證券投資智慧代理人系統-以WebServices為基礎架構

(1)

行政院國家科學委員會專題研究計畫成果報告

證券投資智慧代理人系統-以 Web Services 為基礎架構

研究成果報告(精簡版)

計畫類別：個別型計畫編號： NSC 96-2416-H-151-005- 執行期間： 96 年 08 月 01 日至 97 年 07 月 31 日執行單位：國立高雄應用科技大學金融系計畫主持人：林萍珍共同主持人：柯博昌計畫參與人員：碩士班研究生-兼任助理人員：施志樹碩士班研究生-兼任助理人員：簡偉倫碩士班研究生-兼任助理人員：侯燕伶報告附件：出席國際會議研究心得報告及發表論文處理方式：本計畫涉及專利或其他智慧財產權，2 年後可公開查詢

中華民國 98 年 01 月 20 日

(2)

行政院國家科學委員會補助專題研究計畫

; 成果報告

□期中進度報告

（計畫名稱）

證券投資智慧代理人系統-以 Web Services 為基礎架構

計畫類別：

;

個別型計畫 □ 整合型計畫

計畫編號：NSC 96-2416-H-151 -005

執行期間：

2007 年 8 月 01 日至 2008 年 7 月 31 日

計畫主持人：林萍珍

共同主持人：

計畫參與人員：

施志樹

、

簡偉倫

、

侯燕伶

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

本成果報告包括以下應繳交之附件：

□赴國外出差或研習心得報告一份

□赴大陸地區出差或研習心得報告一份

;

出席國際學術會議心得報告及發表之論文各一份

□國際合作研究計畫國外研究報告書一份

處理方式：除產學合作研究計畫、提升產業技術及人才培育研究計畫、

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

執行單位：國立高雄應用科技大學金融系

(3)

證券投資智慧代理人系統-以 Web Services 為基礎架構

摘要將金融電子市集建構在Web services 的基礎上提供協商系統，以利證券市場買賣雙方都可以進行的協商，本研究基於這樣的研究架構將各種證券投資的議題設計成智慧代理人程式，利用agent 的智慧學習，協助投資協商的進行。亦即應用類神經網路、遺傳演算法、模糊理論設計股票評價、投資組合資金配置、選擇權避險以及違約機率預測等智慧代理人，搜尋最佳的投資機會提供投資人做為決策的參考，創造更高的投資報酬率。本研究內容共分為三個部份：(1)以 Web services 之行動代理人推播技術為基礎，結合資料探勘關聯規則應用於股市投資決策。(2)應用遺傳演算法建構「演化式股票評價」新模型，以最佳化股票價值的合理區間與交易區間。同時結合三次方程內插法，建置最適的非線性資金配置策略，以提高投資效益。(3) 本研究將運用遺傳演算法(Genetic Algorithm,

GA)來挑選投資組合，同時以極值理論(Extreme Value Theory, EVT)評估挑選投資組合的風險值(VaR)。研究貢獻：(1)結合 Web services 與行動代理人推播技術及 Apriori 演算法改善股市投資決策問題；縮短投資期間內，獲得最具價值的決策訊息，加強券商的顧客關係管理，提高其競爭優勢。(2)利用智慧型股票評估系統內含交易成本的交易時機點，具有明顯高獲利能力，表示演化式股票評價本身會依股票特性，自我發展出適合該股票的合理價格與買賣交易時機點進反而提高獲利能力。關鍵字：投資決策、資產配置、遺傳演算法、Web services、極值理論

1.緒論

投資環境與投資工具全球化，使金融市場的波動性和不穩定性日益劇烈，更增添監控及風險管理上的困難，如果不能有效的控管風險將會造成巨大的損失。唯有即時適切獲得投資資訊，才是致富避險關鍵。智慧型代理人是近十年來的新興資訊技術與概念。在這競爭激烈、資訊傳遞快速的金融市場，需要一個具有自我學習、主動反應、溝通協調的智慧型代理人金融資訊系統才能勝任此一任務。近年來全球資訊網路(World Wide Web)及電子商務與應用的快速發展，對金融業造成巨大的機會與衝擊，如何應用新的技術保持競爭優勢是金融業很重要的課題。

本研究目的有三，(1)以 Web services 為基礎依 Apriori 演算法萃取出有價值的證券投資規則，另以行動代理人的推播技術(Vasiu & Mahmoud 2004)，主動即時將投資規則建議推播到無線通訊設備(手機)上，投資人主動選擇訂閱服務頻道客製化證券投資決策系統。(2)結合演化式計算(evolutionary computation)中的遺傳演算法與數值方法的三次方程內插法(cubic spline interpolation)以建構一創新的股票評價與資金分配模型。(3)運用遺傳演算法，來產生投資組合，利用極值理論計算投資組合，對其進行風險值的評估，以取得投資組合的厚尾現象，並再進一步計算加權平均，以作為遺傳演算法的演算依據，使其能產生最佳的投資組合。

(4)

2.文獻探討

發展智慧代理人網路服務架構的證券投資決策系統，應從財務問題到應用技術分別做理論推導與文獻探討。財務問題方面：將討論股票投資策略、股票評價、風險值。在

應用技術方面：探討Web Services、Mobile agent 以及遺傳演算法。以下將針對這些問題

與技術的文獻做一個詳述。投資策略：台灣證券市場成立於民國 49 年，至 69 年才蓬勃發展。市值從 69 年的 219,053 (百萬元)成長到 95 年的 19,376,975；上市家數由 102 家成長到 688 家；大盤指數也由 549.55 成長到 7,823.72。從市值規模來看，成長相當驚人，可見台灣證券市場的運作相當活絡。但就波動度而言，即使是近十年大盤指數成長率的標準差仍有 0.25，此凸顯出要在證券市場獲利，需要承擔不小的風險。金融市場操作極具複雜性，需要依賴電腦運算技術與財務金融決策結合(陳安斌 2005)。技術分析依歷史資料的價與量以圖形分析其走勢，有助於投資人做選擇時的投資策略參考(梁定澎 2003)。股市投資涉及較為複雜而且變化快速的決策問題，股市中存在一些市場變數共變的規則(Hung et al., 1995)，而且資料過於龐大與複雜，分析需要運算快速的演算法協助找出投資策略(Fernando et al., 2000)。但是，仍要靠投資人做最後的彙整依個人投資經驗做投資策略決策的判斷(Tsaih et al., 1998)。

股票評價：股票評價模型企圖從淨值折現(net present value) (Francis et al., 2000;Frankel & Lee, 1998;Bebchuk, 2000)、資產價值(asset appraisal)(Block, 1995;Fama & French, 1995) 、剩餘價值 (residual income)(Bebchuk,2000) 與市場乘數 (price multiples) (Ohlson,1995、2005)等方向評估出企業合理的價值。即使股票評價的研究議題近十年被熱烈的討論，學術界與實務界仍存在有一些爭議。主要原因是每一種評估模型均有其特定的假設前提，實證的結果常出現不一致結論；而缺乏良好的建構技術與堅實的理論基礎也是原因之一；另外不同的模型評估出來的合理價格不同，或許證券的合理價值可能是一個區間而非單一價格。股票評價挑選出投資標的物後，接著是決定資金分配比例。常見的資金分配是依經驗法則、均等分配或是線性分配。這些方法可能會因為人為的錯估或分配方法不當而稀釋報酬或增加投資的風險。風險值：風險值評估方法－包含歷史模擬法(Historical Simulation, HS)、指數加權移動平均法(Exponentially Weighted Moving Average, EWMA)和一般化自我迴歸條件異質變異模型(GARCH)。EWMA 模型較為簡化，導致推估風險值時常發生偏差，並且需假設其

時間序列具有i.i.d(Identical Independent Distribution)的特性；HS 需要大量樣本資料才能

重現資產價格的變化狀況，如果筆數不足則不易求得較佳的結果；GARCH 目前被廣泛應用於評估風險值，然而其估計方法在使用Berndt et al.(1974)提出的 BHHH 最大概似法估計時，其評估係數需先設定初始值，如果初始值不能有效設定，則其估計結果不易找到廣域較佳解(global optima)。這些傳統計量模型的風險值評估方法各具有其優缺點，所估計出的風險值的準確度頗有差異，顯示風險值估計問題存在著不確定性問題，若能結合模糊理論的技術，或許能有效捕抓到更精確的風險值。

