公司危機預測：計量模型與變數選取

Corporate Failure Prediction: Econometric Models and Variable Selection

本文之主要目的在於探討衡量企業危機風險之計量模型中，不同模型之表現與變數選取之影響，同時，本研究提出一套程序用於支援向量機（Support Vector Machine, SVM）之變數選取，研究發現應用此程序可大大提升SVM之準確度與穩定度。研究期間與對象為1995年到2005年底，台灣上市、櫃企業曾被列為全額交割股的公司及用以比較之正常公司。所採用模型為Logit、Z-Score及SVM，另外納入以選擇權評價為基礎的DD值做為輔助變數，而類股間的差異則以虛擬變數（dummy variable）控制。本研究發現在企業危機的預測能力上，不管是完整模型或是變數篩選後模型，SVM均顯著優於Logit及Z-Score。解釋變數的選取明顯影響了模型預測的準確性與穩定性。平均而言，Logit存在過度配適（over-fitting）的問題，而SVM於完整模型同樣有這樣的問題存在，但在經過變數篩選後，可大幅改善這個問題。相較於使用naïve方法預測公司發生危機，以危機前一年的資料建構SVM模型可提升約11%的預測精準度。

The purposes of this paper are to examine the performance of econometric failure-risk models and to study the effect for variables selection. The sample data we considered in this study comprises listed companies in Taiwan Security Exchange Corporation (TSEC) and Over the Counter (OTC) from 1995 to 2005 that have ever been listed on “Full Delivery.” On the other hand, some other companies are also considered and regarded as the “normal companies” to be compared with. The models we used include the Logit model, Z-Score (Altman, 1968) and the support vector machine (SVM). In order to avoid the problem of window dressing for financial variables, we adopt two non-financial variables, namely default distance (DD) and industry dummy variable, in the models. It is found that SVM outperforms Logit and Z-Score in the prediction periods. The over-fitting problem exists in both SVM and Logit models. However, it can be reduced obviously in SVM as the explanatory variables are selected using the method we proposed in this study. Finally, by using the first-year data, the SVM can efficiently increase the forecasting ability compared with the naïve method.