Progressive performance modeling for the strategic determinants of market value in the high-tech oriented SMEs

Int. J. Production Economics 183 (2017) 91–102

Contents lists available at ScienceDirect

Int. J. Production Economics journal homepage: www.elsevier.com/locate/ijpe

crossmark

Progressive performance modeling for the strategic determinants of market value in the high-tech oriented SMEs ⁎

Jooh Leea, , He-Boong Kwonb a b

William G. Rohrer College of Business, Rowan University, 201 Mullica Hill Rd., Glassboro, NJ 08028, USA Hasan School of Business, Colorado State University-Pueblo, 2200 Bonforte Blvd., Pueblo, CO 81001, USA

A R T I C L E I N F O

A BS T RAC T

Keywords: Progressive performance modeling Neural networks R & D intensity Inventory turnover Market value

The purpose of this paper is to present an adaptive performance model using neural networks in scrutinizing the impact of strategic factors on firm performance, especially within high-tech oriented small and medium-sized enterprises (SMEs) in the United States. This paper explores generalized learning of backpropagation neural network (BPNN) in conducting an explanatory and predictive analysis of the strategic determinants of the market value of SMEs. The progressive performance model through BPNN is designed to capture the different and unique significance of strategic determinants for better firm performance by dividing high-tech segments into two performance groups: high performers and low performers. In doing so, this paper introduces a salient BPNN approach for performance modeling and extends the applications of BPNN. Furthermore, efficiency measurement and performance prediction using BPNN adds meaningful value to the literature and highlights the potential advantages of using BPNN. The empirical results demonstrate the successful implementation of the model and clearly distinguish varying patterns at different performance levels, High and Low, which is a significant finding of this study. Overall, sales growth, R & D intensity, and current ratio can be used as major strategic determinants of market value performance of the technology-oriented SMEs.

1. Introduction In a time of increased global competition and economic recession, firms are now more than ever facing the challenge of employing business strategies that may be conducive to their strategic competitive advantages. The role of small and medium enterprises (SMEs) in national economies is increasing, and SMEs' social impact and contributions to the overall industry are significant from the perspectives of employment, wealth creation, innovation, and entrepreneurial leadership. SMEs, as compared to large enterprises, are more stringent in their business performance due to lack of capital and versatile resources. As a consequence, SMEs, especially in the technology oriented industry, strive to minimize slack in their operations. In other words, these SMEs are more conscious of their efficient and effective utilization of limited resources for successful business outcomes (Blackwell et al., 2006; Hsu et al., 2011). In this regard, understanding of key strategic factors and their impact on a firm’s performance is crucial in the decision making process by academicians and practitioners alike, both for short-term profitability and long-term sustainability. The progressive exploration of a firm’s competitive strategic

⁎

determinants for firm performance has become one of the most ongoing topical issues in business and management research. Since market value performance through shareholders’ equity-based finance for SMEs is critical to the fostering of the companies’ survival and growth (Galbraith and Stiles, 1983; Nassr and Wehinger, 2015), it is necessary to identify competitive and distinctive strategic factors in determining market value across different business contexts. But it is not a simple matter to identify the key strategic factors that may significantly affect a firm’s market performance, particularly in SMEs. Although issues may still exist with regard to the strategic determinants of market value, a majority of past studies have been centered on the unique aspect of business strategies of large corporations. Very little attention, until recently, has been paid to the more diverse aspects of strategic determinants of firm performance for SMEs. Therefore, it is questionable whether the traditional notion of the strategic paradigm that might exist in large corporations may be applied to the technologyoriented SMEs. As such, the prospective findings will be noteworthy because they will identify the strategic resource factors that lead to building up more competitive power in given economic contexts, and further provide SMEs with a new paradigm of strategic development and its implica-

Corresponding author. E-mail addresses: [email protected] (J. Lee), [email protected] (H.-B. Kwon).

http://dx.doi.org/10.1016/j.ijpe.2016.10.014 Received 22 March 2016; Received in revised form 21 October 2016; Accepted 22 October 2016 Available online 25 October 2016 0925-5273/ © 2016 Elsevier B.V. All rights reserved.

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

to build a basis for prospective future endeavors. Accordingly, apart from its contributions to laying out a theoretical foundation on market value performance and providing insights into managerial implications, the proposed methodology adds meaningful value to the existing literature. Needless to say, the major findings of this study will be of great significance because they not only discover distinctive strategic factors that are conducive to market performance, but they are also beneficial to top managers who wish to make a better decision to utilize limited resources under competitive market conditions, particularly among SMEs. The remainder of this paper is organized as follows. Section 2 presents a brief introduction to the theoretical significance of the market value of firms and selected strategic variables (exogenous variables) with respect to primary business functions such as finance/ accounting, operations, and marketing. Section 3 provides a methodological foundation of this study, including a brief overview of BPNN. Section 4 describes the design of an empirical model, followed by the results of empirical analysis and managerial implications in Section 5. Conclusions and major findings are presented in Section 6, followed by the contributions of the study and further research implications in Section 7.

tion. It is the conventional notion in today's competitive market that SMEs strive to improve their strategic capabilities and to attain or sustain their competitive market power through diverse driving forces. These forces are mainly R & D-based technology capability (Lee and Habte-Giorgis, 2004; Ito and Pucik, 1993; Zhao and Zou, 2002), salesbased managerial advancement (Guner et al., 2007), return-based financial profit, and cash-based financial strength (Sandner and Block, 2011), credit-based marketing promotion (Guner et al., 2007), and inventory-based lean management (Liberman and Demeester, 1999; Roh and Lee, 2013). More importantly, the enhanced technology strength and capability through more aggressive R & D investment provide the opportunity for companies to improve their market value of equity and increase returns to investors (Chan et al., 1990; Chauvin and Hershey, 1993; Nephilim et al., 2012). Nonetheless, very few studies have empirically examined the strategic significance of the aforementioned factors for a firm’s market value. Accordingly, no previous studies have attempted to further explore the differential impact patterns of strategic factors on overall firm efficiency, particularly with respect to varying performance levels. It is expected that capturing performance profiles of high and low performers and their contrasting impact patterns will enhance our understanding of the influence that key strategic factors have on market value and provide meaningful managerial insights. While high performers can reinforce their efforts in managing key strategic factors for sustainable operations, decision makers in low-performing firms should identify key factors that have to be acted upon in order to advance to the level of their high-performing target peers. Indeed, prudent performance management is a crucial managerial imperative not only for short-term profitability, but also for long-term sustainable operations, regardless of different firm sizes and industry contexts. Thus, firms are facing the challenge of employing business strategies that may be conducive to their strategic competitive advantages. The main question addressed in this study starts with, ‘How do conventional key strategic components (return on assets, R & D intensity, current ratio, sales growth, inventory turnover, and average collection period) influence a firm’s market value in the high-tech oriented SMEs?’ particularly using progressive performance modeling. This study attempts: (1) to examine the strategic significance and effect of selected explanatory factors on a firm’s market performance in a traditional manner; (2) to determine differential performance patterns (i.e., high vs. low performers) based on overall efficiency through predictive segmentation modeling; and (3) to explore strategic determinants on the basis of the different performance level of firms through progressive performance modeling. This study also attempts to explore any potential possibility of performance improvement for lower performers given their current level of resources commitment. In addressing these research initiatives, among others, this study employs backpropagation neural network (BPNN) as a methodological basis for conducting a multitude of explanatory and predictive analyses. With its computational paradigm rooted in the biological neural system, BPNN has its strength in nonlinear functional approximations and is capable of learning a central tendency from limited information even in a changing environment. Indeed, the adaptive and generalized learning property inherent in BPNN provides the sound methodological approach that is required for this unique empirical model, in contrast to conventional statistical methods (e.g., multiple regression) and popular frontier approaches (e.g., data envelopment analysis). Neither of these two methods as a batch processing model is capable of accommodating unseen patterns under changing scenarios due to parametric constraints and deficiency of predictive capacity (Ciampi and Gordini, 2013; Garengo et al., 2005; Hu et al., 1999; Lam, 2004; Rumelhart et al., 1986). Keeping in mind the very sparse previous studies and the lack of a clear theoretical basis regarding the interrelationships of strategic factors in SMEs, this explorative study provides researchers as well as practitioners with a distinct empirical model for progressive analysis

