Determinants of Capital Structure in the UK Retail Industry

download Determinants of Capital Structure in the UK Retail Industry

of 20

description

Determinants of Capital Structure in the UK Retail Industry

Transcript of Determinants of Capital Structure in the UK Retail Industry

  • Seediscussions,stats,andauthorprofilesforthispublicationat:https://www.researchgate.net/publication/264331052

    DeterminantsofcapitalstructureintheUKretailindustry:AcomparisonofmultipleregressionandgeneralizedregressionneuralnetworkARTICLEinINTELLIGENTSYSTEMSINACCOUNTINGFINANCE&MANAGEMENTJULY2012DOI:10.1002/isaf.1330

    CITATIONS4

    READS90

    4AUTHORS,INCLUDING:

    HusseinAAbdouUniversityofHuddersfield22PUBLICATIONS178CITATIONS

    SEEPROFILE

    JohnPointonUniversityofPlymouth47PUBLICATIONS274CITATIONS

    SEEPROFILE

    Allin-textreferencesunderlinedinbluearelinkedtopublicationsonResearchGate,lettingyouaccessandreadthemimmediately.

    Availablefrom:HusseinAAbdouRetrievedon:14January2016

  • DETERMINANTS OF CAPITAL STRUCTURE IN THE UK RETAIL

    In this paper we investigate the determinants of capital structure in the UK retail industry. Key factors

    INTELLIGENT SYSTEMS IN ACCOUNTING, FINANCE AND MANAGEMENTIntell. Sys. Acc. Fin. Mgmt. 19, 151169 (2012)Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/isaf.1330for investigation, which we discuss below in our literature review, are: asset growth, asset structure, taxshields, growth opportunities, operating protability, liquidity, business risk, sales growth, size andtotal protability. In attaining our objectives we investigate different results obtained from applying

    * Correspondence to: Hussein A. Abdou, University of Salford, Salford Business School, Reader in Finance & Banking, TheINDUSTRY: A COMPARISON OF MULTIPLE REGRESSION ANDGENERALIZED REGRESSION NEURAL NETWORK

    HUSSEIN A. ABDOUa*, ANDZELIKA KUZMICb, JOHN POINTONb AND ROGER J. LISTERa

    a University of Salford, Salford Business School, Salford, Greater Manchester, UKb University of Plymouth, School of Management, Plymouth, Devon, UK

    SUMMARYFirms need to rely on different nancing sources, but the question is how capital structure is determined for aparticular industry. Our aim is to undertake an investigation into the factors which determine capital structure inthe UK retail industry. Our initial sample consists of 163 (nal sample: 100) UK retail companies, using data from2000 in order to analyse capital structure from 2002 to 2006. Nonlinear models tend to be unduly neglected incapital structure research, and so we apply generalized regression neural networks (GRNNs), which are comparedwith conventional multiple regressions. We utilize a hold-out sample for the multiple regressions to make themcomparable with the GRNNs. Stability of the data is also conrmed. Our main ndings are: net protability andthe depreciation-to-sales ratio are key determinants of capital structure based on GRNNs, while two morevariables are added in the multiple regressions, namely size and quick ratio; there is strong support for thepecking-order theory; both root-mean-square errors and mean absolute errors are much lower for the GRNNs thanthose for the multiple regressions for overall, training and testing datasets. The potential benet of this research tonancial managers and investors in the UK retail sector is the identication of the overriding role of net protabilityin reducing the nancial risk from high levels of gearing. Copyright 2012 John Wiley & Sons, Ltd.

    Keywords: capital structure; retail; generalized regression neural networks; multiple regression; pecking order

    1. INTRODUCTION

    Relatively few studies, to which a notable exception is Hutchinson and Hunter (1995), investigate thecapital structure of UK retail companies despite the important role of this sector in the UK economy.The retail sector is an advantageous basis for our study: assets, liabilities and income tend to be morevisible and to present relatively fewer valuation problems than complex manufacturing and someservice industries do. There is also more comparability across companies. These factors furtherenhance retails attractiveness as a test bed for capital structure. For non-nancial companies, ingeneral, a number of factors would seem to play a potentially important role in determining capitalstructure. These include growth opportunities, rm size, protability, asset structure, business risk,non-debt tax shields and liquidity. Myers and Majlufs (1984) pecking-order theory indicates apreference for internal retained prots, followed by debt nance and, nally, new equity.Crescent, Salford, Greater Manchester, M5 4WT, UK. Email: [email protected]

    Copyright 2012 John Wiley & Sons, Ltd.

  • risky, in that there may be greater variations in their future value and thus they may face difculties in

    than xed tangible assets, raising a potential conict between the interests of management and share-

    152 H. A. ABDOU ET AL.holders which managers can exploit. Also, Chen (2004) nds that rms characterized by higher growthopportunities tend to have lower values of tangible assets and, as a result, will borrow proportionatelyless. Against Chen, especially at a time of uncertain property values, one must not be dogmatic infavour of tangibility, since well-founded growth opportunities in a stable economy may be worth morethan restricted, immovable, highly specic assets elsewhere. Daskalakis and Psillaki (2005) suggest thatgrowth will push rms into seeking external nancing, as rms with high growth opportunities aremore likely to exhaust internal funds and will require additional capital to support expansion. For in-deed, as Myers and Majluf (1984) have proposed in their pecking order, there should be a preferencefor internal retained prots, followed by debt nance and then new equity. Titman and Wessels(1988) found a negative association between a rms growth opportunities and leverage, whereasaccording to the study by Drobetz and Wanzenried (2006) rms are concerned with future nancingcosts if investment projects fail. Therefore, rms with future growth opportunities are nanced byshort-term borrowing such as greater trade credit and short-term debt. On the other hand, Rajan andZingales (1995) and Frank and Goyal (2000) found growth opportunities positively related to the debtratio, and posited that the debt of a rm grows according to the shortfall between the cost of investmentprojects and retained earnings available to nance such investment, resulting in a demand for externalnance. Another study by Bevan and Danbolt (2004) demonstrated that corporate growth opportunitiesraising debt capital as their future ability to service debt is more uncertain (Bevan and Danbolt, 2002).Furthermore, exploitation of growth opportunities as an asset is more at the discretion of managementgeneralized regression neural networks (GRNNs) compared with the more traditional multiple regres-sion modelling techniques. The two approaches are mutually informative and promise to shed light onwhich factors are the key determinants of the capital structure decision in this sector. Our investigationbridges one of the most contentious aspects of nance theory and one of the crucial decision points innancial management. The theory has been discussed for decades since Modigliani and Miller (1958,1963) initiated the debate on capital structure independence, and nancial constraints weigh more thanever on management as they attempt to dene and handle their nancing. In any industry, nancialdecision-makers need to take cognizance of the determinants of capital structure in their industryfellows, since these may alert them to the most relevant, pressing and opportune constraints. Ourcontribution promises, accordingly, to be both conceptual and practical and will be substantiallyreinforced by our choice of methodology. To the best of our knowledge, neural networks, in particularthe GRNNs, have not been used to investigate the determinants of capital structure in the UK retailindustry. It is expected that the modelling procedures using neural networks will be superior to thetraditional multiple regressions, since they accommodate nonlinearities (e.g. Pao, 2008). The rest ofthis paper is organized as follows: Section 2 reviews relevant literature and develops the formal researchhypotheses; Section 3 is concerned with data sources and methodology; Section 4 reports and analysesresults. We conclude and make recommendations in Section 5.

