the impact of multinationality on firm value: a comparative analysis of machine learning techniques

Decision Support Systems 59 (2014) 127–142

Contents lists available at ScienceDirect

Decision Support Systems

j ourna l homepage: www.e lsev ie r .com/ locate /dss

The impact of multinationality on firm value: A comparative analysis ofmachine learning techniques

Cemil Kuzey a, Ali Uyar a, Dursun Delen b,⁎a Department of Management, Fatih University, Buyukcekmece, Istanbul 34500, Turkeyb Department of Management Science and Information Systems, Spears School of Business, Oklahoma State University, United States

⁎ Corresponding author. Tel.: +1 918 594 8283; fax: +E-mail addresses: [email protected] (C. Kuzey), auya

[email protected] (D. Delen).

0167-9236/$ – see front matter © 2013 Elsevier B.V. All rihttp://dx.doi.org/10.1016/j.dss.2013.11.001

a b s t r a c t
a r t i c l e i n f o
Article history:Received 12 November 2012Received in revised form 31 October 2013Accepted 8 November 2013Available online 16 November 2013

Keywords:Machine learningPredictive analyticsDecision treesArtificial neural networksSensitivity analysisFirm valueMultinationality

In this study, the impact of multinationality (as measured by foreign sales ratio) and fourteen other financialindicators on firm value (characterized by market capitalization and market-to-book ratio) for the period of1997–2011 was investigated using two popular machine learning techniques: decision trees and artificial neuralnetworks.We divided the time period of 1997–2011 into two periods; 1997–2004 and 2005–2011 to investigatethe robustness of results pre- and post-IFRS implementation. To determine the relative importance of factorsas the predictors of firm value, first, a number of classification models are developed; then, the informationfusion based sensitivity analysis is applied to these classification models to identify the ranked order of theindependent variables. Among the independent variables, multinationality was found to determine firm valueonly moderately. In addition to multinationality, other financial characteristics such as firm size (as measuredby natural logarithm of assets), leverage, liquidity, and profitability were consistently found to be affectingfirm value.

© 2013 Elsevier B.V. All rights reserved.

1. Introduction

In recent years, value relevance studies have attracted considerableattention from researchers of diverse backgrounds [1,13,31,38,39,65,87].Those studies generally dealtwith investigating how certainfirm charac-teristics affect firm value. Financial characteristics and non-financialcharacteristics (e.g., voluntary disclosures, adoption of financialreporting standards, auditor type, ownership structure, multinationality,and corporate governance) were among the fundamental dimensionsthat these investigators had focused on.

Beside directly addressing the firm value, a large number of researchstudies focused on analyzing and potentially predicting bankruptcy asa means to identify characteristics (in term of financial ratios) ofsuccessful/unsuccessful firms and their potential values [54]. A simplesearch on the topic “bankruptcy prediction” returns tens of thousandsof studies. A vast majority of these studies differentiate themselvesfrom those of others by using a somewhat unique set of financialcharacteristics and/or employing a different set of prediction models(statistical or machine learning based) [2,42,57,64,72,89]. Thoughmany of these studies are successful in predicting bankruptcy, theyoften fall short on identifying and explaining the characteristics thatcan be used as determinants of the firm value.

1 918 594 [email protected] (A. Uyar),

ghts reserved.

This study aims to address specifically the value relevance ofinternational operations of multinational companies. Previously, somestudies tested the influence of multinationality on firm value albeitmost of them were in developed countries [35]. While these workshave examined the impact of multinationality in various countries,they may not fully capture the extent to which the degree of interna-tionalization affects firm value in emerging/developing countries. Leeet al. [58] argue that the capital markets are less developed and owner-ship concentration is higher in the emerging markets, thus the studiesfor firms headquartered in emerging countries could offer a differentresult for the relationship between international diversification andfirm value. Berrill and Mannella [11] state that there is an increasinginterest in studies that relate to emerging markets and multinationalcorporations. Emerging markets are attracting the attention of thewhole world due to their current high growth rates and potentials forthe future. They are ideal markets for goods and services with largepopulations and increasing incomes [11]. Turkey is one of those emerg-ing countries with its fast growing economy and young population.Thus, this study extends the prior work by examining the impact ofmultinationality along with certain firm characteristics on firm valuein the emerging market context. Moreover, the present paper also dif-ferentiates itself from previous studies in the literature wherein theyutilized data mining and/or text mining in financial reporting areafocusing primarily on detection of financial statement fraud [93–97].However, this study analyzes the impact of multinationality on firmvalue by using data mining technique and a broad set of financial data.

http://crossmark.crossref.org/dialog/?doi=10.1016/j.dss.2013.11.001&domain=f

http://dx.doi.org/10.1016/j.dss.2013.11.001

mailto:[email protected]



http://dx.doi.org/10.1016/j.dss.2013.11.001

http://www.sciencedirect.com/science/journal/01679236

128 C. Kuzey et al. / Decision Support Systems 59 (2014) 127–142

In prior work, regression analysis has been frequently and primarilyused tools to investigate the association of internationalization withfirm value [31,35,58,82]. This study utilizes advanced analysis tech-niques of decision trees and artificial neural networks. The data sampleused in this study includes Turkish non-financial firms over the periodof 1997–2011. Within this time period, Turkish firms are mandated toadopt International Financial Reporting Standards (IFRS). Therefore,we conduct the analysis for the sub-periods of 1997–2004 and 2005–2011; doing so, we aim to examine whether the adoption of IFRS hasany impacted on the firm value characteristics.1

The remainder of the paper is organized as follows. Next sectionprovides the literature review and makes the case for novelty andimportance of this study. Section 3 presents the methodology used forthe study; Section 4 documents the findings of the study. Finally,Section 5 concludes the paper and explains the implications of the study.

2. Literature review

The information of whether multinationality affects firm value is animportant piece of information for decisionmakers for a number of rea-sons. Firstly, this information provides managers with guidance in rela-tion to whether and in which way to expand operations of the businessbeyond the borders of their own home country [31]. Secondly, presentor potential investors of a firm also want to knowwhich characteristicsare value-relevant so as to determine the direction of their investments.Finally, financial analysts also wonder which factors impact firm valueso that they can make the best investment decision on behalf of theircustomers.

Fauver et al. [35] state the motivations behind increase in foreigninvestment as improved communications, lower transaction costs, andincreasingly integrated foreign markets. However, they also argue thatmany firms incur additional costs and risks such as exchange rate, polit-ical instability, the agency costs, and coordination costs. Thus, the keyquestion to be answered is whether the internationalization reducesor increases firm value. In this context, more evidences are needed,since the existing evidence regarding the benefits of internationaldiversification has yielded inconclusive results [35].

Some of the previous studies have dealt with the relationshipbetween multinationality and firm value; however, their findingswere inconsistent. For example, Eckert et al. [31] conducted theirstudy on German firms and concluded that multinationality is not avalue itself, but through either having intangible assets or realizingeconomies of scale. Eckert et al. [31] proved that leverage and sizehave a significant negative impact on shareholder value, whereasprofitability and capital intensity (as a proxy for economies of scaleand as measured by capital expenditures per sales) exert a significantpositive effect. Riahi-Belkaoui [78] conducted a study on U.S. firms andconfirmed that there is a positive relation between the degree ofinternationalization and the market value of the firm. Fauver et al. [35]empirically proved that corporate international diversification isvalue reducing for U.S.firms on average, but has no significant valuationinfluence for German or U.K. firms. Schmid and Walter [82] used thepercentage of sales from non-domestic operations to measure the im-pact of geographic diversification, and they indicated that geographicdiversification is not associated with a significant valuation discount infinancial intermediaries.

Machine learning (ML) as well as artificial intelligence methodolo-gies have been used extensively to handle financial decision makingproblems [5,6,19,21,37,74]. Furthermore, it has been proven that MLalgorithms such as support vector machines (SVM) demonstratedincreasingly important performances in financial time series analysis[48,68].

1 Beginning from 2005, Turkish listed companies have been required to adopt IFRS bythe Capital Markets Board.

ML is the subfield of artificial intelligence concerned with develop-ment of algorithms that allow computer programs to learn from experi-ence [56]. These algorithms are used in a variety of applications.ML algorithms are appropriate in scenarios where the applications in-volve large databases, making it difficult to establish models [66],where financial data sets are large, as in our case. Various studies haveshown that machine learning techniques such as neural networks anddecision tree algorithms can be employed as an alternative method toresolve classification problems instead of the traditional statisticalmethods [6,9,10,14,19]. Traditional statistical methods use restrictiveassumptions such as normality, linearity, and independence amongpredictor variables. Deakin [26] demonstrated that violations of theseimportant assumptions of independent variables frequently happen infinancial data. As a result, these conventional statistical methodsmight produce limitations in terms of validity and effectiveness. Deci-sion trees (DT) algorithms and neural networks (NN) are among themost popular machine learning algorithms. Several of these algorithmssuch as decision trees algorithms, support vector machines (SVM), neu-ral networkswere developed for application in financial and accountingapplications [43,53,66,76]. Therefore, ML algorithms are the mostappropriate for this study.

ML algorithms were employed successfully in some studies thatfocus on firm value. Chaehwan et al. [18] studied dividend policy fore-casting usingML approaches discovering that comparingML algorithmswas one of the most important managerial decisions affecting theprediction of firm value. Also, Chih-Fong et al. [22] used popular ML al-gorithms such as decision trees, genetic algorithms andneural networksto determine the most important features impacting firm value.

In this study, machine learning (ML) techniques were applied as thedata driven approach. While various studies compare the machinelearning algorithms in general [98–101] such as SVMs, neural nets, lo-gistic regression, naive bayes, memory-based learning, random forests,decision trees, bagged trees, and boosted trees, the comparison ofmachine learning algorithms versus classical statistical techniques wasstudied as well. A literature driven argument for the use of MLtechniques in this context, and a comparative analysis of traditional sta-tistical methods versus ML are summarized in Table 1. According toBreiman [102], statistics really starts with data and the main goals ofit: prediction (estimation) and information (detection). Breiman [102]claimed that higher predictive accuracy is associated withmore reliableinformation about the underlying data mechanism, therefore weakpredictive accuracy can lead to questionable conclusions. Moreover,he valued the importance of algorithmicmodels since they can give bet-ter predictive accuracy than data models, and therefore provide betterinformation about the underlying mechanism. In the light of Breiman's[102] valuable study, the ultimate goal in this study is to obtain accurateinformation.

3. Theoretical background

Four theories are proposed to explain the links betweenmultinationality and firm value: the internalization theory [67]; imper-fect world capital markets [33,67]; managerial objectives [67]; and taxavoidance and low-cost inputs [33,67]. According to the first theory, in-ternalization theory, whichwas developed initially by Caves [17], a firmcan enhance its value by internalizing markets for its intangible assets(i.e. superior production skills, managerial skills, marketing abilities,patents, or consumer goodwill). According to this view, internalizationbrings buyers and sellers of information-based assets together [30],and firms can thenmaximize their revenues through selling or licensingtheir assets to firms in other countries [71]. The second theory regardingthe imperfection of world capital marketsmight prevent investors fromoptimally diversifying their portfolios internationally; therefore multi-national firms provide shareholders with an opportunity to diversifytheir investments [67,71]. The third theory, managerial objectives, alsoplays an important role in the internationalization decision, eventually

Table 1Comparison of econometric theory based and data driven approaches.

