運用機器學習提升異常審計公費之估計準確度

Abstract

本研究探討「普通最小平方法(OLS)」與「機器學習模型」於預測審計公費的準確度差異，是否影響異常審計公費之估計，並進而影響其與審計品質之間的關係。針對此一關係，過去文獻並未有一致的結論。本研究發現，預測審計公費時，機器學習模型XGBoost明顯優於OLS；此外，當XGBoost模型加入更多的財務資料並應用特徵工程篩選重要變數後，準確度會再進一步提高。最後，異常審計公費與審計品質之間的關係，研究結果顯示確實會因所選用的審計公費預測模型（OLS或XGBoost）或變數選擇方式（基於先前文獻或特徵工程）的不同，而產出不同準確度的審計公費預測值，進而使該關係有所變化。因此，本研究強調於會計和審計研究中使用預測值時，預測模型和變數選擇方式之重要性。

This study examines the differences in prediction accuracy of audit fees between Ordinary Least Squares (OLS) regression and machine learning models, specifically focusing on their impact on the relationship between abnormal audit fees and audit quality, a topic with mixed evidence in the existing research. The results reveal that the XGBoost, a machine learning model, significantly outperforms OLS regression in predicting audit fees. Further improvements in prediction accuracy are achieved by incorporating additional financial data and applying data-driven feature selection techniques in XGBoost. The association between abnormal audit fees and audit quality also appears to be contingent on the choice of prediction model (OLS vs. XGBoost) and the set of variables used in these models, whether chosen based on prior research or through feature selection. Our findings underscore the critical importance of modeling approaches and variable selection in accounting and auditing research that relies on predicted values.