    2. LITERATURE REVIEWAND FORMATION OF RESEARCH HYPOTHESES

    Growth Opportunities: Agency theory states that rms with growth opportunities increase the poten-tial conict of interest between shareholders and lenders; for example, see Harris and Raviv (1991) andCheng and Shiu (2007). Firms that have higher potential growth opportunities may be regarded as moreCopyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • DETERMINANTS OF CAPITAL STRUCTURE IN UK RETAIL INDUSTRY 153have very little impact on the level of gearing and thus, paradoxically, rms in the UK tend to hold moredebt. For a retail rm it could be argued that sales growth, which reects changes in market share, isimportant as well as asset growth. However, we recognize that the debate is open and we hope thatour ndings will contribute to its progress. We submit two research hypotheses:

    H1a: asset growth is negatively related to gearing.H1b: annual sales growth is negatively related to gearing.

    Some research identies the market-to-book value of equity as a surrogate for growth opportunities,because the book value, unlike the market value, does not anticipate future growth. For example,Drobetz and Wanzenried (2006) used this measure to represent growth and found a negative relationshipwith leverage for their sample of Swiss rms. A problem with this measure is that it may not only betelling a story of growth. For example, a high market-to-book ratio may reect low book values resultingfrom accountings historic cost convention, the depreciation method used or from management andauditors prudence in the face of uncertain valuations. In retail, assets tend to be visible and susceptibleto defensible valuation. We hope, accordingly, that our choice of industry will leave us not undulyvulnerable to such problems. We propose the following hypothesis for UK retail rms:

    H1c: market-to-book value as a proxy for growth opportunities is negatively related to gearing.

    Firm Size: According to trade-off theory, to the extent that they are more diversied, large rms havea higher debt capacity and, thus, lower nancial distress costs than smaller rms do (Cheng and Shiu,2007). Additionally, large rms are not only less risky, but are also likely to have lower asymmetricinformation for creditors than small rms might have (Cheng and Shiu, 2007). Their shareholdersare likely to include sophisticated nancial institutions who will subject them to a degree of scrutiny;this fact will, of itself, reassure the market. The point is even more likely to apply if a rm is largeenough to gure in major nancial indices, since a majority of large institutions perforce include theindex in their portfolios by way of adherence to the capital asset pricing model. Small rms may ndit costly to mitigate information asymmetries with lenders, with such rms being offered less capital athigher rates, consequently discouraging the use of outside nancing (Cassar, 2003). Empirical ndingssupport a signicant positive relationship between debt and rm size (Bennett and Donnelly, 1993;Gatchev et al., 2009). However, some researchers report a negative relationship between debt and rmsize, indicating that small companies, owing to their limited access to the equity market, tend to rely onbank loans, and thus larger rms have substantially less debt than smaller ones do (Titman and Wessels,1988; Hutchinson and Hunter, 1995; Rajan and Zingales, 1995; Ramalho and da Silva, 2009). Whilerecognizing the openness of the debate, we articulate the following research hypothesis based ontrade-off theory:

    H2: size is positively related to gearing.

    Protability: The pecking-order theory indicates that protable rms prefer to nance their projectsrstly using internal funds (such as retained earnings) rather than external nance, secondly using debtnext, and thirdly, owing to transaction costs associated with it, choosing equity as a last resort.Protable rms are expected to generate more internal funds and use less external debt; thus, a rmsleverage is expected to decrease with increasing protability (e.g. Jandik and Makhija, 2001; Panno,2003; Bancel and Mittoo, 2004). Hutchinson and Hunters (1995) ndings for the retail industrysupport this position. The preference for less leverage may furthermore be related to a wish to leaveCopyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • 154 H. A. ABDOU ET AL.property unencumbered to maximize put and call options implicit in opportunities to exibly expandand contract as the merits of locations emerge. Muller (2004) suggested that some rms may refuseto use external debt capital in order to minimize intrusion to their business affairs, and remain in controlof their company so that they welcome the opportunity which protability provides to reduce theirresort to debt.A contrary view is the trade-off model, which states that more-protable rms will use more debt as

    they have ability to access and take on debt, as well as require more debt to take advantage of corporatetax shields (e.g. Cheng and Shiu, 2007) subject to any offsetting effect deriving from personal tax(Miller, 1977). Panno (2003) pointed out that there might be a positive correlation between protabilityand a rms leverage, in contrast to the pecking-order theory, when companies want to take advantageof tax shields. For the retail sector, it might be useful to distinguish between total protability and theoperating prot margin, because of the intensity of competition that can arise over prot margins. Wetest two hypotheses, both consistent with pecking-order arguments; namely:

    H3a: protability is negatively associated with gearing.H3b: operating prot margin is negatively associated with gearing.

    Asset Structure: Agency theory indicates that rms that have more tangible assets tend to havegreater ability to issue debt and, thus, have lower agency costs associated with it (Scott, 1977), espe-cially when these assets are nonspecic and easily tradable, as is to a large measure the case in the retailindustry. Also, according to Scott (1977), lenders require collateral to protect their interests, and rmsunable to provide it are likely to pay higher interest, or will be forced to issue equity instead of debt.However, when the quality of assets is considerable (as in the retail case), it may be that a dynamic vari-able such as protability will kick in as the effective constraint on borrowing even while asset backinghas not yet been fully used. The relevance of asset structure can be seen as supporting a trade-off theory(Deesomsak et al., 2004). Frank and Goyal (2003) argued that rms with intangible assets can expect toget external nancing if these assets create value, such as patents or contractual rights, goodwill, copy-rights and franchise rights, which can be pledged to support debt. However, these values can be losteasily when a rm defaults. Thus, rms with intangible assets are expected to have less debt. Fattouhet al. (2005) provided evidence that tangibility is positively correlated with debt. However, Bevanand Danbolt (2000) found a positive relation only between long-term borrowing and asset structureand a negative relation in the case of short-term debt. This may be because short-term indebtednessreects corporate ability to meet ongoing commitments on borrowing. We submit the followinghypothesis:

    H4: asset structure is positively related to gearing.

    Business Risk: A higher volatility of earnings is common for an uncertain business environment, inwhich observations of management actions tend to be limited and lenders tend to fear agency-relatedexpropriations by equity holders (Jandik and Makhija, 2001). However, the study by Antoniou et al.(2006) provides evidence that leverage and earnings volatility are negatively correlated, implying thatdue to an uncertain environment UK rms avoid entering into long-term commitments involving highlyvolatile earnings. This is in line with the trade-off model, which predicts that rms with higher businessrisk, determined by the variability and uncertainty of its sales and costs, will use less debt to nancetheir projects as high business risk increases the cost of nancial distress (Panno, 2003; Deesomsaket al., 2004; Cheng and Shiu, 2007). Risk-taking companies, in need of new nance, will tend to issueequity rather than debt (Panno, 2003). Yet, Hutchinson and Hunter (1995) identied a positiveCopyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • DETERMINANTS OF CAPITAL STRUCTURE IN UK RETAIL INDUSTRY 155relationship between leverage and business risk in the retail industry, contrary to much of the literature.A relatively predictable stream of taxable prots even in risky times may contribute to the retailphenomenon, as well as other unique characteristics based on corporate positioning associated withexibility of debt policy in difcult circumstances to which Hutchinson and Hunter refer. While notingthat the debate concerning trade-off and pecking order remains inconclusive, we test the followinghypothesis:

    H5: business risk is negatively related to gearing.