(5)

服務的需求者(service requester)、以及介於兩者間的服務中介者(service broker)。透過此架構可達到系統的整合與資源分享的目的，進而創造企業最佳的資訊價值。大型企業在推行資訊系統時較小型公司面臨更大的挑戰，金控公司也是如此。Rabhi et al.(2004)以個

案分析的方式深入企業內部探討資本市場的引用 Web services 的發展概況，包含交易交

換服務、交易資料處理服務以及即時交易監督服務等，對於日後學者要研究的參考。Web

services 結合網格運算(grid computing)提供創新衍生性金融商品的服務，即以網格為基礎的美式選擇權仲介交易為例，提供低成本、高投資報酬與內部報酬率的系統提供金融機構使用(Lan et al. )。應用 Web services 信用卡詐欺偵測(Chiu et al., 2004)，參與的銀行可以分享這些可有發生詐欺的案例於分散式環境以強化其偵測效能與降低財務風險。印度於 1991 年有 25 家銀行 500 個分行聯合採用以結構化財務訊息解決方案在安全的機制下交換訊息，其中介軟體(middleware)於 2004 年改以 Web services 的技術得到相當好的績效 (Radha et al., 2004)。Web services 在金融上的應用最被關心的議題是其安全性，在資本市場上電子金融(e-Finance)的安全問題尤其重要，Yang & Ray(2005)兩位學者提出一個資訊科技的應用安全性問題應從最低階的技術問題到最高階的安全概念包含內控問題均是成功的關鍵。建構一個電子政府（e-government）系統最複雜的部份是金融服務系統。因為金融服務系統需要建立認證安全系統。在建構時要連結金融機構（如國稅局、銀行等）個別的資料交換系統平台。Ｗang et al.(2004)提出以 Multi-tier 與 Web services 建構一個安全、可信任的以及安全的財務服務架構平台，並說明如何藉由此系統的流程引擎（workflow engine）有效的規劃工作、互動式使用者介面以及與第三團體(third party)財務系統整併連結。遺傳演算法：自從達爾文的天擇演化(natural selection)理論提出後，許多學域都深受影響，而遺傳演算法就是Holland 於 1975 年受其啟發而發展的演算法則。遺傳演算法是一種搜尋技術[2]，常被用來解決離散型問題的最佳化或搜尋問題。遺傳演算法屬於演進式的演算法，源自於生物學中的一些行為，例如：遺傳、挑選、交配及突變。遺傳演算法通常是利用電腦模擬的方式實作，而傳統的遺傳演算法是以二元碼來進行編碼並以隨機的方式初始化。之後每個世代都會進行適應值的評價，再從中挑選出適應值較高的個體進行交配或突變並進入下一個世代。在不斷演進之後，所有的個體會逐漸趨於一致，從而得到一個滿意解。由於遺傳演算法流程架構簡單實作容易，加上已有許多發展健全的套裝軟體，因此，遺傳演算法成為近年來十分熱門的人工智慧技術。遺傳演算法採用了自然界中生物與生物之間競爭求生存的觀念，以一組特別的字串模擬各種生物的染色體(chromosome)，根據染色體來計算對環境的適應度(fitness)，在每個世代之間讓各個染色體以隨機的方式進行交配(crossover)與突變(mutation)來產生下一代，而大環境會再根據該染色體的適應度選擇(selection)是否讓其生存，產生「適者生存」的效果，使適應度較佳的染色體能夠有較多的機會將其基因遺傳至子代。這個演化交替的動作會一直持續到達成最終目標（例如事先決定的演化代數）為止。

3.研究方法

本研究內容分三個部份：Web services 之行動代理人結合資料探勘股市投資策略關聯

(6)

規則；非線性演化式股票評價與資金配置模型；投資組合風險值評估使用極值理論模型。以下將分此三個部份說明研究架構。

3.1 Web services 之行動代理人結合資料探勘股市投資策略關聯規則

行動代理人結合資料探勘股市投資策略關聯規則決策系統(簡稱 MAPT_SIDS)架構包括使用者(Mobile investors)、行動代理人層(Mobile agent layer)、商業應用層(Business application layer)、資源層(Resource layer)見圖 1。使用者(如股市投資人)使用手機為使用者用通訊工具與系統溝通，獲取投資股市資訊。行動代理人，有通訊管理人、代理人閘道以及行動服務頻道;在商業應用層有服務管理員及資料探勘搜尋引擎；資源管理層有使用者訂閱資料庫、股價歷史資料庫及類股關聯規則資料庫。圖1 MAPT_SIDS 系統架構圖 3.2 非線性演化式股票評價與資金配置模型 本研究的系統架構如圖2。第一步，由遺傳演算法產生初始化的染色體。第二步，各染色體內容做區性重整，以得知合理股價區間、股價高估區間與股價低估區，同時，為方便下一步三次方程內插法模擬非線性投資交易策略，在股價高估區間與股價低估區分別最適化N 個取樣點，共 2×N 個取樣點。第三步，將解碼後染色體中的 2×N 個取樣點帶入三次方程內插法，求算出非線性區間交易曲線。第四步，依據此一演化後的非線性區間交易曲線模型，套入股價日資料進行交易求算投資報酬，報酬率高的即視為高適應函數值。第五步，依據適應函數值選擇優良的染色體，放進交配池(mating pool)之中。接著進行第六與第七步，交配池之中的染色體進行交配與突變產生新子代。如此循環演化，直到達中止條件為止。

(7)

遺傳演算法最佳化區間股價與資金配置三次方程內插法連接各股價點之資金配置資金配置區間指標模型三次方程內插法產生初始母體合理股價染色體解碼染色體交配計算適應函數選擇優良基因到交配池基因突變產生新子代演化終止？演化完成是否股價高估股價低估 N個取樣點 N個取樣點圖2 股票評價與資金配置架構圖 3.3 投資組合風險值評估使用極值理論模型

本研究也使用牛頓法(Newton Method)及 MATLAB，來計算非線性 GPD(Generalized Paredo Distribution, GPD)函數之機率密度函數。本研究流程以 GA 為基礎架構，提出的六個流程如圖 3。(1)由 GA 的內部運作機制，即：選擇、交配、突變，產生一組新的投資組合。抓取投資組合中各家特定時段之報酬率。(2)將所得之資料以報酬率由小到大排序，再取其前n%之報酬率，並令此 n%區段中的最大值為門檻值，以取得觀察樣本 X 數列。 (3)將觀察樣本代入 GDP，再對所得的估計式取對數概似函數。透過最大概似估計法求取其估計參數β與ξ的估計式，再透過牛頓法(Newton Method)計算參數β與ξ的最佳近似值。(4)將最佳近似值β與ξ代入超越門檻值法之計算式，估計 VaR 值。(5)再以所得之 VaR 值，對所選取的投資組合進行加權平均值。之後再對最後計算出之加權平均值取倒數後，交由遺傳演化法記錄與比較。(6)當遺傳演算法結束後，其產生的投資組合即為最佳投資組合。

(8)

