目標價預測誤差之決定因素

Abstract

以往關於分析師的研究都局限於盈餘預測與投資建議，近幾年，國外開始有研究指出，分析師的目標價預測的確提供了投資評等與盈餘預測以外的資訊，顯示目標價有其重要性。然而，受限於樣本缺乏，所以國內很少有目標價預測誤差的研究。本文利用鉅亨網所提供的券商目標價預測與原始分析報告，並參考Bonini et al.(2006)，利用兩種策略：買進並持有策略、理想世界完美策略，以及兩種目標價預測誤差的衡量方法：過度預測程度、預測的絕對誤差，對兩份樣本進行迴歸檢定。 根據第一部份樣本的結果，發現外資券商的預測誤差的確顯著的小於本國券商，但是在控制了其他因素之後，顯著性降低。而在投資評等方面，強力買進與強力賣出等級的投資建議的預測誤差最大，在投資評等修正方面，降級的預測誤差最小。而從第二部份的結果，發現分析師的經驗越多，預測誤差越小；報告字數越多，則越容易過度自信，因而預測誤差越大，然而，在控制了其他因素之後，顯著性都降低了。而在評價方法則並沒有顯著且ㄧ致的差異。不論在什麼策略下，目標價所隱含的報酬率越高、市場報酬越低，則目標價預測誤差也越大，而且越有過度預測的傾向，這兩個因素是最顯著而且一致的。 為了避免樣本選擇偏誤，或是檢驗結果受到極端值的干擾，本文中也進行的穩健性測試，發現結果並無太大的不同。

Abstract Previous research on analyst reports primarily examines revisions in stock recommendations and earnings forecasts. Recently some foreign studies have found that target price contains information even in the presence of earnings forecasts and stock recommendations. However, domestic research rarely focuses on target price for lack of database. For this paper, I use two strategies and two measures considered by Bonini et al.(2006), and run regressions on two samples from cnYES.com. According to the results of the first sample, foreign brokerages’ errors are significantly smaller than those of domestic brokerages’. However, after controlling for other variables, the errors become significantly lower. The forecasting errors are the largest in strong buy and strong sell and the smallest in downgrade. The results of the second sample show that the forecasting errors made by the analysts are smaller when the analysts have more experiences and when the reports contain fewer words. Similar to the first sample, the errors are significantly lower after controlling for other variables. The different valuation methods do not change the magnitude of the forecasting errors. When the implied returns are higher or the market returns are lower, the forecasting errors are significantly and consistently larger. To avoid sample selection bias or the effect of outliers, this thesis also examines the robustness of the regressions. The results are quite similar to the ones obtained by running simple regressions.