    Non-debt Tax Shield: According to trade-off theory, a major motivation for rms to use debt instead ofequity is to save corporate tax. To the extent that interest payments on debt are benecially deductible asan expense, they reduce the corporate tax bill. In countries with a high corporate tax rate which is notoffset by other taxes, rms are expected to use more deductible external nancing to raise capital thanotherwise (Cheng and Shiu, 2007). However, rms can use alternative non-debt tax shields, such as alegacy of tax losses, international tax planning and, most widely, tax-depreciation (capital allowance)deductions. The higher non-debt tax shield reduces the potential tax benet of debt and hence the abso-lute tax advantage of debt will not apply (Drobetz and Fix, 2005). In accordance with the trade-off the-ory, DeAngelo and Masulis (1980) argued that rms with high non-debt tax shields are less likely toissue debt. Thus, Bennett and Donnelly (1993), Jandik andMakhija (2001) and Drobetz andWanzenried(2006) provided evidence that the non-debt tax shield is negatively related to leverage, but other ndings(Drobetz and Fix, 2005) reveal that the non-debt tax shield is insignicant. It remains a matter of fact thathigh current depreciation, as a proxy for high capital allowances for tax purposes, is likely to beassociated with a current tax loss and a reduced immediate benet from interest deductibility for taxpurposes. Accordingly, in our study we include depreciation/sales and submit the following hypothesis:

    H6: the current depreciation-to-sales ratio is negatively related to gearing.

    Liquidity: Both the quick ratio and the current ratio are serious candidates for measuring a rms abilityto meet its short-term obligations; thus, there should be a positive relationship between the rms liquidityposition and its debt ratio. Firms with high liquidity might support a relatively high debt ratio, due to agreater ability to meet its nancial obligations when they fall due (Panno, 2003), although pecking-ordertheory indicates that rms with a higher liquidity ratio will borrow less (Deesomsak et al., 2004). Panno(2003) also reported that a rms liquidity has a positive effect on the rms borrowing decisions in theUK, and is consistent with a theory of expectations. But, following the pecking-order theory, accordingto which rms resort to debt when internal resources are insufcient, we submit the following hypothesis:

    H7: liquidity is negatively related to gearing.

    Capital Structure Analysis and Techniques: Pao (2008) made a comparison between neural networksand multiple regression analyses in modelling the capital structures of high-tech and traditional corpora-tions in Taiwan. We would suggest that the dual importance of that study is to include the possibilities ofnonlinear effects not captured by the multiple regressions, and to provide conrmatory evidence byusing two methods rather than one. Pao focused on comparing signs between those from the articialneural network (ANN) sensitivities and those of the multiple regression coefcients, which is a validand useful approach. By contrast, in this paper we focus on an impact analysis of the extent towhich eachvariable contributes to capital structure. The reason for this is that the ANN sensitivities, although usefulfor conrming the signs, do not inform us how important each variable is in the overall model. The signCopyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • may be the same as that in the multiple regression model, and the multiple regression may indicate that aparticular variable is very signicant. However, we suggest that when nonlinearities are accounted for,the contribution of a variable that may have been signicant in the multiple regression may play onlya minor role in the ANN. Later in this paper, we demonstrate that this is indeed the case. FollowingPao (2008), who investigated the same topic, albeit not for the UK retail sector but for high-techcompanies in Taiwan, and who found that his nonlinear neural networks outperformed conventionalregressions and achieved better model-tting and predictions, we expect to nd that our neural networksprovide a superior quality of analysis than that of the conventional regressions. The following hypothesisis therefore proposed:

    H8: neural networks provide a better t and predictions than conventional regressions do in analysingcapital structure in the UK retail sector.

    We have identied a research gap in the study of capital structure in the UK retail industry, for, to thebest of our knowledge, we have found that not only have there been a limited number of investigations,but neural networks have been neglected in previous research.

    3. DATA SOURCES AND METHODOLOGY

    Variables Code Measurements

    156 H. A. ABDOU ET AL.Dependent variable

    Y Debt ratio (leverage) LEV Total debt/total assets

    Independent variables

    X1 Growth in assets ASSG [Total assets(t) total assets(t 1)]/total assets(t 1)X2 Assets structure ASST Fixed assets/total assetsX3 Depreciation ratio DPN Current depreciation/total salesX4 Market-to-book value MTBV Share price/book value per shareX5 Operating prot margin OPS Operating prot/total salesX6 Quick ratio QR (Cash, near-cash, marketable securities and debtors)/current liabilitiesX7 Business risk RISK (Standard deviation of the annual earnings before interest and tax, based

    on the current year and the preceding two years)/(mean annual earningsbefore interest and tax over the 3 years)

    X8 Sales growth SGR [Sales(t) sales(t 1)]/sales(t 1)X9 Size SIZE Natural log(sales)X10 Net Protability TPROF Annual earnings before interest, tax and depreciation/total assetsX11 Gross protability

    a GPROF Gross income/sales revenueX12 Current ratio

    a CR Current assets/current liabilities

    aVariables nally discarded in building the models.Table I. List of variables and their measurements3.1. Dataset and Sampling Method

    The dataset for the empirical research is derived from the Datastream database and we collectedaccounting information of 163 UK retail companies between 2000 and 2006. Not all data of retail com-panies could be used in this study, due to missing values or the suspension of a rm. The nal sample is100 companies exactly. Variables are dened in Table 1. Since some variables require more than 1 yearCopyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • elimination of two variables due to multicollinearity, namely gross protability (which is highly

    DETERMINANTS OF CAPITAL STRUCTURE IN UK RETAIL INDUSTRY 157correlated with OPS) and the current ratio (which is highly correlated with QR). The remainingvariables have low collinearity (

  • X1

    Output Summation layer (numerator &

    denominator nodes)Pattern layer (one

    neuron per training case)Inputs

    158 H. A. ABDOU ET AL.Xn

    YX2

    Figure 1. GRNN for a number of independent numeric variables is structured above. The Pattern Layer containsone node for each training case. Each neuron in the pattern Layer computes its distance from the presented case.The values passed to the Numerator and Denominator nodes are functions of the distance and the dependent value.The nodes in the Summation Layer sum its inputs, while the Output node divides them to generate the prediction.The distance function in the Pattern Layer neurons uses smoothing factors and every input has its own smoothingfactor value. With a single input, the greater the value of smoothing factor, the more signicant distant trainingcases become for the predicted value. Training GRNN consists of optimizing factors to minimize the error onautomatically by the software. This analysis is only done on the training dataset, and every independentvariable is assigned a relative variable impact value; these are percentage values and add to 100%.1

    This style of memory-based network was designed by Specht (1991) for regression analysis, butis especially well designed to deal with problems which contain nonlinearities. He showed that hisalgorithm has smooth links between actual values. Tomandl and Schober (2001), extending the workby Specht, demonstrated that their algorithms are robust to changing the values of the parameters,and that the vectors may take on different values. They also discussed the qualities of the gradientsof the regression surfaces.Financial applications have included an investigation into forecasting foreign exchange by which

    Leung et al. (2000) who, after contrasting random walk behaviour with multilayered feed-forwardnetworks, came to advocate neural networks in terms of predictive ability. We run a GRNN for thedeterminants of leverage, for training and testing samples, the former for the years 20022005 andthe latter for 2006. Specically, we conduct a variable impact analysis to assess the relative importanceof each determinant (ASSGit, ASSTit, DPNit, MTBVit, OPSit, QRit, RISKit, SGRit, SIZEit and TPROFit).We report on predictive ability and error rates, namely, Root-Mean-Square Errors (RMSEs) and MeanAbsolute Errors (MAEs), as measures of models accuracies.