2. Conventional key strategic components for market value performance 2.1. Market value-based performance The operational market strength or power of a firm can be gauged by market valuation, which in turn can be traded, thus indicating the overall health and expansiveness of any company, regardless of scale or type of business. In this sense, market value is a standard measure with which to estimate growth as well as the depth and breadth of a firm’s operational strategy. It seems to be a traditional notion that market value can be a firm’s operational barometer, used to control for some strategic resources under uncertain market conditions. The market value approach to corporate performance, which combines the accounting data of firms with their valuation in the financial market, encompasses the overall economic value with a firm’s collective assets (Sandner and Block, 2011). In fact, market value performance indicates not only the combined tangible and intangible asset value a firm could be sold for as a continuing operating business in the stock marketplace, but also the firm’s operating power to generate positive cash flows in determining the value of the firm’s financial securities. As such, it is argued that the market value aspect of a firm’s performance will provide researchers with a better measure of a firm's real economic performance, capturing the most valuable aspect of the firm (Berger, 1995; Hershey, 1985). The importance of major strategic determinants in assessing a firm's market strength favorably for sustainable market value has been the subject of a great deal of research in the area of business management, regardless of the different size and/or type of firms. Although only a few studies have specifically taken the market value approach, the general findings of these studies appear to indicate that core business strategy factors such as R & D intensity, financial profit, sales growth, current ratio, and inventory turnover can be used as major determinants of a firm's market value in SMEs. 2.2. Selected strategic determinants of market value performance 2.2.1. Research and development (R & D) investment In one of the most traditional studies taking the market-based performance approach, Chauvin and Hershey (1993) empirically demonstrated the impact of research and development (R & D) on the market value of the firm. Like financial significance for current cash flows, R & D investment has a significant and positive influence on the market value and growth of the firm (Acs and Audretsch, 1988; Chauvin and Hershey, 1993; Chen et al., 2005; Kafouros and Mario, 2005; Morbey and Reithner, 1990). In particular, Chauvin and Hershey 92

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

2007; Golmohammadi, 2011; Lam, 2004). In order to explore our proposed research through a valid empirical process, we may need a more innovative and integrative analytical tool. In essence, the desired empirical methodology should be capable of the following: (1) abstracting a central tendency of input and output patterns to determine the relative impact of each strategic factor on market value; (2) assessing the relative performance of each firm in terms of efficiency; and (3) categorizing firms into two performance levels (high vs. low performers) through nonlinear segmentation and comparative analysis. In addition to these core requirements, highlighting the progressive performance analysis, the empirical model should possess an adaptive measurement and prediction capability even for unseen inputs. For this series of empirical demands, BPNN is considered one of the most appropriate intelligent analytic tools. The BPNN model used for this study does not require a priori knowledge of distributions, and it excels in nonlinear modeling even when the theoretical basis with regard to associated variables is weak. Additionally, since it is a nonparametric approach, some traditional screening procedures of normality assumption and linearity issues do not hinder BPNN modeling, which is not the case with OLSMR. Furthermore, BPNN is immune to incomplete information and provides better predictive power for unseen data, thereby facilitating whatif tests. Due to its inherent flexibility and adaptability, BPNN is capable of extracting hidden relationships between variables without establishing hypothetical relationships, unlike OLSMR. Plenty of papers support the superior performance of BPNN relative to the parametric OLSMR method with empirical evidences (Amirali et al., 2010; Cavalieria et al., 2004; Ciampi and Gordini, 2013; Das and Datta, 2007; Geng and Man, 2004; Golmohammadi, 2011; Grønholdt and Martensen, 2005; Hu et al., 1999; Lam, 2004; Mazhar et al., 2007; Sexton et al., 2003; Wong and Chan, 2015). In addition to the prominent modeling aspects discussed so far, the desired empirical method further demands the embedded capability to segment firms into high- and low- performance categories, which requires a relative efficiency measurement as a requisite while taking into account the utilization of strategic factors and the relevant performance of each firm. In this capacity, data envelopment analysis (DEA) has been widely used to measure the relative efficiency of firms denoted as decision-making units (DMUs) with its theoretical basis in linear programing. However, as a frontier method, DEA is not capable of abstracting the central tendency of DMUs; hence, it does not provide the relative importance of independent variables. Moreover, its deficiency in prediction capability has been pointed out as an innate weakness, in spite of its rigorous mathematical foundation (Barros and Wanke, 2014; Kwon and Lee, 2015; Wu et al, 2006; Xu and Wang, 2009). In brief, DEA is not adequate to empirically support the proposed research objectives. Overcoming these shortfalls, we explore BPNN for efficiency measurement and performance segmentation tasks. The potential of BPNN to approximate the relative efficiency of DMUs was first introduced by Athanassopoulos and Curram (1996); this was followed by a few other studies (Santín et al., 2004; Santin, 2008; Wang, 2003). Athanassopoulos and Curram (1996), in their comparative analysis, reported the feasibility of BPNN as an alternative to DEA in measuring the technical efficiency of DMUs. Since then, a handful of combined approaches have proved BPNN effective in learning DEA frontiers (Azadeh et al., 2010; Emrouznejad and Shale, 2009; Kwon, 2014; Kwon and Lee, 2015; Pendharkar and Rodger, 2003). However, the implementation of BPNN as a stand-alone technique for efficiency-driven analysis and its effective utilization for progressive modeling is rarely found in the literature up to this date. Furthermore, just a handful of studies have employed BPNN for SME studies, with most of their focus placed on traditional classification problems such as failure predictions (Ciampi and Gordini, 2013; Kim and Sohn, 2010; Williams, 2016), but with notable exceptions that stress the potential of innovative approaches such as neural networks

(1993) argue that well-targeted R & D efforts made by the smallest firms can be viewed as a strategic form of investment in intangible assets that may have a positive impact on further cash flows and can be highly profitable economically. In investigating the effects of intangible knowledge assets like R & D investment on the market value of firms, Sandner and Block (2011) also supported the market value of assets in this regard. Thus, the conventional notion is that firms are striving to improve their technological capabilities and strengths through R & D investment. In fact, R & D investment is one of the most distinctive traditional strategic determinants in generating favorable market growth and value (Asthana and Zhang, 2006; Sambharya and Lee, 2014), particularly in the high-tech oriented SMEs. Sales growth, as one of the most validated indications of the firm’s market performance in SMEs, is used to estimate the potential and sustainable power to survive in a highly competitive market environment, and also to estimate the potential and sustainable power to survive in a competitive technology-oriented market. Sales growth appears to be one of the strategic determinants differentiating organizational capabilities for market power through product innovation (Uhlaner et al., 2013). It also stimulates potential outside investors by providing a good estimate of a firm’s fair market value among technology-based firms (Bertoni et al., 2013). From the perspective of financial profit and strength, return on assets (ROAA) is commonly used to represent a firm’s financial outcome and effective utilization of input resources for market value creation. Angulo-Ruiz et al. (2014) and Sandner and Block (2011) support the inclusion of ROAA in assessing a firm’s market value due to its explanatory nature, which represents short-term operational performance. In addition to traditionally employed financial performance, current (or liquidity) ratio is used as another financial tool to assess whether a firm has adequate cash to pay debts in order to cope with the high risk that SMEs are likely to face in the market. A well planned cash management strategy may lead to sound financial strength and also improve the financial visibility of decision-making information in SMEs (Lee, 2012). Although their empirical evidence is very limited, two operational strategic factors (average collection period and inventory turnover) are also important in determining market value performance. Average collection period is used to represent the ability of a firm to promote marketing power through credit sales and to maintain its cash position (Guner et al., 2007; Roh and Lee, 2013). In fact, SMEs can take advantage of market power through a shorter average collection period than through a marketable credit policy to attract customers. Inventory turnover is also known as one of the driving operations strategies because it is directly linked to the cash generation as well as to performance. However, its strategic impact on market value performance may vary with different industry contexts and performance measures (Gaur et al., 2005; Liberman and Demeester, 1999; Roh and Lee, 2013). 3. Methodological foundation 3.1. Methodological issues in the study Acknowledging the shallow theoretical basis with regard to market performance associated with strategic factors and the present research gap, the introduction of methodologically sound approaches is still a practical necessity, which may usher in further research and thus promote methodological innovation. In this process, a conventional parametric approach such as multiple regression (OLSMR) often reaches its limit constrained by strict statistical assumptions; hence, it falls short in nonlinear approximation problems, and adaptive learning to accommodate what-if scenarios becomes a challenging task. For these reasons, OLSMR has not been a favored choice in nonlinear prediction modeling and comparative performance analysis studies despite its strength in testing the statistical significance of variables under well-defined theoretical assumptions (Das and Datta, 93

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

Input layer

X

Hidden layer

Wh

H

Output layer

Wo

E=

1 ∑ TK −YK 2 K

2

(4)

Despite its excellence in predictive modeling, BPNN is commonly known as a black box-type learning model, which makes explanatory analysis difficult. That is, the transformation process between input and output is not easily interpretable due to its complicated neural structure and the hidden nature of the learning process. However, observation of output responses to varying inputs provides sufficient explanatory power for the comparative impact analysis intended in this study. Therefore, even the shortcoming of BPNN with respect to explanatory analysis neither diminishes the fitness of the model nor reduces the benefits of adopting BPNN as a prime empirical method, as evidenced in prior studies (Marques et al., 2014; Samilienko and Weistroffer, 2010; Sexton et al., 2003; Wong and Chan, 2015). Besides the advantages of the explanatory features of BPNN, a notable benefit of using BPNN in this study lies in exploring its predictive capacity as a methodological basis for efficiency-driven performance categorization and progressive impact analysis, as discussed earlier. In this pilot study, BPNN is expected to detect differential performance patterns with respect to market performance in SMEs and result in intuitive managerial implications.

