BPN暨RN神經網路與向量誤差修正模型對國內債券價格之預測績效

• Main Authors: 紀如龍, Jih, Ru-Long
• Language: 中文
• Published: 國立政治大學
• Subjects: 公債; 殖利率預測; 神經網路; RN模型; BPN模型; 向量誤差修正模型; Government bond; Yield to maturity; Neural network; RN; BPN; VECM
• Online Access: http://thesis.lib.nccu.edu.tw/cgi-bin/cdrfb3/gsweb.cgi?o=dstdcdr&i=sid=%22B2002002748%22.

本研究計畫探討以RN神經網路模型預測國內債券價格的效度。目前一般用於財務預測的神經網路論著主要為BPN模型，惟BPN模型有其限制，所以本研究計畫將（1）分析比較統計計量模型，BPN神經網路，RN神經網路系統對國內公債價格之預測績效。（2）分析不同時期的預測能力，找出景氣和預測變數的關係，同時將比較各個時期統計計量模型和神經網路模型是否同時有效， 抑或有些有效， 有些無效，以探討各工具是否具有互補性或替代性。並探討預測績效是否受到背後經濟環境的影響。 我們研究對象為國內公債，其每日交易資料取樣時間自民國八十一年開始。影響債券價格的因素可拆解成實質利率，預期通貨膨脹率和風險貼水三層面，本研究總體變數之選取，亦循此三項範疇以求周延。 本研究之研究成果對理論及實務應用將有下列三項預期貢獻：（1）比較不同其常的債券在不同景氣狀況下，各不同預測模型的預測效度差異，探討各時期各工具之預測能力，可提供投資實務界對預測工具之選擇，應用與搭配。（2）對債券報酬率預測研究，分析總體變數，利率風險等變數對債券報酬率的影響，可進一步暸解影響債券價格的相關因素及程度。（3）以往神經網路應用在財務預測領域上， 皆以BPN 神經網路為主，此處引進RN神經網路，比較兩者的表現，可提供學術理論界之驗證。 === This research project empirically investigates the accuracy of Reasoning Neural Networks (RN) in forecasting Taiwan's bond prices. We explore (1) the relative predictive abilities of Vector Error Correction Model (VECM), which serve as a representative econometric model, Back Propagation Neural Networks (BPN), which is adopted by most current studies in the application of neural networks in finance, and RN, and (2) th3 potential variations in the three models' predictive power in different phases of economic cycle. Specifically, we aim to study if the three models substitute or complementone another. In addition, we explore the extent to which the relativepredictive abilities of the three models varies with underlying macroecomonic factors. The explanatory variables adopted in this study include all potential drives to (real) risk-free rate, expected inflation rate, and riskspremiums. In this study, we examine the government bond terms to maturity,coupon rate, and prices of government bonds during 1992-1995. This project would contribute to both academic and application researchin the following three aspects : (1) Few, if any , prior study explores whether and how various neuralnetworks and/or eco- nomic models perform under different macro-economicvariables. Our empirical results may indicate an appropriate model ( ormodels ) to improve forecasting of bond prices. (2) This study shows how RN, BPN, and VECM models perform in forecastinggovernment bonds yields to maturity. (3) The BPN model prevails in financial forecasting. Nevertheless, BPNis subject to a few short comings and may thus be a sub-optimal model. This study analyzes if RN is more cost-effective in forecasting bond prices than BPN.