    1It should be noted that the results of the impact analysis are relative to a given net, and a subsequent session with a different typeof net is able to discover how the variable(s) can make a signicant contribution to accurate predictions. For example, in a smallerdataset with a large number of variables, the differences in the relative impact of the variable between trained nets may be morepronounced, and vice versa.

    the training set, and the Conjugate Gradient Descent optimization method is used to accomplish that. The errorused in this net is the Mean Square Error.

    Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • 4. RESULTS AND ANALYSIS

    In Table 2 we set out a comparative statistical evaluation of the dependent variable, leverage (LEV), andthe independent variables. We carried out analysis of variance (ANOVA) tests to determine if the means

    DETERMINANTS OF CAPITAL STRUCTURE IN UK RETAIL INDUSTRY 159differ for different years. This is only the case for asset growth. We performed Fishers least signicantdifference tests and Tamhane tests; the former assume equality of variances, but not the latter. Indeed,the tests designed by Cochran, Bartlett and Levene show some very signicant differences betweenthe standard deviations in different years. Nevertheless, the Tamhane test reveals that the differencesbetween years are not very signicant (except for some moderate signicance for asset growth for pairsof years linked to 2002). The KruskallWallis tests do not disclose very signicant differences betweenthe medians over the different years. At that point we concluded that we had a good dataset.We then proceeded with a conventional multiple regression analysis in order to examine the key fac-

    tors that might determine the leverage in the UK retail sector, as set out in Table 3. We tested the regres-sion residual for autocorrelation using the LjungBox (Q) standardized residuals (for 36 lags) and theBreuschGodfrey Lagrange multiplier (LM) test (seven lags). The tests indicate that autocorrelation isnot present in the residuals. Furthermore, we checked for the presence of heteroskedasticity using thesquared (Q2) standardized residuals (36 lags) and the ARCH test, which is an LM test for ARCH (sevenlags) in the residuals. Again the results show that there is no ARCH effect in the residuals. Furthermore,the augmented DickeyFuller tests demonstrate that for each variable the null hypothesis of a unit rootis rejected in favour of stationarity. These diagnostics are set out in Appendix A.The conventional regression reveals a very signicant quick ratio2 QR (), SIZE (+) and net

    protability TPROF (). These results are consistent with three of our hypotheses: H7 (liquidity),H2 (Size) and H3a (overall protability). Thus, liquidity in the UK retail sector is negatively relatedto gearing, supporting the pecking-order theory according to which rms resort to debt when internalresources are insufcient. This is signicant at the 95% level of condence in the full regression andat the 90% level of condence in the stepwise regression.Size of retail rms is signicantly positively related to gearing, which conrms the trade-off theory,

    at the 99% level of condence in both the full and stepwise regressions. This reverses the earlierndings by Hutchinson and Hunter (1995) to the effect that smaller retail rms used more debt thanthe larger rms did and supports the side of our earlier discussion of size to the effect that large rmsare less risky, suffer less from information asymmetry and benet from the reassuring presence ofinstitutional investors.Overall protability, as measured by TPROF, is signicantly negatively associated with gearing at the

    99% level of condence. This is consistent with the work of Hutchinson and Hunter (1995), who foundthat retail rms use retained prots rather than debt to fund capital investment. Our nding reects thepecking-order theory and our earlier discussion of the theorys implications for the relationship betweenprotability and leverage. Consonant with that theory, management regard protability as a generatorof the retained earnings which they prefer to use as a source of nance. Compared with debt, retainedearnings can be used with less external interference, lower transaction costs and less encroachment onfreedom to expand and contract the asset base. Far from dening the limit of borrowing, protabilitygenerates a fund which increases the room for manoeuvre available to both company and management.Leverage is demoted in exact accordance with pecking-order theory.

    2We found that the quick ratio and the current ratio are highly correlated, but we took the precaution of using current ratio insteadand found that it is not signicant. Also, overall model R2 was lower.Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • TableII.Acomparativestatisticalevaluationof