Y

X: Inputs, H: Hidden outputs, Y: Output Wh: Weights (input-hidden layers) Wo: Weights (hidden-output layer) Fig. 1. Three-layer BPNN structure.

(Abouzeedan and Busler, 2004; Delen et al, 2013). 4. Design of the empirical model 3.2. Structure of BPNN

The proposed empirical model involves three stages of sequential BPNN processes, as depicted in Fig. 2. In this empirical study, the model is designed to investigate the performance patterns of SMEs in high-tech oriented industries from the perspective of the impact of strategic variables on model performance. In particular, this paper aims to explore varying patterns of SMEs in two different performance categories, High and Low, as differentiated by relative efficiency. The empirical model consists of following a sequential process: (1) The first stage is a preliminary analysis to assess the impact of strategic factors on market value performance, and it aims to extract the average performance patterns of all SMEs using OLSMR and BPNN (BPNN 1) for comparative analysis; (2) The second stage is predictive segmentation modeling to categorize SMEs into two different performance levels (BPNN 2); In this capacity, BPNN assesses the relative efficiency of each DMU, and splits DMUs into two efficiency groups, representing high and low performers; (3) Finally, the third stage is progressive performance modeling for further investigation of potential performance determinants for each performance category. This stage requires implementation of two subsequent BPNN models (BPNN_H and BPNN_L) by taking inputs from each performance category segmented by BPNN 2. As a result, the BPNN models capture contrasting performance profiles of each category and unveil its differential impact patterns of strategic factors. In addition to inputfocused explanatory analysis, an adaptive prediction scheme provides output-centered performance improvement options for low performers. In essence, the approach we present provides a methodological breakthrough for efficiency-based performance analysis using BPNN, and lays out an empirical foundation for a series of progressive modeling to support the efficient control of strategic resources and effective management of performance outcomes.

An artificial neural network (ANN) is an intelligent analytical tool inspired by modeling the biological neurons of the human brain with its computational basis in connectionism. BPNN, as one of the most widely used ANN models based on a backpropagation learning algorithm, has a multi-layered structure with hidden layers embedded between the input and output layers (Fausett, 1994; Kourentzes, 2013; Rumelhart et al., 1986). As a nonparametric method, the strength of BPNN resides in its capability to capture nonlinear relationships between input and output variables without a priori knowledge of distributions. The learning process involves the repetitive processes of information feed-forward, error backpropagation, and weight adjustments between neurons in adjacent layers. After the completion of learning, the weights preserve learned information and serve as a key code to map the presented inputs into appropriate classes or levels depending on the characteristics of the problem. In this sense, BPNN learning is an autonomous process to approximating the production functions of given data regulated by hidden neurons. Fig. 1 shows a typical three-layer BPNN structure that can be used for multivariate prediction problems. A single dimensional output ‘Y′ is a nonlinear prediction for input vector (X) through hidden outputs (H), and can be expressed as:

Y = fo (Wo⋅fh (Wh⋅X))

(1)

where, Wh and Wo represent weight vectors interconnecting inputhidden and hidden-output neurons and fh and fo are sigmoid transfer functions applied to net outputs of neurons as determined by the inner product (∙) of the respective input and weight vectors. Accordingly, the output for any neuron k in a hidden or output layer can be determined by the following formulas:

hk = {1 + exp[−(Wh ∙X)]}−1

(2)

5. Empirical analysis and discussion

yk = {1 + exp[−(Wo ∙H)]}−1

(3)

5.1. Sample data and collection For the empirical test, our initial data sample was composed of a cross section of the required variables from small- and medium-sized U.S. firms (SMEs) based on the number of employees being less than 500 in 2014. We extracted all our data from the S & P Research Insight database. Finance-related firms (SIC 6000–6999) were initially eliminated for the generalizability of the results of this study with respect to

The BPNN learning process aims to reduce the minimum sum of squares error; accordingly, at the end of the feedforward process, the network backpropagates calculated error (E) between target (Tk) and actual outputs (Yk) for gradual and iterative weight adjustments (Azadeh et al., 2010; Fausett, 1994; Rumelhart et al., 1986; Wang, 2003). 94

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

Input

Model

Outcomes

SME_ALL

OLSMR

Average pattern OLSMR vs. BPNN

BPNN 1

144 SMEs in High-tech industry

Categorization of SMEs High vs. Low

BPNN 2

Pattern High

SME_HIGH

BPNN_H

SME_LOW

BPNN_L

Pattern Low

Note: 1. OLSMR – Ordinary Least Square Multiple Regression 2. BPNN – Backpropagation Neural Network Fig. 2. Schematic diagram: three-stage BPNN modeling. Note: 1. OLSMR – Ordinary Least Square Multiple Regression. 2. BPNN – Backpropagation Neural Network.

5.2. Description and measurement of variables

Table 1 High-tech industry specification by SIC-codes. Industry classification

SIC

No. of firms

Chemicals and allied products Fabricated metal products Machinery and computer equipment Electrical products Transportation equipment Scientific instruments Tech related business services Total

2800–2899 3300–3499 3500–3599 3600–3699 3700–3799 3800–3899 7370–7379

19 6 21 37 2 49 10 144

First, we employed market value in order to measure the amount for which a firm could be sold as an ongoing business in the marketplace and also to estimate the firm's market power to generate positive cash flows in determining the value of the firm's financial securities. Market value for a firm used in this study is measured by the natural logarithmic value of a company's total market value based on market capitalism (i.e. market value of equity). MRKVAL = Ln (Yearend closing stock price)*(Common shares outstanding). In addition, six potential strategic business factors used in our study are operationalized as follows: Return on assets (ROAA) = Net profit after tax/Total assets. R & D intensity (R & DINT) = Book value of R & D expenditure/ Total sales. Current ratio (CURATIO) = Current assets/Current liabilities. Sales growth (SGROWTH) = (Total sales t – Total sales t-1)/Total sales t-1. Inventory turnover (INVTURN) = Cost of goods sold/Average inventory between t and t-1. Average collection period (ACOLLECT) = [(Accounts receivables)*360]/Total sales.

the high-tech oriented firms. In addition, more firms with over $200 million in sales volume were also eliminated for the purpose of this study. Although there are many different criteria in defining high technology, this study employs four-digit SIC codes (ref. Chauvin and Hershey, 1993; Hadlock et al., 1991), as illustrated in Table 1. After removing additional companies with outlying and missing values in selected variables, 144 high-tech oriented SMEs were finally selected (see Table 1). In order to minimize the floating effect from year-to-year data, all the variables used in this study were measured as five-year aggregated averages for the period 2009 through 2013. Therefore, the potentially biased influence of any single year on the results of this study was avoided.

Table 2 (A) Descriptive statistics and normality tests. Variables

Mean

Std. Dev.

Median

Mode

Min.–Max.

Skewness

Kurtosis

Shapiro-Wilks (p-value)

Market Value (Ln) Return on Assets R & D Intensity Current Ratio Sales Growth Inventory Turnover Avg. Collection Period

3.776 0.261 0.197 3.521 43.182 6.508 59.153

1.29 0.21 0.24 1.94 28.41 5.56 21.23

3.690 0.175 0.098 3.518 36.619 3.476 62.688

3.400 0.015 0.004 2.772 7.096 0.886 6.512

1.770–6.653 0.002–0.843 0.005–1.561 0.899–10.52 7.096–92.67 0.886–9.306 6.512–171.4

0.043 1.170 1.609 1.023 0.480 1.201 1.509

−0.403 1.148 2.182 0.931 0.322 1.210 5.384

0.993 0.894 0.727 0.851 0.923 0.491 0.898

(0.699) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000)

Note: 1. Shapiro-Wilds for normality test. If the significance value of the Shapiro-Wilks test is greater than 0.05, the data is normal. 2. Return on Assets (Rate return on assets), R & D Intensity (= Ratio of R & D expenditures to total sales), Current Ratio (= Ratio of current assets to current liabilities), Sales Growth (Growth rate in sales), Inventory Turnover (= Number of times a company's inventory is sold over a period), Average Collection Period (= Number of days that it takes a company to collect payments owed from its customer or clients after sales), and Market Value (= Natural logarithmic value of a company’s market value).

