科技部補助專題研究計畫成果報告
期末報告
財務危機預測模型之建構: 以統計方法與智慧型技術結合
為例
計 畫 類 別 ： 個別型計畫 計 畫 編 號 ： MOST 102-2410-H-146-003- 執 行 期 間 ： 102 年 08 月 01 日至 103 年 08 月 31 日 執 行 單 位 ： 華夏學校財團法人華夏科技大學資訊管理系 計 畫 主 持 人 ： 陳祐祥 計畫參與人員： 大專生-兼任助理人員：林雅雪 大專生-兼任助理人員：陳國瑞 報 告 附 件 ： 出席國際會議研究心得報告及發表論文 處 理 方 式 ： 1.公開資訊：本計畫可公開查詢 2.「本研究」是否已有嚴重損及公共利益之發現：否 3.「本報告」是否建議提供政府單位施政參考：否中 華 民 國 103 年 10 月 26 日
中 文 摘 要 ： 尋求一個混合預測模型能夠從股票市場公開之財務報表中尋 找出財務艱困的行為，和可以提供投資者一個可靠的技術需 求是很重要的。本研究整合財務比率的專業知識結合軟計算 技術以建構混合公司艱困的預測模型，用以提供早期預警措 施。實證結果顯示本研究所提之混合預測模型，具有卓越預 測績效和應用價值，並可做為以後預測公司破產的可行方案 之一。此外，分析結果所得之管理意涵可做為其他型式的預 測模型於定義有效的財務比率時之參考使用。 中文關鍵詞： 財務艱困、混合程序、財務預測、粗糙集分類器。 英 文 摘 要 ： 英文關鍵詞：
行政院國家科學委員會補助專題研究計畫
行政院國家科學委員會補助專題研究計畫
行政院國家科學委員會補助專題研究計畫
行政院國家科學委員會補助專題研究計畫
■
■
■
■ 成 果 報 告
成 果 報 告
成 果 報 告
成 果 報 告
□
□
□
□期中進度報告
期中進度報告
期中進度報告
期中進度報告
（計畫名稱）財務危機預測模型之建構:以統計方法與智慧型技術結
合為例
計畫類別：■個別型計畫 □整合型計畫
計畫編號：NSC 102－2410－H－146－003－
執行期間： 102 年 08 月 01 日至 103 年 08 月 31 日
執行機構及系所：華夏科技大學 資訊管理系
計畫主持人：陳祐祥
共同主持人：
計畫參與人員：林雅雪、陳國瑞
成果報告類型(依經費核定清單規定繳交)：■精簡報告 □完整報告
本計畫除繳交成果報告外，另須繳交以下出國心得報告：
□赴國外出差或研習心得報告
□赴大陸地區出差或研習心得報告
■出席國際學術會議心得報告
□國際合作研究計畫國外研究報告
處理方式：
除列管計畫及下列情形者外，得立即公開查詢
□涉及專利或其他智慧財產權，□一年□二年後可公開查詢
計畫成果
計畫成果
計畫成果
計畫成果報告
報告
報告
報告內容
內容
內容
內容(
(精簡版
((
精簡版
精簡版)
精簡版
))
)
主持人本次執行的國科會專題研究計畫編號: MOST 102-2410-H-146-003, 已執行完成的研究計畫 成果計有 3 篇 papers，如下: (1) 1 篇 EI paper (Entitled: Building a hybrid prediction model to evaluation of
financial distress corporate, 2014, Vols. 651-653, pp. 1453-1456) 已 接 受 刊 登 在 國 際 期 刊 Applied Mechanics and Materials (AMM)以及另 1 篇 paper (Entitled: Consolidating mathematical statistics and artificial intelligence techniques to building an evoluting hybrid model) 已 於 近 期 投 稿 至 Economic Modelling 國 際 SSCI 等 級 期 刊 和 (2) 1 篇 國 際 研 討 會 論 文 接 受 並 發 表 於 國 際 研 討 會 : The 4th International Conference on Computer Application and System Modeling (ICCASM 2014), Shanghai, China, on August 09-10, 2014。後續並著手投稿計畫，希望再將此 1 篇已發表在國際研討會上之論文，改寫成
適合 SSCI 或 SCI 等級期刊英文論文格式和品質，之後擬再投稿於國際 SSCI 或 SCI 等級期刊。由於投 稿原文有 11000-12000 字(英文)和將近 40 頁次內容，茲因限於篇幅，因此無法將上述已完成論文或刊 登之 3 篇研究論文全數列印出，以下僅就此計畫之其中一個研討會之研究成果論文刊出如下:
Paper title: Building a Hybrid Prediction Model to Evaluation of Financial Distress Corporate
Abstract: Exploring financial distress activity within a listed target of stock markets focused on creating such
prediction models can provide insight into the technological requirements of corporate and the demands placed upon a stock investor in this field. This study integrates professional knowledge to financial ratios into the emerging soft computing techniques for building up a hybrid corporate distress prediction of early warning systems in regarding application fields. Conclusively, the empirical results indicate that the proposed procedure is a great potential alternative of helpful hybrid models to demonstrate its technological merit and application value, and it has increasing the application filings. In terms of managerial implications, the analysis results may be relevant to other types of prediction models seeking to identify financial ratios for the planning processes.
