以總體經濟因素預測股票報酬率—類神經網路與多元迴歸之比較研究

Abstract

證券的風險可分為系統風險與特異風險，其中系統風險主要來自於市場因 素，而特異風險主要來自產業因素及公司因素。若採用投資組合的方式可 減低非系統風險，使總風險能減低至系統風險。本研究以總體經濟因素探 討投資組合報酬率，並依不同的投資組合數目進行投資組合報酬率的預測 。資料蒐集樣本期間自民國七十五年迄至民國八十二年，主要資料取自台 灣經濟新報社之總體經濟資料庫。所使用的研究方法有一般的多元迴歸模 式以及類神經網路模式。主要目的為引用類神經網路於證券總體面的分析 ，有別過去使用技術指標預測股價的方式，並擴大類神經網路所能應用的 範圍，同時比較兩模式在此問題上的預測能力。實證結果發現：一、投資 組合的報酬率預測誤差會隨著投資組合數目的增加而減少。顯示總體經濟 因素較適合投資組合數目大者。可能影響單一證券的因素有很多，而且經 由投資組合可以消去個別證券的風險。二、在預測能力方面，在樣本內的 情形下，類神經網路比多元迴歸模式為優。在樣本外的情形下，兩者在投 資組合數目稀少時，類神經網路的預測能力比多元迴歸模式為差；隨著投 資組合數目的增加，其預測績效逐漸優於多元迴歸模式，顯示類神經網路 可能較能掌握整體趨勢。故類神經網路比較適合應用於投資組合報酬率的 預測分析。 The analysis of a security investment risk basically contains both systemic risk and idiosyncratic risk; the former is caused by market factors and the latter by industry and company factors. As a well known rule that a feasible stock portfolio investment always can minimize idiosyncratic risk and make total risk as low as systemic risk. Thus, this article tried to compare the predictabilities of the multiple regression model and the neural network model, based on the same macro-economic factors. In order to accomplish this purpose, rate of return from different sets were collected. Samples were collected from the dataset of Taiwan Economics Newspaper in the period of 1986 to 1993 and analyzed by using multiple regression model and neural network model. This study not only compared the predictabilities of the two models, but also extended the application of neural network model, which, in contrast to technical indicators, was applied to predict security price by using macro-analysis of security. Three major findings of this study were as follows: 1. The prediction error on rate of return of a stock portfolio decreases with the increment of the number of stocks in a portfolio. This indicates that using macro-economic factors to analyze stock portfolio investment is only good in the large numbers of stock in a portfolio. 2. Although the rate of return from any specific security can be affected by many factors, its risk can be reduced by using portfolio. Finally, it was concluded that neural network technique is a better tool for prediction on the rate of return of a portfolio, which may be due to the fact that neural network model can handle the whole factors parallelly or simultaneously.