A Machine Learning Approach of Measuring Audit Quality: Evidence From China

Abstract

This study develops machine learning models to predict audit quality by using a wide range of data describing detailed characteristics of accounting firms, individual audit partners, and public companies in China. It constructs the surprise score, a new measure of audit quality, calculated as the difference between the predicted probability and the actual value of audit quality proxies (e.g., audit adjustments and non-clean audit opinions). Unlike traditional, observable proxies that measure audit quality indirectly, the proposed measure is driven by unobservable factors directly affecting audit quality. This study also validates the proposed measure by comparing the relative explanatory ability of the surprise score and its corresponding traditional audit quality measure over audit failures proxied by penalties due to accounting violations. Results show that, for the event of non-clean audit opinion, the surprise score is more strongly associated with the occurrence of penalties than traditional audit quality measures. For events of audit adjustments, the surprise score is significantly associated with penalties for the group of companies that received extremely high/low surprise scores, whereas the traditional audit quality measures exhibit no explanatory power for those companies.