Keywords: Financial distress; Hybrid procedure; Financial prediction; Rough set classifier.
1. Introduction
The financial investment is always attracting increased attention for pursuing vigorously the money game by stock investors. Particularly, an effective prediction model, having the advantage of early warning systems (EWSs), becomes critical success factor to investors. Thus, to adequately handle financial status (or called financial health) through proper indicators, such as financial ratios, is a problematic but valuable issue and inspires priority concerns by both practitioners and academicians. To mine the financial ratios available to constructing effective financial distress prediction model is first motivated and emphasized on this study. As for financial ratios, the close relevance of financial statements should be contiguously addressed. The
operating performance of a corporate during a stated period of time is usually reflected in its financial statements, which are summarized or organized from several categories of financial ratios. For the recent decades, such a prediction function of models is incessantly studied in a wide range of fields, including mathematical statistics, artificial machine learning, data mining, pattern recognition, and signal processing.
Previously, prior to artificial intelligence (AI) techniques, statistical methods, such as univariate statistical methods, multiple discriminant analysis, linear probability models, logit analysis, and probit analysis, have been used for solving real-world financial prediction problems [1]. The statistical techniques have shortcomings, including the restrictive assumptions of linear separability, multivariate normality, and independence of the predictive variables [2]. The algorithm of machine learning methods, like neural networks (NNs), support vector machine (SVM), and rough set theory (RST), have been successfully applied to address financial prediction problems [3]. The RST has been widely experienced in the financial distress prediction and performed well applied in the financial fields [4]. This study thus selects rough set classifier as a base of the proposed procedure. This study proposes a hybrid procedure for classifying financial status to maximal the hybrid advantage of models. The main objective of this study aims to test effects of discretization-based feature selection.
2. Related works
This section reviews the related issues of financial prediction for interested parties, including the topics covered in the financial distress and its applications, Financial Ratio, rough set theory, rule extraction method, and rule filter algorithm.
2.1 Financial Distress and its Applications
Financial healthy or unhealthy status of a corporate is an important indicator, not only representing the operating performance but also indicating its potential growth in the future. A bad financial status is unhealthy and is identified as a negative effect of possible hardness, such as distress, failure, crisis, decline, bankruptcy, difficulty, or alarm [5]. Thus, the term of financial distress for resulting in enormous losses on finances is specifically focused on future corporate prediction in this study. The financial distress is a crucial topic when it comes to the inability of a firm to meets its loan obligations.
2.2 Financial Ratio
Typically, financial ratios [5] provide a clear picture of the financial information of a company that is otherwise unclear from raw financial data in evaluation of the overall effectiveness of business management based on quarterly and/or annual sales and investment performance from the four common financial statements of a business, including balance sheet, income statement, cash flow statement, and statement of stockholders' equity. The financial statement analysis includes many of the financial ratios, but finding
representative attribute financial ratios not only can raise prediction accuracy of financial status on a firm but also can make comparisons within the same industry.
2.3 Rough Set Theory
Rough sets (RS) find effective uses in “reducing data or eliminating superfluous data without any information loss, discovering data dependencies, estimating data significance, generating decision algorithms in given data, approximating data classification, discovering similarities or differences from data, discovering data patterns, and discovering cause–effect relationships.” RS has been successfully applied to real-world classification problems, having the advantages [6]: (1) RS data analysis results in information contained in a large number of cases reduced to a model that includes a generalized description of that knowledge, (2) the model is a set of easily understandable decision rules, which normally require no interpretation, (3) each decision rule is supported by a set of real examples, and (4) no additional information is required. The concept of RS can be defined using lower and upper approximations [7].
2.4 Rule Extraction Method
In knowledge discovery of data mining field, the generation of decision rules from a given data is an important concern for interested parties. Rule induction algorithms were first implemented in a Learning from Examples based on Rough Sets (LERS) system [8]. Particularly, the Learning from Examples Module, version 2 (abbreviated as LEM2) algorithm is a part of the LERS system and is one of most popular rule extraction method used in generation of decision rules.
2.5 Rule Filter Algorithm
Typically, some rule-filtering algorithms are applied to reduce the number of distinct decision rule of set and to refine the quality of rules since numerously redundant or ‘poor quality’ decision rules are generated, and these decision rules limit the classification capabilities of the rule set [9]. The rule-filtering solution determines the strength of the rule based on support, consistency, and coverage to calculate the quality indices of rules from a generated rule set.
3. The Proposed Procedure
The computational algorithm of the proposed procedure is divided into eight procedures systematically, as follows:
(1) Selecting the essential attributes,
(2) collecting the data of pre-selected attributes, (3) transforming format,
(4) discretizating data, (5) selecting feature,
(6) generating rule, (7) filtering rule, and (8) evaluating performance.
4. Empirical Case Study Using Financial Data
Experimental Data Set. Examples of a financial data set are extracted from a real database as empirical case
study, which is the Source Information Retrieval Systems (OSIRS) database [10] used to verify the classification performance of the proposed model. The real-world finance database includes balance sheets, income statements, statistical analysis, and ownership and subsidiary information in research and analysis. To implementing the computational processes of the proposed procedure by using the examples in practice is discussed step-by-step in detail, as follows:
(1) Selecting Attributes. A labeled OSIRS data set, extracted from the OSIRS database, is exampled in this
study. The OSIRS data set includes 23 essential (condition and decision) attributes that are initially selected based on the related literature review and expert recommendations from numerous features. The 23 attributes are termed X1-X23, all having continuous data. The first 22 attributes are the conditional-attributes, and the last one (X23) is the decisional-attribute. The decisional-attribute is named ‘Status’ and coded into two categories: ‘T’ (normal) and ‘F’ (distress). For example, the X1 attribute refers to ‘Return on equity ratio’, and the X11 is ‘Fixed assets turnover.’
