結合多屬性決策與資料探勘技術建構企業績效評估與預測之模式

Combining Multi-Attribute with Decision-Making and Data Mining Techniques to Construct an Enterprise Performance Evaluation and Prediction Model

許書綸(Shu-Lun Hsu)

指導教授 ： 張洝源

國立虎尾科技大學/管理學院/工業工程與管理研究所/碩士(2013年)

摘要

財務績效表現是探討公司營運優劣的重要指標，也是投資者在投資企業時所注重的因素。近年來全球金融之衝擊頻傳，2010年的歐債危機更使得台灣企業面臨營運困境。本研究蒐集1996年至2011年間台經院產經資料庫企業季報資料為研究資料庫。此資料庫包含了以下六大指標：報酬率指標、成本指標、每股指標、成長率指標、償債能力指標及經營能力指標，其中更細分為73項準則。但因準則過多，企業較無法有效掌握影響企業績效表現之關鍵因子，且投資者要蒐集所有準則資料分析企業績效上也較困難。因此透過本研究之方法篩選出關鍵影響準則，並深入探討關鍵準則對企業績效表現狀況，以提供企業作為參考之依據。
本研究藉由(1)最大期望演算法及K-Means分群演算法將資料分類；(2)但因評估準則過多，為尋找評估關鍵因子，故利用基因演算結合邏輯迴歸分析及獲取各準則權重，並作為篩選後段準則之依據；(3)透過貝氏理論結合偏最小平方法找出準則間的關聯及顯著性，並獲得潛變量指標權重；(4)利用潛變量指標權重結合理想解類似度順序偏好法探討企業營運績效；(5)最後利用類神經網路作整體的績效分析訓練及測試，以建構績效評估預測系統。
最後結果顯示，在72項準則中，經過兩階段準則篩選作業後淨值/總資產、常續性EPS及總資產週轉率(次)最為重要，因此企業針對重點準則改善時，可以著重在此三項準則進行改善或加強作業。此外TOPSIS分析中也可清楚得知，若以類神經網路98.82%的高準確率之預測模型中，被預測歸類為類別2時，其績效表現0.3470最優，最差者為類別3(0.2594)。因此投資者在投資前進行企業績效分析時，若資料預測類別為2時，則建議投資。

關鍵字

分群演算法 ； 基因演算 ； 邏輯迴歸分析 ； 貝氏網路 ； 偏最小平方法 ； 理想解類似度順序偏好法 ； 類神經網路

並列摘要

Financial performances are important indicators in exploring whether a company has superior or inferior operations, and are also factors that investors notice when they are investing in companies. Recent years have shown frequent global financial impacts, and the 2010 European debt crisis further pushed Taiwanese industries toward operational difficulties. This study collected the corporate quarterly data from the Taiwan Institute of Economic Research for the period between 1996 and 2011 as the database for this research. This database included the following six indicators: rate of return, cost, per share, growth rate, ability to pay debts, and management ability. These indicators can be further divided into 73 criterions. However, because of the excessive number of criterions, corporations cannot effectively grasp key factors that affect corporate performance. It is also difficult for investors to collect all the criterion data to analyze corporate performance. Therefore, methods in this research were used to select key affecting criterions and explore their effects on corporate performance. The results can provide corporations with a reference base.
This study used Expectation-Maximization algorithm and K-means clustering method to classify the data. However, because of the excessive number of evaluation criterions, genetic algorithm was combined with logistic regression to obtain the weight for each criterion (to find the evaluation key factors), which were used as the basis for selecting later stage criterions. Bayesian theory was combined with partial least squares to find the correlation and significance between criterions, and to obtain the indicator weights for the latent variables. The indicator weights for the latent variables were combined with technique for order preference by similarity to ideal solution (TOPSIS) to explore corporate operating performance. Finally, artificial neural network was used for the overall performance analysis training and testing, to construct a performance evaluation forecasting system.
The study results show that of the 72 criterions (after two stage criterion selection), equity/total asset, EPS - net income, and total asset turnover (times) were the most important criterions. Therefore, if corporate managers wish to improve corporate performance, they should focus on these three criterions to improve and strengthen their related operations. In addition, TOPSIS analysis clear shows that if the forecast was categorized as type 2 (using artificial neural network forecast model, which has a 98.82% accuracy rate), its performance was the best. The category with the worst performance was type 3. Furthermore, analyzing from the investors’ perspective, if the pre-investment corporate performance analysis of the data showed the data forecast type as two, then investment is recommended.

並列關鍵字

Expectation-Maximization ； K-Means ； Genetic Algorithm-Logistic Regression ； Bayesian Network ； Partial Least Squares ； Technique for Order Preference by Similarity to Ideal Solution ； Artificial Neural Network

參考文獻

1.尹以修 (2004)，軍事資訊系統計畫建案風險管理之研究，臺灣大學臺灣大學商學研究所學位論文，碩士論文。

7.范姜智勇 (2008)，基因演算法中交配與突變交互作用之研究，國立台北科技大學光電工程系研究所，碩士論文。

8.范維真 (2010)，個人信貸信用風險評分卡模型之探討，淡江大學統計學系碩士班，碩士論文。

9.高子傑、洪僖黛、鄧文雋 (2009)，“台灣地區電子工業公司之財務績效特徵與差異”，危機管理學刊，第6卷，第1期，115-126頁.

10.陳秉洋 (2012)，結合技術指標、籌碼變動及期貨基差建構整合性投資策略─以台灣50指數型基金為例，成功大學成功大學統計學系碩博士班，碩士論文。