AUDIT OPINION PREDICTION BEFORE AND AFTER THE DODD-FRANK ACT USING DATA MINING APPROACHES.

• Author: Kwak, Wikil

INTRODUCTION

In the aftermath of the 2008 financial crisis, the Dodd-Frank Wall Street Reform and Consumer Protection Act (2010) was introduced to bolster the effects of the Sarbanes and Oxley Act (2002) in regulating financial accounting and reporting. However, Nwogugu (2015) proposed that the Dodd-Frank Act has done more harm than good by increasing transaction and compliance costs without creating any significant economic growth. Likewise, Dey and Sullivan (2012) showed that the cost of internal control audit for smaller firms (called non-accelerated filers with market capitalization of less than $75 million) is proportionately higher than for larger firms, suggesting that smaller firms be exempt from the Dodd-Frank Act requirements to avoid or reduce the compliance cost.

In contrast to previous studies, Harp, Myring, and Shortridge (2014) found that the Sarbanes-Oxley Act (2002) resulted in more consistent disclosures across firms. Their results suggest the Sarbanes-Oxley Act was effective in reducing variation in the quality of financial information, thus leading to an improved financial environment. Similarly, He and Yang (2014) found that the effectiveness of audit committees in restricting earnings management and improving financial reporting quality is conditioned on industry regulation. He and Yang's (2014) study suggested that not only the general regulatory changed the implementation of Sarbanes-Oxley Act, the general regulatory pushed the firms to adopt more effective internal monitoring systems, but specific industry-wide regulations also shape the effectiveness of governance structures.

In this study, the authors tested the effects of Dodd-Frank Act (2010) using several data mining approaches (random forecast, decision tree, and the support vector machine). Using financial and market variables (liquidity, profitability, and stock return), the "qualified audit opinion" in the pre-and post-Dodd-Frank Act periods were predicted for each business selected. In Spain, Fernandez-Gamez, Garcia-Lagos, and Sanches-Serrano (2016) used neural networks to predict qualified audit opinions based on their local firm data. Meanwhile in Europe, using a sample of 450 publicly listed, nonfinancial British and Irish data, Kirkos, Spathis, and Manolopoulos (2007) applied decision tree, neural network, and Bayesian belief network for predicting audit opinions. While these are more commonly used data mining tools, the aim of this study is to use more contemporary tools to support improved prediction accuracy in the post-Dodd-Frank Act period using U.S. data. Additionally, the findings indicate that the benefits of the Dodd-Frank Act outweigh the compliance costs.

Empirically, the study tested the relationship between auditor qualification and financial and market variables (liquidity, profitability, and stock return) using data mining tools for the pre-and post-Dodd-Frank Act (2010) periods.

LITERATURE REVIEW

Using a probit model, Dopuch, Holthausen, and Leftwich (1987) predicted qualified auditor opinions and found that they are associated with a firm's financial and market variables. However, most audit opinion studies use traditional regression or logit analysis using U.S. data. Maggina and Angelos (2011) used logit analysis to predict audit opinion using the Athens Stock Exchange data. Burcu and Bengu (2011) used logistic regression to predict audit opinions using Turkesh data. Vichitsarawong and Pornupatham (2015) used regression to study that firms receiving modified opinions have lower earnings persistence. Chen, Martin, and Wang (2013) predicted that managers were motivated to avoid receiving going-concern opinions after their insider trading to reduce the risk of litigation.

Via propensity score matching (PSM) and Rosenbaum Bounds (RB), Peel and Makepeace (2012) demonstrated that higher quality audits are valued by clients, audit fee premiums charged by the Big 4 and the top 4 mid-tier (mid 4) auditors reflect audit quality differentials and that the premiums are relatively insensitive to potential hidden bias.

Using Australian data, Hossain's (2013) study found there were significant and positive links between non-audit service fees and the propensity to issue a going-concern opinion for financially distressed firms post-Corporate Law Economic Reform Program Act (CLERP, 2004). Habib (2013) used several firm-specific financial variables such as firm size, advantage, profitability, and audit-related variables such as the Big 4 audit firm affiliation, industry specialization, partner tenure, audit report lag, and prior qualified audit opinion factors on auditors' decisions to issue modified audit opinions. However, his meta-analysis to predict audit opinion results were not conclusive. Mattens, et al. (2008) proposed support vector machines (SVM) and rule-based classifiers to predict going concern opinions.

Data mining generally can be achieved by association, classification, clustering, and predictions. From the mathematical point of view, binary or induction decision tree, statistics, and neural networks can implement the algorithms of data mining. Kwak, Shi, and Kou (2012) used the Multiple Criteria Linear Programming (MCLP) data mining model using Korean financial data to predict bankrupt firms in the Korean capital market environment. The results of the MCLP approach in their firm bankruptcy prediction study were promising as this approach performs better than traditional multiple discriminant analysis or logistic regression analysis using financial data. This approach may be used in audit prediction studies.

Kirkos, Spathis, and...