(2) Collecting Data. Collect the data of 400 listed stocks (including normal and distress companies) in the
pre-selected 23 attributes in Taiwan into the OSIRS data set for the periods from 2002/01/01 to 2006/12/31 to achieve the verification purpose of the proposed procedure.
(3) Transforming Format. The incorrect values on attributes are eliminated from the OSIRS data set; as a
result, the 1,366 instances in EXCEL format are characterized by the 23 attributes.
(4) Discretizing Data. After the Global discretization, the data discretization information of the 22
condition-attributes is exposed. The 15 attributes, X1-X4, X6-X9, X12-14, and X19-X22, have discretized; that is, the remaining seven attributes are not. For example, the attribute ‘Return on equity ratio’ (X1) has one threshold (0.7719), corresponding to two linguistic terms: A1 ([−∞,0.7719)) and A2 ([0.7719,∞]), which are reacted to the natural language—low and high, respectively.
(5) Selecting Feature. The data discretization results are used as a tool of feature selection. The 15
conditional-attributes, including X1-X4, X6-X9, X12-14, and X19-X22, and one decision attribute X23 are selected as key attributes, and the others are eliminated from the OSIRS data set.
(6) Generating Rule. In the rule learning stage, the 1,366 instances of the selected 15 conditional-attributes
and the one decisional-attribute are randomly split into two sub-datasets: the 66% training set and the remaining 34% testing set. As a result, the original classification accuracy 84.9% is achieved, and a set of 238 rules is created. The top 1 rule is formatted as ‘IF (X1="(-0.12725,∞)") & (X7="(-0.07955,∞)") & (X8="(-∞,-0.0113)") & (X18="(-0.43065,∞) & (X20="(-0.06745,∞)") & (X22="(-0.1275,∞)") Then Status=T’ in the 15 selected attribute samples by the LEM2 algorithm, and its support (or called match) has 110.
(7) Filtering Rule. The rule filtering algorithm is used to decrease the number of decision rules to improve
the rule quality and simplify the knowledge-based prediction system. The rule set with a low support threshold of 1 is removed, which indicates the rule supported by only one real-world example in the experiment data set. After the rule-filtering algorithm, the polished classification accuracy still has 84.9%, but 67 rules are erased. Importantly, only the remaining 171 rules are built as knowledge-based prediction system for assisting interested parties, such as stock investors.
(8) Evaluating Performance. For further comparisons and verification, the proposed procedure is compared
to various AI methods, including traditional RS [7], decision-trees C4.5 [11], multilayer perceptron (MLP) [12], and logistic-regression (LR) [13] for experiments repeated 10 times with the same 66%/34% random sampling. The average accuracy with standard deviation on the 10 tests is accordingly calculated. Table 1 shows the performance comparison on the five AI methods. Clearly, the proposed procedure having 83.7% with standard deviation 2.62% surpasses the other four methods in terms of accuracy in the OSIRS data set; simultaneously, the proposed procedure also has fewer used attributes and number of rules than those of the listed methods from Table 1.
Table 1. Comparison results of different five AI methods in the OSIRS data set
Model Traditional RS C4.5 ANN-MLP LR The proposed procedure
Accuracy 81.2% 76.9% 80.2% 79.1% 83.7%
Used attribute 23 23 23 23 16
Number of rules 209.9 165.0 - - 162.2
5. Conclusions
Exploring activity using financial ratios on the financial status of a company is an interesting issue and attracts priority-concerns. In this era of global economic uncertainty and high-risk, investors are building intelligent forecasting models for classifying the financial status problems. This study has proposed a hybrid corporate distress prediction model based on consolidating professional experience, Global discretization, discretization-based feature selection, the LEM2 algorithm, rough set classifier, and filtering rule algorithm
for classifying the financial status of a company. Particularly, the performance of discretization-based feature selection method is verified in this study. Conclusively, the empirical results of a case study indicate that the proposed procedure is a great potential alternative of helpfully integrated hybrid model to demonstrate its technological merit and application value. In terms of managerial implications, it will be particularly and really valuable that this study further explored suitable application facets in practice and the analytical results applied by both practitioner and academicians fully. The main contribution of this study was conducted to reflect a new trial in the financial field for stock investors and obtained a satisfied result. This study still needs a further improvement on methodology and extensive applications, including (1) various types of decision attributes other than financial distress are used to validate the performance of the proposed procedure, and (2) extra condition-attributes are considered into the proposed procedure.
Acknowledgements
The authors would like to thank the Ministry of Science and Technology of the Republic of China, Taiwan, for financially supporting this research under Contract No. 102-2410-H-146-003.
References
[1] S. Canbas, A. Cabuk and B.S. Kilic: Prediction of commercial bank failure via multivariate statistical analysis of financial structures: The Turkish case. European Journal of Operational Research Vol. 166 (2005), p. 528-546.