Econometric theory based approaches Data driven approaches

It is the traditional field that deals with the quantification, collection, analysis,interpretation, and drawing conclusions from data [103].

It is an interdisciplinary field that draws on computer sciences (data base, artificialintelligence, machine learning, graphical and visualization models), statistics andengineering (pattern recognition, neural networks) [103].

Assumes that the data are generated by a given stochastic data model [102]. Uses algorithmic models and treats the data mechanism as unknown [102].The statistical community has been committed to the almost exclusive use of datamodels.This commitment has led to irrelevant theory, questionable conclusions, and has keptstatisticians from working on a large range of interesting current problems [102].

Algorithmicmodeling, both in theory and practice, has developed rapidly in fields outsidestatistics. It can be used both on large complex data sets and as a more accurate andinformative alternative to data modeling on smaller data sets. Moving away fromexclusive dependence on data models and adopting a more diverse set of tools enable ussolving problems by using the data [102].

Hypothesis testing in order to determine the causes and effects as well as modelinterpretation is critical. Goodness-of-Fit and parameter significance are used formodelselection. [102]. It traditionally concerns itself with analyzing primary data that hasbeen collected to check specific research hypothesis; data can be of an experimentalnature [105].

Predictive accuracy is the focus. Model interpretation is not as important as the predictiveaccuracy [104]. Cross validation of predictive accuracy based on partitioned data sets isused for model selection [102]. It can only concern itself with secondary data collected forother reasons (e.g., analyzing company data that comes from a data warehouse); the datais typically of an observational nature [105].

The sampling of a massive database cannot be analyzed with the traditional statisticalsampling theory tools [105]. The size of the data set and the data are initially collectedfor experimental design is important topic in traditional statistics [103].

The aim of the data driven techniques (ML, datamining) is to analyze greatmasses of dataand carrying out sampling. Sampling is necessary since accessing/analyzing the wholedatabase can be impossible for many applications because of computer efficiency issues[105]. Experimental design is usually irrelevant to DM [103].

There are not available analytical methods to be developed in the statistics field [105]. Appropriate analytical methods are needed to be developed sincemany databases do notlead to the classic forms of statistical data organizations (e.g., data from internet) [105].

Asymptotic analysis, sometimes criticized as being irrelevant. Traditional statisticsemphasizes the mathematical formulation and validation of a methodology, and viewssimulations and empirical or practical evidence as a lesser form of validation. Theemphasis on rigor has required proof that a proposed method will work prior to its use[103].

Asymptotic analysis becomes very relevant. Computer science and machine learning useexperimental validation methods. Mathematical analysis of the performance of astatistical algorithm is not feasible in a specific setting, but becomes so when analyzedasymptotically. When size becomes extremely large, studying performance bysimulations is also not feasible. It is therefore in settings typical of DM problems that as-ymptotic analysis becomes both feasible and appropriate. [103].

The visualization tools of statistics are usually not calibrated for the size of the data sets[103].

Visualization of the data and its structure, aswell as visualization of the conclusions drawnfrom the data, are central theme [103].

Decision makers formulate the hypothesis and it is confirmed on the basis of sampleevidence. Statistical validation technique provides elements to confirm or disprove thehypotheses formulated by the decision maker, according to a top-down analysis flow[106]

Learning models are capable of playing an active role by generating predictions andinterpretations which actually represent new knowledge available to the users. Theanalysis flow has a bottom-up structure. It is hard to formulate a priori meaningful andwell-founded hypotheses when faced with large amounts of data [106].

129C. Kuzey et al. / Decision Support Systems 59 (2014) 127–142

impactingfirmvalue.Morck and Yeung [67] argue that internationaliza-tion might leave more room for managers to act in their own interests,at the expense of investors' interests. Moreover, Denis et al. [30] arguethat diversification increases the complexity of organizations, leadingto higher coordinating costs for each organization. These factors maycontribute to reducing the firm value of multinationals as opposed touninationals. Finally, under the fourth theory, internationalization mayimpact firm value favorably due to the opportunities provided for taxavoidance and low cost inputs [67]. Denis et al. [30] support this viewstating that a multinational firm has an opportunity to relocate itsproduction to another country where production costs are lower, andit has the ability to reduce its tax liabilities due to the differencesbetween tax systems in other countries.

4. Research methodology

Fig. 1 pictorially illustrates the analytic process employed in thisstudy. The main tasks included obtaining the raw dataset, pre-processing the data, splitting the data into ten randomly selectedcross-validation folds, developing predictive models for each fold,assessing and comparing the predictive accuracy of different machinelearning techniques, and finally generating and consolidating sensitivityresults of the predictive models.

4.1. Data

The dataset used in this studywas obtained from the Financial Infor-mation News Network (FINNET). FINNET has the largest financialdatabase in Turkey, providing a variety of financial data, software, andWeb-based analysis tools to their members. Even though the FINNETdata is rich in content, it had a variety of data problems; demanding athorough process of data cleaning and pre-processing. Proper pre-processing of data is perhaps the most important step in any analyticsstudy [73]. The pre-processing activities included (i) identification and

imputation/elimination of missing data/records, (ii) identification andinvestigation of anomalies and outliers, and (iii) transformation ofnominal values to proper representations for predictive modeling.

The initial dataset for the study consisted of all listed Turkish non-financial companies for the time period of 1997 to 2011. The numberof unique cases/records retrieved from the database was 5835. Afterthe analyses of the data for missing values, 1452 cases had a largenumber of missing values for critical financial indicators; thereforethey were eliminated. There were also 38 cases with unexplainablylarge values (identified as outliers), which were also eliminated fromthe dataset. After the data pre-processing, the final dataset consistedof 4347 cases. The final dataset included proper values for all financialindicators for the years 1997 to 2011.

4.2. Variables

In this study, the endogenous and exogenous variables (Table 2)were not employed in terms of classical econometricmodeling perspec-tive, rather they were used in terms of data driven contemporarymachine learning perspective as endogenous (dependent, response,predicted or target) variable and exogenous (independent, explanatory,predictor or source) variables. Consequently, the variable “Firm Value”was employed as the endogenous variables and the size, leverage,sales growth, capital expenditure, profitability, asset structure & growthrate, and liquiditywere used as exogenous variables. The following tableillustrates the literature in which some of the similar endogenousand exogenous variables were employed. In addition, Table 3 lists andbriefly defines all of the independent and dependent variables (financialindicators) collected, consolidated and used in this study.

What follows are brief descriptions and inclusion justifications(as per the published literature) of the data categories that theseindependent variables belong to. These variables' association withfirm value has been widely investigated in previous firm value studies.

Fig. 1. The process map.

Table 3The list of financial indicators (predictors) included in the study.

Multinationality

Foreign sales ratio Foreign sales ÷ Total sales

SizeLnAssets Natural logarithm of total assets

LeverageLeverage ratio Total debt ÷ Total assetsFinancial debt ratio Total financial debt ÷ Total debt


4.2.1. MultinationalityThere are two competing theoretical arguments about international-

ization of firms; opponents of multinationality argue that it reducesfirm value. They base their arguments on the additional costs that afirm incurs while operating abroad due to environmental, cultural,political challenges/differences [31]. On the other hand, proponents ofmultinationality support the idea that it enhances firm value in a varietyofways. Riahi-Belkaoui [78] states that investors consider international-ization as a hidden or unbooked asset of a firmwhich is not reflected onthe balance sheet. Kim andMathur [49] argue that geographic diversifi-cation can increase firm value by economies of scale, location-specificadvantages, increased operational flexibility, and synergy effects.

Table 2The use of endogenous and exogenous variables in the literature.

Endogenous variable:Firm value

Denis et al. [30], Eckert et al. [31], Olsen and Elango[71], Schmid and Walter [82],

Exogenous variables: Size: Ammanna et al. [3], Bae et al. [8],Connelly et al. [23], Erickson et al. [32],Faleye [34], Fauver et al. [35], Konijn et al. [52], Makand Kusnadi [62], Pramborg [75], Uyar and Kılıç [87],Wu [92],

Leverage: Ammanna et al. [3], Bae et al. [8],Connelly et al. [23,32][34,52,62,75,87],

Sales growth: Hiraki et al. [41], Mak and Kusnadi [62],Uyar and Kılıç [87], Wu [92],

Capital expenditure: Ammanna et al. [3], Connelly et al. [23]Faleye [34], Fauver et al. [35], Konijn et al. [52], Makand Kusnadi [62], Pramborg [75].

Profitability Ammanna et al. [3], Bae et al. [8], Connelly et al. [23],Faleye [34], Pramborg [75],Pramborg [75], Uyar and Kılıç [87],

Asset structure & growth rate Mak and Kusnadi [62]Liquidity Ammanna et al. [3], Pramborg [75],

Sales growthSales growth rate (Salest − Salest − 1) ÷ Salest − 1

Asset turnover rate Sales ÷ Total assets

Capital expenditureCapital expenditure (Δ Long-term assets + Depreciation &

amortization) ÷ Total assets

ProfitabilityReturn on assets Net income ÷ Total assetsNet profit growth rate (Net incomet − Net incomet − 1) ÷ Net incomet − 1

Net profit margin Net income ÷ Sales

Asset structure & growth rateAssets growth rate (Total assetst − Total assetst − 1) ÷ Total assetst − 1

Long-term assets ratio Long-term assets ÷ Total assets

LiquidityQuick ratio (Current assets − Inventory) ÷ Current liabilitiesCash ratio Cash and cash equivalents ÷ Total assetsCash conversion cycle (Accounts

receivable ÷ Sales) ∗ 365 + (Inventories ÷ Cost ofgoods sold) ∗ 365 − (Accounts payable ÷ Cost ofgoods sold) ∗ 365

Dependent variables(market values)

Market capitalization Share price × The number of shares outstanding.Market-to-book ratio Market capitalization ÷ Total book value


Formerly, multinationality was measured using the ratio of foreignsales to total sales (FSTS) and/or the ratio of foreign assets to total assets(FATA)2 primarily. Bae and Noh [7] used two criteria: foreign sales ratio(foreign sales divided by total sales) and foreign tax ratio (foreignincome taxes divided by total income taxes). However, FSTS has beenwidely used as a measure of MNCs because other data are less easilyavailable [7]. Due to the widespread usage and data availability, weused FSTS to measure multinationality.

4.2.2. SizeDenis et al. [30] argued that diversified firms are likely to be substan-

tially larger than domestic firms, further claiming that the former grouphas greater capital market values than the latter one. Jensen [44] claimsthatmanagers have incentives to enlarge theirfirmsbeyond the optimalsize, with growth increasing their power by increasing the resourcesunder their control. Thus, agency costs arise which might influence afirm's value unfavorably. Konijn et al. [52] assumed a negative relationbetween firm value and size; supported by other studies such asAmmanna et al. [3], Pramborg [75], Bae et al. [8], and Erickson et al.[32] who also found a negative correlation between size and firmvalue. On the other hand, Mak and Kusnadi [62], Uyar and Kılıç [87],and Wu [92] found a positive relation between size and firm value.Faleye [34] and Connelly et al. [23] found an insignificant relation forsize and firm value. Fauver et al. [35] found that size is significant andnegative for firms in Germany, but is significant and positive for firmsin the U.K. and the U.S. To measure size, in this study we utilized thenatural logarithm of total assets.