95

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

significance of strategic factors for market performance through the linear regression analysis. In addition to the descriptive statistics, Table 2-B presents correlations among all the variables and reveals a strong and consistent relationship across the primary variables of the study. The correlation matrix illustrates the interrelationship among all the exogenous strategic variables (ROAA, R & DINT, CURATIO, INVTURN, ACOLLECT and SGROWTH) and the endogenous variable (MRKVAL). Overall, most of the strategic variables except inventory turnover (INVTURN) and average collection period (ACOLLECT) are positively correlated with the market value of the firm at the 0.01 level.

Table 2 (B) Correlations.a Variables 1 2 3 4 5 6 7

Return on Assets R & D Intensity Current Ratio Sales Growth Inventory Turnover Average Collection Period Market Value (Ln)

1

2

3

4

5

0.294*** −0.047 0.022 0.074

0.219** 0.107 0.022

0.028 −0.151*

−0.010

−0.090

−0.010

−0.128

−0.058

0.022

0.210**

0.388***

0.324**

0.664***

0.092

6

−0.030

Note: Two-tailed tests. a n =144. * P < 0.05. ** P < 0.01. *** P < 0.001.

5.4. Preliminary analysis: OLSMR vs. BPNN As a first step in assessing the impact of strategic factors on market performance of SMEs, an OLSMR analysis was conducted and compared with BPNN outputs prior to exploring the characteristics of SMEs in two different categories, high and low performance. In fact, the OLSMR analysis was conducted to serve as a basis for comparison with the outputs of BPNN. Both models generated valuable information on the overall patterns of SMEs in terms of average performance. As presented in Table 3, our OLSMR model is highly and significantly adequate to provide meaningful information about the degree of explained variation of market value (R2=0.6382; F=40.27, p < 0.001). As proposed, most strategic factors except average collection period (ACOLLECT) are statistically significant (at least at 5% level of significance) and positively associated with market value (MRKVAL) in the high-tech oriented SMEs. In addition, there are no signs of multicollinearity problems since none of the variance inflationary factors (V.I.F.s) for any explanatory variables is greater than 1.5 (i.e. The V.I.F.s for our employed explanatory variables are between 1.015 and 1.182). As a rule of thumb, the suggested cutoff for a multicollinearity problem is that the V.I.F. of each variable should not exceed 10 (Gujarati and Porter, 2009; Hair et al., 2010a, 2010b). Following the OLSMR analysis, BPNN was employed, utilizing the same data set. For this experiment, the data set was split into a 7:3 ratio for training and test subsets, where test data was used to prevent over-fitting of the model during the network training process. The BPNN and OLSMR models were both built on the same number of DMUs, 144 SMEs, thus facilitating a comparison of the two different methods. The resulting BPNN model has a 6-10-1 structure with 10 hidden neurons. Then, a comparative prediction analysis was conducted for both OLSMR and BPNN models by using various performance metrics such as correlation between actual and predicted value (Pearson R), mean absolute error (MAE), mean absolute percentage error (MAPE), and the number of DMUs within 20% (and 25%)

5.3. Descriptive statistics, normality tests, and correlations As a preliminary step toward empirical analysis, Table 2-A and B provides descriptive statistics and correlations for our variables. The descriptive statistics with an inclusion of skewness and kurtosis for normality tests in addition to means, medians, modes, standard deviations, and the minimum and maximum values of each variable is presented to show the feasibility of our empirical model. Although the primary purpose of this study is not to test and/or prove the statistical inferences in determining the strategic factors for market performance, it may be conducive to the adoption of BPNN modeling to assess the consequence of normality assumption of variables on the interpretation of the empirical results (Hair et al., 2010a, 2010b). As presented in Table 2, only the transformed market value for performance measure (Skewness=0.043, Kurtosis=−0.403, the SharpiroWilks=0.993 and P > 0.05) is reasonably normally distributed, whereas all strategic variables are relatively peaked and positively skewed. As such, the normality of all strategy variables is strongly rejected by the Shapiro-Wilks test and the data significantly deviate from a normal distribution (P < 0.001). Due to non-normality on the employed strategic variables, we also examine the linear relationship between market value and other strategic factors through scatterplots and one way ANOVA (ref. linearity test through SPSS). But only four out of the six strategic variables (i.e. Return on assets, R & D intensity, Sales growth, and Current ratio) show a strong evidence of linearity at the 0.001 level. Nevertheless, the assessment of the descriptive statistics appears to be enough to ensure that all values for each of the variables are valid for examining our empirical models. As with all reasonably large samples of 100 or more cases, the significant level of skewness may not be as important as its actual size and underestimation of variance associated with positive kurtosis may be negligible (Hair et al., 2010a, 2010b; Tabachnick and Fidel, 2013). Quite differently from the standardized OLSMR analysis, artificial neural network models do not rest on the general assumption of normality or linearity for the relationships between the dependent variable and independent variables (Sharma et al., 2003). In this study, the empirical implementation of BPNN is our main interest, particularly in discovering the strategic determinants of market performance. It is believed that neural network models are preferable to linear regression models, particularly when there exists some evidence of nonnormality or nonlinearity in financial and economic variables (Altay and Satman, 2005; Qi, 1999; Sharma et al., 2003). However, an empirical implementation of BPNN along with an OLSMR model is an ideal choice for flexible and preferable modeling with nonsymmetrical variables in determining key strategic factors for different level of performance. Accordingly, despite advantages of neural networks in neutralizing the severity of nonnormality and nonlinearity problems, our study initially explored the statistical

Table 3 Results of OLS multiple regression (OLSMR) analysis. Variables

(Constant) Return on assets R & D intensity Current ratio Sales growth Inventory turnover Average collection period

Unstandardized β

Std. ε

1.300 0.700 1.147 0.149 0.028 0.021 0.003

(0.273) (0.268) (0.302) (0.028) (0.002) (0.009) (0.003)

Stand. β

t-value

0.143 0.212 0.293 0.634 0.127 0.056

4.761 2.613 3.794 5.408 12.246 2.429 1.080

Sig. Level

V.I.F.

*

1.127 1.182 1.113 1.015 1.030 1.033

** *** *** *

R2 = 0.6382; Adj. R2 = 0.6223 ; F-Ratio = 40.269*** Note: V.I.F. indicates Variance Inflationary Factor to detect multicollinearity problems. * P < 0.05. ** P < 0.01. *** P < 0.001.

96

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

Table 4 Predictive performance: OLSMR vs. BPNN.

Table 5 Predictive performance: BPNN-2.

Model

Data

R

MAE

MAPE

< 20% error

< 25% error

Records

Model

Data

R

MAE

MAPE

< 20% error

< 25% error

Records

OLSMR BPNN 1 (610-1)

All Train Test All

0.797 0.849 0.870 0.869

0.596 0.516 0.520 0.518

16.9% 16.6% 16.7% 16.6%

66.0% 75.0% 84.0% 75.0%

75.0% 75.0% 86.4% 84.0%

144 100 44 144

BPNN 2 (148-1)

Train Test Valid. All

0.886 0.882 0.890 0.886

0.450 0.487 0.500 0.480

14.6% 13.9% 15.8% 14.6%

77.6% 76.0% 71.4% 76.4%

89.4% 80.0% 81.0% 84.0%

98 25 21 144

Note: MAE (Mean abs. error), MAPE (Mean abs. percentage error).

Note: MAE (Mean abs. error), MAPE (Mean abs. percentage error).

prediction errors. Table 4 summarizes the prediction performance of OLSMR and BPNN (titled BPNN-1). Both models show high prediction performance, but BPNN demonstrates superior outcomes across all performance metrics, which is in line with the extant literature. Although both OLSMR and BPNN models can approximate models using a curve-fitting approach, BPNN produced better results without a priori information on input distributions. In addition to the predictive analysis of both models, the impact of independent variables on the output was also assessed. In this measure, relative variations of BPNN output in accordance with input changes were measured as a partial derivative of an input to the output. These results were then compared with the OLSMR output, the normalized β (standardized regression coefficients). Fig. 3 illustrates the comparative marginal impact of strategic variables on the market value of the firm. Interestingly enough, the relative impact of these variables shows similar patterns in both models. As in the OLSMR experiment, BPNN identifies sales growth (SGROWTH) as the most impactful strategic variable, followed by current ratio (CURATIO) and R & D intensity (R & DINT). Inventory turnover (INVTURN) and return on assets (ROAA) are less impactful, and average collection period (ACOLLECT) shows minimal impact, which is consistent with OLSMR in that ACOLLECT is the only insignificant variable. Even though BPNN is not designed directly to determine the statistical significance of variables like OLSMR, it can detect the relative impact of input variables by observing the network output as a response to varying inputs. Accordingly, the shortcomings of BPNN neither disrupt nor diminish its capacity, as demonstrated in this analysis. Rather, it provides a sound basis for further explanatory and predictive exploration of unknown impact patterns and further progression to the strategic performance analysis.

