股票報酬可預測性在金融市場一直是很熱門的議題，而當中Campbell and Yogo (2006) 提出的Bonferroni Q-檢定一直是很常用並且廣用的方法。但在近年有學者Phillips (2014) 提出它並不全然可行，因此在某些情況下，Bonferroni Q-檢定過程中的信賴區間取得必須有不一樣的方法。我們定義一個干擾參數的估計值，並且設立一個界線來區別不同信賴區間使用的時機，藉此完善整個可預測性的Bonferroni Q-檢定。而我們也利用蒙地卡羅來逐步驗證我們綜合性的檢定方法。除此之外，一般的可預測性檢定都具有常態假設，但這並不符合實際金融 資料。我們知道經濟與金融資料普遍具有高持續性以及厚尾的情形，所以我們把重點著重在具有接近單根的情況，並且把常態假設放寬至厚尾假設，甚至變異數不存在來做可預測性檢定，如此一來更符合現實資料的分析。我們也從蒙地卡羅的結果得到，我們綜合性的檢定方法在厚尾假設下可行並且精確。而我們也根據美國金融市場股票報酬的實證得到，Earning-Price Ratio 具有可預測S&P 500指數報酬的性質，另外所有選取的解釋變數皆具有高持續性以及厚尾的特性，由此反觀，相對於以往皆是常態假設，我們的厚尾假設是更符合真實資料的。關鍵詞: 可預測性; Bonferroni 檢定; 單根; 信賴區間; 厚尾 ;The stock returns predictability is always a popular issue in the financial market, and the Bonferroni Q-test proposed by Campbell and Yogo (2006) has been a common and general method. But in recent years, the researcher in Phillips (2014) proposed that this method is not always valid in some situations. Therefore, we need to use the different way to get the relative confidence intervals which are needed in Bonferroni Q-test procedure. We then define an estimator of the nuisance parameter and set a boundary to distinguish the time when to use the different confidence intervals, and from this, we can complete the whole predictive Bonferroni Q-test. Then we use Monte Carlo to progressively verify our composite testing method. Beyond that, the general predictive tests usually have a normal assumption, this assumption is not satisfied the practical financial data. We all know that economic and financial data have high persistent and heavy tail, so we focus on the case that the data are near unit root. And we relax the normal restriction to a heavy-tailed assumption even infinite variance to do the predictive test so that the analysis is more corresponding to the real data. We also can have the result from Monte Carlo that our composite method is valid and precise under the heavy-tailed assumption. According to the empirical analysis using the U.S. equity data, we find reliable evidence for predictability of the earnings-price ratio, and the other predictor all have high persistence and heavy-tailed property. From the empirical results, we can conclude that unlike the normal assumption in the test before, our heavy-tailed assumption in this predictive test is more corresponding to the data.