4.2.3. LeverageDeterminants of leverage, which is measured by total debt to total

assets or total debt to total equity ratios, are explained by two theoriesin the financial literature; being the trade-off theory and the peckingorder theory [60]. DeAngelo and Masulis [27] proposed the trade-offtheory which purports to set a theoretical optimum level of debt for afirm, in which the amount of tax savings due to additional borrowingis offset by an increase in the cost of distress. The pecking order theorysuggests that firms have an order of preference when using financingsources [69]. According to this theory, firms prefer internal financingto debt, short-term debt over long-term debt, and any debt outside ofequity [16,60]. Leverage can be negatively (e.g., riskiness, debt over-hang) or positively (e.g., disciplining role) associated with firm value[52]. Some researchers found that leverage is negatively correlatedwith firm value [32,34,52]. On the other hand, Mak and Kusnadi [62],and Bae et al. [8] found a positive relation between leverage and firmvalue. Wu [92] found an insignificant relationship. Ammanna et al. [3]found both a negative and a positive relation depending on changingspecifications. Pramborg [75], Uyar and Kılıç [87], and Connelly et al.[23] could not find a significant relation. In the analysis, we used twoleverage ratios; the ratio of total debt to total assets and the ratio offinancial debt to total debt.

4.2.4. Sales growthSales growthmeasureswhether resources are used efficiently or not.

Current growth rate in sales indicates future growth rates. The expecta-tions of investors for growth are reflected in share prices; the higher theexpectations for growth, the more a firm has value. Thus, we assume apositive impact of sales growth on firm value. Mak and Kusnadi [62]and Hiraki et al. [41] found evidence of a positive relation betweensales growth and firm value. Wu [92] and Uyar and Kılıç [87] found aninsignificant relationship. In addition to sales growth in the currentyear relative to previous year's sales revenue, we used asset turnoverrate to measure sales generating ability of firms using assets.

2 Eckert et al. [31] andRiahi-Belkaoui [78] used both FSTS and FATA; Riahi-Belkaoui [77]used FSTS.

4.2.5. Capital expenditureThe influence of capital expenditures might be primarily explained

by two conflicting approaches. According to the first approach (i.e. thetraditional view), managers are expected to act in the best interests oftheir shareholders, and are assumed to undertake capital projectswhich generates a positive net current value, thereby maximizingtheir stockholders' wealth [46]. In this way, a positive associationbetween capital expenditures and firm value is predictable. On the con-trary, agency theory [45] suggests that managers maymake investmentdecisions to fulfill their own self-interests even though at the expense ofstockholders, leading to a decrease in firm value. According to earlierstudies, capital expenditure has a significant positive effect on firmvalue [3,23,34,35,52,62,75]. Thus, in this study we assume a significantinfluence of capital structure on firm value.

4.2.6. ProfitabilityThe signaling theory explains the relationship between profitability

and firm value. Profitability is a primary performance indicator that isclosely followed by investors and creditors. For the former group, itindicates dividend payment ability (signaling theory), and is alsoindicative of future earnings as pointed out by Wu et al. [91] (signalingtheory), for the latter group, this demonstrates the debt and interestpaying ability of the firm (signaling theory). Therefore, profitability isassumed to impact firm value. A profitable firm is likely to trade at apremium compared to a less profitable one [75]. Faleye [34], Pramborg[75], Bae et al. [8], Uyar and Kılıç [87], Connelly et al. [23], andAmmannaet al. [3] have all found a significant positive impact of profitability onfirm value. For this variable, we used three ratios; return on assets, netprofit margin, and net profit growth rate.

4.2.7. Asset structure & growth rateAgency theory arises out of the conflicts of interest between princi-

pals and agents. Asset growth is a way of demonstratingwhere the pro-vided capital has been invested, and whether it has been investedproperly in order to bring an appropriate rate of return. In otherwords, growth in assets assistsmonitoring of investments, thusmitigat-ing agency costs, much as a governance mechanism. Mak and Kusnadi[62] investigated the impact of asset tangibility on firm value, andfound insignificant relation between fixed asset ratio and firm value.To test the association of this variable with firm value, we used assetgrowth rate and the ratio of long-term assets to total assets in this study.

4.2.8. LiquidityLiquidity is a “two-edged sword”3: keeping liquid assets on hand

creates advantages and disadvantages for the firm as well as its inves-tors. Holding sufficient liquid assets prevents financial distress anddefault risk, enabling the payment of short-term liabilities. However,stockpiling more than sufficient liquid assets could mean that thecapital provided by investors is not utilized to the best advantage,since liquid assets are generally considered to bring lower returnscompared to long-term assets. This has a strong influence on firmvalue. Pramborg [75] could not prove a statistically significant associa-tion between liquidity (as measured by current ratio) and firm value.However, Ammanna et al. [3] found a significant positive associationbetween cash ratio (Cash/Assets) and firm value. Three liquidity ratioswere used in the analysis; quick ratio, cash ratio, and cash conversioncycle.

4.3. Cross validation

Cross validation is a recently popularized technique to estimate theaccuracy of a predictivemodel's performance inpractice. It is sometimescalled rotation estimation and the aim of the technique is to assess how

3 This phrase was previously used by some researchers (see [70]).


the result of an analysis will generalize to an independent data set. 10-fold cross validation is widely used since the empirical studies demon-strated that 10 was an optimal number of folds [51]. In the 10-foldcross validation, the data set is randomly split into 10 mutuallyexclusive subsets of approximately equal sizes. The model is trainedfirst and tested 10 times. Each time, the model is trained on 9 folds(as a combined training data that includes 90% of the total dataset)and tested on the remaining 1 fold. The cross validation estimatesof the overall accuracy of a model are evaluated by averaging the 10individual accuracy measures [73] as shown in Eq. (1):

CVA ¼ 1k∑k

i¼1Ai ð1Þ

where, “CVA” stands for cross validation accuracy, k is the number offolds (here k = 10), and A is the accuracy measure.

Table 4Confusion (coincidence) matrix.

Predicted

Actual Unsuccessful SuccessfulUnsuccessful True Negative False PositiveSuccessful False Negative True Positive

4.4. Decision tree algorithms

Decision trees are commonly used machine learning methods indata mining. There are two main types of decision tree classifica-tions: classification tree analysis and regression tree analysis. Deci-sion trees are becoming increasingly popular for data miningbecause they are easy to understand and interpret, require littledata preparation, handle numerical and categorical data, and theyperform very well with a large data set in a short time. Decisiontrees produce excellent visualizations of results and their relation-ships. Although there are many specific decision tree algorithms,the ID3, C4.5, C5.0, CART, and CHAID and QUEST algorithms are themost commonly used.

Chi-squared Automatic Interaction Detector (CHAID) is anextremely effective statistical technique developed by Kass [47]. Itsmain use is for segmentation, or tree growing. CHAID is a decisiontree technique based on adjusted significance testing. It can beused for predictions in the same way as for regression analysis andclassification as well as detecting interaction between variables.Differing from other decision tree techniques, CHAID can producemore than two categories at any level in the tree; therefore it is nota binary tree method.

C5.0 was developed by Quinlan [76]. It offers a number of improve-ments on C4.5: it is significantly faster than C4.5; it is more memory ef-ficient than C4.5; it creates a considerably smaller decision tree whileproducing similar results; it boosts the trees, improving them and creat-ing more accuracy; it makes it possible to weight different attributesand misclassification types; and, it automatically winnows the data tohelp reduce noise. As a result, it improves the objectivity and precisionof the decision tree classification algorithm.

Classification and Regression Trees (CART)were established by [15].CART is a binary decision tree algorithm capable of processing continu-ous or categorical predictor or target variables. It works recursively:data is partitioned into two subsets to make the records in each subsetmore homogeneous than in the previous subset; the two subsets arethen split again until the homogeneity criterion or some other stoppingcriteria is satisfied. The same predictor field may be used many times inthe tree. The ultimate aim of splitting is to determine the right variableassociatedwith the right threshold tomaximize the homogeneity of thesample subgroups.

The Quick, Unbiased, Efficient Statistical Tree (QUEST) algorithm is arelatively new binary-split decision tree algorithm for classification indatamining [61]. It is similar to the CART Tree algorithm [15]. However,there are someminor differences. For instance, QUEST employs an unbi-ased variable selection method, uses imputation for dealing with miss-ing values instead of surrogate splits, and handles categorical variableswith many categories.

4.5. Neural network analysis

Artificial Neural Networks (or simply Neural Networks) are analytictechniques that were inspired from biology; the basic element of themis a neuron (types of cells found in human brain). The neurons are orga-nized into layers: input, hidden, and output. It operates as the nervoussystem operates. The neural network model accepts many inputs,sums them, usually applies non-linear transfer functions, and generatesthe results. Neural networks are capable ofmodeling very complex non-linear functions [40]. Multilayer perceptron structure of neural net-works was employed in this study. The training of this structure usesback propagation of error method based on generalized delta rule[79]. Information in the form of input fields feeds forward through thenetwork to generate a prediction from the output layer for each recordin the network during training procedure. This prediction is comparedto the recorded output value for the training record, and the differencebetween the predicted and actual output is propagated backwardthrough the network to adjust the connection weights to improve theprediction for similar patterns [84]. In this study, a multi-layeredperceptron (MLP) type feed-forward neural network architecture isused. The network had one hidden layer with 18 processing elements.The network is trained using the back-propagation learning method.The ANN parameters such as number of hidden layers, number of pro-cessing elements, learning and momentum rates, number of epochs,are all determined with numerous experimentations.

4.6. Performance measurements of prediction models

The performance of models used in predicting binary (two-group)outcomes is measured by using a confusion matrix (see Table 4). Aconfusion matrix (also known as coincidence matrix) contains valuableinformation about the actual and predicted classifications created by theprediction model [50]. It is important to use a variety of performancecriteria to evaluate the learning methods [98,99]. For purposes of thisstudy, we used well-known performance measures such as OverallAccuracy, AUC (area under the ROC curve), Recall and F-measure. Allof these measures were used to evaluate each model in the study,after which the models were compared on the basis of the proposedperformance measurements.

4.6.1. Overall Accuracy (AC)Accuracy is defined as the percentage of records that are correctly

predicted by the model. It is also defined as being the ratio of correctlypredicted cases to the total number of cases (see Eq. (2)).

Accuracy ¼ TPþ TNTPþ TNþ FP þ FN

ð2Þ

4.6.2. PrecisionPrecision is defined as the ratio of the number of True Positive

(correctly predicted cases) to the sum of the True Positive and theFalse Positive.

4.6.3. RecallRecall is also known as the Sensitivity or True Positive rate. It is

defined as the ratio of the True Positive (the number of correctlypredicted cases) to the sum of the True Positive and the False Negative.

4 Some researchers preferred to use the term “intangible assets” [20,63] rather than“off-balance-sheet assets”. Not all intangible assets are absent in the balance sheets. Someintangibles,whichmeet recognition criteria according to accounting standards (see IAS 38Intangible Assets), are recognized as assets on the balance sheets, while at the same timesome cannot be, such as employee skills, innovation ability and so on.


4.6.4. F-measureF-measures take the harmonic mean of the Precision and Recall

Performance measures. Therefore, it takes into consideration both thePrecision and the Recall Performance as being important measurementtools for these calculations, as shown in Eq. (3) [90].

F‐measure ¼ 2� Precision� RecallPrecisionþ Recall

ð3Þ

4.6.5. SpecificityThis is also known as the True Negative Rate (TN). It is defined as the

ratio of the number of the True Negative to the sum of the TrueNegativeand the False Positive.