[2] A. Ravi, H. Kurniawan, P.N.K. Thai and P. Ravi Kumar: Soft computing system for bank performance prediction. Applied Soft Computing Vol. 8 (2008), p. 305-315.
[3] H.M. Chung and K. Tam: A comparative analysis of inductive learning algorithms. International Journal of Intelligent Systems in Accounting, Finance and Management Vol. 2 (1993), p. 3-19.
[4] F.E.H. Tay and L. Shen: Economic and financial prediction using rough sets model. European Journal of Operational Research Vol. 141 (2002), p. 641-659.
[5] P. Ravi Kumar and V. Ravi: Bankruptcy prediction in banks and firms via statistical and intelligent techniques–A review. European Journal of Operational Research Vol. 180 (2007), p. 1-28.
[6] R. Slowinski and C. Zopounidis: Application of the rough set approach to evaluation of bankruptcy risk. Intelligent Systems in Accounting, Finance and Management Vol. 4 (1995), p. 27-41.
[7] Z. Pawlak: Rough sets: Theoretical aspects of reasoning about data, Kluwer Academic Publishers (1991).
[8] J.W. Grzymala-Busse: LERS—A system for learning from examples based on rough sets. Handbook of Applications and Advances of the Rough Sets Theory (1992), p. 3-18.
[9] Y.S. Chen and C.H. Cheng: Hybrid models based on rough set classifiers for setting credit rating decision rules in the global banking industry. Knowledge-Based Systems Vol. 39 (2013), p. 224-239.
[10] The Open Source Information Retrieval Systems (OSIRS) database. Retrieved on April 20, 2014 from http://newmops.tse.com.tw/.
[11] J.R. Quinlan: C4.5: Programs for machine learning. CA: Morgan Kaufmann, San Mateo (1993).
[12] S.Haykin: Neural networks, a comprehensive foundation (2nd edition). Section 2.13, p. 84-89; Section 4.13, p. 208-213. Prentice-Hall, Upper Saddle River New Jersey (1999).
國科會補助專題研究計畫項下出席國際學術會議心得報告
國科會補助專題研究計畫項下出席國際學術會議心得報告
國科會補助專題研究計畫項下出席國際學術會議心得報告
國科會補助專題研究計畫項下出席國際學術會議心得報告
日期： 103 年 10 月 26 日計畫編號
MOST 102－2410－H－146－003－
計畫名稱
財務危機預測模型之建構: 以統計方法與智慧型技術結合為例
出國人員
姓名
陳祐祥
服務機構
及職稱
華夏科技大學 資訊管理系
會議時間
(1)
103 年 07 月 03
日至 103 年 07
月 05 日
(2)
103 年 08 月 09
日至 103 年 08
月 10 日
會議地點
(1) Osaka, Japan
(2) Shanghai, China
會議名稱
(中文) (1) 2014 年企業與資訊國際研討會
(2) 第四屆 2014 年計算機應用及系統模式國際研討會
(英文) (1) The 2014 International Conference on Business and
Information (BAI 2014) (on July 03-05)
(2) The 4th International Conference on Computer Application and
System Modeling (ICCASM 2014) (on August 09-10)
發表論文
題目
(中文) (1) 應用重複抽樣方法解決不平衡類別問題於醫療服務產業
(英文) (1) A bootstrap method of addressing unbalanced class problems in
the healthcare service
(中文) (2) 建構混合預測模型於財務艱困公司之評估
(英文) (2) Building a hybrid prediction model to evaluation of financial
distress corporate
一、參加會議經過
本次研究計畫之主持人共出席2次國際研討會:在(1)日本大阪和(2)中國上海分別舉辦的這 2場國際學術研討會議，第一次是在日本大阪7月03-05日的2014年企業與資訊國際研討 會，第二次則為在中國上海8月09-10日的第四屆2014年計算機應用及系統模式國際研討會 議。總共發表的這2篇國際研討會論文，題目分別為(1) 應用重複抽樣方法解決不平衡類 別問題於醫療服務產業和(2) 建構混合預測模型於財務艱困公司之評估。在這2次前後共 10多天(7月03日-7月09日和8月08-12日)的參加會議行程中，行程滿滿而且收獲良多，茲將 這2次會議的主要內容和經過分述如下:2 7/03 桃園國際機場搭機直飛日本大阪國際機場並至BAI 2014會場報到並領取資料開幕式 及註冊。 7/04 參與和聆聽Keynote speaker專題演講以及發表一篇研討會論文並加討論。 7/05 參加和聆聽Keynote speaker專題演講，並藉機與國際學者熱烈互動。 7/06-08 參訪日本熊本市區及其周邊景點等。 7/09 準備返回行程及返抵台北。 8/08 桃園國際機場搭機直飛中國上海浦東國際機場。 8/09-8/10 參加ICCASM 2014會議開幕式、報到領取會議相關資料及發表一篇研討會論文。 8/11 參訪中國上海市區及其周邊景點等。 8/12 收拾行李準備返回台北行程。
二、與會心得