圖3 GA 為基礎的極值理論之投資組合風險值系統架構

4. 實證結果與分析

實證結果將分三個部份說明：Web services 之行動代理人結合資料探勘股市投資策略關聯規則；非線性演化式股票評價與資金配置模型；投資組合風險值評估使用極值理論模型。 4.1 Web services 之行動代理人結合資料探勘股市投資策略關聯規則 本實驗收集9 家企業股價資料，見表 1 中 A 到 I 欄。觀察日期為 2005 年 4 月 25 日到2005 年 5 月 6 日。表中發現 D(晶元光電)上漲時，G(佰鴻工業)同時上漲的機率為 75%。反之，分析其下跌趨的關聯性，實驗也發現當D 股票下跌 4 次當中，G 股票也隨著一起下跌，兩檔股票的關聯性頗高。表1 事後股價漲跌資料 Date\ID A B C D E F G H I 0425 0.85 -1.3 0.1 -1.3 -0.22 -3 -1.8 -0.55 -0.2 0426 0.3 0.15 0.6 1.35 0.1 -1.5 0.6 0.3 0 0427 -0.8 -0.75 -0.7 -0.85 -0.12 -0.5 -1.6 -0.4 -1.4 0428 0 -0.1 0.7 0 0.07 3 0.7 0.1 -2 0429 -2.2 -1.25 -1.3 -2.5 -0.15 -2.5 -0.1 0.05 -0.6 0503 0.6 -0.95 0 0.6 0.13 3 -0.2 0.45 4 0504 -0.6 -2.1 0 -0.1 -0.13 -2.5 -0.4 -0.5 -1.1 0505 2.75 -1.1 1.3 3.15 0.2 3 2.35 1.7 1.1 0506 1.85 0.65 1.9 2.95 0.03 1 1.55 0.25 2.5 4.2 非線性演化式股票評價與資金配置模型 當大盤處於空頭走勢時，演化式股票評價模型的投資績效優於大盤，原因在於空頭市場時本模型的交易策略是股價愈低，持有股票的資金配置將會非線性遞增，分散持有成本；當大盤處於盤整走勢時，股票落入演化式股票評價模型的合理股票區，不會進行交易，直到股價低於低估區時才會進行買入動作，高於高估區時方會進行賣出的動作，有利於減少交易次數與成本並提高報酬績效。當走勢處理一般多頭時本模型的績效仍優於大盤，只有在大多頭時大盤的買入持有策略才能有稍好的績效。 4.3 投資組合風險值評估使用極值理論模型 利用歷史模擬法估算出的風險值，其成功率均能達到信賴水準，但在 99%的信賴水準下，可投資家數較95%的信賴水準多，且個別股票的成功率也較 95%來得高，顯示出

(9)

發現，兩種信賴水準的成功率，皆並未達到應有的信賴水準，但均己相當接近，且可看出兩者所預估的投資股票數，差距不大，但99%的信賴水準依舊較優於 95%，表示雖然 EWMA 是以資料時間的遠近，來決定對未來估算風險值的影響程度，但在 99%的信賴水準下，依舊會估算較高的風險值，來減少可能發生的損失。極值理論的結果發現，95% 及 99%的信賴水準，其成功率均達到信賴水準，且近乎百分之百成功，而比較兩者的成功率，95%的成功率較高，表示極端值選取的範圍較大，其估計的數值較多，提高了準確度。歷史模擬法及指數加權移動平均法，兩者估算出的風險值，均是以信賴水準來取出特定的數值，再加以計算，並未考慮到此一數值以下的範圍，而超越門檻值法是屬極值理論取值方法的一種，因估算的數值較多，所以較為嚴謹，能夠估計出更為精確的風險值。

5. 結論

Web services 為基礎的行動代理人之股市投資決策系統，可讓使用者利用探勘後的關聯規則，分析出各類股上漲或下跌的關聯性，建立關聯規則資料庫。Web services 行動代理人之推播技術，主動即時推播合乎使用者投資需求的投資策略到投資人的無線裝置上。投資人可在最短時間內獲得對投資獲利有幫助的資訊。本研究第二部份，以遺傳演算法＂適者生存＂的特性，結合極值理論模型，來預測最佳的風險值，當每組投資組合產生後，找出對應的股票的報酬率，計算出過去一段時間的風險值後，與實際的報酬相比較，判斷其成功與否，並以成功率作為投資組合的Fitness 值的依據，讓遺傳演算法決定該如何產生下一代，此演算法的好處，是能在執行後，產生最佳的投資組合，節省找尋最佳投資組合的時間。本研究第三部份，利用智慧型股票評估系統內含交易成本的交易時機點，具有明顯高獲利能力，表示演化式股票評價本身會依股票特性，自我發展出適合該股票的合理價格與買賣交易時機點能提高獲利能力。另外，當大盤處於空頭時，演化式股票評價模型具有分散持股成本；盤整走勢時，股價需落入低估區與高估區才有交易行動，此將會減少交易次數與降低交易成本。因此，空頭與盤整走勢最能表現本模型的獲利能力；當走勢處理於大多頭時大盤的買入持有策略才能有稍好的績效。整體而言，演化式股票評價的投資績效優於大盤。感謝國科會給予經費補助(NSC 96-2416-H-151 -005)，使得本計劃得以將研究成果表示於國際期刊，已發表於到國際刊的文章條列如下：

1. Wen-Hsiung Wu, Po-Chang Ko, Ping-Chen Lin & Ming-Hua Su(2008). Applying Mobile Agent to a Mobile Stock Intermediary Services System Development. International

Journal of Smart Home, 2(2), 1-12.

2. 林萍珍，柯博昌，田育任，「非線性演化式股票評價與資金配置模型」，管理與系統，第十七卷，第四期，2010 年 10 月。(TSSCI).

3. Ping-Chen Lin & Po-Chang Ko(2010). Portfolio Value-at-Risk Forecasting with GA-based Extreme Value Theory. Expert Systems with Applications, 40(3), (Accepted) (SCI)

參考文獻

[1] 陳安斌，「新金融實驗教學之-財務金融資訊系統與投資管理」，新陸出版社，台北，

(10)

[2] 梁定彭，「決策支援系統與企業智慧」，智勝出版社，台北，2002。

[3] Bebchuk, L. A. .Using option to divide value in corporate bankruptcy. NBER Working

Paper 2000, No. 7614.

[4] Berndt, E., Hall, B., Hall, R., & Hausman, J. (1974). Estimation and inference in nonlinear structural models. Annals of Economic and Social Measurement, 3, 53-665. [5] Block, F. E. (1995). A study of price to book relationship. Financial Analysts Journal, 51,

63-73.

[6] Chiu C. C. & Tsai C. Y. (2004). A Web services-based collaborative scheme for credit card fraud detection. In Proceedings of IEEE International Conference on

e-Technology, e-Commerce and e-Service, 28-31, 177-181.

[7] Fama, E. F. & French,K. R.(1995). Size and book-to-market factors in earnings and returns. Journal of Finance, 50, 131-155.

[8] Fernando, F.R., Christian, G.M., & Simon, S.R. On the Profitability of Technical Trading Rules Based on Artificial Neural Networks: Evidence from the Madrid Stock Market.

Economics Letters, 200.

[9] Francis, J. Olsson, P., & Oswald, D. R. (2000). Comparing the accuracy and explainable of dividend, free cash flow, and abnormal earnings equity value estimates. Journal of

Accounting Research, 38, 45-70.

[10] Frankel, R. & Lee, C. M., (1998). Accounting valuation, market expectation, and cross-sectional stock returns. Journal of Accounting and Economics, 25, 283-319. [11] Hung, S.Y., Liang, T.P., & Liu, V.W. (1995). Integration Arbitrage Pricing Theory and

Artificial Neural Networks to Support Portfolio Management. Proceedings of ISDSS, Hong Kong.

[12] J. Holland (1975). Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control and Artificial Intelligence. University of

Michigan Press.

[13] Lan R., Yang F. & Zheng S. Financial service innovation based on grid computing. In

Proceedings of IEEE International Conference on Services Systems and Services Management, 2, 1081-1084.

[14] Ohlson, J. (1995). Earning, book value and dividends in equity valuation.

Contemporary Accounting Research, 18(1), 661-687.

[15] Ohlson, J. (2005). On accounting based valuation formulae. Review of Accounting

Studies, 10, 323-345.

[16] Rabhi F. A., Dabous F. T. & Yu H.(2004). In Proceedings of IEEE International

Conference on e-Technology, e-Commerce and e-Service, 28-31, 38-41.

[17] Radha V. Gulati V. P. & Thapar R.(2004). Evolution of Web services approach in SFMS – a case study. In Proceedings of IEEE International Conference on Web

(11)

[19] Vasiu,L. & Mahmoud, Q. H.(2004). Mobile Agents in Wireless Devices. IEEE

Computer,37(2), 104-105.

[20] Wang Z. Y., Qi D. Y. & Qiu S. S.(2004). Design of finance service platform of e-government architecture in one-station service infrastructure. In Proceedings of the

Third International Conference on Machine and Cybernetics, Shanghai, 26-29,

3029-3034.