4.7. Sensitivity analysis (predictor's importance)

“Cause and effect” relationship between the dependent (output) andindependent (input) variables of a predictionmodel is often determinedby “sensitivity analysis” in machine learning algorithms [25]. Sensitivityanalysis aims to measure the importance of predictor variables. It iscommonly used to identify and focus on the more important variablesand to ignore or drop the least important ones. They are related to theimportance of each variable in making a prediction, not necessarilywhether the prediction itself is accurate. The variance of predictiveerror is arrived at by dropping one predictor variable at a time, andobserving the performance of the remainder. A variable is consideredmore important than another if it increases the variance, compared tothe complete model containing all the variables [84]. Predictor impor-tance is determined by evaluating variance reduction of the targetattributable to each predictor (see Eq. (4)). Predictors are rankedaccording to the sensitivity measure defined as [85]:

Si ¼Vi

V Yð Þ ¼V EðY jXið ÞÞ

V Yð Þ ð4Þ

where Y is the target (dependent variable), Xj(j = 1,…,k) are predictors(independent variables). V(Y) is the unconditional output variance. Theexpectation operator E represents an integral over X−i; that is, over allfactors but Xi. The variance operator V denotes a further integral overXi. Predictor importance of ith variable is then computed as the normal-ized sensitivity (see Eq. (5)).

PIi ¼SiXkj¼1

Sj

ð5Þ

It is shown that Si is the proper measure of sensitivity to rank thepredictors in order of importance for any combination of interactionand non-ortogonality among predictors [81].

4.8. Information fusion-based sensitivity analysis

There are various definitions of information fusion in the literature.A substantial amount of research has been dedicated to problemsconcerning how to combine data from various sources. It is alsoknown as data fusion. “Data fusion is a process that combines data andknowledge from different sources with the aim of maximizing the use-ful information content, for improved reliability or discriminant capabil-ity, while minimizing the quantity of data ultimately retained” [86]. Inthis study, obtained predictions are the data or information, “predictionmodels” are the sources, and combining the predictions is the process offusion. Studies have shown that combining predictions (fusion) revealsmore accurate and more robust results [28,83].

Each decision tree model generated variable importance scores foreach independent variable. The combination of these predictionmodels

is called information fusion-based sensitivity analysis, and is recom-mended because it produces more accurate, robust models [36].

Each of the predictionmodels produced somewhat different predictorimportant values. An information fusion-based sensitivity analysis wasperformed to combine these values into a common representation. Therelative variable importance score produced by each decision tree modelwas normalized by using Eq. (6) below. They were then aggregated intoa single set of importance numbers and are represented in a tabularform (the normalized variable importance scores were combined usingEq. (7)) [29]. Essentially, the normalized score of each independent vari-able was multiplied by the normalized weight value for each predictionmodel and finally, these multiplied scores were added together to find asingle combined (fused) relative importance value for each variable.

PInew ¼ PI−PImin

PImax−PIminð6Þ

PIn fusedð Þ ¼ w1PI1n þw2PI2n þ…þwmPImn ð7Þ

PI represents the relative predictor importance score that wasinitially produced by the individual model.

wi represents the normalizedweight values for eachmodel. Thisrepresents the importance of models and is proportional totheir predictive powers.

M represents the number of prediction models (m = 5 in thisstudy)

N represents the number of variables (n = 15 variables in thisstudy)

These fused sensitivity scores were presented as bar-charts to visu-ally illustrate the relative importance of the independent variablesfrom the highest (most important) to the lowest (least important) forpredicting (contributing to the prediction of) the dependent variable.

5. Results

The impact of multinationality (as measured by foreign sales ratio)and fourteen other financial indicators on market capitalization andmarket-to-book ratio for the period of 1997–2011 was investigatedusing decision tree and neural network algorithms. Although a greatmany previous studies used market-to-book ratio as a proxy of the firmvalue [7,12,55,59,88,92], other studies utilized market capitalization as amarket value indicator [4,24,87]. As a result, we decided to use both var-iables as dependent variables. Market value (i.e. market capitalization) iscalculated bymultiplying the numbers of shares outstanding by the sharepricewhich represents the price investors are willing to pay to buy or sellthe stock. Thus, the market value is open to fluctuation depending uponchanges in share price which is determined in the market place. Rustet al. [80] claimed that the market value of firms depended largely ontheir growth prospects and sustainability of profits. However, bookvalue is the value of the organization as reflected in the firm's financialstatements which are prepared in accordance with accounting and/or fi-nancial reporting standards and laws. Market value and book value aredifferent in that the former is forward-looking and the latter is retrospec-tive [80]. Malighetti et al. [63] puts it succinctly “the market values thecompany as a going concern”. Rust et al. [80] explained the market-to-book gap by off-balance-sheet assets4 (i.e. market-based and intellectualproperty) and by an excess or lack of investor enthusiasm [80].


Beginning from 2005, Turkish firms were mandated to startimplementing IFRS. Thus, we divided the time period of 1997–2011into two periods; 1997–2004 and 2005–2011 to investigate the robust-ness of results pre- and post-IFRS implementation. There were fifteeninputs and two outputs for the formulation of this study. Theoutput var-iables were market capitalization and market-to-book ratio, bothrepresenting the firm value. The input variables were asset turnoverrate, assets growth rate, capital expenditure, cash conversion cycle,cash ratio, financial debt ratio, foreign sales ratio, leverage ratio,LnAssets, long-term assets ratio, net profit growth rate, net profitmargin, quick ratio, return on assets, and sales growth rate. The outputsvariables were: Market capitalization and market-to-book ratio.

The dependent variables as outputs were entered into themodels asbinary variables. These output variables represent the firm values fromtwo different perspectives. Central tendencymeasure (statistical mean)values were employed as split criterion for creating binary output vari-ables: the class with a performance score above the mean values wasrated as 1 (successful) and the class with a performance score belowthe mean values was rated as 0 (unsuccessful). In this study, the cases(successful and unsuccessful) were unbalanced. The data set was bal-anced as to be approximately 50% (successful) and 50% (unsuccessful).It is advisable to correct imbalances in datasets to conform to specifiedtest criteria [89]. Many modeling techniques have trouble withskewed/biased data (i.e., suppose that a data set has only two values:Yes and no; 90% “yes”, 10% “no”) since they tend to learn only theoutcome with high percentage (“yes”) and ignore outcome with lowpercentage (“no”). Models have a better chance of finding patternsthat distinguish the two groups if the data are well balanced withapproximately equal numbers of outcomes [85].

5.1. Prediction results for market capitalization

A total of five classification models were included: C5.0/DT, CART,QUEST/DT, CHAID/DT and Multi Layered Perceptron/Neural Networkmodels. Also, in order to obtain more accurate and more robust predic-tions, in addition to the individual models, ensemble models were alsodeveloped. The obtained prediction results for each model were basedon hold-out test data using 10-fold cross validation methodology: 10different models were trained and tested, each time using a differentmutually exclusive 10% of the total dataset as the hold-out/test sample.The testing results were combined and used for comparison of theprediction models.

Table 5 illustrates the comparison of the prediction models in termsof performances between 2005 and 2011. As the results show, the C5.0decision treemodel outperformed the other individual models in termsof overall accuracy rate with almost 91%. CHAID and CART decision treemodels performed equally well with almost 86% while Neural NetworkandQUEST performed 84% in overall accuracy rate. The ensemblemodelwas the combination of these individual models which performedbetter than almost all of the models except C5.0. It was expected thatthe ensemble model demonstrates high predictive ability compared tothe individual models. The other performance measures: sensitivity,

Table 5Prediction results for market capitalization (2005–2011).

Accuracy (AC) Sensitivity/True PositiveRate/Recall (TP)

Specificity/TrueNegative rate (TN)

C5.0 0.9074 0.8632 0.9518C&R Tree 0.8629 0.7905 0.9355QUEST 0.8349 0.7743 0.8958CHAID 0.8658 0.8664 0.8651Neural Network 0.8440 0.7931 0.8951Ensemble 0.8801 0.8392 0.9212

specificity, precision, F-measure and AUC also indicate that the C5.0decision tree model performed best, and ensemble model was thesecond highest prediction model. Even though the order of the otherprediction models was changed depending on the performancemeasure, they demonstrate consistency by being close to each otherwith around or over 80% prediction rate.

Table 6 shows the confusion (coincidence) matrix constructed fromthe test data samples. It provides overall accuracy rates as well as per-class accuracy rates for each individual models and the ensemblemodel (linear combination of the individual models). The results indi-cate that prediction accuracy for the successful class was higher thanthe prediction rate for the unsuccessful class in all individual modelsas well as in the ensemble model. Successful prediction rate was thehighest in the C5.0 model with almost 95% significant accuracy rate,while CART model predicted the successful class with almost 93% andensemble model predicted with almost 92% significant accuracy rates.Therefore, these models predicted successful companies in terms ofmarket capitalization with over 90% prediction rate. Also, C5.0, CHAID,and ensemble models predicted unsuccessful companies with almost87% and 85% accuracy rate, while the others also predicted unsuccessfulcompanies in terms of market capitalization around 80% accuracy rate.

The same five data mining models as well as the ensemble model(linear combination of the individual models) were also conducted todetermine the outcome of market capitalization between 1997 and2004. Tables 7 and 8 indicate the performance measurements and con-fusionmatrices of themodels from 1997 to 2004. It shows that C5.0wasthe best performing model, while CHAID, CART and Neural Networkwere the next leading performing models. The ensemble modelperformed very well after C5.0. These results show consistency withthose taken between 2005 and 2011. Again, C5.0 and ensemble modelspredicted successful firms with almost 90% accuracy rate, while CHAID,CART, Neural Network and QUEST models predicted successful classwith over 80% accuracy rate. The ability of the included predictionmodels to predict successful and unsuccessful firms demonstratedconsistency in both time periods: 1997–2004 and 2005–2011.

5.2. Sensitivity analysis results for market capitalization

In order to determine the relative predictor importance of the inputs(independent variables), model-specific sensitivity analysis as well asinformation fusion basedmulti-model sensitivity analysis was conduct-ed. Each of the five models created a somewhat different predictor im-portance scores. The relative predictor importance values generatedby each model initially was normalized (using Eq. (6)). The normalizedscores of each model were multiplied by weight values of each model(using Eq. (7)) and these multiplied values were then added togetherin order to find a single fused (combined) predictor importance scorefor each independent variable (see Table 9). Table 9 represents sensitiv-ity analysis values for market capitalization between 2005 and 2011. Toillustrate visual representation of the fused predictor importancevalues of independent variables in the order of importance level, a bar-chart was created using the aggregated sensitivity values (see Fig. 2).

False Positiverate (FP)

False Negativerate (FN)

Precision (P) F-measure Area undercurve (AUC)

0.0482 0.1368 0.9473 0.9033 0.9590.0645 0.2095 0.9249 0.8524 0.9150.1042 0.2257 0.8818 0.8246 0.8560.1349 0.1336 0.8658 0.8661 0.9440.1049 0.2069 0.8837 0.8360 0.9080.0788 0.1608 0.9145 0.8752 n/a

Table 6Confusion matrices of the models based on 10-fold cross validation test data (market capitalization, 2005–2011).