    selected

    variables

    LEV

    ASSG

    ASST

    DPN

    MTBV

    OPS

    QR

    RISK

    SGR

    SIZE

    TPROF

    Count

    479

    461

    479

    479

    444

    478

    476

    434

    460

    478

    478

    Mean

    0.245

    0.191

    0.470

    0.063

    1.778

    0.268

    1.087

    1.323

    0.260

    11.85

    0.042

    Standarddev

    0.533

    1.196

    0.212

    0.120

    17.45

    3.237

    5.309

    35.22

    2.626

    2.570

    0.710

    ANOVAF-ratio

    0.541

    2.486**

    0.644

    1.207

    1.161

    1.532

    1.080

    1.200

    1.210

    0.426

    0.746

    Fishersleastsignicant

    difference

    test

    2002

    2003

    0.030

    0.515**

    0.0040

    .004

    1.794

    0.812

    0.169

    0.468

    0.115

    0.115

    0.083

    2002

    2004

    0.095

    0.062

    0.0360

    .016

    0.835

    0.860

    0.122

    9.598

    0.072

    0.176

    0.115

    2002

    2005

    0.080

    0.170

    0.031

    0.011

    5.131

    0.998**

    0.128

    0.785

    0.696

    0.383

    0.021

    2002

    2006

    0.023

    0.127

    0.035

    0.018

    1.180

    1.033***1

    .141

    0.205

    0.010

    0.402

    0.105

    2003

    2004

    0.066

    0.453**

    0.0320

    .011

    0.959

    0.048

    0.047

    9.131

    0.043

    0.061

    0.033

    2003

    2005

    0.051

    0.346**

    0.027

    0.015

    3.337

    0.186

    0.041

    0.317

    0.581

    0.268

    0.103

    2003

    2006

    0.006

    0.388**

    0.031

    0.023

    0.614

    0.221

    1.310

    0.263

    0.105

    0.287

    0.022

    2004

    2005

    0.015

    0.108

    0.005

    0.026

    4.296

    0.138

    0.007

    8.814

    0.624

    0.207

    0.136

    2004

    2006

    0.072

    0.065

    0.001

    0.034**

    0.345

    0.173

    1.262

    9.394

    0.062

    0.226

    0.010

    2005

    2006

    0.057

    0.0424

    0.004

    0.008

    3.951

    0.035

    1.069

    0.580

    0.686

    0.020

    0.126

    CochransCtest

    0.586***

    0.921***

    0.221

    0.581***

    0.374***

    0.937***

    0.948***

    0.986***

    0.974***

    0.226

    0.826***

    Bartlettstest

    1.685***

    5.238***

    1.004

    1.954***

    1.878***

    8.768***

    7.930***

    30.36***

    14.05***

    1.003

    2.890***

    Levenestest

    0.618

    2.794**

    0.552

    1.109

    2.727**

    1.514

    1.118

    0.957

    0.998

    0.312

    0.826

    Tamhane

    testa

    2002

    2003

    0.030

    0.515

    0.0040

    .004

    1.794

    0.812

    0.169

    0.468

    0.115

    0.115

    0.083

    2002

    2004

    0.095

    0.062

    0.0360

    .016

    0.835

    0.860

    0.122

    9.598

    0.072

    0.176

    0.115

    2002

    2005

    0.080

    0.170*

    0.031

    0.011

    5.131

    0.998

    0.128

    0.785

    0.696

    0.383

    0.021

    2002

    2006

    0.023

    0.127*

    0.035

    0.018

    1.180

    1.033

    1.141

    0.205

    0.010

    0.402

    0.105

    2003

    2004

    0.066

    0.453

    0.0320

    .011

    0.959

    0.048

    0.047

    9.131

    0.043

    0.061

    0.033

    2003

    2005

    0.051

    0.346

    0.027

    0.015

    3.337

    0.186

    0.041

    0.317

    0.581

    0.268

    0.103

    2003

    2006

    0.006

    0.388

    0.031

    0.023

    0.614

    0.221

    1.310

    0.263

    0.105

    0.287

    0.022

    2004

    2005

    0.015

    0.108

    0.005

    0.026

    4.296

    0.138

    0.007

    8.814

    0.624

    0.207

    0.136

    2004

    2006

    0.072

    0.065

    0.001

    0.034

    0.345

    0.173

    1.262

    9.394

    0.062

    0.226

    0.010

    2005

    2006

    0.057

    0.042

    0.004

    0.008

    3.951

    0.035

    1.069

    0.580

    0.686

    0.020

    0.126

    KruskalW

    allis

    medianteststatistic

    0.620

    7.663

    3.144

    3.435

    6.443

    1.264

    1.075

    1.753

    3.799

    1.225

    3.802

    *,**and***denotestatistically

    signicant

    differencesatthe10%,5%

    and1%

    levelrespectively.

    a The

    Tamhane

    Test,w

    hich

    assumes

    unequalvariances,gave

    differentmean-difference

    results,w

    hilst,Fishers

    leastsignicant

    difference

    testassumes

    equalvariances.

    The

    sampleconsistsof

    a500rm

    -yearobservationpaneldata-setfrom

    theUKretailindustry.T

    hedataarederivedfrom

    theDatastreamdatabase

    fortheyears2002

    to2006.

    LEV=Debtratio(Leverage),A

    SSG

    =GrowthinAssets,ASST

    =AssetsStructure,D

    PN=DepreciationRate,MTBV=Market-to-BookValue,O

    PS=OperatingProtM

    argin,

    QR=Quick

    Ratio,RISK=BusinessRisk,SGR=Sales

    Growth,S

    IZE=Size;TPR

    OF=NetProtability.

    160 H. A. ABDOU ET AL.

    Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • Table III. Model1 (overall sample)

    Model analysis R1 Step. R1 GRNN1

    DETERMINANTS OF CAPITAL STRUCTURE IN UK RETAIL INDUSTRY 161Coeff. Coeff. Variables impact analysis

    Constant 0.0552 0.1278* ASSG 0.0073 2.1846ASST 0.0208 1.2325DPN 1.7119*** 1.6017*** 24.1530MTBV 0.0003 1.5099OPS 0.0799*** 0.0155*** 4.9051QR 0.0051** 0.0041* 1.7875RISK 0.0003 1.5817SGR 0.0058 1.4041SIZE 0.0182*** 0.0255*** 0.7072TPROF 0.6219*** 0.5634*** 60.5343

    100.00Further analytical results (diagnostic criteria)F-ratio 167.95*** 247.54*** R2 80.53 72.56 R2 adj. 80.05 72.27 Good prediction (%) 36.69Std. dev. of abs. errors 0.1199RMSE 0.2390 0.2758 0.1571MAE 0.1569 0.1771 0.1015

    *, **and ***denote statistically signicant differences at the 10%, 5% and 1% level respectively.The sample consists of a 500 rm-year observation panel dataset from the UK retail industry. The data are derived from the Data-On the other hand, depreciation ratio DPN (+) and operating prot margin3 OPS (+) are signicantlypositive at the 99% level of condence. It is initially surprising and out of tune with some earlier work(e.g. Bennett and Donnelly, 1993; Jandik and Makhija, 2001; Drobetz and Wanzenried, 2006) that thedepreciation ratio should coincide with greater gearing. However, it remains the case that other empir-ical results, for example by Drobetz and Fix (2005), nd that non-debt tax shields do not have any effecton the level of leverage of a rm. Our nding is consistent with the fact that the retail industrys assetsare notably acceptable as security for debt to the extent that they tend to be nonspecic. Thisencourages borrowing, since the terms will be relatively favourable. Furthermore, the industry is lessvulnerable to operational gearing than manufacturers, and this enhances borrowing capacity.Perhaps the leverage policies in the UK retail sector are affected by the fact that for tax purposes their

    buildings do not qualify as industrial buildings and lose tax allowances as a result. They can claimcapital allowances on their eets of vehicles and on their equipment to the extent that these are pur-chased rather than leased. Despite theoretical developments in nance over several decades beingdevoted to tax considerations, work on tax shield theory is inconclusive and our ndings suggest thatleverage in the retail industry does not appear to be tax driven. The tax cartdoes not drive the capitalstructure horse. But since newer assets attract higher depreciation, the results are neverthelessconsistent with the view that more recent asset acquisitions have been funded by debt.

    stream database for the years 20022006. ASSG: Growth in Assets; ASST: Assets Structure; DPN: Depreciation Rate; MTBV:Market-to-Book Value; OPS: Operating Prot Margin; QR: Quick Ratio; RISK: Business Risk; SGR: Sales Growth; SIZE: Size;TPROF: Net Protability; RMSE: Root-Mean-Square Error; MAE: Mean Absolute Error.

    3We found that operating prot margin and gross protability are highly correlated, but we took the precaution of using grossprotability instead and found that it is not signicant. Also, overall model R2 was lower.

    Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • The results for operating prot margin and net protability give conicting evidence on the peckingorder, in that hypothesis H3b (operating prot margin) is rejected because of the incorrect sign, whilsthypothesis H3a (overall protability) is not rejected. We submit that the support lent to pecking order by