exhibits comparable performance across different data subsets (training- test-validation). In this learning paradigm, the network should deal with the ‘generalization’ and ‘memorization’ dilemma (Pendharkar and Rodger, 2003). In other words, the trained network should be capable of extracting general patterns of data instead of memorizing specifics of individual information in order to prevent over-fitting and to accommodate new or unseen inputs. In this experiment, comparable performance on the unseen validation subset is an indication of the achieved generalization of the designed network. This model (BPNN-2) plays a crucial role in differentiating our selected SMEs into two groups: High (or above average) and Low (below average) performers. By using multiple inputs and outputs, the BPNN model forms a nonlinear separation surface, which is an approximation of a central tendency or a general pattern of the given data set. Consequently, a positive prediction error between actual and prediction value indicates high performance, while a negative prediction error indicates low performance. Accordingly, the discrepancy between actual and prediction values provides a basis for categorizing SMEs into high and low performers as shown in Fig. 4. Besides categorization of DMUs, the prediction result can determine a rank and a relative efficiency score of each DMU (Azadeh et al., 2010; Emrouznejad and Shale, 2009: Santín et al., 2004). For DMUJ, the ratio of actual (ya_J) and prediction (ynn_J) values indicates relative performance under multiple input-output settings. Stated in a different manner, ya_J / ynn_J represents performance of DMUJ, or efficiency as a scalar representation of performance designated as the efficiency score (ES). Accordingly, the relative value of ES provides the ranked order of the DMUs under evaluation. The efficiency scores are normalized to represent DMUs with maximum efficiency as 1. Mathematically, the efficiency of DMUJ can be expressed as follows:

5.5. Predictive segmentation modeling for differential performance

⎞−1 ⎛ ESJ = ( ya J / ynn J )• ⎜M ax( ya i / ynn i) ⎟ ⎠ ⎝ i

(5)

In this experiment, the BPNN model identified 73 High and 71 Low SMEs, demonstrating the centralized learning capacity of BPNNs. Fig. 5 displays the efficiency patterns of DMUs determined by BPNN 2. This figure contrasts High and Low SMEs with the average efficiency of 0.757 and 0.576 for each category separated by a cutoff efficiency of 0.670 (the highest score from Low SMEs). This experiment highlights the potential advantages of using BPNN for efficiency analysis and natural performance segmentation without necessarily devising a partitioning mechanism or relying on an arbitrary categorization of DMUs.

Output

Strategic variables

Following up the comparative analysis of OLSMR and the initial BPNN model (BPNN 1), the superior predictive power of BPNN prompted further refinement of the model for the prediction experiment in generating the performance patterns of SMEs in different performance categories. As a predictive segmentation model for use in categorizing the selected firms into two performance groups, the input preprocessing function of the software package (NeuralWare, 2013) was utilized, and the data set was split into training (68%), test (17%), and validation (15%) subsets. The model (BPNN-2) has a 14-8-1 structure, and the training result in Table 5 shows an improvement across all performance measures as compared to the initial BPNN model (BPNN-1). As shown in Table 5, the predictive performance SGROWTH CURATIO

High (Above average)

R&DINT INVTURN ROAA

BPNN

ACOLLECT

0%

20%

40% 60% Normalized importance

80%

OLSMR

Low (Below average)

100%

Fig. 3. Comparison of marginal significance of strategic variables for market value.

Input

Fig. 4. Graphical representation of BPNN frontier.

97

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

High

Low

Avg_High

60%

Avg_Low

Prediction error

Relative efficiency

1.0

0.7

0.3 1

21

41 61 81 101 DMUs (Sorted by efficiency)

121

High

Low

All DMUs

30% 0% 1

21

41

61

81

101

121

141

-30% DMUs (sorted by error scale)

141

-60% Fig. 6. Centralized learning and prediction error.

Fig. 5. BPNN efficiency of DMUs.

5.6. Progressive performance modeling: high vs. low performers Performance

5.6.1. Robustness of the model For further investigation of strategic determinants according to the different performance levels, additional BPNN models (BPNN-H/L) were implemented. As clearly shown in Table 5, the BPNN-2 model produced promising results, providing a sound methodological basis for further elaboration of the method. This experiment in particular aims to probe the differential performance patterns of high and low performers and extract their relative impact profiles of strategic factors. In fact, this is a meaningful research agenda, but one that has been neglected so far partly due to a lack of proper methodology. In this stage of the research advancement, average collection period, which showed a minimal impact in BPNN and no significance in OLSMR, was eliminated to improve the parsimony of the model. For each category of high and low SMEs, new BPNN models (BPNN-H and BPNN-L) were implemented, and their noteworthy performance is summarized in Table 6. Indeed, the result shows a high correlation coefficient (above 0.938) and low MAPE (less than 8.6%), with the majority of the DMUs (more than 95%) presenting less than 20% error in both models. In addition, comparable performance measures across different data sets represent generalized network learning in both models as discussed earlier, and the results based on segmented data show high-precision prediction performance as evidenced by error terms. By building models on the subset of high and low SMEs separated by a prior BPNN model (BPNN-2), both BPNN-H and BPNN-L models take on inputs of less fluctuation. In other words, the subsets of data preserve better monotonicity, which is an important aspect for the stable learning of BPNN and a consequent robust model (Archer and Wang, 1993; Pendharkar, 2005). Thus, the two later models learned the central tendency of SMEs in high and low performance. Fig. 6 is a visual representation of the improvement in terms of prediction errors as compared to the results of BPNN-2, which was previously built on all SMEs in both performance categories. The figure clearly exhibits the centralized learning property of BPNNs, and also shows smaller error obtained from each performance segment, with most of the DMUs at less than 20% error. The empirical analysis presented so far can be summarized as a sequential BPNN process that analyzes the average performance of

R

MAE

MAPE

< 20% error

< 25% error

Records

BPNN-H (High) 10-7-1

Train Test All

0.940 0.933 0.938

0.301 0.340 0.313

8.6% 7.2% 7.7%

96.1% 95.6% 95.9%

100.0% 100.0% 100.0%

51 22 73

BPNN-L (Low) 10-7-1

Train Test All

0.978 0.978 0.978

0.159 0.188 0.168

5.4% 6.6% 5.8%

95.9% 95.5% 95.8%

98.0% 100.0% 98.6%

49 22 71

0.7

OLSMR

BPNN-1

BPNN-2 BPNN-H Experiments

BPNN-L

Fig. 7. Visual summary of experiment results.

SMEs and the subsequent advancement into the hierarchical analysis of two performance patterns. Fig. 7 summarizes the aggregated performance of all the experiment models presented thus far in terms of correlation coefficient (R) and normalized MAPE. A noteworthy aspect is the remarkable improvement in performance and predictive power revealed by BPNN-H and BPNN-L. 5.6.2. Explanatory impact analysis The promising outcomes of sequential BPNN models indicate the successful estimation of the production functions of SMEs capturing complex relationships of input-output variables. In this black box-type learning mechanism, network responses to the changing inputs provide an effective means to observe the functional relationships of variables. Fig. 8 presents the relative importance of variables in both models (BPNN-H and BPNN-L), which represents output variations on the uniform change of individual input. The result clearly demonstrates the fruitful outcomes of the proposed neural network approach as evidenced by the differing impact of strategic variables within two different performance categories, High and Low. Admitting that sales growth (SGROWTH) is the most impactful input in both performance categories, a notable distinction between high and low performers can be observed from the relative importance of R & D intensity (R & DINT). In low-performing SMEs, R & D intensity (RNDINT) is the third most important strategic variable after sales growth (SGROWTH) and current ratio (CURATIO). However, in highperforming SMEs, R & D intensity is a crucial factor comparable to sales growth (94%), in contrast to 32% in low-performing SMEs. Furthermore, the result explicitly shows increasing returns on R & D intensity in high-performing SMEs over low-performing SMEs (e.g., 1.93% vs. 1.32%), indicating a more than 46% impact on output. It is a 3.5% Impact on output

Data

MAPE (norm.)

0.3

Table 6 Predictive performance: BPNN-H & BPNN-L. Model

Pearson R

1.0

2.5% 1.5% 0.5% -0.5% Low High

SGROWTH 3.45% 2.05%

R&DINT 1.32% 1.93%

CURATIO 1.35% 1.00%

ROAA 0.88% 0.50%

Fig. 8. Relative importance of variables.

Note: MAE (Mean abs. error), MAPE (Mean abs. percentage error).

