馬氏-田口系統之特性探討及其於信用評分之應用

Abstract

田口方法已廣泛地被應用在許多工程方面以改善產品及製程的品質。馬氏–田口系統(Mahalanobis-Taguchi System, MTS)是田口博士提出的新方法，它結合品質工程原理及馬氏距離，透過建構多變量量表來進行診斷及預測以提供決策之參考。 本研究首先探討MTS之特性，藉由改變各種實驗條件，包括馬氏空間、抽樣策略及訊號雜音比等，以評估MTS在分類之績效，並且利用紫羅蘭資料進行MTS特性之說明，以擴展此方法未來之發展與應用。其次，本研究將MTS應用在信用評分模式上，信用評分模式已廣泛被使用在信用決策上，主要可減少信用分析的時間與成本，雖然目前已存在不少信用評分模式，但傳統的模式並沒有考慮雜音對模式之影響，本研究以MTS建構一信用評分模式，並利用國內某銀行之實例驗證模式之可行性。此外，針對MTS處理類別資料之缺點，本研究亦提出一整合羅吉斯迴歸與田口方法之修正模式，以改善其績效；並將MTS與修正模式之實證結果與決策樹以及鑑別分析進行比較，發現MTS及其修正模式在信用評分的領域上可獲得很好的應用。 最後，本研究也針對MTS的幾個議題及未來的研究方向進行說明。

Taguchi Methods have been successfully applied to many engineering application to improve the performance of product and process. The Mahalanobis-Taguchi System (MTS) developed by Dr. Taguchi is a relatively new technology for diagnosis and forecasting using multivariate data. MTS is based on the principles of quality engineering and Mahalanobis distance (MD). This technology is aimed at providing a better prediction for multivariate data through the construction of a multivariate measurement scale. The results of MTS can be a reference for making decisions. For development and applications of MTS, to understand the features is very important. In this dissertation, first, we study the features of MTS, and evaluate the performance of MTS. Through changing the conditions of experiments including Mahalanobis Space, Sampling and Signal to Noise ratios, the performance of MTS was evaluated. Iris data was used to illustrate the performance of MTS. Secondly, the credit scoring system based on MTS is proposed. Credit scoring is widely used to make credit decisions, to reduce the cost of credit analysis and to make decisions fast. There are many credit scoring techniques, however, traditional credit scoring models do not consider the influence of noises. The effectiveness of the proposed MTS approach is demonstrated by real case data from a large Taiwanese bank. In addition, we propose a modified approach that is combined logistic regression (LR) with Taguchi’s approach for improving the shortcoming of categorical data of MTS. Results are benchmarked against two traditional methods including decision tree and linear discriminant analysis. Due to the results of the above, both MTS and modified approach (LRTM) can be applied to credit scoring system. Finally, several important issues regarding to the MTS and the suggestions of future research are summarized in the conclusion.

Key words: Mahalanobis-Taguchi System (MTS), Mahalanobis distance, orthogonal array, Signal-to-Noise ratio, Logistic regression, Credit scoring