    162 H. A. ABDOU ET AL.our nding for protability is more reliable than our conicting nding for operating prot margin,since not only does the former have more economic signicance in terms of value creation, but it is alsothe more reliable generator of accumulating retained earnings.Some other of the hypotheses are rejected:H1a (asset growth, which is not signicant),H1b (annual sales

    growth, which is not signicant), H1c (market-to-book value, which is not signicant),H4 (asset structure,which is not signicant) and H5 (business risk, which is not signicant, unlike that of the earlier work byHutchinson and Hunter (1995) on the UK retail sector, who, on appearing to generate a positive relation-ship, acquiesced in the idiosyncrasy of their nding). Market-to-book value (MTBV), as a surrogate forgrowth opportunities, is negatively related to leverage, although not sufciently signicant. The sign isconsistent with agency theory and is congruent with Titman and Wessels (1988) and Jandik and Makhija(2001). Nevertheless, others nd the contrary (e.g. Bevan and Danbolt, 2000; Drobetz and Wanzenried,2006), and indeed others detect no relationship (e.g. Bevan and Danbolt, 2004).The GRNNs4 approach does not require stationarity tests, which makes it an attractive methodology

    compared with normal regression. In Table 3 we set out our results from the GRNN which allows fornonlinearities. The principal importance of using a neural network is the variable impact analysis, for itreveals that TPROF and DPN are very important and in that order, representing 60.5% and 24.2% re-spectively of the total impact on leverage. Hence, the GRNN methodology lends very strong supportfor H3a (overall protability), and conrms the role of a pecking order in the retail sector. Furthermore,the positive OPS coefcient in the conventional regression was a problem for a pecking order, but thelow variable impact of OPS of less than 5% in the GRNN can reduce our concern about the earlierseemingly troublesome result for this variable. Thus, the use of the GRNN has resolved a conict inthe interpretation of the results. Such resolution is of direct practical importance to a nancial deci-sion-maker for whom misplaced emphasis on relatively unimportant capital structure criteria could leadto costly nancing decisions which might directly impinge on corporate stability.In Table 4, we set out the results of using the data for 20022005 to build the models (training data)

    and the 2006 data to test them. The conclusions are the same as for the whole sample, except that thequick ratio is weakly signicant (at the 90% level of condence) in the full regression and does notappear in the stepwise regression. Table 5 addresses diagnostics. The RMSE and MAE of the trainingand testing samples of the stepwise regression (Model R2) are similar to each other (suggesting stabilitybetween the training and testing periods), and similar to the RMSE of the stepwise regression forthe whole sample period (Model R1). However, it can be observed that the RMSEs and MAEs forthe GRNNs are much smaller than those for the conventional multiple regressions. It follows thathypothesis H8 is supported; namely, that neural networks provide a better t and predictions thanconventional regressions in analysing capital structure in the UK retail sector.Having established that there appears to be some prima facie evidence of stability within the time-

    frame, we then use random samples to run the GRNNs, taken from the whole time period, either to buildor to test the models, as shown in Appendix B. Whereas we used the same observations for Model1 andModel2 for training and testing samples under both conventional regression models and GRNNs, inAppendix B we set out the samples that had been randomly selected by the software across differentyears for training and testing purposes.

    4The maximum default running time in GRNN is 2 h. However, in building our models, the running time was less than a minutewith a total number of trials of 96 and 64 for overall samples and training samples respectively.Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • The mean good prediction rate from Sample1 (67%/33% for training/testing) is not signicantly dif-ferent from that of Sample2 (80%/20% for training/testing) based on Fishers least signicant differencetests. Neither are the means of Sample1 and Sample3 (90%/10% for training/testing) signicantly dif-ferent from each other, as shown in Appendix C. However, the means from Sample2 and Sample3 at

    Table IV. Model2 (training 20022005/testing 2006, samples)

    Model analysis

    R2 Step. R2 GRNN2

    Training Testing Training Testing Training Testing

    Coeff. Coeff. Variables impact analysis

    Constant 0.0256 0.1751** ASSG 0.0130 1.8891 ASST 0.0545 0.8110 DPN 1.7006*** 1.6206*** 24.7080 MTBV 0.0004 1.4202 OPS 0.0690*** 0.0150*** 6.2025 QR 0.0306* 2.3708 RISK 0.0002 1.5693 SGR 0.0059 1.5713 SIZE 0.0196*** 0.0287*** 0.7284 TPROF 0.6242*** 0.5725*** 58.7295

    100.00Further analytical results (diagnostic criteria)

    F-ratio 178.30 300.26 R2 85.03 76.30 R2 Adj. 84.55 76.05 Good prediction % 35.08 29.35Std. dev. of abs. errors 0.0998 0.1637RMSE 0.2326 0.3998 0.2814 0.2959 0.1355 0.2054MAE 0.1568 0.1961 0.1829 0.1741 0.0917 0.1240

    *, **, and ***denote statistically signicant differences at the 10%, 5% and 1% level respectively.The sample consists of a 500 rm-year observation panel dataset from the UK retail industry. The data are derived from theDatastream database for the years 20022006. ASSG= Growth in Assets, ASST= Assets Structure, DPN=Depreciation Rate,MTBV=Market to Book Value, OPS=Operating Prot Margin, QR=Quick Ratio, RISK=Business Risk, SGR=Sales Growth,SIZE= Size, TPROF=Net Protability, RMSE=Root Mean Square Error, and MAE=Mean Absolute Error.

    Table V. Comparing the efciencies of different models

    Diagnostic criteria Model1 Model2

    Overall sample Training Testing

    R1 Step. R1 GRNN1 R2 Step. R2 GRNN2 R2 Step. R2 GRNN2

    RMSE 0.2390 0.2758 0.1571 0.2326 0.2814 0.1355 0.3998 0.2959 0.2054

    DETERMINANTS OF CAPITAL STRUCTURE IN UK RETAIL INDUSTRY 163MAE 0.1569 0.1771 0.1015 0.1568 0.1829 0.0917 0.1961 0.1741 0.1240

    The sample consists of a 500 rm-year observation panel dataset from the UK retail industry. The data are derived from theDatastream database for the years 20022006. R1: regression model for overall sample; R2: regression model for the trainingsample covering from 2002 to 2005 and for the testing sample covering 2006; RMSE=Root Mean Square Error, andMAE=MeanAbsolute Error.Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • KruskallWallis tests. However, overall, and for the purposes of this analysis, the results for the analysis

    the 95% condence level. Firm size in particular, which is very signicant in the multiple regression

    164 H. A. ABDOU ET AL.analysis, contributes less than 1% of the total variable impact. Third, the GRNNs provide more efcientmodels in terms of lower error rates. Future research could usefully pursue our topic into the unusualperiod of the nancial crisis and from retail into other industries and across countries.

    ACKNOWLEDGEMENTS

    We thank the anonymous referees and the editor for helpful comments and suggestions. Any remainingerrors are ours.of the means of the good prediction rates suggest that different randomly selected models can display areasonable degree of stability within this dataset, although some of the randomly selected models hadmuch better prediction rates than the predetermined 20022005 training sample used earlier.

    5. CONCLUSION

    Our objectives in this paper have led us to investigate the key factors which determine capital structurein the UK retail industry. Within this investigation we apply GRNNs to add insights which areunavailable with conventional multiple regressions. The multiple regression analyses full an importantrole in identifying the signs and signicances of the respective variables. But addition of the variableimpact analysis of the neural networks enables us to account for nonlinearities. The training and testingsamples reveal that later results are consistent with those of earlier years. We make a statisticalevaluation of different neural networks samples. Although some randomly selected models couldimprove the prediction rates, overall there is reasonable stability in the data.Our main result is that the pecking-order theory holds for the UK retail sectors capital structure.