[21] Yang W. & Ray P.K.(2005). Evaluation methodology for the security of e-finance systems. In Proceedings of IEEE International Conference on e-Technology, e-Commerce and e-Service, 29, 120- 123.

(12)

計畫成果自評

本計劃請就研究內容與原計畫相符程度、達成預期目標情況、研究成果之學術或應用價值、是否適合在學術期刊發表作一綜合評估。 1. 研究內容與原計畫相符程度 原計劃各年度與執行的研究內容相符程度說明如下: 1. 原第一年主題：智慧代理人在Web Service架構下的選擇權避險與違約機率預測。本年度計劃執行重點先以(1)Web services為架構發展一個投資人主動選擇訂閱服務頻道客製化證券投資決策系統[1]。使用到的技術為Apriori演算法萃取出有價值的證券投資規則，另以行動代理人的推播技術，主動即時將投資規則建議推播到無線通訊設備(手機)上。其次，發展(2)演化式多重組合羅吉斯迴歸模型-應用於信用評等[2]。提出演化式多重組合羅吉斯迴歸模型，每一個等級有個別的羅吉斯迴歸模型，可依使用者需求設定不同等級的評等模型，並且模型的門檻值與預測變數是藉由遺傳演算法以非線性方式做最佳化，以此研究模型建立一套演化式多重組合羅吉斯迴歸信用評等系統。（因為報告內容篇幅的限制，此研究成果僅在此簡要陳述，論文內文如附件）。 2. 原第二年主題：證券評價之模糊區間模型：以遺傳演算法為基礎。研究重點為非線性演化式股票評價與資金配置模型[3]，應用遺傳演算法建構「演化式股票評價」新模型，以最佳化股票價值的合理區間與交易區間。同時結合三次方程內插法，建置最適的非線性資金配置策略，以提高投資效益。 3. 原第三年主題：最適資源配置學習為基礎之新類神經網路模型。本年度依計劃進行，研究重點在於資源配置類神經網路配置模型應用於投資組合最佳化[4]，主要目的是改善現有的類神經網路模型應用於投資組合資產配置決策上，無法最佳化輸出層神經元做為個別資產資金配置比例以及其比例總合為100%。本研究提出配置型類神經網路模型，在無任何假設及限制條件下，求解投資者面臨不確定因素影響及有限自有資金情況下之投資組合個別資產資金配置比例，並且保證配置比例總合為100%。並且考量個別資產預期報酬與風險、個別資產間報酬相互影響關係及投資者本身的風險趨避程度(risk averter)，（因為報告內容篇幅的限制，此研究成果僅在此簡要陳述，論文內文如附件）。另外，本年度也發展演化式極值理論評估投資組合風險值[5]，此研究運用遺傳演算法挑選投資組合，同時以極值理論(Extreme Value Theory, EVT)評估挑選投資組合的風險值(VaR)。

2. 達成預期目標情況

預計申請三年計劃通過1 年，除原三個研究主題已達成預期目標，另外增加兩個，共分

成5 個研究子題進行研究也完成，超過預期目標。

(13)

(1) Web services為架構發展一個投資人主動選擇訂閱服務頻道客製化證券投資決策系統：整合網路技術與資料探勘應用在金融投資決策議題上，結合金融與資訊議提供投資人最佳即時的投資決策資訊。學術與應用價值有1利用行動代理人推播投資決策；2 關聯規則應用於股價關聯分析；3加強證券商之顧客關係管理。

(2) 演化式多重組合羅吉斯迴歸模型-應用於信用評等：以GA及MCLR (Combinatorial Logistic Regression Model) 模型整合來建構出的演化式多重組合羅吉斯迴歸信用評等系統。學術與應用價值有1提出多重組合羅吉斯迴歸，改善單一羅吉斯迴歸模型在同一條S曲線上決定門檻值的限制，不論依均等法或經驗法則在同一條曲線上決定門檻值均有不對稱性的問題；2不同的評等等級應該有各自的門檻值，亦即每一個等級可以擁有各自的羅吉斯迴歸與其所屬的門檻值；3結合GA最佳化的能力，協助LR挑選變數以及決定門檻值；4採用Basel II驗證方式，除預測效力外，進一步加入模型穩定性與同質性的験證，強化評等的效能。 (3) 非線性演化式股票評價與資金配置模型：建構一個演化式股票評價模型結合非線性的資金配交易策略，利用遺傳演算法最佳化合理的股價區間與交易區間，並結合三次方程內插法以非線性方式規劃最適交易區間的資金配置。經由多項實驗以驗證本研究模型的績效與適合性。學術與應用價值有1合理的股價區間：過去的傳統股票評價文獻的假設前題與實證結果存在一些爭議與不一致現象。相同的評估模型所計算的評估值在不同的證券市場可能產生不同的研究結果，此意謂著股票的合理價格可能是區間而非單點；2模型沒有假設前題：一般評價方法專注於模型萃取，多數評價模型有使用上的假設前題(例如：資料要符合常態分配)，本研究採用計算智慧的非線性特性改善計量模型對資料分配的假設限制；3整合交易策略與資金配置：多數文獻專注於股價挑選或買賣策略，甚少結合買賣策略與資金配置，易造成買賣時機出現，卻不知該如何配置投資資金的情形。本研究發展的演化式股票評價模型，當股價進入超跌或超漲的股價區間時，會建議採用買入或賣出交配策略，同時會算出適合的資金配置部位，提高投資的便利性與有效性；非線性資金配置：相關研究限於考量當股價偏離合理區間時，以線性方式做適當的資金配置，欠缺彈性。股票市場具有資料量龐大、資訊超載及訊息快速變化等特性。 (4) 資源配置類神經網路配置模型應用於投資組合最佳化：傳統類神經網路理論很少提到資源配置模型，因此無法處理投資組合中個別資產資金比例問題，且使輸出層神經元輸出個別資產資金配置比例總合為100%，故本研究以此為研究方向，藉由數學模型的推導，利用輸出層神經元預測輸出值與神經元目標輸出值間之差異量求解個別權重值修正模型之學習率修正權重值，訓練出能達成本研究目的之配置型類神經網路，應用於投資組合中個別資產資金配置權重最佳化問題。學術與應用價值有1提出以配置型類神經網路模型，創新傳統類神經網路模型可最佳化輸出結果總合為1；2本研究提出的配置型類神經網路模型於投資組合資金配置，對於股票市場的擇時交易策略的決策機制與證券投資具有創新性的意義；3投資組合是很多財務管理議題的基礎，例如企業公司理財、基金管理公司、外資法人等機構重要的投資決策問題。因此，建立一套有效、穩健的投資組合資金配置最佳化機制為輔助，方能制定最佳的投資決策。 (5) 演化式極值理論評估投資組合風險值：本研究將運用遺傳演算法(Genetic Algorithm) 來挑選投資組合，同時以極值理論(EVT)評估挑選投資組合的風險值(VaR)。財務資料

(14)

通常不是常態分配，大部份具有厚尾現象，極值理論使用極限分配可有效捕捉財務資料的厚尾現象，較能專注於資料部的變化，精確估算風險值。極值理論主要是針對尾部的極端值(或稱稀少事件)，進行風險的評估，極值理論的優點是，原始資料並不假設使用何種分配，只是針對資料經過適當的配置後，其尾部極端事件的處理，而本研究是以極值理論中，兩大類的極端事件選取法的其中之一－越超門檻值法，選取一定範圍的尾部極端事件，因此不同的門檻值，會影響風險值的評估。 4. 是否適合在學術期刊發表 研究成果之學術期刊發表依上述5 個研究子題說明如下，將註明已發表、已接受或審稿中：

[1] Wen-Hsiung Wu, Po-Chang Ko, Ping-Chen Lin, Ming-Hua Su, “Applying Mobile Agent to a Mobile Stock Intermediary Services System Development”, International Journal of Smart Home, Vol. 2, Iss. 2, pp.1-12, April 2008(已發表)。

