應用分類元系統於財務危機預警之研究

Abstract

財務危機預警對於公司的內部或外部利害關係人，一直是個重要的議題。早期學者多運用區別分析、logistic迴歸模型、或是probit迴歸模型等統計方法建立財務危機預警模型。然而，近年來許多已研究證實，諸如類神經網路(NNs)之人工智慧方法學對於分類問題(ex.預測公司財務危機)有較優異的表現。即使有些學者運用NNs預測財務危機得到有效的結果，但由於對資料之敏感性的關係，因而不易建構適當的架構；且在使用模型時，無法提供清楚解釋結果的能力，造成使用之不易。 本研究的主要目的為應用XCS分類元系統來建置公司財務危機預警模型。由於XCSR模型(一種XCS分類元系統的延伸模型)結合了增強式學習(Reinforcement learning)與演化式計算(Evolutionary computation)，因此具備優良的預測能力。且模型中的規則對於預測的結果具可讀性，公司的利害關係人因而較容易了解預測的結果。 經由本研究的實證，結果顯示XCSR模型的預測能力將顯著優於比較的logistic迴歸模型，以精確度而言，XCSR模型高達86.8%，logistic迴歸模型只有79.9%。另外，文中亦針對XCSR所得之規則，與logistic迴歸模型做一討論與比較。

The prediction of financial distress is an important and active topic since it is critical to all stakeholders both internal and external to the company. Earlier studies of financial distress prediction used statistical approaches such as multiple discriminant analysis, logistic regression and probit model. Recently, however, several studies have demonstrated that artificial intelligence methodology such as neural networks (NNs), has the superior abilities on classification problems. Even though some of the studies using NNs to the prediction of financial distress have reported its usefulness, there are still several drawbacks in developing and using these models. The sensitivity of financial data would affect building an appropriate model and the learning results could not be read comprehensibly.

The purpose of this paper is to propose XCS classifier systems approach and illustrate how the XCSR model (one model extended from XCS classifier systems) can be applied to financial distress. The exploitation of reinforcement learning and evolutionary computation constitutes a considerably advantage for the XCSR model to provide the superiorly predictive ability. Also, the obtained regularities are a means of easily understanding for the stakeholders of a firm.

The results obtained with the XCSR model showed to be significantly superior to those obtained from the benchmark model (the logistic regression model). The XCSR model has a better accuracy, it is 86.8% accuracy compared to logistic regression model, which only has 79.9% accuracy. Moreover, the extracted regularities were discussed for the increased understanding when comparing to the logistic regression model.