應用企業年報中之文字資訊於盈餘預測之研究-以美國為例

Abstract

盈餘預測長久以來廣為投資者、債權人、財務報表使用者等大眾所關注之焦點。雖然過去已有許多學者針對會計盈餘預測進行研究，但所使用的盈餘預測模型大都是以財務報表的數字資訊為基礎來預測企業未來盈餘。量化的財務資訊雖然較為客觀，但是因其為歷史訊息的特性，相較於質化的文字資訊，卻反而不具攸關性。因此這些僅含量化資訊之預測模型其準確度往往不甚理想。年報中富含的文字訊息包含了企業管理當局對於未來的展望，彌補量化資訊的不足，對於預測企業未來的盈餘應有相當助益。 本研究以美國之企業年報為樣本，運用內容分析法，量化年報中的文字資訊，並將其加入以量化資訊為基礎之會計基礎盈餘預測模型，建立本研究之模型。本研究預期年報中文字資訊所呈現的正負面觀感，與企業未來盈餘具有關聯性；因此加入文字資訊變數後之盈餘預測模型，預測準確度能有所提升。同時當年報中文字資訊所呈現的正負面觀感較模糊不清、不具一致性時，代表未來狀況不確定程度較高，因此會計基礎盈餘預測模型之預測能力會較低。 本研究實證結果發現，文字資訊與企業未來盈餘不具顯著關聯性。本研究模型之預測能力與會計基礎模型相較，也並無顯著改善。然而當年報中文字資訊所呈現的正負面觀感較不明確時，會計基礎盈餘預測模型之預測準確性確實較低。

The prediction of future earnings has always been a major issue to investors, creditors, and financial statements users. Many studies have been devoted to earnings prediction; however, the performance of their prediction models is not satisfactory. These models are based largely on the quantitative financial accounting information contained in annual reports. Although quantitative information is more objective, it is based on historical data. On the other hand, textual information of annual reports usually is more forward-looking, thus presumably is more relevant to future earnings prediction. Using the annual reports of US-listed companies as samples, this study conduct content analysis to quantify the textual information contained in the reports. We then incorporate the textual information into the accounting-based prediction model to form our prediction model. We presume that sentiment revealed by the textual information possesses information about future earnings. Therefore, prediction accuracy can be improved by adding the sentiment variable. In addition, we expect that prediction accuracy of the accounting-based models will be lower if sentiment uncertainty is higher. The results show that there is no significant relationship between textual information and future earnings. Compared to the traditional accounting-based model, our model does not improve prediction accuracy. However, when the sentiment uncertainty is higher, prediction accuracy of the accounting-based model does perform worse.