[2] 柯博昌、林萍珍、游俊忠，演化式多重組合羅吉斯迴歸模型-應用於信用評等，資訊管理學報(審稿中) (TSSCI)。

[3] 林萍珍, 柯博昌, 田育任, 「非線性演化式股票評價與資金配置模型」, 管理與系統，第十七卷，第四期，2010 年 10 月(已接受) (TSSCI)。

[4] Po-Chang Ko, Ping-Chen Lin, “Resource Allocation Neural Network in Portfolio Selection”, Expert Systems With Applications, Vol. 35, Iss. 1-2, pp.330-337, July-August 2008(已發表) (SCI)。

[5] Ping-Chen Lin and Po-Chang Ko, “Portfolio Value-at-Risk Forecasting with GA-based Extreme Value Theory”, Expert Systems with Applications, Vol. 40, Iss. 3, 2010(已接受) (SCI)。

(15)

International Journal of Smart Home Vol. 2, No. 2, April, 2008

Applying Mobile Agent to a Mobile Stock Intermediary Services

System Development

Wen-Hsiung Wu1, Po-Chang Ko1, *Ping-Chen Lin2, Ming-Hua Su1 1 Dept. of Information Management, 2 Institute of Finance and Information

National Kaohsiung University of Applied Sciences {whwu,cobol,lety}@cc.kuas.edu.tw

Abstract

Due to the radical changing of the global economy, a more precise stock valuation helps providing important judgment principles to decision-makers and investors. With the advent of the third-generation (3G) or future forth-generation (4G) Internet, the mobile commerce (M-Commerce) will become increasingly important. In addition, the mobile stock investment decision support system attracts great interests for professionals, such as stockholders, bondholders, financial analysts, governmental officials, and even the general public, recently. This study introduces a Mobile Agent-based Stock Intermediary Services System (MASISS) framework based on the mobile agent perspective to provide ubiquitous and seamless transaction activities for financial institutions. It also helps customers to make a more precise decision in the current intense commercial competition environment. For building distributed enterprise systems, The MASISS framework is developed in an integration of J2ME and J2EE environment with cross-platform portability, a huge server-side and client-side deployment base, and coverage for most W3C standards..

Keywords: Mobile agent, Mobile stock intermediary services, Stock investment decision, Wireless handheld devices.

1 Introduction

Investment decision making plays an increasing important role in the knowledge economy age, because the global economy has radically changed. It helps many professionals, such as stockholders, bondholders, financial analysts, governmental officials, and even the general public, to make a more precise decision in the current intense commercial competition environment. However, in the face of a rapidly fluctuating market and a vast amount of information to be digested, the information asymmetry makes it hard for investors to make precise and real-time decisions. Now that the Internet has become ubiquitous, most securities companies and investment organizations provide their own e-stock intermediary service systems for the benefit of investors. These systems allow investors to obtain real-time information, conveniently place orders over the Internet and thereby provide economic benefit. Even the widespread use of wireless handheld devices (including mobile phones and PDAs) has gradually moved mobile stock investment into the limelight in recent years. The use of mobile services not subject to restrictions of time and space is better able to meet investors' needs. As a consequence, as far as securities companies and investors are concerned, the use of mobile stock intermediary services to provide stock information and implement transactions has become a necessity.

*Corresponding author: Tel.: +886-7-3814526 ext.7508. E-mail address: [email protected] (Ping-Chen Lin).

(16)

From the point of view of the market, some companies – such as E-Ten Information System Company – are trying to provide mobile stock intermediary services. Nevertheless, its main functions provided by these services are identical with those of e-stock intermediary service systems, which include market trends, individual stock trends, category rankings, and technical analysis. While these functions offer the advantages of time information, real-time order placement, and mobility, they make no attempt to provide advanced individualized services via agent-based intermediary services. As far as securities companies are concerned, the establishment of a mobile agent-based intermediary service to serve investors can not only disperse transaction volume and reduce bandwidth use, but also provide the advantages of cross-platform compatibility, autonomy, and interdependence (Picco, 2001; Manvi and Venkatardm, 2004).

The characteristics of agents have recently started to attract great attention in financial fields. For instance, Chen and Liao (2005) used an agent-based stock market model to analyze the relationship between stock volume and price. Nevertheless, fewer researches have applied the features of mobile agents to the mobile stock investment decision-making environment. Furthermore, although most existing commercial mobile stock intermediary services provide at least prototype functions, they still fail to actively provide individualized services. For example, an investor may only be concerned about the associations between the rising prices of stocks in a certain category. Mobile stock intermediary services must therefore take investors' specific needs into account.

Based on the above considerations, this study proposes a framework for a "mobile agent-based stock intermediary services system" (MASISS). This study also develops MASISS in a JEME and J2EE environment with cross-platform portability to build a huge server-side and client-side deployment based enterprise system. We also assess MASISS by using the example of association between stocks in a single category. MASISS not only can provide a reference for securities companies developing mobile stock service functions but also provide individualized assistance to investors making mobile stock investment decisions.

2 Intelligent Mobile Agents with Mobile Commerce Applications

Agents are autonomous objects created for dynamic and distributed applications that are responsible for executing designated tasks (Woodridge and Jennings, 1995). Use of agents and knowledge can improve transactions between information suppliers and customers, but place a heavy workload on the intermediary and thus limit the number of participants that can be served by the intermediary. In contrast, mobile agents can solve the heavy workload problem, and can be transported on different systems after being executed asynchronously and autonomously, carrying with them their program code, current state of execution, and any data obtained. Furthermore, mobile agents can communicate with one another (Chess et al., 1995; Picco, 2001).

The main characteristics of mobile agents include: (1) Remote distributed processing: The distributed processing program to be executed is sent to other computers for execution. The program code is compressed before transmission, communications volume is minimized, and the burden on computer and network is distributed. (2) Asynchronous processing: When the transmitting server sends the program to a receiving computer, the two ends do not need to stay in continuous communication; mobile agents can be executed independently on a

(17)

receiving computer, which sharply reduces network traffic. (3) Simplification of distributed program design: As long as a run time system is installed on every computer, all computers will possess a mobile and autonomous processing ability. As a result, mobile agents possess the advantages of disconnected operation, less network traffic, roaming ability in a heterogeneous environment, support for electronic commerce, ease of development, personalization and high flexibility, and real time application (Lange and Oshima, 1999; Picco, 2001; Cao et al., 2004; Manvi and Venkatardm, 2004).

Because mobile agents possess the foregoing characteristics and advantages, they are being used by researchers in many fields for an increasingly wide range of applications. Mobile agents were formerly used mainly for network management, monitoring, and scheduling, including in heterogeneous network management and integration systems (Du et al., 2003) and dynamic scheduling systems (Hiroyuki et al., 1997). In e-commerce environments, mobile agents have been applied to trading, brokerage, auction, and e-marketplace functions, including the developments of MAGENT, BROKERAGE, and Nomad systems (Dasgupta et al., 1999; Jung and Jo, 2000; Sandholm and Huai, 2000; Du et al., 2005). We have found that a small number of researchers have applied mobile agents in the applications of financial decision-making. Specifically, Chen and Liao (2005) proposed an agent-based stock market model, and investigated linkage between stock returns and trading volume.

Mobile commerce has become an important trend in recent years. Quah and Lim (2002) suggested that we are carrying out research into the use of mobile agents in wireless handheld devices because they will overcome wireless network limitations such as small bandwidth and intermittent disconnection problems. As mentioned in the previous section, agents possess autonomy and several other advantages. We have found that the mobile agent concept has been applied to web search, tourist guide, and auction applications in relevant studies, and is developed as the Search Sweep, Gulliver's Genie, and MoRAAS systems (Zerfiridis and Karatza, 2004; O'Grady et al., 2005; Shih et al., 2005). This shows that mobile agents can be readily applied to mobile commerce issues.

In summary, we have found that few studies have applied the features of mobile agents to mobile stock investment decision-making. Therefore, we try to develop a mobile agent-based intermediary service system for both securities companies and mobile investors in this study.

3 Framework of MASISS

This study proposes a "mobile agent-based stock intermediary services system" (MASISS) based on the concept and features of mobile agents, and taking the mobile stock investment decision-making environment into consideration. The MASISS framework consists of mobile agent layer, business application layer, and resource layer shown in Figure 1. The mobile agent layer includes communication manager, agent gateway, and mobile server channel. The business application layer consists of service manager and a data mining engine. The resource layer includes a user subscriber database, a stock price historical database, and a category stock association rule database. The following sections explain various layers of MASISS.

