006_029評論

Report
2013第五屆海峽兩岸會計學術研討會
暨 現代會計論壇
運用詞頻分析技術分析XBRL
財務報表中之附註揭露
—以投資性不動產為例
作
者:黃世銘、林聖訓及周玲儀
評論人:陳叡智
2013年11月28日
研究動機
2010年導入XBRL,要求企業以XBRL
申報財務報表及與附註揭露,可提高財
務資訊的透明度及品質。
2013年開始採用IFRS,因強調專業判
斷,管理當局裁量空間增加,可能造成
財報的操弄。
2
研究目的
以文字探勘技術及詞頻分析法則擷取
附註揭露資訊,檢視異常詞頻或擁有共
同特徵之企業,降低檢閱財報之難度。
由異常詞頻與低頻詞頻檢視附註,分
析裁決性科目(投資性不動產)是否因裁
量權變動(IFRS)而產生操弄的風險。
3
研究流程
蒐集XBRL財報:有投資性不動產科目者361
家,2013Q1有12家於轉換日將部分資產轉列為投
資性不動產,並採公允價值認定成本。
資料前置作業:擷取會計科目及附註;斷詞及詞
頻計算;建立資料庫。
以Zipf’s Law詞頻分析分析資料:比較理論值
及實際值,以偵測異常詞頻(區間估計、Z檢定)及
低頻詞頻(下方切斷點/gap參數、相似度比較)。
系統展現:異常詞頻分析圖、個別公司詞頻分布
圖及表、低頻詞頻分析圖、符合程度公司列表。
4
研究結果
附註揭露的品質
重大會計政策附註偏於制式化。
不同會計政策卻有一致性的財報附註。
附註誤植:粗心的事務所及客戶。
5
優點
以XBRL報表格式的特性,使用Zipf’s
法則建立詞頻分析機制,可辨認財務報
表附註揭露之異常資訊。一來提升資料
處理的效率,二來有詞頻分析的理論基
礎。
由異常或低頻詞頻分析,發現事務所
附註揭露品質的問題(制式化或誤植附
註)。
6
建議
檢視研究目的是否達成?
說明XBRL在本研究扮演的角色。
系統實測與分析可提供更多的操作說明或
分析過程。例如:實驗目標一異常詞頻偵
測分析,可提供圖7的分析圖搭配表2說
明。
可以轉列至投資性不動產的金額(重大性)
比較揭露差異,而非以事務所比較揭露差
異。
7
建議
選擇建材營造業進行分析的原因為何?
(29家vs5家 or 361家vs12家)
低頻門檻(gap值1.5-4)如何決定?本研
究設置gap為4的原因為何?
資料量小,實驗結果為通則或例外?
例如:附註誤植,因為誤植所以造成低
頻。
8
2013第五屆海峽兩岸會計學術研討會
暨 現代會計論壇
An Examination on the Consistency
between Textual and Numerical
Information in Financial Reports:
A Cross-Country Comparison
Authors: Chi-Chun Chou, C. Janie Chang, and
Wei-Ta Chiang
Commentator: Jui-Chih Chen
2013/11/28
Motivation
 The consistency of quantitative and qualitative
data (Kloptchenko, et al., 2004)
 Quantitative data reflect past performance,
qualitative data may have contained messages
about future performance of a company .
Qualitative data can help indicate or reveal
insiders’ moods and anticipations(Kloptchenko,
et al., 2004).
 The comparison of US, Taiwan and China listed
companies in semiconductor industry
 Financial ratio vs text analytics technique
 Developed market vs developing market
10
Data Sources
Quantitative Data
US
Taiwan
China
YCharts
TEJ
SINA
MD&A
SEC Edgar
database
Status of
operations
MOPS
TWSE
Director’s
report
SINA
33
39
30
290
(203/87)
311
(217/94)
225
(157/68)
Changes in three
financial ratios: ROE,
RT, CR
Qualitative Data
No of Companies
No of Reports
(Training/Testing)
Periods: 2002-2010 annual report
11
Methodology
Data Treatment
Quantitative Data
Qualitative Data
K-means clustering algorithm
“Good” vs “Poor”
TFIDF analytics techniques
“Positive” vs “Negative”
Data Analysis
Qualitative Positive
Data
Negative
12
Quantitative Data
Good
Poor
Fair
Exaggeration
Pessimistic
Fair
Research Findings
 Companies in China have a highest tendency
to exaggerate and overstate about their
performance.
 They were more inclined to hide negative
information while promoting positive
information.
 The effect of “disclosure discipline” on the
report consistency
 The difference in language, culture or
reporting practices may have more
implications in this research.
13
Suggestions
 The choices of three financial ratios and
relationships to the textual contents.
 The accuracy of the word segmentation system.
 The possibilities of sentiment differentiation
(positive words vs negative words) between
countries and cultures (sentiment analysis,
Loughran and McDonald, 2011).
 From Table 11-14, it indicates the companies in
China may be more realistic because its fairness
ratio (76.47%) is the highest amongst three
countries.
14
Suggestions
 The differences of individual comparisons
between three ratios and textual contents may
result from the main concepts of textual
contents and local reporting practices.
Country
United
States
Taiwan
China
15
Ratios
ROE
RT
CR
ROE
RT
CR
ROE
RT
CR
Fairness
43.68%
51.73%
48.28%
40.43%
45.75%
47.87%
77.94%
61.76%
50.00%
Exaggeration
4.60%
5.75%
3.45%
23.40%
18.09%
17.02%
22.06%
38.24%
50.00%
Pessimism
51.72%
42.53%
48.28%
36.17%
36.17%
35.11%
0.00%
0.00%
0.00%
Suggestions
The possibilities of using regression
analysis to test the relationships between
financial ratios and textual contents. Also, it
may be easier to test whether the textual
contents are more related to the financial
performance of the current year or the
following year.
16

similar documents