Model type Unsuccessful (0) Successful (1) Overall accuracy Per-class accuracy

C5.0 Unsuccessful (0) 1461 74 Correct 2792 90.74% 87.38%Successful (1) 211 1331 Wrong 285 9.26% 94.73%Sum 1672 1405 3077

C&R Tree Unsuccessful (0) 1436 99 Correct 2655 86.29% 81.64%Successful (1) 323 1219 Wrong 422 13.71% 92.49%Sum 1759 1318 3077

QUEST Unsuccessful (0) 1375 160 Correct 2569 83.49% 79.80%Successful (1) 348 1194 Wrong 508 16.51% 88.18%Sum 1723 1354 3077

CHAID Unsuccessful (0) 1.328 207 Correct 2664 86.58% 86.57%Successful (1) 206 1.336 Wrong 413 13.42% 86.58%Sum 1534 1543 3077

Neural Network Unsuccessful (0) 1374 161 Correct 2597 84.40% 81.16%Successful (1) 319 1223 Wrong 480 15.60%. 88.37%Sum 1693 1384 3077

Ensemble Unsuccessful (0) 1414 121 Correct 2708 88.01% 85.08%Successful (1) 248 1294 Wrong 369 11.99% 91.45%Sum 1662 1415 3077


The y-axis shows financial indicators as the independent variables whilex-axis shows the predictor importance score for each indicator. Accordingto Fig. 4, LnAssets was the most important predictor in determining themarket capitalization between 2005 and 2011, while leverage ratio wasthe second the most important predictor. Also, cash ratio, quick ratio,net profit margin and return on assets financial indicators were thefollowed leading variables on market capitalization. It is noteworthy torecite that LnAssets variable was the most important variable in all fiveindividual models (Table 9).

Fig. 3 illustrates the information fusion based sensitivity analysisresult of market capitalization from 1997 to 2004. As the result indicate,

Table 7Prediction results for market capitalization (1997–2004).



C5.0 0.9061 0.9016 0.9105C&R Tree 0.8138 0.7352 0.8915QUEST 0.7764 0.6768 0.8749CHAID 0.8281 0.7944 0.8614Neural Network 0.8050 0.7896 0.8203Ensemble 0.8540 0.8008 0.9066

Table 8Confusion matrices of the models based on 10-fold cross validation test data (market capitaliz

Model type Unsuccessful (0) Successful (1

C5.0 Unsuccessful (0) 1150 113Successful (1) 123 1127Sum 1273 1240

C&R Tree Unsuccessful (0) 1126 137Successful (1) 331 919Sum 1457 1056

QUEST Unsuccessful (0) 1105 158Successful (1) 404 846Sum 1509 1004

CHAID Unsuccessful (0) 1088 175Successful (1) 257 993Sum 1345 1168

Neural Network Unsuccessful (0) 1036 227Successful (1) 263 987Sum 1299 1214

Ensemble Unsuccessful (0) 1145 118Successful (1) 249 1001Sum 1394 1119

LnAssets were the most important factor, while return on assets,leverage ratio, assets growth rate, cash ratio were the followed financialcharacteristics in predicting the outcome of market capitalization. Thisresult is consistent with the result from 2005 to 2011; the variableLnAssets was the most important variable in both analyses. In bothFigs. 2 and 3, we see that firm size, as measured by LnAssets, is a domi-nant variable affecting firm value. This finding supports the empiricalevidences provided by prior studies. Although a few studies found insig-nificant relation for size and firm value [23,34], many previous studiesfound either a significant negative correlation [3,8,32,52,75] or a signif-icant positive association between the two variables [62,87,92].




0.0895 0.0984 0.9089 0.9052 0.95300.1085 0.2648 0.8703 0.7971 0.84600.1251 0.3232 0.8426 0.7507 0.81600.1386 0.2056 0.8502 0.8213 0.91200.1797 0.2104 0.8130 0.8011 0.87400.0934 0.1992 0.8945 0.8451 n/a

ation, 1997–2004).

) Overall accuracy Per-class accuracy

Correct 2277 90.61% 90.34%Wrong 236 9.39% 90.89%

2513Correct 2045 81.38% 77.28%Wrong 468 18.62% 87.03%

2513Correct 1951 77.64% 73.23%Wrong 562 22.36% 84.26%

2513Correct 2081 82.81% 80.89%Wrong 432 17.19% 85.02%

2513Correct 2023 80.50% 79.75%Wrong 490 19.50% 81.30%

2513Correct 2146 85.40% 82.14%Wrong 367 14.60% 89.45%

2513

Table 9Aggregated sensitivity analysis results of market capitalization (2005–2011).

C5.0 CHAID C&R Tree QUEST Neural Network PI (fused)

Asset turnover rate 0.0316 0.0000 0.0000 0.0107 0.0000 0.0376Assets growth rate 0.0373 0.0000 0.0160 0.0107 0.0164 0.0705Capital expenditure 0.0000 0.0000 0.1192 0.0038 0.0433 0.1426Cash conversion cycle 0.0325 0.0000 0.0160 0.0107 0.0264 0.0745Cash ratio 0.0000 0.0061 0.0586 0.0107 0.2752 0.2970Financial debt ratio 0.0941 0.0000 0.0586 0.0000 0.0358 0.1661Foreign sales ratio 0.0000 0.0053 0.0110 0.0107 0.0930 0.1015Leverage ratio 0.0178 0.0417 0.0160 0.0286 0.3164 0.3570LnAssets 1.0000 1.0000 1.0000 1.0000 1.0000 4.3149Long-term assets ratio 0.0044 0.0000 0.0160 0.0107 0.0990 0.1103Net profit growth rate 0.0160 0.0094 0.0160 0.0107 0.0621 0.0978Net profit margin 0.0103 0.0000 0.0160 0.0107 0.2859 0.2734Quick ratio 0.1334 0.0736 0.0766 0.0107 0.0178 0.2747Return on assets 0.0807 0.0213 0.0891 0.0107 0.0937 0.2566Sales growth rate 0.0124 0.0131 0.0009 0.0107 0.0261 0.0543


5.3. Prediction results for market-to-book ratio

In this experiment, market-to book ratio was chosen as the depen-dent variables. These same inputs (independent variables) were usedto determine market-to-book ratio output variable. Table 10 revealedthe performance measurements of the prediction models from 2005

Fig. 2. Illustration of 10-f

Fig. 3. Illustration of Neural Ne

to 2011. According to overall accuracy rates, C5.0 was the bestperforming model with approximately 85%, while the ensemblemodel performed with 81% rate as the second best prediction model.It is expected that the ensemble model usually produced better accura-cy rates than the individual models. CART and Neural Networks werethe next best prediction models with 74% and 73% respectively. QUEST

old cross validation.

twork used in this study.

image of Fig.�2

image of Fig.�3

Table 10Prediction results for market-to-book ratio (2005–2011).






C5.0 0.8544 0.8251 0.8842 0.1158 0.1749 0.8782 0.8508 0.9340C&R Tree 0.7418 0.6550 0.8297 0.1703 0.3450 0.7956 0.7185 0.7940QUEST 0.6964 0.5557 0.8388 0.1612 0.4443 0.7773 0.6481 0.7110CHAID 0.6797 0.7688 0.5895 0.4105 0.2312 0.6546 0.7071 0.7350Neural Network 0.7284 0.7070 0.7500 0.2500 0.2930 0.7411 0.7237 0.8130Ensemble 0.8109 0.7476 0.8750 0.1250 0.2524 0.8582 0.7991 N/A


and CHAID decision tree models had the lowest prediction rate with al-most 70% and 68% respectively. Besides overall accuracy rate perfor-mance measurement, precision, F-measure, AUC, sensitivity andspecificity also indicated that C5.0 was the outperforming predictionmodel while the ensemble model was the next best performing modelfor predictingmarket-to-book ratio between 2005 and 2011. Successfuland unsuccessful firms' predictions by themodels were indicated usingconfusion matrix in Table 11 between 2005 and 2011. As the resultsshows, C5.0 and ensemble models predicted successful firms with 88%and 86% accuracy rates respectively, they were the leading models.CART decision tree model predicted successful cases with almost 80%as the third best model, while QUEST and Neural Network predicted itwith 78% and 74% accuracy rate. CHAID was the weakest model forpredicting successful firms with 65%. According to the obtained results,unsuccessful firms were predicted very well with the C5.0 decision treemodel and Ensemblemodel with 83% and 77% accuracy rate respective-ly. Except the QUEST decision tree model, all other prediction modelspredicted the unsuccessful class with over 70% accuracy rate.

Table 11Confusion matrices of the models based on 10-fold cross validation test data (market-to-book

Model type Unsuccessful (0) Successfu

C5.0 Unsuccessful (0) 1443 189Successful (1) 289 1363Sum 1732 1552

C&R Tree Unsuccessful (0) 1354 278Successful (1) 570 1082Sum 1924 1360

QUEST Unsuccessful (0) 1369 263Successful (1) 734 918Sum 2103 1181

CHAID Unsuccessful (0) 962 670Successful (1) 382 1270Sum 1344 1940

Neural Network Unsuccessful (0) 1224 408Successful (1) 484 1168Sum 1708 1576

Ensemble Unsuccessful (0) 1428 204Successful (1) 417 1235Sum 1845 1439

Table 12Prediction results for market-to-book ratio (1997–2004).

Accuracy (AC) Sensitivity/TruePositive Rate/Recall (TP)


Fa

C5.0 0.7998 0.7027 0.8952 0.1C&R Tree 0.7120 0.6912 0.7325 0.2QUEST 0.6767 0.5725 0.7792 0.2CHAID 0.7348 0.7305 0.7389 0.2Neural Network 0.6901 0.6446 0.7349 0.2Ensemble 0.7766 0.7805 0.7728 0.2

Prediction models' performance measurements as well as per-classaccuracy rates of successful and unsuccessful firms' performance resultsare shown in Tables 12 and 13 for the period 1997–2004. The obtainedresults reveal strong consistency with the result from 2005 to 2011.Again, C5.0 and the ensemble models were the outperformed modelswith approximately 80% and 78% overall accuracy rates in the ordergiven. C5.0 and ensemble models outperformed in sensitivity, specifici-ty, precision and F-measure performance measures. The same modelsdemonstrated a strong predictive ability in previous experiments aswell. In terms of per-class accuracy rates, C5.0 and the ensemblemodelspredicted successful firmswith 87% and 77% respectively, while the restof the models revealed strong accuracy rates with above 70% accuracyrates. For unsuccessful firms' prediction, the ensemble and C5.0 modelsdemonstrated high predictive results with 78% and 75% accuracy rates.While CHAID and CART decision tree models predicted unsuccessfulfirms with above 70% accuracy rates, Neural Network and QUESTmodels showed weak prediction results with 68% and 65% accuracyrates in terms of market-to-book ratio.

ratio, 2005–2011).

l (1) Overall accuracy Per-class accuracy

Correct 2806 85.44% 83.31%Wrong 478 14.56% 87.82%

3284Correct 2436 74.18% 70.37%Wrong 848 25.82% 79.56%

3284Correct 2287 69.64% 65.10%Wrong 997 30.36% 77.73%

3284Correct 2232 67.97% 71.58%Wrong 1052 32.03% 65.46%

3284Correct 2392 72.84% 71.66%Wrong 892 27.16% 74.11%

3284Correct 2663 81.09% 77.40%Wrong 621 18.91% 85.82%

3284

lse Positive rate (FP) False Negativerate (FN)


048 0.2973 0.8684 0.7768 0.8580675 0.3088 0.7177 0.7042 0.7650208 0.4275 0.7184 0.6372 0.7000611 0.2695 0.7336 0.7320 0.8270651 0.3554 0.7052 0.6735 0.7280272 0.2195 0.7717 0.7761 n/a

Table 13Confusion matrices of the models based on 10-fold cross validation test data (market-to-book ratio, 1997–2004).