    Indeed, the GRNN indicates that net protability (TPROF) accounts for 60.5% overall and 58.7% forthe training sample. Both the full regressions and the stepwise regressions had led us to understand thatthe protability-related variables gave conicting recommendations. The total protability supported anancial pecking-order hypothesis, yet the operating prot margin suggested the opposite. How couldwe resolve this dilemma? Fortunately, the GRNN analyses demonstrate their strength by clarifying thisseeming contradiction. The variable impact analyses of the GRNNs conrm a major role for totalprotability (TPROF) in the capital structure levels adopted by UK retail rms, vis--vis a very minorrole for the operating-prot-to-sales variable (OPS). So, pecking order is strongly supported, regardlessof very signicant probability values for competing independent variables in the multiple regressions.This illustrates the superiority of GRNN over traditional linear techniques in modelling complexfunctions. GRNN has served the practical purpose of focusing a decision-makers attention where itis most productive and away from less relevant and potentially misleading foci.Other important indications are rst that the capital structure in the UK retail industry is not tax

    driven. Second, the GRNNs reveal that two variables, namely TPROF and depreciation ratio (DPN),are much more important than the others, whilst the multiple regressions identify ve very importantcapital structure determinants: TPROF, DPN, SIZE, and OPS at the 99% condence level and QR atthe training stage are signicantly different from each other at the 95% level of condence. Neverthe-less, the variances are similar at the training stage as per the Cochrane, Bartlett and Levene tests.On testing, the sample means in each respective pair are not different from each other at the

    prescribed levels of condence, although there are still some weak differences in medians as per theCopyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • APPENDIX

    A:DIAGNOST

    ICTESTS

    (a)Equation(dependent:LEV;skew

    ness:1.1561;kurtosios:10.8914)

    Test:

    Q(7)

    Q(21)

    Q2(7)

    Q2(21)

    BGLM

    JBARCH

    Statistic:

    0.0400

    0.0090

    0.0640

    0.011

    0.8288

    1174.90

    0.00096

    P-value:

    0.2980

    0.5100

    0.7930

    0.9270

    0.5638

    0.0000

    0.97520

    (b)Stationarity

    tests

    Variable

    ADFstatistic

    P-value

    LEV

    21.2622

    0.0000

    ASSG

    18.2427

    0.0000

    ASST

    15.3030

    0.0000

    DPN

    21.9761

    0.0000

    MTBV

    20.6169

    0.0000

    OPS

    75.2919

    0.0000

    QR

    20.9294

    0.0000

    RISK

    19.3444

    0.0047

    SGR

    20.5994

    0.0000

    SIZE

    12.4298

    0.0000

    TPROF

    20.8184

    0.0000

    Notation:

    Q=Ljung-Box

    standardised

    residualsforgivenlags;BGLM

    =Breusch-Godfrey

    LagrangeMultipliertestfor7lags;JB

    =Jarque

    -Beranorm

    ality

    test;ARCH=

    LagrangemultipliertestforARCHfor7lags

    intheresiduals;andADF=augm

    entedDickey-Fullertest.

    DETERMINANTS OF CAPITAL STRUCTURE IN UK RETAIL INDUSTRY 165

    Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • APPENDIX

    B:ANALY

    SIS

    OFDIFFERENTNEURALNETWORKS

    SAMPLESIZES

    M1

    M2

    M3

    M4

    M5

    M6

    M7

    M8

    M9

    M10

    Testingsamples

    Sam

    ple 1

    Goodpre.

    42.750

    26.810

    26.810

    23.910

    34.780

    34.780

    31.150

    33.330

    39.130

    23.190

    St.dev.

    0.7440

    0.1466

    0.1036

    0.1342

    0.7503

    0.7553

    0.1473

    0.4456

    0.1305

    0.1828

    RMSE

    0.7618

    0.1969

    0.1532

    0.1907

    0.7717

    0.7789

    0.1889

    0.4714

    0.1700

    0.2301

    MAE

    0.1638

    0.1315

    0.1129

    0.1355

    0.1803

    0.1904

    0.1182

    0.1537

    0.1089

    0.1398

    Sample 2

    Goodpre.

    24.100

    44.580

    44.580

    42.170

    40.960

    48.190

    46.990

    44.580

    37.350

    33.730

    St.dev.

    0.1600

    1.0930

    0.9597

    0.1455

    0.9537

    0.9530

    0.1157

    0.1432

    0.1620

    0.0765

    RMSE

    0.2058

    1.1300

    0.9838

    0.1756

    0.9736

    0.9738

    0.1381

    0.1784

    0.1996

    0.1326

    MAE

    0.1295

    0.2841

    0.2164

    0.0983

    0.1957

    0.2003

    0.0753

    0.1063

    0.1167

    0.1083

    Sam

    ple 3

    Goodpre.

    50.000

    50.000

    26.190

    35.710

    59.520

    23.810

    35.710

    42.860

    26.190

    14.290

    St.dev.

    0.1240

    0.1297

    0.1344

    0.1452

    1.3290

    0.1076

    0.1641

    0.1154

    0.1636

    0.9990

    RMSE

    0.1507

    0.1648

    0.1885

    0.1867

    1.3630

    0.1667

    0.2057

    0.1474

    0.2229

    1.0490

    MAE

    0.0857

    0.1016

    0.1322

    0.1173

    0.3057

    0.1273

    0.1240

    0.0917

    0.1514

    0.3191

    Trainingsamples

    Sample 1

    Goodpre.

    76.700

    53.050

    46.950

    40.860

    72.760

    68.100

    55.170

    69.530

    100.00

    36.910

    St.dev.

    0.0180

    0.0782

    0.0924

    0.1201

    0.0290

    0.0369

    0.0692

    0.0373

    0.0000

    0.0960

    RMSE

    0.0204

    0.0993

    0.1181

    0.1526

    0.0359

    0.0457

    0.0908

    0.0475

    0.0000

    0.1326

    MAE

    0.0095

    0.0613

    0.0735

    0.0941

    0.0210

    0.0270

    0.0588

    0.0293

    0.0000

    0.0915

    Sam

    ple 2

    Goodpre.

    35.330

    98.500

    98.200

    85.030

    73.650

    67.070

    87.720

    99.100

    52.100

    35.630

    St.dev.

    0.1121

    0.0001

    0.0002

    0.0121

    0.0287

    0.0355

    0.0069

    0.0000

    0.0761

    0.1263

    RMSE

    0.1527

    0.0001

    0.0002

    0.0134

    0.0351

    0.0450

    0.0079

    0.0000

    0.0951

    0.1630

    MAE

    0.1037

    0.0000

    0.0000

    0.0056

    0.0202

    0.0277

    0.0039

    0.0000

    0.0571

    0.1031

    Sam

    ple 3

    Goodpre.

    63.730

    37.600

    43.730

    35.200

    68.530

    36.000

    88.270

    60.270

    34.670

    34.930

    St.dev.

    0.0446

    0.1043

    0.0694

    0.1185

    0.0344

    0.1241

    0.0000

    0.0504

    0.1189

    0.1067

    RMSE

    0.0573

    0.1346

    0.1030

    0.1554

    0.0427

    0.1617

    0.0000

    0.0660

    0.1573

    0.1455

    MAE

    0.0360

    0.0852

    0.0762

    0.1005

    0.0253

    0.1037

    0.0000

    0.0426

    0.1030

    0.0989

    Notation:

    M1,M

    2..

    .M10aredifferentsamples

    thathadbeen

    random

    lyselected

    bythesoftwareacross

    differentyearsfortraining

    andtestingpurposes.R

    MSE=RootMean

    SquareError,and

    MAE=MeanAbsoluteError.Sam

    ple 1(67%

    /33%

    fortraining/testing);Sam

    ple 2(80%

    /20%

    fortraining/testing);and

    Sam

    ple 3(90%

    /10%

    fortraining/testing).