98

INVTURN -0.39% 0.02%

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

significant finding obtained from this empirical study. R & D intensity not only has a significant strategic influence on the market value of the firm in SMEs, but also has a positive impact on long-term performance. This supports the idea that R & D investment can be considered one of the most significant strategic determinants of market performance through technology advancement and scientific innovation leading to sustainable competitive advantage (Morbey and Reithner, 1990; Tubbs, 2007; Verma and Sinha, 2002; Wang et al., 2013). It is evident that R & D intensity has more significant and positive effects on market value of high-performing firms than on low-performing firms in the SMEs under study. In contrast to previous studies of large corporations (Gaur et al., 2005; Roh and Lee, 2013), our findings show that inventory turnover (INVTURN) seems to be the least impactful strategic factor, followed by return on assets (ROAA) across SMEs, regardless of different performance. Even though inventory turnover shows minimal effect on firm performance, the present approach captures a contrasting impact on two different performance groups: negative in low-performing SMEs, but marginally positive in high-performing SMEs. Overall, the results recall previous research findings that the inventory turnover alone cannot properly evaluate inventory performance of a firm (Gaur et al., 2005), and that even the lean inventory itself is not likely to have a significant impact on performance (Eroglu and Hofer, 2011). Indeed, despite the fact that inventory is one of the key resources for lean management targets, it has been argued that the impact of inventory on firm performance appears to be different depending on firm characteristics across industries (Dedeke and Eroglu, 2015). Since smaller firms are naturally more concerned with competition than larger corporations, managerial efforts to improve inventory management should be made through the synergistic advantages of operational tools such as accurate market prediction, just-in-time (JIT), and supply chain management in SMEs.

Table 7 Potential improvement for low SMEs. EFF. tier (DMUs)

Below 0.4 ( 2) Below 0.5 ( 6) Below 0.6 (27) Below 0.7 (36) Average (71 DMUs)

Current performance (Low)

Target performance (High*)

Eff. (Avg.)

MKTVAL

MKTVAL

Improvement

Remark

0.323

0.904

2.613

191%

Desired Efficiency

0.474

2.362

3.646

56%

0.544

2.949

4.073

43%

0.632

3.696

4.369

21%

0.576

3.221

4.146

37%

0.757*

involves relative comparisons with peer entities based on available strategic resources and expected strategic outcomes commonly represented by selected input-output variables. Accordingly, integrative performance assessment of firms demands input-centered explanatory analysis and output-focused predictive analysis as a dual process following the careful selection of strategic resources and key output measures. Most conventional approaches, however, have dealt with partial process either through significance analysis of inputs using parametric statistical methods or through optimal output measures relying on nonparametric extreme point approaches. Under this scheme, managerial utility is narrow in scope: At best, managers can either assess statistical significance of strategic factors on average performance or determine optimal outputs for firms, thus limiting benefits to managers in a practical sense. Overcoming these shortfalls, the proposed streamlined approach provides practical advantages to managers as a holistic performance analysis model. First, managers can be aware of performance positioning of a firm, above or below average, determined by relative efficiency as a proxy for resource utilization of firms. The efficiency-based segmentation takes both strategic resources and outputs into account as compared to arbitrary segmentation using output-only measures. Second, based upon a segmentation approach, managers perceive the comparative impact of strategic resources based on the output between two performance groups. Accordingly, managers can set prudent future goals through selective commitment of rare resources. As presented in this study, comparatively higher impact of R & D intensity on the above-average (High) performers is a clear indication of the strategic importance of R & D, which sparks managerial attention for belowaverage performers. Third, in addition to perceiving the relative impact of resources, managers can set desired performance goals to advance to the above-average performance for committed resources. Indeed, the ability to set actionable above-average performance in contrast to unreachable best practice might be a significant and practical decision aid for managers in high-tech oriented SMEs, especially under a rapidly changing competitive environment. As a result, managers are equipped with multifaceted strategic tools for efficient resource control and effective goal setting.

5.6.3. Predictive target setting An intriguing question that arises from the input-focused impact analysis concerns how much improvement needs to be made for lowperforming SMEs to achieve performance comparable to that of highperforming peers, given the same amount of committed resources. Besides being conscious of differential impact factors, the setting of target performance through efficient utilization of strategic resources is another crucial challenge for managers in high-tech SMEs. In this managerial advancement, BPNN provides an adaptive means to support what–if tests, especially for underperforming SMEs (Low) in pursuit of above-average performance (High). Utilizing efficiency scores previously obtained by BPNN2, the low-performing SMEs were further broken down into the varying efficiency (EFF) tiers (below 0.4, 0.5, 0.6 and 0.7), and desired performance outcomes (market value) for each tier were predicted by running the BPNN-H model. Table 7 summarizes the predictive experiment results. As shown in the table, SMEs in lower tiers need to achieve more improvement, with a decreasing rate of necessary improvement for higher tiers. On average, the BPNN-H model suggests a 37% potential improvement of market value for low-performing SMEs (average efficiency of 0.576) in order to achieve performance comparable to that of high-performing SMEs (average efficiency of 0.757). In addition to measuring efficiency as depicted in Fig. 5, the BPNN-H model estimates desired target performance for less efficient SMEs. Taking advantage of an innovative methodological advancement, this study not only explores differential impact patterns of key strategic factors in SMEs, but also presents an intuitive improvement option to maximize the firm’s interest that may be achieved through the efficient utilization of scarce strategic assets.

6. Discussion of major findings In an attempt to explore the crucial factors of market performance, this study demonstrates the successful implementation of progressive sequential modeling using artificial neural networks. Further, the proposed empirical model clearly distinguishes varying performance patterns of two different strategic groups (i.e. above-and below-average performers) within the high-tech oriented SMEs. This study extensively

5.7. Managerial implications Strategic decision making is of a comparative nature and typically 99

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

2005). To the best of the authors’ knowledge, this paper is the first attempt to use BPNN as a standalone method to conduct a series of multi-task which includes efficiency measurement, segmentation of strategic groups, variables impact analysis, and output predictions. From a performance-benchmarking aspect, an efficiency-based segmentation approach using BPNN provides a sound basis for promoting above-average performance as opposed to conventional superioritydriven best practice benchmarking (Cox et al., 1997; Dai and Kuosmanen, 2014). In the midst of best-driven benchmarking avenue, flexible and achievable improvement mechanisms supported by sound methodology can provide managers with actionable and achievable targets for improvement. In this sense, the proposed study potentially contributes to the benchmarking literature. In addition to methodological advancement and subsequent potential for decision support, this study pioneers investigation of strategic determinants of SMEs in terms of their relative importance and potential impact on market value, especially by exploring different performance patterns in two performance groups. Indeed, managerial awareness on key strategic factors (e.g., R & D) in conjunction with the ability to set achievable target goals will provide a two-pronged strategic tool for managers in allocating internal resources to optimize external market performance. In the future, this study may be extended to include large enterprises with a bigger sample size for the generalization of the proposed methodology. The utilization of additional key variables is worth exploring. In addition, the development of a theoretical basis with regard to these variables will result in meaningful contributions to the literature. It may also be beneficial for future research to perform analyses of large corporations through predictive exploration of performance patterns in different levels of performance categories, and then to identify whether the significant impact on various dimensions of firm performance varies across different industry contexts, particularly with respect to R & D activities. Follow-up studies should further investigate how R & D and other potential strategic factors (particularly inventory leanness) can be conducive to a firm’s competitive success in today’s ever-changing global markets.

explores the explanatory and predictive power of a standalone BPNN in discovering the relative impact of strategic resources on the market value performance through an integrative implementation of the salient empirical model. As previously indicated, few studies have empirically attempted to investigate the impact of conventional strategic factors including R & D intensity on different levels of performance. Given that very limited findings are available in the extant literature, rather than developing a whole new theory, this study attempts to provide a new perspective into the exploration of strategic determinants of firm performance in terms of prospective impact patterns under varying performance levels. The knowledge of varying impact patterns of core resources and their predictive outcomes can provide new insights into utilizing the firm's strategic resources within SMEs. Our study supports the notion that sales growth, R & D intensity, current ratio, and return on assets are most likely to be the major strategic determinants of the market value of SMEs in all performance levels. R & D intensity shows a very comparable impact on market performance to sales growth in the case of above-average performers. Accordingly, R & D intensity can be used as one of the most critical strategic factors of market value performance in the technology oriented SMEs, particularly in high-performing firms and always with potential for further improvement. Indeed, among other things, this pilot study reveals a greater and more significant effect of R & D intensity on the market value of high-performing SMEs than on that of low-performing SMEs, thus reinstating strategic significance of R & D investment in SMEs. The vivid contrast of R & D effects on the market performance of two different strategic groups of SMEs is one of the most noteworthy observations of this study. The findings of our study are in line with prior studies which stress a crucial role of R & D investment in increasing innovative capability that may impact longterm performance and stock-holder return in the high-tech oriented business environments (Eberhart et al., 2004; Kraiczy et al., 2015; Mudambi and Swift, 2014; Wang et al., 2013). Noticeably, inventory turnover does not show any significant impact on the market value of SMEs. Even without a general consensus on the effects of inventory, this is particularly in contrast to the findings of prior studies in which inventory turnover ratio has been significantly associated with market value (Roh and Lee, 2013) and financial profitability in the small-sized firms (Koumanakos, 2008; Shin et al., 2015). The major findings of our study will provide both academicians and practitioners with new insights into the strategic direction toward managerial advancement in developing crucial components of a firm’s economic performance, including sustainable market value, in the technology-oriented SMEs and even larger corporations. Notwithstanding certain limitations in the selection of impactful variables and performance measures, and also in the supporting literature, the sequential BPNN modeling introduced in this study can be extended to explore the strategic paradigm of efficiency determination of firms under the competitive market environment.