(18)

3.1 Mobile agent layer

This agent layer, supported by Tahiti, of IBM’s Tokyo laboratory, provides an application interface to model complicated agent behavior. It can create, clone, retract, dispose, and dispatch mobile agent. The agent layer has two major classes, AgletProxy and AgletContext. The AgletProxy is a proxy server that sends the client requests to the remote server and brings back the results after completing the assignment. The communication between agents can only pass through the designated functions. Hence the Agletproxy is like a protective umbrella that provides the transparency of objects. The AgletContext provides an implementation environment for the Tahiti platform. When a mobile investor’s request is sent to a remote Tahiti server, it is decomposed into several message streams by its own AgletContext of the local Tahiti server and then sent to the AgletProxy of the remote Tahiti server. The message streams are then recomposed by the AgletContext of the remote Tahiti server. Figure 2 shows the detailed process of agent activities.

In addition, the mobile agent layer is further divided into three modules: communication manager, agent gateway, and mobile server channel. These modules are described as below.

Communication manager: The communication manager is a software agent situated between wireless equipment and a wired network. The communication manager manages all system connections and communication with agent gateways. The system will automatically generate a user profile data record whenever a user uses a MIDlet download service channel via the communication manager, and will use Java Servlets and a Common Gateway Interface (CGI) to record the identity of the user profile and communicate the agent gateway's attributes and address.

Agent gateway: The agent gateway provides an interface with the communication manager and establishing and updating user profiles. The agent gateway saves data to the user profiles database. In order to maintain security, the user is not permitted to directly access the user profiles database. The agent gateway can also implement the mobile agent platform, and this platform can be implemented effectively in a heterogeneous server environment – which is a key feature of mobile agents.

Figure 1. Mobile Agent-Based Stock Intermediary Service System (MASISS) Framework

(19)

Figure 2. The Detailed Process of Agent Activities

Mobile server channels: Mobile server channels satisfy users' multifaceted investment needs, such as analysis of arbitrary stocks, market analysis, intelligent stock selection, and category linkage analysis. A mobile server channel relies on the agent gateway to perform identification and accept user needs.

3.2 Business application layer

The business application layer is situated between the mobile agent layer and resource layer. It consists of service manager to provide services to investors, and data mining engine to find optimal association rule.

Service manager: The service manager uses J2EE architecture to connect mobile agent layer, resource layer's subscriber database, and business application layer's data mining engine (Fig. 3). The J2EE architecture mainly consists of Web Container, EJB (Enterprise JavaBeans) Container, and JDBC (Java Database Connectivity). The Web Container includes JSP, Servlet, and JavaBean, to interact with the mobile agent and EJB Container. EJB container includes eSession Bean, Entity Bean, and Message Driven Bean. As a consequence, the service manager can push association rules obtained by the data mining engine to an investor's wireless handheld device at the client layer via the mobile agent in the mobile agent platform.

Data mining engine: The data mining engine mainly provides association rules function and application. The focus of the MASISS uses Apriori algorithm (Agrawal and Srikant, 1994) shown below to find optimal association rules.

1. Determine the minimum support threshold.

2. Delete those candidate itemsets appearing less often than an itemset with minimum support threshold, and generation of frequent itemsets.

3. Join the mutual products in accordance with the generated set to create larger candidate itemsets. Loop Step 2 until no candidate itemset can be generated.

4. Find the itemset with the best confidence value among high frequent itemsets, and derive the optimal association rules.

Figure 4 shows an example. Assume the minimum support threshold is 2 and the trading database (TDB) contains four items of trade data, as well as five "candidate itemsets" {A}, {B}, {C}, {D}, and {E}. The degree of support (C1) of these five itemsets is obtained. The four "frequent itemsets" (L1) - {A}, {B}, {C} and {E} are obtained after deleting {D} with less than the minimum support threshold. Join the mutual products of L1 to generate "candidate itemsets" (C2) and yield their degree of support. Pruning all itemsets among C2 with less than the minimum degree of support will leave four "frequent itemsets" (L2) -

ContextA _Dispatch() ContextB

Clone() Class Retract() Aglet Class Create() Aglet Dispose() Secondary Storage

(20)

{A,C}, {B,C}, {B,E}, and {C,E}. Finally, all frequent itemsets (L3) are obtained by repeating the foregoing steps.

Figure 3. Multi-layer Business Application Architecture

Figure 4. Application Data Mining Engine Employing Apriori Algorithm Procedures

3. 3 Resource layer

The resource layer mainly manages the resources stored in databases to support various applications and interactions in business application layer. It includes subscriber database, stock price historical database and application rule database.

Subscriber database

The subscriber database accesses user subscription channel data and user profile records. The business application layer's service manager transmits association rules derived from the records in this database to the subscribing user.

Stock price historical database

Stock price historical database accesses various raw stock market data. This raw data is the basic data source used for association rules analysis by the business application layer's data mining engine.

Mobile Agent Layer Business Application Layer

JSP Servlet JavaBean

Web Container EJB Container

eSession Bean Entity Bean Message Driven Bean Subscriber Database JDBC JDBC driver JAXP J2EE Connector Mobile Agent Platform

Service Manager Resource Layer

Data Mining Engine

Business Application Layer

Service Manager

Apply Apriori Algorithm for Data Mining

Resource Layer Association Rule Database Stock Price Historical Database Minimum Support Threshold = 2

(21)

Association rule database

The association rule database takes the subscriber database and stock price historical database as its data sources. This database compiles an association rule data warehouse from the enterprise's heterogeneous databases distributed at various locations. It then integrates, unifies, and summarizes information from various heterogeneous data sources, and generates data marts or data cubes. The database further uses Online Analytical Processing (OLAP) to perform effective online query and analysis. Finally, the business application layer's data mining engine performs analysis using the Apriori algorithm, and returns the resulting patterns or relationships to the association rule database or provides them to the business application layer's service manager for query purposes (Fig. 5).

Figure 5. Association Rule Database Employing a Data Warehouse Architecture

4. System Evaluation of MASISS

The MASISS assessment process consisted of: (1) Obtaining and organizing the raw stock price data, (2) Using Apriori algorithm to find association rules, (3) Receiving information on investor's mobile phone, (4) Testing of feasibility of association rules. The following is a detailed description of these steps.

4.1 Obtaining and organizing the raw stock price data

Raw stock price data was obtained from "Intelligence Winner 2000" system of Infotimes Company in Taiwan. The selected stock category was "mobile phone parts and components concept stocks", which contained nine stocks 2393 - Everlight Electronics Co., Ltd. (A), 2402 - Ichia Technologies Inc. Technology (B), 2439 - Merry Electronics (C), 2448 - Epistar Corp (D), 2457 - Phihong (E), 3007 - Greenpoint (F), 3031 - Bright LED (G), 6168 - Harvatech (H), and 6285 - Wistron (I). The selected time period was from March 22, 2005 to April 22, 2005; data frequency was set as one day, and there was a total of 23 data sets and 22 stock price change data sets (Table 1). Then, find the stocks that rose for each day (Table 2). Any rising items in Table 2 (column I) that were blank or had only one transaction record were deleted because we only considered the condition of "concurrently rising" records.