    166 H. A. ABDOU ET AL.

    Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • APPE

    NDIX

    C:COMPA

    RATIVESTA

    TISTICALEVALUATIONOFDIFFERENTNEURALNETWORKSSAMPLESBASEDONGOODPREDICTIONRATES

    Different

    neuralnetworksamples

    Training(goodpredictionrates)(%

    )Testing(goodpredictionrates)(%

    )

    Count

    3030

    Average

    (mean)

    61.84

    36.27

    Standarddeviation

    22.53

    10.29

    ANOVAF-ratio

    2.940*

    2.082

    Fishersleastsignicant

    difference

    test

    Sam

    ple 1

    Sam

    ple 2

    11.2264

    9.05625

    Sam

    ple 1

    Sam

    ple 3

    11.7138

    4.76192

    Sam

    ple 2

    Sam

    ple 3

    22.9402**

    4.29433

    CochransCtest

    0.4647

    0.6774***

    Bartlettstest

    1.0363

    1.2866**

    Levenestest

    0.5835

    3.4227**

    Tamhane

    testa

    Sam

    ple 1

    Sam

    ple 2

    9

    .05625**

    Sam

    ple 1

    Sam

    ple 3

    4

    .76192

    Sam

    ple 2

    Sam

    ple 3

    4.29433

    KruskalW

    allis

    medianteststatistic

    5.1174*

    5.0139*

    *,**

    and***denotestatistically

    signicant

    difference

    at10%,5%

    and1%

    levelrespectively.

    aThe

    Tamhane

    test,w

    hich

    assumes

    unequalvariances,gave

    differentresults,whilstFishers

    leastsignicant

    difference

    testassumes

    equalvariances.

    DETERMINANTS OF CAPITAL STRUCTURE IN UK RETAIL INDUSTRY 167

    Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • 168 H. A. ABDOU ET AL.REFERENCES

    Antoniou A, Guney Y, Paudyal K. 2006. The determinants of debt maturity structure: evidence from France,Germany and the UK. European Financial Management 12: 161194.

    Bancel F, Mittoo UR. 2004. Cross-country determinants of capital structure choice: a survey of European rms.Financial Management 33: 103132.

    Bennett M, Donnelly R. 1993. The determinants of capital structure: some UK evidence. The British AccountingReview 25: 4059.

    Bevan AA, Danbolt J. 2000. Dynamics in the determinants of capital structure in the UK. http://www.accn.gla.ac.uk/ (accessed 15 May 2010).

    Bevan AA, Danbolt J. 2002. Capital structure and its determinants in the UK a decomposition analysis. AppliedFinancial Economics 12: 159170.

    Bevan AA, Danbolt J. 2004. Testing for inconsistencies in the estimation of UK capital structure determinants.Applied Financial Economics 14: 5566.

    Cassar G. 2003. The nancing business start-ups. Journal of Business Venturing 19: 261283.Chen JJ. 2004. Determinants of capital structure of Chinese-listed companies. Journal of Business Research 57:1341135.

    Cheng S, Shiu C. 2007. Investor protection and capital structure: international evidence. Journal of MultinationalFinancial Management 17: 3044.

    Daskalakis N, Psillaki M. 2005. The determinants of capital structure of the SMEs: evidence from Greek and theFrench rms. http://www.univ-orleans.fr (accessed 18 June 2010).

    DeAngelo H, Masulis RW. 1980. Optimal capital structure under corporate and personal taxation. Journal ofFinancial Economics 8: 329.

    Deesomsak R, Paudyal K, Pescetto G. 2004. The determinants of capital structure: evidence from the Asia Pacicregion. Journal of Multinational Financial Management 14: 387405.

    Drobetz W, Fix R. 2005. What are the determinants of capital structure? Evidence from Switzerland. Swiss Journalof Economics and Statistics 141: 71113.

    Drobetz W, Wanzenried G. 2006. What determines the speed of adjustment to the target capital structure?The Journal of Applied Financial Economics 16: 941958.

    Enke D, Thawornwong S. 2005. The use of data mining and neural networks for forecasting stock market returns.Expert Systems with Applications 29: 927940.

    Fattouh B, Scaramozzino P, Harris L. 2005. Capital structure in South Korea: a quantile regression approach.Journal of Development Economics 76: 231250.

    Frank MZ, Goyal VK. 2000. Testing the pecking order theory of capital structure. Journal of Financial Economics67: 217248.

    Frank MZ, Goyal VK. 2003. Capital structure decisions. http://repository.ust.hk (accessed 18 May 2010).Gatchev V, Spindt PA, Tarhan V. 2009. How do rms nance their investments? The relative importance of equityissuance and debt contracting costs. Journal of Corporate Finance 15: 179195.

    Harris M, Raviv A. 1991. The theory of capital structure. Journal of Finance 46: 297355.Hutchinson RW, Hunter LR. 1995. Determinants of capital structure in the retailing sector in the UK. InternationalReview of Retail, Distribution & Consumer Research 5: 6378.

    Jandik T, Makhija K. 2001. Empirical evidence on determinants of capital structure. Advances in FinancialEconomics 6: 143167.

    Leung MT, Chen A-S, Daouk H. 2000. Forecasting exchange rates using general regression neural networks.Computers and Operations Research 27: 10931110.

    Miller M. 1977. Debt and taxes. Journal of Finance 32: 261275.Modigliani F, Miller M. 1958. The cost of capital, corporation nance and the theory of investment. AmericanEconomic Review 48: 261297.

    Modigliani F, Miller M. 1963. Corporate income taxes and the cost of capital: a correction. American EconomicReview 53: 433443.

    Muller E. 2004. Benets of control, capital structure and company growth. ftp://ftp.zew.de/ (accessed 24May 2010).

    Myers SC, Majluf NS. 1984. Corporate nancing and investment decisions when rms have information thatinvestors do not have. Journal of Financial Economics 13: 187221.Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf

  • Palisade Corporation. 2011. Neural Network Tools. Version 1.0. New York: Palisade Corporation.Panno A. 2003. An empirical investigation on the determinants of capital structure: the UK and Italian enterprise.Applied Financial Economics 13: 97112.

    Pao H-T. 2008. A comparison of neural network and multiple regression analysis in modeling capital structure.Expert Systems with Applications 35: 720727.

    Rajan RG, Zingales L. 1995. What do we know about Capital Structure? Some evidence from international data.Journal of Finance 50: 14211460.

    Ramalho JJS, da Silva JV. 2009. A two-part fractional regression model for the nancial leverage decisions ofmicro, small, medium and large rms. Quantitative Finance 9: 621636.

    Scott Jr JH. 1977. Bankruptcy, secured debt, and optimal capital structure. Journal of Finance 32: 119.Specht DF. 1991. A general regression neural network. IEEE Transactions on Neural Networks 2: 568576.Titman S, Wessels R. 1988. The determinants of capital structure choice. Journal of Finance 43: 119.Tomandl D, Schober A. 2001. A modied general regression neural network (MGRNN) with new, efcient trainingalgorithms as a robust black box-tool for data analysis. Neural Networks 14: 10231034.

    DETERMINANTS OF CAPITAL STRUCTURE IN UK RETAIL INDUSTRY 169Copyright 2012 John Wiley & Sons, Ltd. Intell. Sys. Acc. Fin. Mgmt. 19: 151169 (2012)DOI: 10.1002/isaf