運用倒傳遞類神經網路建構電子業財務危機預警模型

Abstract

上市公司陸續發生財務弊案後，很多人不禁要問：「企業舞弊掏空與財務危機是否可以預測？舞弊掏空與財務危機爆發前是否有蛛絲馬跡可尋呢？」因此，本研究試著運用類神經網路之優點，建立財務危機預警模型，在企業面臨財務危機前，從財務指標中找出可能出狀況的公司，提醒投資者趁早出脫手中持股，避免日後嚴重虧損。此外，並從危機公司危機前後的股價走勢，評估本研究模型是否有達到提早預警之效果。 考量不同產業都有其特殊的經營環境，財務指標與財務危機的模式自然會不相同，故選擇針對電子業單一產業別做財務危機預測。本研究利用資料平移法使財務變數值介於0~1 之間，並且自創出一套過濾影響力不顯著財務指標變數方法，以提高正確預測率。 本研究實驗結果發現：(1)使用危機前二季之季報，無法及時有效地預測四月份發生財務危機企業，故建議使用危機前三季之季報作為輸入資料。(2)危機前三季之季報公佈日至危機發生日後一周股價平均跌幅大於危機前四季，故並非越早發現危機公司，投資者會得到較好的效果。(3)無法從類神經網路模型中得知自變數與因變數之間關係，故難以分析型一誤差與型二誤差形成原因。(4)運用本研究模型，輸入危機前三季之季報資料，正確預測率為90%，股價平均跌幅為34.95%，可達到提早預警效果。

Enterprises encountered financial crisis one after another. Many people can't help asking a question, is there a way to forecast financial crisis or something wrong for a enterprise? Is there any clue detected before the events happen? Therefore, this research tries to adopt the advantage of neural network to build a financial crisis forecast model, expect to find out the companies which are battling financial crisis by financial indicates and remind investors to sell stock before the crises occur. Besides, it will compare stock price trend of before and after crisis to judge whether the model is really workable or not. It is proposed to choose electronic industry only for the forecast model. Every industry has its special management environment, the categories of financial crises and input financial indicates shall not be all the same. This research uses a data translation to make the values of financial indicates between 0 and 1, and also create a new method to filter out trivial variables to get higher ratio of prediction. The result and findings of the experiments are as following, (i) If the previous two quarter report is adopted, the forecast model can’t timely find out the company which meets financial crisis in April. Suggest adopting the previous three quarter report for input data. (ii) The average stock price decline ratio for the announcement date of the previous three quarter report is higher than the previous four quarter one. So, it is not to find out the company which meets financial crisis earlier and the investors will get better result. (iii) It can’t analyze the reasons for type I and type II errors because neural network is a black box. (iv) The forecast model is really workable. If the previous three quarter report is adopted, accuracy is 90% and stock price decline ratio is 34.95%.