Evaluation of the going‐concern status for companies: An ensemble framework‐based model

• DOI: http://doi.org/10.1002/for.2653
• Published date: 01 July 2020
• Date: 01 July 2020

RESEARCH ARTICLE

Evaluation of the going-concern status for companies: An

ensemble framework-based model

Yu-Feng Hsu

1

| Wei-Po Lee

2

1

Department of Accounting and

Information Technology, National Chung

Cheng University, Chiayi, (ROC), Taiwan

2

Department of Information

Management, National Sun Yat-sen

University, Kaohsiung, (ROC), Taiwan

Correspondence

Yu-Feng Hsu, Department of Accounting

and Information Technology, National

Chung Cheng University, Chiayi, Taiwan

(ROC).

Email: ben.yufenghsu@gmail.com

Funding information

Ministry of Science and Technology,

Taiwan, Grant/Award Number: MOST

106-2410-H-128-034

Abstract

Issuing a going-concern opinion is a difficult and complex task for auditors.

The auditors have to take into account different critical factors in order to

make the right decision based on information obtained from the auditing pro-

cess. This study adopts the so-called “random forest”approach (based on the

ensemble method) to assist auditors in making such a decision. To investigate

the corresponding effect of the proposed approach, we conduct a series of

experiments and a performance comparison. The results show that the random

forest method outperforms the baseline methods in terms of the accuracy rate,

ROC area, kappa value, type II error, precision, and recall rate. The proposed

approach is proven to be more accurate and stable than previous methods.

KEYWORDS

artificial neural networks, decision tree approach, ensemble method, going-concern prediction,

random forest method, support vector machine method

1|INTRODUCTION

The going-concern opinion is a well-known and funda-

mental concept in the accounting and auditing domain.

The Statement on Auditing Standards (Ellingsen, Pany, &

Fagan, 1989) states that auditors are required to assess

whether their clients can survive without questions about

their ability to remain in operation. Auditors need to

issue an audit opinion based on the financial condition of

the audited company. This financial status report is cru-

cial to investors and stockholders. They rely heavily on

the audit report and expect it to reflect the real financial

situation. When a company encounters a financial crisis

without receiving a going-concern report in advance, it is

recognized as an audit failure (K. C. W. Chen & Church,

1992; Geiger & Raghunandan, 2002; McKeown,

Mutchler, & Hopwood, 1991). Although this evaluation

process is essential, it is difficult and complicated.

A variety of approaches have been taken in past studies

to overcome the difficulty in making such decisions. Both

statistical methods (e.g., logistic regression and multiple

discriminant analysis; see Altman, 1982; Bell & Tabor,

1991; Dopuch, Holthausen, & Leftwich, 1987; Koh, 1991;

Levitan & Knoblett, 1985; Menon & Schwartz, 1987;

Mutchler, 1985) and machine learning techniques

(e.g., artificial neural network, support vector machine, and

decision tree methods; see Anandarajan & Anandarajan,

1999; Koh & Low, 2004; Martens, Bruynseels, Baesens, Wil-

lekens, & Vanthienen, 2008) have been used to build a

going-concern prediction model to assist in decision mak-

ing, including some suggested representative variables (data

attributes) that can be implemented during the auditing

process. Although these methods are shown to be useful in

the experiments presented, most of them are composed of a

single major model. However, as some of the single major

models usually operate under limitations including normal-

ity and linearity assumptions, financial data usually violate

these assumptions (Deakin, 1972); moreover, it re mains

unclear whether a combined model can deliver better per-

formance in going-concern prediction issue, and the data

attributes (i.e., financial indicators) will obtain equally

importance in the prediction process. Also, it is unknown

Received: 13 December 2017 Revised: 23 July 2019 Accepted: 3 January 2020

DOI: 10.1002/for.2653

Journal of Forecasting. 2020;39:687–706. wileyonlinelibrary.com/journal/for © 2020 John Wiley & Sons, Ltd. 687

whether these methods will perform equally well in the real

complicated financial world a company faces, especially

during a subprime crisis. Thus, in addition to prediction

accuracy, in this study we emphasize stability and robust-

ness with regard to a minimal type II error rate, which is

treated as equally important in the computer model for

assisting with going-concern decision making for practical

applications.