本專題研究計畫共參加 2 次國際學術研討會，會議期間共為五天，分別包括(1) 07/03~07/05 日本大阪和(2)08/09~08/10 中國上海。第一次是:(1)參加在日本大阪舉辦的國際學術研討 會，會議期間共 07/03~07/05 三天，是由 International Business Academics Consortium (iBAC) 和 Academy of Taiwan Information Systems Research (ATISR)單位主辦，主要參與國家有Malaysia, New England, Japan, Israel, Mauritius, Iran, Sri Lanka, Australia, Turkey, Singapore, Taiwan 和 United States 等國家或區域，議題範圍有 Accounting, Economics, Electronic Commerce, Entrepreneurship, Financial and Banking, Health Care Administration, Human Resource, Information System and Technology, Management Information System, Operations Management, Research Methods, 和 Other Relevant Topics 等，共有超過數百位各國學者與
業界專家與會，口頭論文發表近數百篇，並有多場專題演講。透過此次與會學者專家的意 見交流，相當廣泛多元，對相關問題之實務與研究都有深一層的探討。此次藉由參加研討 會機會和國際學者從事經驗交流，以建立個人學術界的人脈。第二次是(2)在中國上海舉 行的國際研討會，是由 Computer Science and Electronic Technology International Society 和
TTP Press 單位共 同協辦，商討在 Computer Science, Electronics Engineering, System Modeling and Automation 的相關技術 事 件。相關議題 包含有 : 如 Computer Science, Electronics Engineering, System Modeling, Automation, Networking, Communication and Other topics 等，希望建立一個提供國際上科學家 Scientists, 工程師 Engineers, 學術單位 Educators 和學生們 Students 共用平台以討論上述議題。此次國際研討會主要的參與國家 包括有 Korea、UK、USA、India、Canada、Libya、Australia、Hong Kong、Malaysia、Iran 和 Taiwan (ROC)等多個國家或區域人員，主要以計算機科學與應用和相關技術為主，共 計 2 天有數十位各國學者與業界專家與會，口頭論文發表近數十篇。在參與的這 2 場國際 研討會議中藉由論文發表的機會，可以實際和來自不同國家的國際學者從事學術的經驗交 流。除了大家可以多方參加討論外，所報告的議題內容也是多元化的，因此對主持人個人 未來於相關軟計算技術等研究上必有相當大的助益。這 2 次的相關議題無論是理論或實務 技術多相當廣泛多元化，對目前世界比較受到各國關心的問題也都能受到熱烈的探討與研 究。最重要的，個人希望透過這 2 次與會專家和學者間的意見交流和交換名片，能對個人
未來研究有進一步的幫助；此外，主要希望可以幫助提高我國在國際間的學術研究之知名 度，輔以建立個人學術界的國際人脈。因此主持人非常感謝科技部提供此 2 次出國經費， 讓此次日本大阪和中國上海之 2 次行程，實質收穫和受益良多，除可以提昇個人未來研究 能力，更增加國家和學術界之曝光率與能見度，可謂收獲滿滿。
三、考察參觀活動
在第一次的會議行程中參訪了大阪區域和京都周邊景點，大阪在日本類似於我們的高雄， 是比較屬於工業都市，地方較大，人口也不少，更是政令指定都市之一。由於鄰近於有日 本古都之稱的京都，因此交通都很便利。總歸來說，日本不愧為已開發國家，街道乾淨， 人民非常守法和彬彬有禮，汽車都會自動禮讓行人，處處可見文明大國的風範；確實，她 們有些地方是值得我們學習和稱讚的。此外，他們在許多古文化建築的保留上確實也很用 心，一些具歷史上特殊意義的建築都被完整保留或保護。在處處可見日本國家在開發現代 建設時卻也能兼顧傳統文化之用心。印象更深刻的是日本人的實事求是精神，相當值得我 們國人學習與推廣的。在第 2 次的行程中順道參訪了中國大陸的上海，上海果然是中國的 第 2 大城，人口就有 2 千多萬人，到街道上或景點走，只有一句話可以形容:就是擁擠， 到處除了人還是人。不過，他們的地鐵(相當於我們的捷運)建構得倒是相當完整，只是車 箱人擠人罷了。綜合而言，而且他們的文化素質還是比不上台灣的，在人文上應該還落後 台灣 1-20 年。上海的國父孫中山紀念館和博物館等，館內蒐藏了很多懷舊文化和歷史淵 源背景，值得一再探究的。大陸的五千年悠久歷史，確實是有很多地方值得我們挖掘。四、建議
主持人相當感謝科技部對本人這 2 次的參與國際研討會議相關經費大力支援，使得個人 才可以有此機會進行日本大阪和中國上海之學術行程，並得以有這麼好的機會可以參加 這 2 場的國際學術會議論文發表；相信這樣務實的學術經驗，不管對個人往後的研究計 畫會有較深遠影響和啟發或增進我們國家的國際學術地位都會有正面影響；因此非常建 議 貴部能夠繼續支持和補助國內學者參加此類這麼有意義的國際學術研討會議活動，加 以積極鼓勵國內研究人員參與發表其研究成果於國際學術會議上，除可促進與各國學者 之互動外，更期望能增加國內學術界在國際上的研究水平、整體研究品質與能力和國際 地位，並且間接增加台灣國民外交之能見度及研究領域之深度和廣度，善盡做為國家一 份子之責任。五、攜回資料名稱及內容
此 2 次國際研討會攜回資料計有 BAI 2014 和 ICCASM 2014 之 Program book 和所有發表 者之發表順序表以及研討會論文集光碟一片，相關內容包括有此 2 次會議的所有會議議 程、作者、與所有論文摘要。
六、其他
參加 2 次國際研討會會議之部份照片: (1) 參加中國上海國際研討會部份照片
4
(2) 參加日本大阪國際研討會部份照片
附錄
附錄
附錄
附錄、
、
、研討會論文
、
研討會論文
研討會論文
研討會論文
(註: 本次專題研究計畫在前後共計10多天的兩次國外參與會議之行程內(包括日
本大阪和中國上海)共發表2篇國際研討會論文，因限於篇幅，以下僅列出其中一篇
中國上海研討會之全文，如下：)
Title: Building a Hybrid Prediction Model to Evaluation of Financial Distress
Corporate
Abstract: Exploring financial distress activity within a listed target of stock markets
focused on creating such prediction models can provide insight into the technological requirements of corporate and the demands placed upon a stock investor in this field. This study integrates professional knowledge to financial ratios into the emerging soft computing techniques for building up a hybrid corporate distress prediction of early warning systems in regarding application fields. Conclusively, the empirical results indicate that the proposed procedure is a great potential alternative of helpful hybrid models to demonstrate its technological merit and application value, and it has increasing the application filings. In terms of managerial implications, the analysis results may be relevant to other types of prediction models seeking to identify financial ratios for the planning processes.
