Predicting corporate bankruptcy based upon
marketing variables using Machine Learning.
Rick Bos
S2281872 R.bos.14@student.rug.nl Supervisor: Dr. Bhattacharya Date: 18-06-2018 University of Groningen Faculty of Economics and BusinessAbstract:
For the past several decades the prediction of bankruptcy has been a topic of great interests to researchers, financial institutions and managers. Simultaneously, despite decades of research, most bankruptcy prediction methods consist solely of financial ratios applied in a MDA model. However, machine-learning methods have been found to significantly
outperform the MDA model in terms of accuracy when predicting bankruptcy. Moreover, it is time to look beyond financial ratios and also take into consideration other variables. In this study these other variables are marketing variables, as marketing investments and assets can significantly affect cash flows and profitability, which in turn affects profitability. This was confirmed in this study. While financial ratios can achieve high accuracy rates when
Acknowledgements
I am highly grateful to my thesis supervisor, Dr. Bhattacharya, who has provided supportive guidance and was supportive in helping me learn more about machine learning methods and corporate bankruptcy while writing this thesis. I am also grateful to the University of
Table of contents
1. Introduction 4
2-5. Theoretical background
2. Bankruptcy Studies
2.1 Accounting-based models.
2.2 Accounting and market-based models
7 8 3. Machine Learning Methods
3.1 Boosting 3.2 Random Forest
3.3 Support Vector Machines 3.4 ANN
10 10 11 12 4. The effect of Marketing and R&D on Bankruptcy
Prediction.
• 4.1.1 The effect of Advertising on Bankruptcy Prediction
• 4.1.2 The Risk Reducing Effect of Brands • 4.1.3 The Risk Reducing Effects of Customer
Satisfaction
• 4.1.4 Hypothesis development
• 4.2 The effect of R&D on Firm Risk and Bankruptcy Prediction 15 16 18 19 19
1. Introduction
Since the 1960s economists, accountants and financial researchers have been attempting to predict more accurately the likelihood of bankruptcy. A bankruptcy can be defined as the “ formal process whereby debtors who cannot meet their obligations sign over all of
their assets. “ (Gc. 2018). This becomes increasingly relevant, as since the 1960s the amount of corprate bankrupties has substantially increased according to the American Bankruptcy Institute (ABI. 2000). So far, the role of marketing in forecasting to the likelihood of bankruptcy has not been addressed within bankruptcy prediction research. This is not surprising, as marketing has largely been considered purely a short-term expense within the field of finance (Day and Fahey. 1998), and could thus not aid in predicting a long-term future. Nevertheless, there are clear indicators that marketing can have an effect on predicting risk within firms. In order for marketing to get a “seat at the table” and part take in corporate decision making, it is increasingly important for marketing to be connected to financial rewards and risks (Lukas et al. 2005). Hence, this study wishes to combine the risk of bankruptcy with marketing, as there are clear, theoretical indicators that marketing may significantly affect the financial risk of firms. Therefore, the aim of this study is to analyse the effect of marketing variables on (the prediction of) bankruptcy.
A large share of current bankruptcy prediction is based on the work of Altman (1968), which uses a multiple discriminant analysis (MDA) on the basis of accounting ratios. The MDA method is a multivariate technique for classification used for classification. Despite being introduced in the 1960s, the MDA method as proposed by Altman is still one of the major methods of predicting bankruptcy (Charitou et al. 2004). Nevertheless, more recent research has shown that the MDA method is an unreliable method in predicting bankruptcy (Alaka et al. 2018), as it has a high amount of type I errors (Kim. 2011). Moreover, recent research has shown that the most reliable methods for predicting bankruptcy are artificial intelligence tools, such as support vector machines (SVM), which can give significantly more accurate predictions (Alaka et al. 2018; Kim. 2011). Machine learning methods have been found to have significantly higher accuracy rates than MDA (Min and Lee. 2005). As generally machine learning methods will outperform traditional models in predicting
bankruptcy (Barboza et al. 2017), this study will make use of machine learning techniques in order to forecast bankruptcy, such as random forest and SVM.