Taking into account the above, this paper proposes a

newapproachwithanensembleframeworkthatmakesuse

of multiple base models to obtain a better synthesized deci-

sion than those of other constituent models. We choose the

random forest algorithm to implement the idea and conduct

extensive experiments for performance evaluation. The

observations show that the new model synthesized from a

variety of different base models not only offers better predic-

tions but also more stable decision making. Moreover, we

also perform a series of analyses and discussion of the quan-

titative experimental results.

The remaining part of this paper is constructed as fol-

lows: Section 2 reviews the going-concern related research

work, and introduces the concept of the random forest and

ensemble framework. Section 3 describes the research

framework, including the data sets, variables used, related

statistical methods, machine learning techniques, and ran-

dom forest parameters. Section 4 presents the experimental

outcomes of the different methods used in the tests. In the

Section 5 some conclusions are offered.

2|BACKGROUND AND

LITERATURE REVIEW

The issue of making a going-concern judgment has

attracted increasing attention from both academics and

practitioners over the last 30 years. Various methods have

been proposed to tackle this problem, which can be catego-

rized into two types: traditional methods and machine

learning techniques. Using traditional methods, Mutchler

(1985) chose six important variables related to auditing as

independent variables and applied multiple discriminant

analysis (MDA) to a history data set to build a prediction

model for the going-concern judgment. Koh employed the

logistic regression (LR) technique with six financial ratios

that had been referenced in a previous study to construct a

going-concern prediction model (Koh, 1992). In contrast to

the traditional methods are the machine learning tech-

niques used to predict a going-concern opinion. For exam-

ple, Anandarajan and Anandarajan (1999) first applied

artificial neural networks (ANN) and MDA to the going-

concernissueandthenproposedanexpertsystem(ES)to

predict the decision. They compared the prediction accuracy

obtained with ANN, MDA, and ES, and found that ANN

surpassed the others. Lenard, Alam, and Booth (2000) used

a fuzzy clustering (FC) technique and a hybrid model that

included an M-estimator discriminant method and the GC

advisor developed in their algorithm to produce a going-

concern opinion. They showed that their hybrid model had

a high prediction accuracy and offered useful information

to auditors in making a going-concern decision. In addition,

Koh and Low adopted three techniques—LR, ANN, and

decision trees (DT)—to establish the going-concern model

(Koh & Low, 2004). Their experiments showed that DT had

the best performance and offered an intuitional method to

demonstrate the decision-making process. Martens et al.

(2008) tried to find the shortest path of a decision process

and found a technique that provided better performance.

They employed support vector machine (SVM), LR and ant

colony (AC) algorithms to perform the going-concern pre-

diction and then compared the prediction accuracy of these

methods with DT. The results showed that, among the three

methods, SVM and LR provided better prediction accuracy.

They also discovered that, although the AC algorithm had a

shorterpaththantheDTapproachforfiguringoutthe

going-concern status, it had the worst prediction accuracy

among the four described methods. A common characteris-

tic of the above studies is that they only focus on looking

for the best method rather than assembling models to

improve prediction accuracy.

The most relevant machine learning method is the

ensemble model, derived from the “wisdom of crowds”con-

cept (Surowiecki, 2004). This method has in fact been dem-

onstrated by the machine learning community to show that

better performance can be obtained by using a combination

of multiple classification techniques for prediction than any

single one (Frosyniotis, Stafylopatis, & Likas, 2003; Tin

Kam, Hull, & Srihari, 1994). When using an ensemble

method to create a model, the target data are separated into

a collection of training data sets and the testing data set.

Each of the former is used to train a classifier, which will

then be tested on the latter data set to verify the

corresponding performance. When all classifiers are trained

and tested, the model starts to work and each classifier will

individually produce an output in response to an input

datum. The model then collects all these outputs to generate

the final outcome. There are many ensemble methods, such

as bagging, boosting, Bayesian, Borda count, and majority

voting; all show good performance (e.g., see G. Y. Chen &

Kégl, 2010; Kim, Kim, & Lee, 2003; J. Sun, Jia, & Li, 2011;

West, Dellana, & Qian, 2005). The random forest algorithm,

based on the bagging method, has shown superior perfor-

manceinsomeapplicationdomains(see,e.g.,Catal&Diri,

2009; Chan & Paelinckx, 2008; Cutler et al., 2007; Diaz-

Uriarte, 2007; Genuer, Poggi, & Tuleau-Malot, 2010;

Lundström & Verikas, 2013; Rodriguez-Galiano, Ghimire,

Rogan, Chica-Olmo, & Rigol-Sanchez, 2012; Sakiyama

688 HSU AND LEE