6
1. Introduction
The financial investment is always attracting increased attention for pursuing vigorously the money game by stock investors. Particularly, an effective prediction model, having the advantage of early warning systems (EWSs), becomes critical success factor to investors. Thus, to adequately handle financial status (or called financial health) through proper indicators, such as financial ratios, is a problematic but valuable issue and inspires priority concerns by both practitioners and academicians. To mine the financial ratios available to constructing effective financial distress prediction model is first motivated and emphasized on this study. As for financial ratios, the close relevance of financial statements should be contiguously addressed. The operating performance of a corporate during a stated period of time is usually reflected in its financial statements, which are summarized or organized from several categories of financial ratios. For the recent decades, such a prediction function of models is incessantly studied in a wide range of fields, including mathematical statistics, artificial machine learning, data mining, pattern recognition, and signal processing.
Previously, prior to artificial intelligence (AI) techniques, statistical methods, such as univariate statistical methods, multiple discriminant analysis, linear probability models, logit analysis, and probit analysis, have been used for solving real-world financial prediction problems [1]. The statistical techniques have shortcomings, including the restrictive assumptions of linear separability, multivariate normality, and independence of the predictive variables [2]. The algorithm of machine learning methods, like neural networks (NNs), support vector machine (SVM), and rough set theory (RST), have been successfully applied to address financial prediction problems [3]. The RST has been widely experienced in the financial distress prediction and performed well applied in the financial fields [4]. This study thus selects rough set classifier as a base of the proposed procedure. This study proposes a hybrid procedure for classifying financial status to maximal the hybrid advantage of models. The main objective of this study aims to test effects of discretization-based feature selection.
2. Related works
This section reviews the related issues of financial prediction for interested parties, including the topics covered in the financial distress and its applications, Financial Ratio, rough set theory, rule extraction method, and rule filter algorithm.
2.1 Financial Distress and its Applications
Financial healthy or unhealthy status of a corporate is an important indicator, not only representing the operating performance but also indicating its potential growth in the future. A bad financial status is unhealthy and is identified as a negative effect of possible hardness, such as distress, failure, crisis, decline, bankruptcy, difficulty, or alarm [5]. Thus, the term of financial distress for resulting in enormous losses on finances is specifically focused on future corporate prediction in this study. The financial distress is a crucial topic when it comes to the inability of a firm to meets its loan obligations.
Typically, financial ratios [5] provide a clear picture of the financial information of a company that is otherwise unclear from raw financial data in evaluation of the overall effectiveness of business management based on quarterly and/or annual sales and investment performance from the four common financial statements of a business, including balance sheet, income statement, cash flow statement, and statement of stockholders' equity. The financial statement analysis includes many of the financial ratios, but finding representative attribute financial ratios not only can raise prediction accuracy of financial status on a firm but also can make comparisons within the same industry.
2.3 Rough Set Theory
Rough sets (RS) find effective uses in “reducing data or eliminating superfluous data without any information loss, discovering data dependencies, estimating data significance, generating decision algorithms in given data, approximating data classification, discovering similarities or differences from data, discovering data patterns, and discovering cause–effect relationships.” RS has been successfully applied to real-world classification problems, having the advantages [6]: (1) RS data analysis results in information contained in a large number of cases reduced to a model that includes a generalized description of that knowledge, (2) the model is a set of easily understandable decision rules, which normally require no interpretation, (3) each decision rule is supported by a set of real examples, and (4) no additional information is required. The concept of RS can be defined using lower and upper approximations [7].
2.4 Rule Extraction Method
In knowledge discovery of data mining field, the generation of decision rules from a given data is an important concern for interested parties. Rule induction algorithms were first implemented in a Learning from Examples based on Rough Sets (LERS) system [8]. Particularly, the Learning from Examples Module, version 2 (abbreviated as LEM2) algorithm is a part of the LERS system and is one of most popular rule extraction method used in generation of decision rules.
