應用智慧型決策支援系統於審計公費訂價之研究

A Knowledge-Based Decision Support System for Audit Pricing

車詠淩(Yong-Ling CHE)

指導教授 ： 王維康

元智大學/管理學院/會計學系/碩士(2005年)

摘要

隨著經濟環境的演變，台灣的審計市場，在供給及需求上，產生很大的變化；「供給」與「需求」，意指人們在市場經濟中相互交易時所表現出來的行為，其價格之多寡，視所處市場狀態、供需雙方主觀意願而定。依據供需觀點而言，審計公費之訂價，理論上，應由會計師事務所（供給面）及受查者（需求者）雙方，依合理客觀基礎訂定。
　　綜觀國內外學者對於審計公費決定因素之研究，多以建立一多元迴歸模型為估計基礎，隨著電腦科技之發達、產業環境之快速變化，本研究運用傳統迴歸分析(Ordinary Least Square，OLS)與支向機迴歸分析(Support Vector Regression，SVR)，並透過決策支援系統(Decision Support System，簡稱DSS)，結合資料庫、模式庫及知識庫之建構，得以建立一較完整、便利的審計公費訂價基礎。
本研究發現91年及92年共363筆資料中，在準確度方面，SVR有200筆高於OLS，約占55%，額外163筆雖然OLS高於SVR，但其準確度差異並不大；在穩定性方面，SVR預測值與實際值之差異較OLS小，其穩定性優於OLS，整體而言，SVR無論在準確度或穩定性方面，皆優於OLS。且使用不同的SVR模型或改變其核心函數，對於SVR預估公費準確度並無太大影響；由此可知，SVR在預估公費主要是由其所包括的變數所影響，且在與OLS比較後，我們可以知道使用非線性迴歸在公費估算可以較線性公式準確，因此，可以推論這些變數對於公費的影響力並不是依線性比例，未來如欲建立更精確的公費訂價模型，非線性的模型是值得考量的。

關鍵字

審計公費、知識型決策支援系統、迴歸模型、支向機模型

並列摘要

The audit fee market in Taiwan has changed dramatically over the past few years in both supply and demand sides due to the changing economic conditions. Theoretically, the supply side (accounting firms) and demand side (clients) behavior determines the reasonable audit fee. However, in practice, there are numerous factors that are known to affect audit fee. Previous research on determinants of audit fee is based mostly on linear regression model. This provides a foundation for using the ordinary least square (OLS) regression model to forecast audit fee.
In additional to the well-known OLS regression model, this study incorporated the support vector regression (SVR) model into a decision support system (DSS) to build a framework for audit fee pricing system. The system integrates a database for known audit fee records, a model library and a knowledge-base to facilitate a transparent and convenient process in deciding audit fee.
This study has also used audit fee data from 2002 to 2003 to validate the system. Preliminary findings indicate that using leave one out cross validation, among the 363 audit fee records, SVR outperformed OLS regression in 200 cases. For those cases when OLS are better, the difference between predictions of both models is not significant. In addition, the results are robust against using different kernels in SVR. Therefore, one conclusion can be made from this research is that non-linear models are better in describing the audit fee process as opposed to the widely adopted linear models.