References Abouzeedan, A., Busler, M., 2004. Typology analysis of performance models of small and medium-size enterprises (SMEs). J. Int. Entrep. 2 (1/2), 155–177. Acs, Z.J., Audretsch, D.B., 1988. Innovation in large and small firms: an empirical analysis. Am. Econ. Rev. 78 (4), 678–690. Altay, E., Satman, M.H., 2005. Stock market forecasting: artificial neural network and linear regression comparison in an emerging market. J. Financ. Manag. Anal. 18 (2), 18–33. Amirali, M., Bright, J., Bird, K., 2010. A comparison of artificial neural networks and multiple regression in the context of research on personality and work performance. Organ. Res. Methods 13 (3), 540–561. Angulo-Ruiz, F., Donthu, N., Prior, D., Rialp, J., 2014. The financial contribution of customer oriented marketing capability. J. Acad. Mark. Sci. 42 (4), 380–399. Archer, N., Wang, S., 1993. Application of the back propagation neural network algorithm with monotonicity constraints for two-group classification problems. Decis. Sci. 24 (1), 60–75. Asthana, S., Zhang, Y., 2006. Effect of R & D investments on persistence of abnormal earnings. Rev. Account. Financ. 5 (2), 124–139. Athanassopoulos, A.D., Curram, S.P., 1996. A comparison of data envelopment analysis and artificial neural networks as tools for assessing. J. Oper. Res. Soc. 47 (8), 1000–1016. Azadeh, A., Saberi, M., Anvari, M., 2010. An integrated artificial neural network algorithm for performance assessment and optimization of decision making units. Expert Syst. Appl. 36 (8), 5688–5697. Barros, C., Wanke, P., 2014. Insurance companies in Mozambique: a two-stage DEA and neural networks on efficiency and capacity slacks. Appl. Econ. 46 (29), 3591–3600. Berger, A., 1995. The Profit-structure relationship in banking-tests of market-power and efficient-structure hypotheses. J. Money Credit Bank. 27 (2), 404–431. Bertoni, F., Colombo, M.G., Grilli, L., 2013. Venture capital investor type and the growth mode of new technology-based firms. Small Bus. Econ. 40, 527–552. Blackwell, P., Shehab, E.M., Kay, J.M., 2006. An effective decision-support framework for implementing enterprise information systems within SMEs. Int. J. Prod. Res. 44 (17), 3533–3559. Cavalieria, S., Maccarrone, P., Pinto, R., 2004. Parametric vs. neural network models for the estimation of production costs: a case study in the automotive industry. Int. J. Prod. Econ. 91 (2), 165–177. Chan, S.H., Martin, J., Kensinger, J., 1990. Corporate research and development

7. Contributions of the study and further directions This paper makes several contributions to the existing literature through its unique empirical modeling. From the methodological perspective, the innovative approach exploits centralized learning property of BPNN for nonlinear segmentation of two performance groups based upon efficiency determination. Whilst DEA has been commonly used for efficiency measurement as a frontier approach, the literature shows rare evidence of using neural networks for measuring relative efficiency of firms despite earlier hints at the potential capability of BPNN for that purpose (Athanassopoulos and Curram, 1996; Santin, 2008; Wang, 2003). Up to this date, limited studies have presented combined DEA-BPNN approaches to take advantage of efficiency measurement and prediction capability, respectively, from each method (Azadeh et al., 2010; Kwon and Lee, 2015; Pendharkar, 100

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

linkages in the Japanese automotive industry. Manag. Sci. 45, 466–495. Marques, A., Lacerda, D.P., Camargo, L.R., Teixeira, R., 2014. Exploring the relationship between marketing and operations: neural network analysis of marketing decision impacts on delivery performance. Int. J. Prod. Econ., 178–190. Mazhar, M., Kara, S., Kaebernick, H., 2007. Remaining life estimation of used components in consumer products: life cycle data analysis by Weibull and artificial neural networks. J. Oper. Manag. 25 (6), 1184–1193. Morbey, G.K., Reithner, R.M., 1990. How R & D affects sales growth, productivity and profitability. Res. Technol. Manag. 33 (3), 11–14. Mudambi, R., Swift, T., 2014. Knowing when to leap: transitioning between exploitative and explorative R & D. Strateg. Manag. J. 35 (1), 126–145. Nassr, I.K., Wehinger, G., 2015. Opportunities and limitations of public equity markets for SMEs. OECD J. Financ. Mark. Trends 1, 1–36. Nephilim, M., Boubakerb, S., Lakhalc, F., 2012. Ownership structure, voluntary R & D disclosure and market value of firms: the French case. Int. J. Bus. 17 (2), 126–138. NeuralWare, 2013. Neuralware Predict Manual. Pittsburg, PA. Pendharkar, P., 2005. A data envelopment analysis-based approach for data preprocessing. IEEE Trans. Knowl. Data Eng. 17 b (10), 1379–1388. Pendharkar, P.C., Rodger, J.A., 2003. Technical efficiency-based selection of learning cases to improve forecasting accuracy of neural networks under monotonicity assumption. Decis. Support Syst. 36 (1), 117–136. Qi, M., 1999. Nonlinear predictability of stock returns. Using financial and economic variables. J. Bus. Econ. Stud. 17 (4), 419–429. Roh, J.J., Lee, J., 2013. Does lean inventory lead to firm performance? An international comparison between the US and Japanese manufacturers. J. Distrib. Sci. 11 (7), 23–30. Rumelhart, D.E., Geoffrey, E.H., Ronald, J.W., 1986. Learning representation by back propagating errors. Nature 323, 533–536. Sambharya, R.B., Lee, J., 2014. Renewing dynamic capabilities globally: an empirical study of the world's largest MNCs. Manag. Int. Rev. 54 (2), 137–169. Samilienko, S., Weistroffer, R., 2010. Improving the relative efficiency of revenue generation from ICT in transition economies: a product life cycle approach. Inf. Technol. Dev. 16 (4), 279–303. Sandner, P.G., Block, J.H., 2011. The Market value of R & D, patents and trademarks. Res. Policy 40 (7), 969–985. Santin, D., 2008. On the approximation of production functions: a comparison of artificial neural networks frontiers and efficiency techniques. Appl. Econ. Lett. 15 (8), 597–600. Santín, D., Delgado, F.J., Valiño, A., 2004. The measurement of technical efficiency: a neural network approach. Appl. Econ. 36 (6), 627–635. Sexton, R., Sriram, R., Etheridge, H., 2003. Improving decision effectiveness of artificial neural networks: a modified genetic algorithm approach. Decis. Sci. 34 (3), 421–442. Sharma, J., Kamath, R., Tuluca, S., 2003. Determinants of corporate borrowing: an application of a neural network approach. Am. Bus. Rev., 63–73. Shin, S., Ennis, K.L., Spurlin, W.P., 2015. Effect of inventory management efficiency on profitability: current evidence from the U.S. manufacturing industry. J. Econ. Econ. Educ. Res. 16 (1), 98–106. Tabachnick, B.G., Fidel, L.S., 2013. Using Multivariate Statistics Sixth edition. Pearson Education, New York, NY. Tubbs, M., 2007. The relationship between R & D and company performance. Res. Technol. Manag. 50 (6), 23–30. Uhlaner, L.M., Stel, A., Duplat, V., Zhou, H., 2013. Disentangling the effects of organizational capabilities, innovation and firm size on SME sales growth. Small Bus. Econ. 41, 581–607. Verma, D., Sinha, K., 2002. Toward a theory of project interdependencies in high tech R & D environment. J. Oper. Manag. 20, 451–468. Wang, C., Lu, Y., Huang, C., Lee, J., 2013. R & D. Productivity, and market value: an empirical study from high-technology firms. Omega 41 (1), 143–155. Wang, S., 2003. Adaptive non-parametric efficiency frontier analysis: a neural-networkbased model. Comput. Oper. Res. 30 (2), 279–296. Williams, D.A., 2016. Can neural networks predict business failure? Evidence from small high tech firms In the UK. J. Dev. Entrep. 21 (1), 1–17. Wong, T., Chan, A., 2015. A neural network-based methodology of quantifying the association between the design variables and the users' performance. Int. J. Prod. Res. 53 (13), 4050–4067. Wu, D., Yang, Z., Liang, L., 2006. Using DEA-neural network approach to evaluate branch efficiency of a large Canadian bank. Expert Syst. Appl. 31 (1), 108–115. Xu, X., Wang, Y., 2009. Financial failure prediction using efficiency as a predictor. Expert Syst. Appl. 36, 366–373. Zhao, H., Zou, S., 2002. The Impact of industry concentration and firm location on export propensity and intensity: an empirical analysis of Chinese manufacturing firms. J. Int. Mark. 10 (1), 52–71.