Altman (1968) and Ohlson (1980). Nevertheless, more recent research such as Hillegeist (2004) and Shumway (2001) has shown that non-accounting variables can significantly contribute to predicting bankruptcy. Nevertheless, the role of marketing in predicting
bankruptcy has so far been largely ignored. It could be considered somewhat remarkable that marketing has been ignored within bankruptcy prediction research, considering that research has long establisehd a connection between marketing and long-term profitability Luo and Jong (2012), as well as reduce cash flow variability Srivastana et al. (1998). Therefore, there seems to be reasonable support that marketing variables could have an effect on the
likelihood of bankruptcy.
Moreover, research has shown that that marketing based assets, such as brands,
advertising and customer satisfaction can significantly affect earnings, liabilities and a firm’s market value, all of which are predictorrs of bankruptcy (Altman. 1968). For example, increased customer satisfaction may significantly affect the cash flow variability of a firm (Gruce and Rego. 2005), which in turn affects the market-to-book ratio (Barnes. 2002). Moreover, Rego et al. (2009) found that brands can significantly affect both the systematic and unsystematic risk of firms. As firms can reduce risk around earnings, increase
profitability and increase the market-to-book ratio through marketing activity, it is expected that marketing can significantly, negatively affect the likelihood of bankruptcy. Hence, marketing based investments as well as marketing based assets could significantly affect the likelihood of bankruptcy
Furthermore, it is known that marketing and R&D are strongly connected to market share. Commonly these types of investments have been perceived as barriers of entry. Hence, advertising may become an increasingly important tool in order to maintain or gain market share (Kessides. 1986). Similarly, firms with a larger market share are often more likely to be able to handle the, often very large, risks associated with R&D investments. Hence,
throughout this study it is assumed that there will be a significant, moderating effect of market share on the effectiveness of marketing variables.
To conclude, this study aims to address two problems currently prevalent in a majority of bankruptcy prediction research: the use of statistically weak methods and the lack of focus on non-accounting variables. The new variables introduced are the effect of R&D and
marketing and R&D investments, and make long-term marketing budgets and investments more justifiable.
2.1 Accounting-based Models
One of the primary studies that gained a large amount of recognition when attempting to predict bankruptcy was (Altman.1968). Like many bankruptcy prediction studies in the time, the methodology used was that of multiple discriminant analysis (MDA) on the basis of accounting variables which can be found in table 1.1. (Altman. 1968). Till date the Altman (1968) paper is considered the most cited papers in the field of bankruptcy prediction models. Nevertheless, the MDA technique was considered flawed by Ohlson (1980) as it was was too statistically restrictive, and by Zmijweski (1984) because it used non-random sampling. For Ohlson and Zmijweski logistic and probit model was considered to be more appropriate. However, the Ohlson (1980) model did not find a significant improvement over the Altman (1968) model. Additionally, the Zmijewski (1984) model opted to use a probit model using random sampling, as opposed to non-random sampling prior models had applied. However, Zwmijewski found his results to be relatively similar to those of non-random sampling. Hence, it is not surprising that a majority of studies still employ MDA when predicting failure (Charitou et al. 2004).
Nevertheless, the MDA model is not considered the most ideal technique in the comparison study from Alaka et al. (2018), which found MDA to be the least accurate bankruptcy prediction method. This was mainly attributed to the fact that the presence of a Type I error occurs significantly in the employment of the MDA method. A similar effect was found in Kim (2011), which confirmed that one of the major problems with MDA is the frequency of Type 1 errors in comparison to other bankruptcy prediction methods. Moreover, the MDA model assumes variables to have a multivariate normality, which is hardly found in practice (Wilson and Sharda. 1994).
Additionally, traditional failure models such as those previously discussed are only able to predict a short forecasting horizon (Lin et al. 2014). The longer the horizon of forecasting, the less success traditional models will have in predicting bankruptcy (Jardin, 2015). Such
findings were also confirmed in Kouki and Elkhaldi (2009), which made a direct comparison between MDA models, logit models, and neural networks, and found MDA and logit models are most powerful at predicting two to three years.
