在資訊科技不斷進步的環境下,各種投資的金融商品亦不斷推陳出新,而債券型基金更已普遍成為保守投資者重要的理財工具之一。雖然債券型基金為一較保守的投資工具,但是仍然存在著一定的風險程度,對於極保守的投資人產生了一份不安感,為了滿足這類投資者的需求,信評式債券基金孕育而生。 信評式債券基金由於限制較多,相對風險低報酬亦有限,然而要如何在低風險低報酬的商品中挑選及預測出表現較好的基金,透過通用迴歸類神經網路(General Regression Neural Network ,GRNN)的運用,建構一預測績效的模式。 GRNN是從機率類神經網路(Probability Neural Network,PNN)所演變而來,主要應用在預測及控制上,可用來建立連續變數之函數關係,無論迴歸問題為線性或非線性均可用GRNN來解決。 本論文採用通用迴歸類神經網路(GRNN)的模式,針對國內已通過信用評等的債券型基金共15檔,進行實證研究,利用各檔基金民國91年1月至民國92年12月共24個月的投資組合比率為研究資料,找出最佳的平滑參數,並利用此學習模型對各檔基金績效進行預測之應用。 為驗證GRNN在此類研究標的預測成效,研究中並採用多元線性迴歸分析同時進行預測,結果顯示GRNN的均方誤差低於多元線性迴歸,故依本研究結果而言GRNN在樣本數及變數不多的情況下,對於債券型基金的績效預測亦能有不錯的表現。
Under the environment with constantly advanced information technology, all kinds of investment finance articles enormously vary, among which bond foundation has become one of the critical investment tools for conservative investors. Although bond foundation is a conservative investment tool, it still has a certain degree of risk, which makes very conservative investors uneasy. For the purpose of satisfying such investors’ needs, the bond foundation for credit appraisal(信評式債券基金) has been created. The bond foundation for credit appraisal(信評式債券基金)relatively has lower risk but limited rewards due to its more restrictions. However, through the application of General Regression Neural Network (GRNN), a forecast effect mode can be constructed to ensure that foundations with better effects can be selected and forecasted among lower risks and less rewards of articles. GRNN, which is developed from Probability Neural Network(PNN), is primarily applied to forecast and control, and used to set up a function relation with successive variables. Either linear or non-linear of regression issues can be settled by using GRNN. The mode of general regression neural network is adopted in this essay. The total of 15 bond foundations has been testified and analyzed. The investment combination rates of each foundation by 24 months from January 2002 to December 2003 are used as research data to find the best smooth parameter and to forecast application on each foundation effect by using this learning mode. In order to testify forecast effects of GRNN on this research, multiple regression analysis has been forecasted at the same time. The result reveals that the even equation difference of GRNN is lower than multiple regression analysis. This research shows that under less samples and variables, GRNN still has an excellent performance of the forecast of bond foundation effects.