(22)

Table 1. Summarized Stock Price Changes Table 2. Data Summary

ID A B C D E F G H I ID List of Item_IDs 0323 1.6 0.35 0.9 3 -0.15 3 2.25 2.3 0 0323 ABCDFGH 0324 0.6 0.3 -0.4 -0.25 0.3 8 -0.25 0.1 -2.3 0324 ABEFH 0325 -0.6 1 2 -0.05 0.2 2.5 0.4 0 -3.3 0325 BCEFG 0328 -0.15 0.9 5.6 0.9 0.15 0 1.9 0.1 -0.4 0328 BCDEGH 0329 -1.65 -2.1 -1.9 -0.8 -0.55 -6 -1 -0.8 -2.5 0329 0330 -0.1 0.15 -1.2 -0.2 0.1 0.5 -0.65 0 -2.4 0330 BEF 0331 1 0.05 -5.8 1.1 0.1 2 2.45 1.5 4.7 0331 ABDEFGHI 0401 0.55 -0.7 -2.8 0.2 -0.15 -0.5 0.6 -0.05 1.1 0401 ADGI 0404 -0.45 -0.5 0.1 -0.4 0.05 -1 -2.2 -1 0 0404 CE 0406 0.6 -0.35 1 0.7 -0.15 0 0.6 2.45 -1.2 0406 ACDGH 0407 1.55 -2.05 -1.5 2.5 -0.05 -4 0.4 -0.45 -0.5 0407 ADG 0408 0.55 -0.9 -2 0 0.1 2.5 0.9 -0.4 2.3 0408 AEFGI 0411 0.1 0.65 -0.5 0.1 0 -3.5 1.2 0 -1.8 0411 ABDG 0412 -0.3 -0.1 1.2 -0.5 0 -0.5 0 -0.45 0.1 0412 CI 0413 1.05 -0.3 0.9 2.4 -0.15 2 0.45 0.6 0.4 0413 ACDFGHI 0414 -0.05 0.05 -1.1 -0.1 -0.1 2 0.15 -0.3 -0.7 0414 BFG 0415 0.25 0.05 -0.7 0.1 -0.35 0 0.3 -1 -0.3 0415 ABDG 0418 -1.65 -1.6 -1.8 -1.4 -0.65 -7 -2.05 -2.5 -5 0418 0419 1.1 0.4 1.5 1 0.25 -0.5 1.3 0.1 1.5 0419 ABCDEGHI 0420 -1.75 -2.45 -0.5 -3.5 -0.22 -7 -2.55 -2.35 -4 0420 0421 0.45 0.2 -0.3 0.4 -0.03 4 0.85 -0.75 3.3 0421 ABDFGI 0422 0.25 -0.2 0.7 0 -0.11 0.5 -0.65 -0.15 1.3 0422 ACFI

(23)

4.2 Using Apriori algorithm to find association rules

Use of the Apriori algorithm to perform data mining yielded the final itemsets listed in Fig. 6. Table 3 specifies all association rules on the basis of the minimum support and confidence threshold. As an example, the meanings of the optimal association rules in Table 3 – "D => G" and "A D => G" - are as follows:

Association rule "D => G": The probability of stock G simultaneously rising whenever stock D rises is 100% during the time period.

Association rules "A D => G": The probability of stock G simultaneously rising whenever stock A and stock D both rise is 100% during the time period.

Table 3. Support and Confidence of Association Rules

Association Rules

Support Confidence Association Rules Support Confidence A ⇒ D 52.6% 76.9% A⇒ D ∪ G 52.6% 96.9% D ⇒ A 52.6% 90.9% D ⇒ A ∪ G 52.6% 90.9% A ⇒ G 57.9% 84.6% G⇒ A ∪ D 52.6% 71.4% G ⇒ A 57.9% 78.6% A ∪ D ⇒ G 52.6% 100% D ⇒ G 57.9% 100% A ∪ G ⇒ D 52.6% 90.9% G ⇒ D 57.9% 78.6% D ∪ G ⇒ A 52.6% 90.9%

4.3 Receiving information on investor's mobile phone

Investors can obtain stock analysis services by subscribing to service channels, and can regularly obtain reports on analysis of stock category association rules. If an investor wishes to obtain information on some other stock category, he/she can enter the "stock category analysis" function (Fig. 7 (a)), input the stock category code, and select a data period (Fig. 7 (b)). When the service manager receives the request, it will first determine whether there is data in the association rule database. If there is no relevant data, the service manager will immediately use data mining engine to perform association rules analysis, and send the results of analysis back to the investor's mobile phone via a mobile agent (Fig. 7 (c)).

Figure 7. Operating Interface of Investor’s Mobile Phone

4.4 Testing of feasibility of association rules

This study used the association rule "D => G" to test feasibility. The time period was from April 25 to May 6. The price changes of these nine stocks are shown in Table 4. Four valid

(24)

data items that remain after data items for stock D's failure to rise were deleted. Three of these four data items comply with the results of our analysis. This result indicates that the probability of 3031 rising when 2448 rises is 75%. We can also see that stock G uniformly fell whenever stock D fell, which proves that there is a high degree of association between the prices of stock D and stock G.

Table 4. Subsequent Stock Price Change Data

ID A B C D E F G H I 0425 0.85 -1.3 0.1 -1.3 -0.22 -3 -1.8 -0.55 -0.2 0426 0.3 0.15 0.6 1.35 0.1 -1.5 0.6 0.3 0 0427 -0.8 -0.75 -0.7 -0.85 -0.12 -0.5 -1.6 -0.4 -1.4 0428 0 -0.1 0.7 0 0.07 3 0.7 0.1 -2 0429 -2.2 -1.25 -1.3 -2.5 -0.15 -2.5 -0.1 0.05 -0.6 0503 0.6 -0.95 0 0.6 0.13 3 -0.2 0.45 4 0504 -0.6 -2.1 0 -0.1 -0.13 -2.5 -0.4 -0.5 -1.1 0505 2.75 -1.1 1.3 3.15 0.2 3 2.35 1.7 1.1 0506 1.85 0.65 1.9 2.95 0.03 1 1.55 0.25 2.5 5. Conclusions

Mobile stock investment decision is already an inevitable trend in recent year. It provides important judgments to help investors to make investment decisions in the mobile commerce environments. However, seldom researches are focused on the intermediary services system for both stock investors and securities companies based on mobile agent perspective. This study proposes a Mobile Agent-Base Stock Intermediary Services System (MASISS) framework, which allows mobile agents to operate independently and asynchronously, and significantly reduces the bandwidth usage over the network. Hence, it provides the advantages of autonomy and independence to the stock investors and stock companies. This is particularly useful if the agents are implemented under a heterogeneous environment; the framework allows the minimum system integration effort to implement the agent-based system. Furthermore, this study shows the ease of aggregating investors’ stock order for decreasing transaction cost by using intermediary mobile agents. The framework allows mobile investors to possess the mobility and save a great deal of time and money, if comparing our framework with using mobile push personalized information matching based on the mobile users requirements via publish and subscribe channel to wireless handheld devices. The mining engine also effectively extract investment rule to provide decision makers the most reliable information. Finally, the Java-based mobile agent system can be implemented on heterogeneous platforms and interoperate with cross-platform applications via wireless handheld devices in practice.

References

[1] G. P. Picco, Mobile agents: an introduction, Microprocessors and Microsystems 25 (2001) 65-74. [2] S. S. Manvi, P. Venkataram, Applications of agent technology in comm.: a review, Comp. Comm.

27 (2004) 1493–1508.

[3] D. B. Lange, M. Oshima, Seven good reasons for mobile agents, Comm. of the ACM 42(3) (1999) 88–99.

(25)

[4] J. Cao, X. Wang, S. K. Das, A framework of using cooperating mobile agents to achieve load sharing in distributed web server groups, Future Generation Comp. Sys. 20 (2004) 591–603. [5] R. Agrawal, R. Srikant, Fast algorithms for mining association rules, Proc. of 20th International

Conf. on Very Large Data Bases, Santiago, Chile, (1994) 487-499.

[6] T. C. Du, E. Y. Li, E. Wei, Mobile agents for a brokering service in the electronic marketplace, Decision Support Systems 39 (2005) 371-383.

[7] S. H. Chen, C.C. Liao Agent-based computational modeling of the stock price-volume relation, Information Sciences 170 (2005) 75-100.

[8] M. J. Wooldrige, N. R. Jennings, Agent theories, architectures and languages: a survey, Lecture Notes in Computer Science 890 (1995) 1-39.

[9] D. H. Shih, S. Y. Huang, D. C. Yen , A new reverse auction agent system for m-commerce using mobile agents, Computer Standards & Interfaces 27 (2005) 383-395.

[10] T. Sandholm, Q. Huai, Nomad: mobile-agent system for an internet-based auction house, IEEE internet Computing (2000) 80-86.

[11] P. Dasgupta, N. Narasimhan, L. E. Moser. P. M. Melliar-Smith, MAgNET: mobile agents for networked electronic trading, IEEE Trans. on Knowledge and Data Engineering 11 (4) (1999) 509-525.