expenditures and share value. J. Financ. Econ. 26 (2), 255–276. Chauvin, K.W., Hershey, M., 1993. Advertising, R & D expenditures and market value of the firm. Financ. Manag. 22 (4), 128–140. Chen, M.-C., Cheng, S.-J., Hwang, Y., 2005. An empirical investigation of the relationship between intellectual capital and firms' market value and financial performance. Journal Intellect. Cap. 6 (2), 159–176. Ciampi, F., Gordini, N., 2013. Small enterprise default prediction modeling through Artificial Neural Networks: an empirical analysis of Italian small enterprises. J. Small Bus. Manag. 51 (1), 23–45. Cox, J., Mann, L., Samson, D., 1997. Benchmarking as a mixed metaphor: disentangling assumption of competition and collaboration”. J. Manag. Stud. 34 (2), 285–314. Dai, X., Kuosmanen, T., 2014. Best-practice benchmarking using clustering methods: application to energy regulation. Omega 42 (2), 179–188. Das, P., Datta, S., 2007. Exploring the non-linearity in empirical modelling of a steel system using statistical and neural network models. Int. J. Prod. Res. 45 (3), 699–717. Dedeke, A., Eroglu, C., 2015. Does inventory leanness impact performance? Industry Week. July 7. 〈http://www.industryweek.com/lean-six-sigma/does-inventoryleanness-impact-performance〉 Delen, D., Zaim, H., Kuzey, C., Zaim, S., 2013. A comparative analysis of machine learning systems for measuring the impact of knowledge management practices. Decis. Support Syst. 54 (2), 1150–1160. Eberhart, A.C., Maxwell, W.F., Siddique, A.R., 2004. An examination of long-term abnormal stock returns and operating performance following R & D increases. J. Financ. 59 (2), 623–649. Emrouznejad, A., Shale, E., 2009. A combined neural network and DEA for measuring efficiency of large scale data sets. Comput. Ind. Eng. 56 (1), 249–254. Eroglu, C., Hofer, C., 2011. Lean, leaner, too lean? The inventory-performance link revisited. J. Oper. Manag. 29 (4), 356–369. Fausett, L., 1994. Fundamentals of Neural Networks: Architectures, Algorithms, and Applications. Prentice Hall, Englewood Cliffs, NJ. Galbraith, C., Stiles, C., 1983. Firm profitability and relative firm power. Strateg. Manag. J. 4 (3), 237–250. Garengo, P., Biazzo, S., Bititci, U.S., 2005. Performance measurement systems in SMEs: a review for a research agenda. Int. J. Manag. Rev. 7 (1), 25–47. Gaur, V., Fisher, M.L., Raman, A., 2005. An econometric analysis of inventory turnover performance in retail services. Manag. Sci. 51 (2), 181–194. Geng, C., Man, W., 2004. Implementing neural networks for decision support in direct marketing. Int. J. Mark. Res. 46 (2), 235–254. Golmohammadi, D., 2011. Neural network application for fuzzy multi-criteria decision making problems. Int. J. Prod. Econ. 131, 490–504. Grønholdt, L., Martensen, A., 2005. Analyzing customer satisfaction data: a comparison of regression and artificial neural networks. Int. J. Mark. Res. 47 (2), 121–130. Gujarati, D.N., Porter, D.C., 2009. Multicollinearity: what Happens If the Regressors Are Correlated? Basic Econometrics 5th ed.. McGraw-Hill Irwin, New York, 320–364. Guner, B., Lee, J., Habte-Giorgis, B., 2007. Strategic orientation and export performance, a three-country study of manufacturing firms. J. Transnatl. Manag. 12 (2), 61–77. Hadlock, P., Hecker, D., Gannon, J., 1991. High technology employment: another view. Mon. Labor Rev., 26–30. Hair, J.F., Jr., Black, W.C., Babin, B.J., Anderson, R.E., 2010a. Multivariate Data Analysis Seventh edition. Pearson Education Inc, New Jersey, 204–205. Hair, J.F., Black, W.C., Babin, B.J., Anderson, R.E., 2010b. Multivariate Data Analysis 7th ed.. Pearson Education, New York. Hershey, M., 1985. Market structure and market -value. J. Bus. 58 (1), 89–98. Hsu, C., Tan, K.C., Laosirihongthong, T., Leong, G.K., 2011. Entrepreneurial SCM competence and performance of manufacturing SMEs. Int. J. Prod. Res. 49 (22), 6629–6649. Hu, M., Zhang, G., Jiang, C., Patuwo, E., 1999. A cross-validation analysis on neural network out-of-sample performance in exchange rate forecasting. Decis. Sci. 30 (1), 197–216. Ito, K., Pucik, V., 1993. R & D spending, domestic competition and export performance of japanese manufacturing firms. Strateg. Manag. J. 14 (1), 61–75. Kafouros, Mario, I., 2005. R & D and productivity growth, evidence from the UK. Econ. Innov. New Technol. 14 (6), 479–497. Kim, H.S., Sohn, S.Y., 2010. Support vector machines for default prediction of SMEs based on technology credit. Eur. J. Oper. Res. 201 (3), 838–846. Koumanakos, D., 2008. The effect of inventory management on firm performance. Int. J. Product. Perform. Manag. 57 (5), 355–369. Kourentzes, N., 2013. Intermittent demand forecasts with neural networks. Int. J. Prod. Econ. 143 (1), 198–206. Kraiczy, N.D., Hack, A., Kellermanns, F.W., 2015. CEO innovation orientation and R & D intensity in small and medium-sized firms: the moderating role of firm growth. J. Bus. Econ. 85 (8), 851–872. Kwon, H., 2014. Performance modeling of mobile phone providers: a DEA-ANN combined approach. Benchmark.: Int. J. 21 (6), 1120–1144. Kwon, H., Lee, J., 2015. Two-stage production modeling of large U.S. banks: a DEAneural network approach. Expert Syst. Appl. 42, 6758–6766. Lam, M., 2004. Neural network techniques for financial performance prediction: integrating fundamental and technical analysis. Decis. Support Syst. 37, 567–581. Lee, B. 2012. Financial Guide for Small and Medium Enterprises. CPA Australia. Singapore. Available at 〈http://www.cpaaustralia.com.au/documents/FinancialGuide-for-SME.pdf〉 Lee, J., Habte-Giorgis, B., 2004. Empirical approach to the sequential relationships between firm strategy, export activity, and performance in U.S. manufacturing firms. Int. Bus. Rev. 13 (1), 141–201. Liberman, M.B., Demeester, L., 1999. Inventory reduction and productivity growth,

Dr. Jooh Lee a Professor of Management and Entrepreneurship at Rowan University. He earned his Master’s degree at Colorado State and his PhD at the University of Mississippi. His main research areas of interest are focused on strategic and operational linkages between performance and strategy, particularly with respect to corporate sustainability, technology capabilities, corporate reputation, and executive compensation across countries. He has published over 100 scholarly articles including more than 60 publications in refereed journals and received more than dozens of professional accolades. He was formerly a holder of the John B. Campbell Professorial Chair at Rowan. Dr. He-Boong Kwon is an Associate Professor of Management at Colorado State University-Pueblo, USA. Dr. Kwon holds a Doctoral and Master’s degree in Electrical Engineering from the Florida Institute of Technology. He was enrolled in an MBA

101

Int. J. Production Economics 183 (2017) 91–102

J. Lee, H.-B. Kwon

International Journal of Logistics Management, Benchmarking: An International Journal, International Journal of Productivity and Quality Management, and Expert Systems with Applications. His research interests include performance modeling, operations strategy, and technology innovation.

program from Pacific Lutheran University, and completed the AACSB AQ Bridge to Business Program from the University of Toledo. He has more than 20 years of experience in operations and quality management in his tenure in the Korean Air Force. His research has been published in journals including European Journal of Innovation Management, International Journal of Procurement Management,

102