Model type Unsuccessful (0) Successful (1) Overall accuracy Per-class accuracy

C5.0 Unsuccessful (0) 1111 130 Correct 1969 79.98% 75.37%Successful (1) 363 858 Wrong 493 20.02% 86.84%Sum 1474 988 2462

C&R Tree Unsuccessful (0) 909 332 Correct 1753 71.20% 70.68%Successful (1) 377 844 Wrong 709 28.80% 71.77%Sum 1286 1176 2462

QUEST Unsuccessful (0) 967 274 Correct 1666 67.67% 64.94%Successful (1) 522 699 Wrong 796 32.33% 71.84%Sum 1489 973 2462

CHAID Unsuccessful (0) 917 324 Correct 1809 73.48% 73.60%Successful (1) 329 892 Wrong 653 26.52% 73.36%Sum 1246 1216 2462

Neural Network Unsuccessful (0) 912 329 Correct 1699 69.01% 67.76%Successful (1) 434 787 Wrong 763 30.99% 70.52%Sum 1346 1116 2462

Ensemble Unsuccessful (0) 959 282 Correct 1912 77.66% 78.16%Successful (1) 268 953 Wrong 550 22.34% 77.17%Sum 1227 1235 2462


5.4. Sensitivity analysis results for market-to-book ratio

The same procedure for sensitivity analysis was employed formarket-to-book ratio dependent variable. The obtained relative impor-tance values were normalized and then they were fused using informa-tion fusion-based sensitivity analysis order to combine these values as asingle predictor importance value in both time intervals: 1997–2004and 2005–2011. These values aggregated into a tabular form and thenthey were illustrated in bar-charts. Table 14 and Fig. 4 represent sensi-tivity analysis results from 2005 to 2011 for prediction market-to-book ratio. As results indicate, leverage ratio was the most importantfactor in all models; therefore, it is the most significant financial ratiofor prediction market-to-book ratio. Also, return on assets, LnAssets,and quick ratio were the followed most important predictors in orderfor determining market-to-book ratio. Assets growth rate, foreignsales ratio, and sales growth rate were the least important predictorsfor market-to-book ratio (Figs. 6 and 7).

Table 15 and Fig. 5 show the sensitivity analysis results between1997 and 2004 for predicting market-to-book ratio. After obtainingthe linear combination (fused) scores for each predictor of each modelin the table, the values were illustrated on a bar-chart. According tothe fused sensitivity analysis results, leverage ratiowas themost impor-tant predictor, while asset turnover rate, LnAssets, and Return on assets

Table 14Aggregated sensitivity analysis results of market-to-book ratio (2005-2011).

C5.0 CHAID C&R T

Asset turnover rate 0.0000 0.0647 0.000Assets growth rate 0.0019 0.0382 0.000Capital expenditure 0.0000 0.1204 0.112Cash conversion cycle 0.4007 0.0479 0.000Cash ratio 0.0864 0.0929 0.000Financial debt ratio 0.3405 0.0000 0.000Foreign sales ratio 0.0355 0.0280 0.000Leverage ratio 1.0000 1.0000 1.000LnAssets 0.3529 0.4271 0.209Long-term assets ratio 0.1928 0.0000 0.000Net profit growth rate 0.2684 0.2002 0.015Net profit margin 0.0000 0.0000 0.045Quick ratio 0.5519 0.0382 0.000Return on assets 0.5426 0.3009 0.458Sales growth rate 0.0841 0.0739 0.000

were the next important variables for determining market-to-bookratio. These indicate consistency as well with the sensitivity result be-tween 2005 and 2011. In both time intervals, leverage ratio came outas the most important predictor.

6. Conclusion and implications

The impact of multinationality (as measured by foreign sales ratio)and fourteen other financial indicators on market capitalization andmarket-to-book ratio for the period of 1997–2011 was investigatedusing decision tree and neural network algorithms. We dividedthe time period of 1997–2011 into two periods; 1997–2004 and2005–2011 to check the robustness of results pre- and post-IFRSimplementation.

The influence of multinationality and other financial indicators wereinvestigated on market capitalization for the period of 2005–2011,which also covers IFRS implementation period. We found that themost important predictor on market capitalization was firm size asmeasured by natural logarithm of assets. This variable was significantlydistinguished from other financial indicators. The other variables whichhave been found to have explanatory power on market capitalizationwere leverage, liquidity (i.e. cash ratio and quick ratio), and profitability(i.e. net profit margin and return on assets). Among the independent

ree QUEST Neural Network PI (fused)

0 0.0488 0.3703 0.34770 0.0488 0.0000 0.06160 0.1357 0.1407 0.36190 0.0488 0.2623 0.60000 0.0488 0.2488 0.35220 0.0488 0.0362 0.35140 0.0488 0.1294 0.17750 1.0000 1.0000 3.70079 0.0243 0.6965 1.27180 0.0000 0.2814 0.36981 0.0488 0.0299 0.43238 0.0488 0.5814 0.49140 0.0000 0.5423 0.89257 0.4260 0.8607 1.93200 0.0488 0.0931 0.2239

0.0376

0.0543

0.0705

0.0745

0.0978

0.1015

0.1103

0.1426

0.1661

0.2566

0.2734

0.2747

0.2970

0.3570

4.3149

Asset Turnover Rate

Sales Growth Rate

Assets Growth Rate

Cash Conversion Cycle

Net Profit Growth Rate

Foreign Sales Ratio

Long-Term Assets Ratio

Capital Expenditure

Financial Debt Ratio

Return on Assets

Net Profit Margin

Quick Ratio

Cash Ratio

Leverage Ratio

LnAssets

Fig. 4. Graphical representation of sensitivity analysis result for market capitalization (2005–2011).


variables included in the study, multinationality was found to deter-mine market value only moderately. When the analysis was conductedpre-IFRS implementation for the period of 1997–2004, as the resultsindicate, firm size was again the leading predictor of market value.Other variables that were important predictor for this periodwere prof-itability (i.e. net profit margin and return on assets), leverage, liquidity(i.e. cash ratio), and asset growth rate. In this period, multinationalitywas among the least important variables. Thus, based on the limits ofthe study, we may deduce that the obtained results were robust andconsistent for the two periods even though there were minor differ-ences. These results, based on the usage of market capitalization as adependent variable, do not provide implications regarding growthopportunities for firms. They basically reveal that the larger the firmsize, as measured by sales revenues, the higher the market value.

Furthermore, we scaled market value by book value (market-to-book ratio) to investigate the impact of multinationality and other

0.0132

0.0366

0.0511

0.0925

0.1119

0.1391

0.1505

0.1549

0.1940

0.1982

0.3121

0.3294

0.4311

0.5546

Sales Growth Rate


Foreign Sales Ratio


Capital Expenditure

Asset Turnover Rate

Quick Ratio



Net Profit Margin

Cash Ratio

Assets Growth Rate

Leverage Ratio

Return on Assets

LnAssets

Fig. 5. Graphical representation of sensitivity analys

financial characteristics of firms on market value relative to bookvalue. We have found that leverage ratio was the most important indi-cator in both periods 1997–2004 and 2005–2011. Moreover, firm size,profitability, and liquidity were the next leading indicators for market-to-book ratio in both periods 1997–2004 and 2005–2011.

The results have important capital market implications. Internation-al activities of Turkish firms do not have asmuch importance as expect-ed on the firm value. In other words, investors may choose to considersome other firm characteristics rather than multinationality, and there-fore they may have a greater impact on share prices and eventually onthe firm value. As pointed out in the literature part, there is no consen-sus on the effect of multinationality on firm value; while some studiesproved that it increases firm value, whereas some others proved thatit reduces firm value. If market-to-book ratio is considered to be anindicator of growth opportunities, we can say that multinationalityhas little effect on firm's growth opportunities. Therefore, our finding

4.1293

is result for market capitalization (1997–2004).

image of Fig.�5

0.0616

0.1775

0.2239

0.3477

0.3514

0.3522

0.3619

0.3698

0.4323

0.4914

0.6000

0.8925

1.2718

1.9320

3.7007

Assets Growth Rate

Foreign Sales Ratio

Sales Growth Rate

Asset Turnover Rate


Cash Ratio

Capital Expenditure



Net Profit Margin


Quick Ratio

LnAssets

Return on Assets

Leverage Ratio

Fig. 6. Graphical representation of sensitivity analysis result for market-to-book ratio (2005–2011).


corroborates the arguments and findings of those prior studies that wecould not find a strong support for the effect of multinationality on firmvalue. Other than multinationality, some financial characteristics wereconsistently found to affect bothfirmvalue variables (market capitaliza-tion and market-to-book ratio), and in both periods (i.e., 1997–2004and 2005–2011). Our research yielded that the similar predictive vari-ables, such as firm size, profitability, and leverage, are playing a leadingrole on market-to-book ratio before and after IFRS adoption. Firm sizehas been found as an important indicator of firm value. It might bedue to the number of shares issued by firms and/or price of their shares.In addition, size influences the growth expectations of firms. Profitabil-ity is one of the most important variables for investment decisions.Investors consider it while buying, holding or selling shares of a firm,and also analysts follow profitability of firms closely while making

0.2205

0.2742

0.3138

0.3295

0.4072

0.4388

0.5242

0.5654

0.5768

0.6167

0.77

Sales Growth Rate



Net Profit Margin

Capital Expenditure

Quick Ratio

Cash Ratio


Foreign Sales Ratio

Assets Growth Rate


Return on Assets

LnAssets

Asset Turnover Rate

Leverage Ratio

Fig. 7. Graphical representation of sensitivity analys

investment recommendations to and decision on behalf of their cus-tomers. Both profitability and leverage are two significant variablesthat play important roles in terms of the growth opportunities offirms, since both are essential sources of funding. Profitability is aninternal source of funding generated as a result of operations, whileleverage is an external source of funding applied in instances wherethe internal source of financing is insufficient to finance growth. More-over, debt paying ability of a firm is important for its investors since itshows how a company is able to keep itself away from financial riskand bankruptcy. Debt level impacts profitability of a company due tothe interest burden that they have to face from borrowing. Thus, liquid-ity and leverage are two important determinants of firm value. Whilethe former measures short-term debt paying ability of a firm, the lattermeasures long-term solvency of a firm.

68

0.9836

1.4540

2.3497

2.5884

is result for market-to-book ratio (1997–2004).

Table 15Aggregated sensitivity analysis results of market-to-book ratio (1997–2004).

C5.0 CHAID C&R Tree QUEST Neural Network PI (fused)

Asset turnover rate 1.0000 0.5534 1.0000 0.1776 0.4509 2.3497Assets growth rate 0.0784 0.0847 0.0794 0.0000 0.6308 0.6167Capital expenditure 0.0202 0.0121 0.0794 0.0000 0.4718 0.4072Cash conversion cycle 0.0405 0.3215 0.2034 0.0000 0.5265 0.7768Cash ratio 0.2181 0.1886 0.0794 0.0000 0.2242 0.5242Financial debt ratio 0.0000 0.0495 0.0794 0.2678 0.0000 0.2742Foreign sales ratio 0.3426 0.0000 0.3259 0.0000 0.1025 0.5768Leverage ratio 0.4845 0.5061 0.6492 1.0000 1.0000 2.5884LnAssets 0.4902 0.2179 0.6166 0.0709 0.6012 1.4540Long-term assets ratio 0.0392 0.0000 0.3909 0.0456 0.3258 0.5654Net profit growth rate 0.1621 0.0000 0.0059 0.0000 0.2606 0.3138Net profit margin 0.0278 0.1407 0.0000 0.0000 0.2954 0.3295Quick ratio 0.1157 0.0377 0.0794 0.0000 0.3797 0.4388Return on assets 0.1391 1.0000 0.0434 0.0451 0.1103 0.9836Sales growth rate 0.0025 0.0000 0.0794 0.0000 0.2346 0.2205


References

[1] F.A. Alali, P.S. Foote, The value relevance of internationalfinancial reporting standards:empirical evidence in an emerging market, Int. J. Account. 47 (2012) 85–108.

