本研究利用民國81年至83年間，臺灣證券交易所八項分類指數及36家樣本公司的日資料，測試以OLS估計市場模式（market model）是否符合簡單線性迴歸的基本假設，並提出一套估計市場模式之程序。經測試結果，發現在三項基本假設中（隨機干擾項常態性、同質及無序列自我相關），以OSL估計實證的模型，在隨機干擾項常態性及同質之基本假設上，有普遍違反的情況。特別是時間序列同質的假設方面，絕大部份的迴歸方程式均不符合隨機干擾項同質的假設，顯示有關自我迴歸條件異質（ARCH）的存在非常普遍。此時由於OLS估計出之迴歸係數□、□除了仍是不偏及一致估計量外，其他的統計特性，包括BLUE、有效性、近似有效性，均未具備。□、□之估計變異數，一般而言是偏誤的，而且傳統的檢定及信賴區間亦是無效的。本研究的發現，對於CAPM的相關研究，例如事件研究（event studies）、市場異常現象（market anomalies）及市場效率性的研究，均具深遠涵意。日後應用到相關模式時，若以OLS估計係數，宜先行測試基本假設，而不宜逕行假設所建構之模式均已符合，以免造成推論之錯誤。此外，本研究發現規模與異質（Breusch-Pagan檢定下）可能有關，較大的規模較有異質的趨勢。最後，本研究以水泥窯製類指數為例（同時違反隨機干擾項常態性、同質及無自我相關的假設），用已發展之估計方法逐步修正ARCH異質、序列自我相關、非常態性及異質，結果發現經修正後之殘差皆符合各項基本假設。這一套估計市場模式（或其他線性模式）的程序，應是一項嚴謹的學術研究，所不能忽視的！

This paper employs daily trading data during 1992 to 1994, of 8 industry indexes and of 36 sample stocks traded on Taiwan Stock Exchange, to investigate if the market model (estimated by OLS method), which is popularly used in the financial area, follows or violates the basic assumptions of normal linear regression model. This paper also provides a estimation procedure for market model under such conditions that basic assumptions are violated. The results shows that among the three basic assumptions (normality, homoskedasticity and nnonautocorrelation of the stochastic disturbance), most of the empirical models violate the basic assumptions of normality and homoskedasticity, especially the assumption of homoskedasticity against autoregressive conditional heteroskedasticity. It should be noted that under heteroskedasticity, the least squares estimates are not efficient and the estimates of the variances are also biased, thus invalidating the tests of significance. Our findings provide strong implications to CAPM related research, such as event studies, market anomalies and market efficiency. Besides it seems that capital sizes are correlated to the violation of heteroskedasticity (under Breusch-Pagan tests), and it seems that the larger the capital size, the easier to violates the assumption of homoskedasticity. Finally, by taking the Cement Industry Index as example, which simultaneously violate the basic assumptions of normality, homoskedasticity and nonautocorrelation, this paper demonstrates a procedure for correctly estimating the regression coefficients. The suggesting estimation procedure should not be neglected by a rigorous academic research.