2.5 Rule Filter Algorithm
Typically, some rule-filtering algorithms are applied to reduce the number of distinct decision rule of set and to refine the quality of rules since numerously redundant or ‘poor quality’ decision rules are generated, and these decision rules limit the classification capabilities of the rule set [9]. The rule-filtering solution determines the strength of the rule based on support, consistency, and coverage to calculate the quality indices of rules from a generated rule set.
3. The Proposed Procedure
The computational algorithm of the proposed procedure is divided into eight procedures systematically, as follows: (1) Selecting the essential attributes, (2) collecting the data of pre-selected attributes, (3) transforming format, (4) discretizating data, (5) selecting feature, (6) generating rule, (7) filtering rule, and (8) evaluating performance.
8
4. Empirical Case Study Using Financial Data
Experimental Data Set. Examples of a financial data set are extracted from a real database as
empirical case study, which is the Source Information Retrieval Systems (OSIRS) database [10] used to verify the classification performance of the proposed model. The real-world finance database includes balance sheets, income statements, statistical analysis, and ownership and subsidiary information in research and analysis. To implementing the computational processes of the proposed procedure by using the examples in practice is discussed step-by-step in detail, as follows:
(1) Selecting Attributes. A labeled OSIRS data set, extracted from the OSIRS database, is
exampled in this study. The OSIRS data set includes 23 essential (condition and decision) attributes that are initially selected based on the related literature review and expert recommendations from numerous features. The 23 attributes are termed X1-X23, all having continuous data. The first 22 attributes are the conditional-attributes, and the last one (X23) is the decisional-attribute. The decisional-attribute is named ‘Status’ and coded into two categories: ‘T’ (normal) and ‘F’ (distress). For example, the X1 attribute refers to ‘Return on equity ratio’, and the X11 is ‘Fixed assets turnover.’
(2) Collecting Data. Collect the data of 400 listed stocks (including normal and distress companies)
in the pre-selected 23 attributes in Taiwan into the OSIRS data set for the periods from 2002/01/01 to 2006/12/31 to achieve the verification purpose of the proposed procedure.
(3) Transforming Format. The incorrect values on attributes are eliminated from the OSIRS data
set; as a result, the 1,366 instances in EXCEL format are characterized by the 23 attributes.
(4) Discretizing Data. After the Global discretization, the data discretization information of the 22
condition-attributes is exposed. The 15 attributes, X1-X4, X6-X9, X12-14, and X19-X22, have discretized; that is, the remaining seven attributes are not. For example, the attribute ‘Return on equity ratio’ (X1) has one threshold (0.7719), corresponding to two linguistic terms: A1 ([−∞,0.7719)) and A2 ([0.7719,∞]), which are reacted to the natural language—low and high, respectively.
(5) Selecting Feature. The data discretization results are used as a tool of feature selection. The 15
conditional-attributes, including X1-X4, X6-X9, X12-14, and X19-X22, and one decision attribute X23 are selected as key attributes, and the others are eliminated from the OSIRS data set.
(6) Generating Rule. In the rule learning stage, the 1,366 instances of the selected 15
conditional-attributes and the one decisional-attribute are randomly split into two sub-datasets: the 66% training set and the remaining 34% testing set. As a result, the original classification accuracy 84.9% is achieved, and a set of 238 rules is created. The top 1 rule is formatted as ‘IF (X1="(-0.12725,∞)") & (X7="(-0.07955,∞)") & (X8="(-∞,-0.0113)") & (X18="(-0.43065,∞) & (X20="(-0.06745,∞)") & (X22="(-0.1275,∞)") Then Status=T’ in the 15 selected attribute samples by the LEM2 algorithm, and its support (or called match) has 110.
(7) Filtering Rule. The rule filtering algorithm is used to decrease the number of decision rules to
support threshold of 1 is removed, which indicates the rule supported by only one real-world example in the experiment data set. After the rule-filtering algorithm, the polished classification accuracy still has 84.9%, but 67 rules are erased. Importantly, only the remaining 171 rules are built as knowledge-based prediction system for assisting interested parties, such as stock investors.
(8) Evaluating Performance. For further comparisons and verification, the proposed procedure is
compared to various AI methods, including traditional RS [7], decision-trees C4.5 [11], multilayer perceptron (MLP) [12], and logistic-regression (LR) [13] for experiments repeated 10 times with the same 66%/34% random sampling. The average accuracy with standard deviation on the 10 tests is accordingly calculated. Table 1 shows the performance comparison on the five AI methods. Clearly, the proposed procedure having 83.7% with standard deviation 2.62% surpasses the other four methods in terms of accuracy in the OSIRS data set; simultaneously, the proposed procedure also has fewer used attributes and number of rules than those of the listed methods from Table 1.