[2] E. Alfaro, N. García, M. Gámez, D. Elizondo, Bankruptcy forecasting: an empirical com-parison of AdaBoost and neural networks, Decis. Support. Syst. 45 (1) (2008) 110–122.

[3] M. Ammanna, D. Oesch, M.M. Schmid, Corporate governance and firm value: inter-national evidence, J. Empir. Finance 18 (2011) 36–55.

[4] O.A. Anam, A.H. Fatima, A.R.H. Majdi, Effects of intellectual capital informationdisclosed in annual reports on market capitalization: evidence from BursaMalaysia, J. Hum. Resour. Costing Acc. 15 (2) (2011) 85–101.

[5] M. Anandarajan, P. Lee, A. Anandarajan, Bankruptcy prediction of financiallystressed firms: an examination of the predictive accuracy of artificial neuralnetworks, Int. J. Intell. Syst. Account. Finance Manag. 10 (2001) 69–81.

[6] A. Atiya, Bankruptcy prediction for credit risk using neural networks: a survey andnew results, IEEE Trans. Neural Networks 12 (4) (2001).

[7] S.C. Bae, S. Noh, Multinational corporations versus domestic corporations: acomparative study of R&D investment activities, J. Multinatl. Financ. Manag. 11(2001) 89–104.

[8] S.C. Bae, T.H. Kwon, J.W. Lee, Does corporate diversification by business groups cre-ate value? Evidence from Korean chaebols, Pac. Basin Financ. J. 19 (2011) 535–553.

[9] R. Barniv, A. Agarwal, R. Leach, Predicting the outcome following bankruptcy filing:a three-state classification using neural networks, Intell. Syst. Acc. Financ. Manag. 6(1997) 177–194.

[10] T. Bell, Neural nets or the logit model? A comparison of eachmodel's ability to pre-dict commercial bank failures, Intell. Syst. Acc. Financ. Manag. 6 (1997) 249–264.

[11] J. Berrill, G. Mannella, Are firms from developed markets more international thanfirms from emerging markets? Research in International Business and Fi-nance2012. (http://dx.doi.org/10.1016/j.ribaf.2012.04.002).

[12] B. Black, W. Kim, The effect of board structure on firm value: a multiple identifica-tion strategies approach using Korean data, J. Financ. Econ. 104 (2012) 203–226.

[13] B.S. Black, I. Love, A. Rachinsky, Corporate governance indices and firms' marketvalues: time series evidence from Russia, Emerg. Mark. Rev. 7 (2006) 361–379.

[14] J. Boritz, D. Kennedy, Effectiveness of neural networks types for prediction of busi-ness failure, Expert Syst. Appl. 9 (1995) 503–512.

[15] L. Breiman, J.H. Friedman, R.A. Olshen, C.J. Stone, Classification and RegressionTrees, Chapman & Hall/CRC, New York, 1984.

[16] G. Cassar, S. Holmes, Capital structure and financing of SMEs: Australian evidence,Account. Finance 43 (2003) 123–147.

[17] R.E. Caves, International corporations: the industrial economics of foreign invest-ment, Economica 38 (149) (1971) 1–27.

[18] W.Chaehwan,K. Jinhwa,K.B. Jae,Usinggenetic algorithmbasedknowledge refinementmodel for dividend policy forecasting, Expert Syst. Appl. 39 (2012) 13472–13479.

[19] C. Charalambous, A. Charitous, F. Kaourou, Comparative analysis of artificial neural net-workmodels: application inbankruptcyprediction,Ann.Oper. Res. 99 (2000)403–425.

[20] R. Charles, C.R. Hulten, X. Hao,What is a company reallyworth? Intangible capital andthe “market to book value” puzzle, NBER Working Paper Series, Working Paper145482008. (Retrieved from http://www.nber.org/papers/w14548 on 08.01.2013).

[21] M.C. Chen, S.H. Huang, Credit scoring and rejected instances reassigning throughevolutionary computation techniques, Expert Syst. Appl. 24 (4) (2003) 433–441.

[22] T. Chih-Fong, L. Yu-Hsin, C.Y. David, Determinants of intangible assets value: thedata mining approach, Knowl.-Based Syst. 31 (2012) 67–77.

[23] J.T. Connelly, P. Limpaphayom, N.J. Nagarajan, Form versus substance: the effect ofownership structure and corporate governance on firm value in Thailand, J. Bank.Financ. 36 (2012) 1722–1743.

[24] M. Darrough, J. Ye, Valuation of loss firms in a knowledge-based economy, Rev.Acc. Stud. 12 (2007) 61–93.

[25] G. Davis, Sensitivity analysis in neural net solutions, IEEE Trans. Syst. Man Cybern.19 (1989) 1078–1082.

[26] B.E. Deakin, A discriminant analysis of predictors of business failure, J. Account. Res.167–179 (1976).

[27] H. DeAngelo, R. Masulis, Optimal capital structure under corporate and personaltaxation, J. Financ. Econ. 8 (1980) 5–29.

[28] D. Delen, A comparative analysis of machine learning techniques for student reten-tion management, Decis. Support. Syst. 49 (2010) 498–506.

[29] D. Delen, A. Oztekin, L. Tomak, An analytic approach to better understanding andmanagement of coronary surgeries, Decis. Support. Syst. 52 (2012) 698–705.

[30] D.J. Denis, D.K. Denis, K. Yost, Global diversification, industrial diversification, andfirm value, J. Financ. 57 (5) (2002) 1951–1979.

[31] S. Eckert, M. Dittfeld, T. Muche, S. Rässler, Does multinationality lead to value en-hancement? An empirical examination of publicly listed corporations fromGermany, Int. Bus. Rev. 19 (2010) 562–574.

[32] J. Erickson, Y.W. Park, J. Reising, H.-H. Shin, Board composition and firm valueunder concentrated ownership: the Canadian evidence, Pac. Basin Financ. J. 13(2005) 387–410.

[33] V.R. Errunza, L.W. Senbet, The effects of international operations on the marketvalue of the firm: theory and evidence, J. Financ. 36 (2) (1981) 401–417.

[34] O. Faleye, Classified boards, firm value, and managerial entrenchment, J. Financ.Econ. 83 (2007) 501–529.

[35] L. Fauver, J.F. Houston, A. Naranjo, Cross-country evidence on the value of corporateindustrial and international diversification, J. Corp. Finance 10 (2004) 729–752.

[36] C.M. Fuller, D.P. Biros, D. Delen, An investigation of data and text mining methodsfor real world deception detection, Expert Syst. Appl. 38 (2011) 8392–8398.

[37] S.J. Grice, T.M. Dugan, The limitations of bankruptcy prediction models: some cau-tions for the researcher, Rev. Quant. Finan. Acc. 17 (2001) 151–166.

[38] F.A. Gul, J. Tsui, D.S. Dhaliwal, Non-audit services, auditor quality and the value rel-evance of earnings, Account. Finance 46 (2006) 797–817.

[39] O.A.G. Hassan, G. Giorgioni, P. Romilly, D. Power, The value-relevance of disclosure: ev-idence from theemerging capitalmarket of Egypt, Int. J. Account. 44 (1) (2009)79–102.

[40] S. Haykin, Neural Networks and Learning Machines, 3rd ed. Prentice Hall Publish-ing, Englewood Cliffs, New Jersey, 2008.

[41] T. Hiraki, H. Inoue, A. Ito, F. Kuroki, H.Masuda, Corporate governance andfirmvalue inJapan: evidence from 1985 to 1998, Pac. Basin Financ. J. 11 (2003) 239–265.

[42] C.W. Holsapple, J. Wu, An elusive antecedent of superior firm performance: theknowledge management factor, Decis. Support. Syst. 52 (1) (2011) 271–283.

[43] Z. Huang, H. Chen, C.-J. Hsu, Credit rating analysis with support vector machine andneural networks: amarket comparative study,Decis. Support. Syst. 37 (2004) 543–558.

[44] M.C. Jensen, Agency costs of free cash flow, corporate finance, and takeovers, Am.Econ. Rev. 76 (2) (1986) 323–329.

[45] M.C. Jensen, W.H. Meckling, Theory of the firm: managerial behavior, agency costsand ownership structure, J. Financ. Econ. 3 (4) (1976) 305–360.

[46] C.-H. Jiang, H.-L. Chen, Y.-S. Huang, Capital expenditures and corporate earnings: evi-dence from the Taiwan Stock Exchange, Manag. Financ. 32 (11) (2006) 853–861.

[47] G. Kass, An exploratory technique for investigating large quantities of categoricaldata, Appl. Stat. 29 (2) (1980) 119–127.

[48] K.J. Kim, Financial time series forecasting using support vector machines,Neurocomputing 55 (1/2) (2003) 307–319.

[49] Y.S. Kim, I. Mathur, The impact of geographic diversification on firm performance,Int. Rev. Financ. Anal. 17 (2008) 747–766.

[50] R. Kohavi, F. Provost, Glossary of terms. Editorial for the special issue on applica-tions of machine learning and the knowledge discovery process, Mach. Learn. 30(2–3) (1998) 271–274.

[51] R. Kohavi, A study of cross-validation and bootstrap for accuracy estimation andmodel selection, The proceedings of the 14th International Conference onAI(IJCAI), Morgan Kaufmann, San Mateo, CA, 1995. 1137–1145.

[52] S.J.J. Konijn, R. Kräussl, A. Lucas, Blockholder dispersion and firm value, J. Corp.Finance 17 (2011) 1330–1339.

[53] K. Kumar, S. Bhattacharya, Artificial neural networks vs. linear discriminant analy-sis in credit ratings forecast, Rev. Acc. Finance 5 (3) (2006) 216–227.

[54] P.R. Kumar, V. Ravi, Bankruptcy prediction in banks and firms via statistical andintelligent techniques — a review, Eur. J. Oper. Res. 180 (1) (2007) 1–28.

[55] Y. Kusnadi, Do corporate governance mechanisms matter for cash holdings andfirm value? Pac. Basin Financ. J. 19 (2011) 554–570.

[56] P. Langley, Elements of Machine Learning, Morgan Kaufmann, San Francisco, 1996.[57] K.C. Lee, I. Han, Y. Kwon, Hybrid neural network models for bankruptcy predic-

tions, Decis. Support. Syst. 18 (1) (1996) 63–72.

http://refhub.elsevier.com/S0167-9236(13)00261-3/rf0005








































http://www.nber.org/papers/w14548
















































































[58] K.T. Lee, C.-W. Hooy, G.-K. Hooy, The value impact of international and industrialdiversifications on public-listed firms in Malaysia, Emerg. Mark. Rev. 13 (3)(2012) 366–380.