[12] J.J. Jung, G. S. Jo, Brokerage between buyer and seller agents using constraint satisfaction problem models, Decision Support Systems 28 (4) (2000) 293-304.

[13] K. G. Zerfridis, H. D. Karaza, Brute force web search for wireless devices using mobile agents, J. of Sys. and Software 69 (2004) 195-206.

[14] M. J. O’Grady, G. M. P. O’Hare, C. Sas, Mobile agents for mobile tourists: a user evaluation of Gulliver’s Genie, Interacting with Comp. 17 (2005) 343-366.

[15] T. C. Du, E. Y. Li, A. P. Chang, Mobile agents in distributed network management, Comm. of the ACM 46 (7) (2003) 127-132.

[16] D. Chess, C. Harrison, A. Mobile agents: are they a good idea? IBM Research Division, research report, (1995).

[17] J. T. S. Quah, G. L. Lim, Push selling-multicast messages to wireless devices based on the publish/subscribe model, Electronic Commerce Research and Applications 1 (2002) 235-246. [18] T. Hiroyuki, K. Koji, I. Yoshihide, Y. Kenji, A Takayoshi, WorkWeb system-multi-workflow

management with a multi-agent system, Proc. of ACM SIGGROUP Conf. on Supporting Group Work: the Integration Challenge, Phoenix, Arizona, USA, (1997) 299 – 308.

(26)

(27)

國際

演化式多重組合羅吉斯迴歸模型-應用於信用評等

摘要

間各國銀行都積極以新巴塞爾協定為主要指導原則進行信用風險管理改革，尋求信用風險控管的最佳策略。新巴塞爾協定建議銀行採用內部評等法自建授信系統以減少人為錯誤帶來的作業損失，並且能夠快速正確處理授信放款。銀行在建置內部評等系統時決定其優劣主要因素在於預測模型建構以及預測變數的篩選，羅吉斯迴歸是目前被廣泛應用於違約風險預測的模型。傳統羅吉斯迴歸僅適用於解決分成兩類的預測問題，雖然累積羅吉斯迴歸模型可以分成多類，但是累積羅吉斯的S曲線通常以等距或經驗法則切割門檻值做分等，當違約機率產生變動時會造成等級變動的不對稱現象。因此，本研究提出演化式多重組合羅吉斯迴歸模型，每一個等級有個別的羅吉斯迴歸模型，可依使用者需求設定不同等級的評等模型，並且模型的門檻值與預測變數是藉由遺傳演算法以非線性方式做最佳化，以此研究模型建立一套演化式多重組合羅吉斯迴歸信用評等系統。另外，其目標函數是以新巴塞爾協定建議的驗證方法對本研究模型進行違約預測力、評等穩定度以及等級同質性三方面的模型驗證。實驗結果發現，信用風險違約評等的預測效力以及等級同質性方面明顯優於台灣經濟新報的信用風險指標；多期違約時間點的實證中，反應出近期的財報與公司治理等相關資訊的揭露對模型具有較佳的預測效力；代表穩定性指標的移轉矩陣率會隨著使用者設定的評等級數增加而呈下降的常態現象；財務變數的獲利能力與每股收益以及公司治理的董監報酬、持股、財測與管理者異動等是影響評等模型預測效力的重要因素；關鍵字：演化式多重組合羅吉斯迴歸模型、遺傳演算法、新巴塞爾資本協定、 信用評等、違約機率。

(28)

Evolutionary Multiple Combinatorial Logistic Regression

Model Applied in Credit Rating

ABSTRACT

The serious financial issues, such as Asian Financial Crisis, Subprime Mortgage Crisis, occurred recently. The real economy may suffer from credit crunches as results of the financial crises are not self-evident and bank inadequate management. Due to one of major profit of bank is loan growth, especially in the enterprise loans, it is important to manage and evaluate corporate financial risk effectively. Basel II published in June 2004 was a well-known international initiative that requires banks to have a more risk sensitive framework. It establishes regulatory expectations for credit risk through the Internal Ratings Based (IRB) approach, which allows banks to assess the key risk drivers as the primary capital calculation. In statics, the logistic regression is only suitable for probabilistic binary classification, but it cannot provide multiple classifications. Although cumulative logic regression (CLR) introduces a multi-class algorithm, it is had to decide the thresholds in CLR. In this paper, we propose an evolutionary MCLR credit rating system (EMCRS), which uses evolutionary approach to optimize multiple combinatorial logistic regression models. We use GA to estimate non-stationary time-series data with dynamic nonlinear searching capabilities. Finally, the EMCRS would be verified by (1) capability of predicting default rate (e.g. KS, ROC, CAP), (2) rating stability (e.g. TM) and (3) grade homogeneity (e.g. CIER). The experimental results demonstrate that EMCRS has better capability to predict the enterprise default rate than TEJ. It is reasonable that the rating stability will decrease if the number of rating increases. The profitability, earning per share and management factors are critical to evaluate the performance of EMCRS.

Keywords: Multiple Combinatorial Logistic Regression Model, Genetic Algorithm, Basel II、Credit Rating、Default Probability。

(29)

壹、緒論

近年來國際間的金融重大事件頻傳，例如亞洲金融風暴與美國次貸風暴等，這些風暴形成的原因與銀行金融機構的授信業務風險管理有著重要的關連，銀行若資本計提(指依授信戶的信用評等等級給予不同的風險權數，進而計算其應計提的資本)不足，即可能發生重大損失。大部份銀行主要獲利來源為放款業務，而且多數都屬於企業放款，故企業違約風險的管理顯得格外重要。新巴塞爾協定(Basel II)所定義的信用風險是由貸方所產生的一種潛在損失風險；信用風險可經過客觀且系統化的方式轉換成信用等級，使得授信人員可以根據信用等級，快速且正確的制訂放款決策。Shin 與 Han(2001)指出銀行過去常仰賴授信人員的主觀評估信用評等，易造成人為判斷標準不一，使企業授信風險不能有效管理。 Basel II 協定提到銀行若要有效控管信用風險必須系統化成為符合 Basel 規範的風險管理系統。Basel II 推薦的信用風險衡量方法是內部評等法(Internal Rating Based Approach, IRB)。IRB 法允許銀行依據本身策略需求，建置較具自

主性和彈性的信用風險評等系統，頗獲各國金融機構推崇。在建置IRB 系統過

程中，銀行需考量違約機率(Probability of Default, PD)、違約損失率(Loss Given Default, LGD)、違約暴險額(Exposure at Default, EAD)以及有效到期期間 (Maturity, M)這些因素。其中，PD 為 IRB 的主要核心，故銀行更為重視對 PD 的評估預測。決定信用風險評等系統績效的關鍵因素在於建構評等模型以及選擇預測變數，羅吉斯迴歸(Logistic Regression, LR)是目前被普遍應用在建置違約預測系統的統計模型。LR 特性允許估計值落於正、負無限大之間，再透過 Logit 函數轉換成值域 0 到 1 的機率值，藉此門檻值協助判斷違約事件發生的程度。LR 模型在應用上放寬對變數的假設限制，允許使用質化或量化變數型態；自變數不須符合常態分配或任何機率分配假設。傳統統計模型使用線性組合方式來挑選變數，例如逐步分析、因素分析等。在變數的挑選會以次序加入或剔除的方式來組合，依特定順序選擇變數的方式不易找到最佳預測變數組合。除了變數選擇是建構模型的重點外，模型門檻值的決定也是預測分類模型的重點。門檻值是將風險程度相同的授信戶做歸類的判斷依據，門檻值的決定也將會影響模型預測分類的效力。本研究目的之一是藉由遺傳演算法挑選出在企業財務體質、公司自治以及總體環境等變數中最具預測效力的因素，提昇預測企業的信用風險等級的準確性。各家銀行的文化和特性不同，授信策略也會有所差異，故我們加入依銀行授信評等等級的設定來建立一套準確、有效的信用風險評等管理資訊系統，亦

證券投資智慧代理人系統-以WebServices為基礎架構

行政院國家科學委員會專題研究計畫 成果報告