應用企業年報中之文字資訊於盈餘預測之研究

Abstract

盈餘預測向來為學術界與實務界所感興趣的議題，一直以來許多研究不停嘗試為開發出更好的預測模型，而目前這些盈餘預測模型大多只包含了存在於財務報表中或是由財務報表衍生出來的數量性會計資訊。但由於會計資訊只能反映出企業在過去年度的表現，因此這些只含數量性會計資訊模型的預測能力也常是差強人意；相反地，年報中的文字訊息則包含了許多前瞻性的資訊，而這些訊息對於盈餘預測可能會是更加有幫助的。 本研究試圖探討企業年報之文字資訊對於盈餘預測是否有所幫助，欲建立一結合由管理階層透過年報文字所傳達之正負面觀感和數量性會計資訊的盈餘預測模型。由於包含了這些具有前瞻性的資訊，本研究預期此一預測模型能較過往之模型有更高之預測能力，同時亦推測當年報文字所呈現之正負面觀感模糊不明確的時候，由於未來狀況不確定程度較高，則會計基礎盈餘預測模型之預測誤差將會提高。 藉由比較本研究模型與隨機漫步模型、會計基礎模型和分析師預測之預測誤差，發現本研究所提出之模型在預測誤差上較隨機漫步模型與會計基礎模型具有顯著之改善，雖在預測準確度上略遜於分析師預測，但研究顯示其差異並未達顯著。同時亦發現會計基礎模型在年報文字態度呈現顯露較為曖昧、不明確的情況下，預測能力也會有下降的情況。

Earnings prediction has always been a major interest to both the academia and the practitioners. Numerous studies have attempted to develop models for better prediction. The majority of these models incorporate only quantitative accounting data contained in or derived from financial statements. However, since accounting data reflects only past performance, the prediction power of these models is not very satisfactory. In contrast, textual information in annual reports contains lots of future-oriented information. This type of information could be more useful for earnings prediction. This study attempts to investigate whether textual information in annual reports is useful for earnings prediction. We build a prediction model that incorporates the positive/negative sentiment as conveyed by the management through the textual information contained in annual reports. We posit that the model can have more prediction power by including such information. We also posit that prediction error of the accounting-based model will be large if the sentiment is more ambiguous, because the future is more uncertain. By comparing the prediction error of our model against those of random walk model, accounting-based model, and analysts’ forecasts, we find that our model is significantly better than the random walk model and the accounting-based model. Even though our model is inferior to analysts’ forecasts, the difference is not significant. We also find that the accounting-based model has less prediction power when the sentiment is more ambiguous.