[59] R. Lensink, R. van der Molen, Does group affiliation increase firm value for diversifiedgroups? New evidence from Indian companies, J. Empir. Finance 17 (2010) 332–344.

[60] J. Lo'pez-Gracia, F. Sogorb-Mira, Testing trade-off and pecking order theoriesfinancing SMEs, Small Bus. Econ. 31 (2) (2008) 117–136.

[61] W.Y. Loh, Y.S. Shih, Split selection methods for classification trees, Stat. Sin. 7(1997) 815–840.

[62] Y.T. Mak, Y. Kusnadi, Size really matters: further evidence on the negative relation-ship between board size and firm value, Pac. Basin Financ. J. 13 (2005) 301–318.

[63] P. Malighetti, M. Meoli, S. Paleari, R. Redondi, Value determinants in the aviationindustry, Transp. Res. E 47 (2011) 359–370.

[64] A. Martín-Oliver, V. Salas-Fumás, IT assets, organization capital and market power:contributions to business value, Decis. Support. Syst. 52 (3) (2012) 612–623.

[65] A. Mínguez-Vera, J.F. Martín-Ugedo, Does ownership structure affect value? Apanel data analysis for the Spanish market, Int. Rev. Financ. Anal. 16 (2007) 81–98.

[66] T. Mitchell, Machine Learning, McGraw Hill, New York, 1997.[67] R.Morck, B. Yeung,Why investors valuemultinationality, J. Bus. 64 (2) (1991)165–187.[68] S. Mukherjee, E. Osuna, F. Girosi, Nonlinear prediction of chaotic time series using

support vector, Proceedings of the IEEE Workshop on Neural Networks for SignalProcessing, IEEE Publishing, Amelia Island, FL, 1997, pp. 511–520.

[69] S.C. Myers, The capital structure puzzle, J. Financ. 39 (1984) 575–592.[70] S.C. Myers, R.G. Rajan, The paradox of liquidity, Q. J. Econ. 113 (3) (1998) 733–771.[71] B. Olsen, B. Elango, Do multinational operations influence firm value? Evidence

from the triad regions, Int. J. Bus. Econ. 4 (1) (2005) 11–29.[72] D.L. Olson, D. Delen, Y. Meng, Comparative analysis of data mining methods for

bankruptcy prediction, Decis. Support. Syst. 52 (2) (2012) 464–473.[73] D.L. Olson, D. Delen, Advanced Data Mining Techniques, Springer, 2008.[74] P.C. Pendharkar, A threshold-varying artificial neural network approach for classi-

fication and its application to bankruptcy prediction problem, Comput. Oper. Res.32 (10) (2005) 2561–2582.

[75] B. Pramborg, Derivatives hedging, geographical diversification, and firm marketvalue, J. Multinatl. Financ. Manag. 14 (2004) 117–133.

[76] J. Quinlan, C4.5: Programs for Machine Learning, Morgan Kaufmann, San Mateo,CA, 1993. (MA: MIT Press).

[77] A. Riahi-Belkaoui, The effects of the degree of internationalization on firm perfor-mance, Int. Bus. Rev. 7 (1998) 315–321.

[78] A. Riahi-Belkaoui, The degree of internationalization and the value of the firm:theory and evidence, J. Int. Account. Audit. Tax. 8 (1) (1999) 189–196.

[79] D.E. Rumelhart, J.L. McClelland, PDP research group, Parallel distributed processing:explorations in themicrostructure of cognition, vol. IMIT Press, Cambridge, MA, 1986.

[80] R.T. Rust, T. Ambler, G.S. Carpenter, V. Kumar, R.K. Srivastava, Measuringmarketingproductivity: current knowledge and future directions, J. Mark. 68 (2004) 76–89.

[81] A. Saltelli, S. Tarantola, F. Campolongo, M. Ratto, Sensitivity Analysis in Practice— AGuide to Assessing Scientific Models, JohnWiley, 2004.

[82] M.M. Schmid, I. Walter, Geographic diversification and firm value in the financialservices industry, J. Empir. Finance 19 (2012) 109–122.

[83] G. Seni, J. Elder, R. Grossman, EnsembleMethods in DataMining: Improving Accuracythrough Combining Predictions, Morgan and Claypool Publishers, San Rafael, C., 2010

[84] SPSS, Clementine12 User Manual, Chicago, IL. 2007.[85] SPSS, IBM SPSS Modeler User Manual, Chicago, IL, 2012.[86] A. Starr, M. Desforges, Strategies in data fusion— sorting through the tool box, Pro-

ceedings of European Conference on Data Fusion, 1998.[87] A. Uyar, M. Kılıç, Value relevance of voluntary disclosure: evidence from Turkish

firms, J. Intellect. Cap. 13 (3) (2012) 363–376.[88] B. Villalonga, R. Amit, How do family ownership, control, and management affect

firm value? J. Financ. Econ. 80 (2006) 385–417.[89] R.L. Wilson, R. Sharda, Bankruptcy prediction using neural networks, Decis.

Support. Syst. 11 (5) (1994) 545–557.[90] I.H. Witten, E. Frank, Data Mining: Practical Machine Learning Tools and

Techniques, second ed. Elsevier, San Francisco, 2005.[91] H. Wu, N. Fargher, S. Wright, Accounting for investments and the relevance of

losses to firm value, Int. J. Account. 45 (2010) 104–127.[92] H.-L. Wu, Can minority state ownership influence firm value? Universal and con-

tingency views of its governance effects, J. Bus. Res. 64 (2011) 839–845.[93] P. Ravisankar, V. Ravi, G. Raghava Rao, I. Bose, Detection of financial statement

fraud and feature selection using data mining techniques, Decis. Support. Syst.50 (2011) 491–500.

[94] E.W.T. Ngai, Yong Hu, Y.H. Wong, Yijun Chen, Xin Sun, The application of data min-ing techniques in financial fraud detection: a classification framework and an aca-demic review of literature, Decis. Support. Syst. 50 (2011) 559–569.

[95] Wei Zhou, Gaurav Kapoor, Detecting evolutionary financial statement fraud, Decis.Support. Syst. 50 (2011) 570–575.

[96] Sean L. Humpherys, Kevin C. Moffitt, Mary B. Burns, Judee K. Burgoon, William F.Felix, Identification of fraudulent financial statements using linguistic credibilityanalysis, Decis. Support. Syst. 50 (2011) 585–594.

[97] Fletcher H. Glancy, Surya B. Yadav, A computational model for financial reportingfraud detection, Decis. Support. Syst. 50 (2011) 595–601.

[98] Rich Caruana, Alexandru Niculescu-Mizil, An empirical comparison of supervisedlearning algorithms, Proceedings of the 23rd international Conference on Machinelearning (ICML '06), ACM, New York, NY, USA, 2006, pp. 161–168.

[99] Rich Caruana, Nikos Karampatziakis, Ainur Yessenalina, An empirical evaluation ofsupervised learning in high dimensions, Proceedings of the 25th International

Conference on Machine learning (ICML '08), ACM, New York, NY, USA, 2008,pp. 96–103.

[100] Eric Bauer, Ron Kohavi, An empirical comparison of voting classification algorithms:bagging, boosting, and variants, Mach. Learn. 36 (1–2) (July 1999) 105–139.

[101] ThomasG.Dietterich, Dietterich, approximate statistical tests for comparing supervisedclassification learning algorithms, Neural Comput. 10 (7) (Oct. 1998) 1895–1923.

[102] L. Breiman, Statistical modeling: the two cultures, Stat. Sci. 16 (3) (2001) 199–231.[103] Y. Benjamini, M. Leshno, Statistical methods for data mining, in: O. Maimon, L.

Rokach (Eds.), Data Mining and Knowledge Discovery Handbook, 2nd ed., Spring-er, New York, 2010, pp. 523–541.

[104] J.H. Friedman, Data mining and statistics: what's the connection? Comput. Sci. Stat.29 (1) (1998) 3–9.

[105] P. Giudici, Data mining model comparison, in: O. Maimon, L. Rokach (Eds.), DataMining and Knowledge Discovery Handbook, 2nd ed., Springer, New York, 2010,pp. 641–655.

[106] C. Vercellis, Business Intelligence: Data Mining and Optimization for Decision Mak-ing, John Wiley, U.K., 2009.

Dr. Cemil Kuzey is an Assistant Professor in the Departmentof Management at Fatih University in Istanbul, Turkey.Dr. Kuzey is primarily teaching Operation Research andStatistics topics for Social Sciences. He acquired his Ph.D.degree in Business Administration in the Department ofQuantitative Analysis at Istanbul University, Turkey. Amonghis academic pursuits, he took several graduate coursesat the Ontario Institute for Studies in Education, Universityof Toronto, Canada. His research interests are related toOperation Research, Data Mining, and Business Intelligence.

Dr. Ali Uyar is an Associate Professor of Accounting andFinance in the Department of Business Administration atFatih University in Istanbul, Turkey. He received his Ph.D. inAccounting and Finance from Marmara University, Turkeyin 2007. He teaches cost accounting, managerial accounting,and financial accounting. His research interests includeman-agement accounting practices and corporate reporting. Hisresearch papers have been published in various internationaljournals, such as Managerial Auditing Journal, InternationalJournal of Hospitality Management, Journal of IntellectualCapital, Pacific Accounting Review, International Journalof Accounting, Auditing and Performance Evaluation, Inter-national Journal of Quality and Reliability Management,
African Journal of Business Management, Business and
Economics Research Journal, Eurasian Journal of Business and Economics, InternationalResearch Journal of Finance and Economics, The TQM Journal.

Dr. Dursun Delen is the William S. Spears Chair in BusinessAdministration and Professor of Management Science andInformation Systems in the Spears School of Business atOklahoma State University (OSU). He received his Ph.D. inIndustrial Engineering and Management from OSU in 1997.Prior to his appointment as an Assistant Professor at OSU in2001, he worked for a privately-owned research and consul-tancy company, Knowledge Based Systems Inc., in CollegeStation, Texas, as a research scientist for five years, duringwhich he led a number of decision support and other infor-mation systems related research projects funded by federalagencies, including DoD, NASA, NIST and DOE. His researchhas appeared in major journals including Decision Support
Systems, Decision Sciences, Communications of the ACM,
Computers and Operations Research, Computers in Industry, Journal of Production Oper-ations Management, Artificial Intelligence in Medicine, Expert Systems with Applications,among others. He recently published four books: Advanced Data Mining Techniques withSpringer, 2008; Decision Support and Business Intelligence Systems with Prentice Hall,2010; Business Intelligence: A Managerial Approach, with Prentice Hall, 2010; andPractical Text Mining and Statistical Analysis for Non-structured Text Data Applications,with Elsevier, 2012. He is often invited to national and international conferences forkeynote addresses on topics related to Data/Text Mining, Business Intelligence, DecisionSupport Systems, and Knowledge Management. He served as the general co-chair forthe 4th International Conference on Network Computing and Advanced InformationManagement (September 2–4, 2008 in Soul, South Korea), and regularly chairs tracksand mini-tracks at various information systems conferences. He is the associateeditor-in-chief for International Journal of Experimental Algorithms, associate editor forInternational Journal of RF Technologies and Decision Analytics, and is on editorial boardsof six other academic and technical journals. His research and teaching interests are indataand textmining, decision support systems, knowledgemanagement, business intelligenceand enterprise modeling.






































































































Unlabelled image

Unlabelled image

the impact of multinationality on firm value: a comparative analysis of machine learning techniques

Documents