財務危機發生頻繁，許多公司皆因財務危機的發生瀕臨破產、下市，導致員工及投資人莫大的損害，也造成經濟市場的恐慌。財務危機的評估具複雜性，因此，如何能提早預測財務危機的發生，對於建立一個有效且準確的財務危機預測模型就相當重要。研究樣本取自於臺灣經濟新報資料庫(TEJ)，研究樣本為2011年至2021年台灣有發生財務危機之上市櫃及下市櫃公司，並以1:1及1:2進行配對，變數方面包含20個財務變數及5個非財務變數。本研究以隨機森林與決策樹C5.0篩選出重要變數，再利用卷積神經網路(CNN)及循環神經網路(RNN)建立有效之財務危機預測模型。實證結果顯示，在1:1樣本中，以隨機森林搭配卷積神經網路在預測準確率最高，平均準確率達92.11%；在1:2樣本中，仍以隨機森林搭配卷積神經網路為最佳之模型，平均準確率達83.11%。
In recent years, financial crises have occurred frequently, leading many companies to face bankruptcy and delisting due to financial distress. This has resulted in significant losses for employees and investors, causing panic in the economic market. The evaluation of financial crises is complex, making it crucial to develop an effective and accurate financial crisis prediction model. The research sample for this study was obtained from the Taiwan Economic Journal (TEJ) database, consisting of listed and delisted companies in Taiwan that experienced financial crises from 2011 to 2021. The sample was matched in a 1:1 and 1:2 ratio, and the variables included 20 financial variables and 5 non-financial variables. In this study, random forest and decision tree C5.0 were employed to identify important variables, and a combination of Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) was used to establish an effective financial crisis prediction model. The empirical results showed that in the 1:1 sample, the random forest model combined with CNN achieved the highest prediction accuracy with an average accuracy rate of 92.11%. In the 1:2 sample, the random forest model combined with CNN remained the best model, with an average accuracy rate of 83.11%.