Table 1. Comparison results of different five AI methods in the OSIRS data set
Model Traditional RS C4.5 ANN-MLP LR The proposed procedure
Accuracy 81.2% 76.9% 80.2% 79.1% 83.7%
Used attribute 23 23 23 23 16
Number of rules 209.9 165.0 - - 162.2
5. Conclusions
Exploring activity using financial ratios on the financial status of a company is an interesting issue and attracts priority-concerns. In this era of global economic uncertainty and high-risk, investors are building intelligent forecasting models for classifying the financial status problems. This study has proposed a hybrid corporate distress prediction model based on consolidating professional experience, Global discretization, discretization-based feature selection, the LEM2 algorithm, rough set classifier, and filtering rule algorithm for classifying the financial status of a company. Particularly, the performance of discretization-based feature selection method is verified in this study. Conclusively, the empirical results of a case study indicate that the proposed procedure is a great potential alternative of helpfully integrated hybrid model to demonstrate its technological merit and application value. In terms of managerial implications, it will be particularly and really valuable that this study further explored suitable application facets in practice and the analytical results applied by both practitioner and academicians fully. The main contribution of this study was conducted to reflect a new trial in the financial field for stock investors and obtained a satisfied result. This study still needs a further improvement on methodology and extensive applications, including (1) various types of decision attributes other than financial distress are used to validate the performance of the proposed procedure, and (2) extra condition-attributes are considered into the proposed procedure.
10
Acknowledgements
The authors would like to thank the Ministry of Science and Technology of the Republic of China, Taiwan, for financially supporting this research under Contract No. 102-2410-H-146-003.
References
[1] S. Canbas, A. Cabuk and B.S. Kilic: Prediction of commercial bank failure via multivariate statistical analysis of financial structures: The Turkish case. European Journal of Operational Research Vol. 166 (2005), p. 528-546.
[2] A. Ravi, H. Kurniawan, P.N.K. Thai and P. Ravi Kumar: Soft computing system for bank performance prediction. Applied Soft Computing Vol. 8 (2008), p. 305-315.
[3] H.M. Chung and K. Tam: A comparative analysis of inductive learning algorithms. International Journal of Intelligent Systems in Accounting, Finance and Management Vol. 2 (1993), p. 3-19. [4] F.E.H. Tay and L. Shen: Economic and financial prediction using rough sets model. European
Journal of Operational Research Vol. 141 (2002), p. 641-659.
[5] P. Ravi Kumar and V. Ravi: Bankruptcy prediction in banks and firms via statistical and intelligent techniques–A review. European Journal of Operational Research Vol. 180 (2007), p. 1-28.
[6] R. Slowinski and C. Zopounidis: Application of the rough set approach to evaluation of bankruptcy risk. Intelligent Systems in Accounting, Finance and Management Vol. 4 (1995), p. 27-41.
[7] Z. Pawlak: Rough sets: Theoretical aspects of reasoning about data, Kluwer Academic Publishers (1991).
[8] J.W. Grzymala-Busse: LERS—A system for learning from examples based on rough sets. Handbook of Applications and Advances of the Rough Sets Theory (1992), p. 3-18.
[9] Y.S. Chen and C.H. Cheng: Hybrid models based on rough set classifiers for setting credit rating decision rules in the global banking industry. Knowledge-Based Systems Vol. 39 (2013), p. 224-239.
[10] The Open Source Information Retrieval Systems (OSIRS) database. Retrieved on April 20, 2014 from http://newmops.tse.com.tw/.
[11] J.R. Quinlan: C4.5: Programs for machine learning. CA: Morgan Kaufmann, San Mateo (1993). [12] S.Haykin: Neural networks, a comprehensive foundation (2nd edition). Section 2.13, p. 84-89;
Section 4.13, p. 208-213. Prentice-Hall, Upper Saddle River New Jersey (1999).
科技部補助計畫衍生研發成果推廣資料表
日期:2014/10/26科技部補助計畫
計畫名稱: 財務危機預測模型之建構: 以統計方法與智慧型技術結合為例 計畫主持人: 陳祐祥 計畫編號: 102-2410-H-146-003- 學門領域: 作業研究／數量方法無研發成果推廣資料
102 年度專題研究計畫研究成果彙整表
計畫主持人：陳祐祥 計畫編號：102-2410-H-146-003- 計畫名稱：財務危機預測模型之建構: 以統計方法與智慧型技術結合為例 量化 成果項目 實際已達成 數（被接受 或已發表） 預期總達成 數(含實際已 達成數) 本計畫實 際貢獻百 分比 單位 備 註 （ 質 化 說 明：如 數 個 計 畫 共 同 成 果、成 果 列 為 該 期 刊 之 封 面 故 事 ... 等） 期刊論文 0 0 100% 研究報告/技術報告 0 0 100% 研討會論文 0 0 100% 篇 論文著作 專書 0 0 100% 申請中件數 0 0 100% 專利 已獲得件數 0 0 100% 件 件數 0 0 100% 件 技術移轉 權利金 0 0 100% 千元 碩士生 0 0 100% 2 位大專生兼任助 理 博士生 0 0 100% 博士後研究員 0 0 100% 國內 參與計畫人力 （本國籍） 專任助理 0 0 100% 人次 期刊論文 1 2 100% 另 1 篇論文已投 稿 至 SSCI 等 級 國際期刊 研究報告/技術報告 0 0 100% 研討會論文 1 1 100% 篇 論文著作 專書 0 0 100% 章/本 申請中件數 0 0 100% 專利 已獲得件數 0 0 100% 件 件數 0 0 100% 件 技術移轉 權利金 0 0 100% 千元 碩士生 0 0 100% 博士生 0 0 100% 博士後研究員 0 0 100% 國外 參與計畫人力 （外國籍） 專任助理 0 0 100% 人次其他成果