Unfortunately, as the 1980s moved forward, the predictable qualities of the Altman and Ohlson model became lessened as due to rising corporate debt levels the statistical value of debt in predicting bankruptcy became less (Begley et al. 1994). This finding is also supported
in which they were developed. Moreover, it is suggested that the latter two models are more suited for predicting financial distress rather than bankruptcy (Grice and Dugan. 2001). Lastly, a considerable flaw of these accounting based methods is that the ratios used within these models are derived from searching through a sample of bankrupt and non-bankrupt firms, which causes the models to become sample specific (Agarwal and Taffler. 2008). Moreover, accounting statement are based upon the going-concern principle, which means that the assumption is the business will remain in business, and they are designed to measure past performance rather than future performance (Hillegeist et al. 2004). Hence, accounting variables may not be as valuable in predicting bankruptcy as suggested in prior bankruptcy prediction models (Hillegeist et al. 2004).
Comparison of Financial Ratios in predicting Bankruptcy in Traditional Models
Altman (1968) Ohlson (1980) Zmijewski (1984)
Working capital/Total assets ln (Total assets/GNP Price-level index)
Net income/Total assets Retained earnings/Total assets Total liabilities/ Total assets Total liabilities/Total assets Earnings before interest and taxes/total
assets
Working capital/ Total assets Current assets/ Current liabilites Market value equity/Book value of total
debt
Current liabilities/Current assets Sales/Total sales Net income/Total assets
Funds provided by operation/Total liabilities
Table 1.1 A comparison of financial ratios in predicting bankruptcy in traditional models.
2.2 Accounting and Market based Models
Hence, as the prior discussed accounting based models became less useful after the 80s, the demand grew for models which rely upon more than accounting variables. which
according to (Wu et al. 2010). Hence, these two models will be briefly elaborated upon below.
Shumway (2001) used a hazard model using multiple periods and both accounting and market variables, and found that his model outperformed the original Altman (1968) model. This model was considered more suitable for bankruptcy prediction in the 90s (Wu et al. 2010). These findings were also confirmed by Chava and Jarrow (2004), which found that Shumway (2001)’s model outperformed the Altman (1968) and Ohlson (1980) model. Additionally, in Hillegeist (2004) the intention was to make a comparison between the accounting based measures of Altman (1968) and Ohlson (1980) with a market-based Black-Scholes-Merton option pricing model, which included both market-based and accounting variables. It was found that their method yielded better results than Altman (1968) and Ohlson (1980).
3. Machine Learning Methods for Bankruptcy Prediction.
The methods for predicting bankruptcy so far, primarily logistic regression and MDA, have thus been proven incapable of achieving high accuracy rates on current data, or on data samples different from those based upon which the model has been designed. This is where machine learning methods, such as boosting, random forest and support vector machines can make a significant difference. In the following the section the choice for these different methods of predicting bankruptcy will be briefly elaborated on.
3.1 Boosting:
Boosting is an algorithm based on bagging and includes the process of incrementally building an ensemble of models, by incorporating the misclassifications of weaker models when creating a new model.The most common method of boosting is AdaBoosting, which is a method introduced by Freund and Schaphire (1997).
There has been some research, which has already shown the value of boosting with regards to the prediction of bankruptcy. For example, in (Alforo et al. 2008), it was acknowledged that Artificial Neural Networks (ANN) is a common method of predicting bankruptcy, and compared it to boosting as an alternative method, which was found to outperform ANN. This could be due to the fact that ANN suffers from the problem of data multimodality, which boosting does not have (Barboza et al. 2017).
Moreover, other research has shown the value of boosting in predicting corporate bankruptcy. For example, in Zhou and Liu (2017) it was shown that AdaBoost provided highly accurate predictions even in the cases of missing data. Additionally, in (Barboza et al. 2017) and (Basso et al. 2017) it was also shown that AdaBoost provides significantly more accurate predictions for bankruptcy than other methods, such as LR, MDA and ANN. Hence, it could be concluded that boosting is highly suitable for the prediction of bankruptcy, which is why it will be implemented in this study.
3.2 Random Forests:
as CART first introduced in (Breiman et al. 1984). RF generates multiple trees based upon the bootstrapping method, each one of these trees receives a classification and ‘vote’, and the random forest then chooses the classification that received the most ‘votes.’
The RF method is a popular method, because it is relatively robust and allows for the presence of outliers and noise within the data set (Yeh et al. 2014). However, as a result one of the main flaws of random forests is that they may be prone to overfitting (Zhang. 2017), as deep decision trees may be associated with low bias and high variance. Nevertheless, whether or not a tree is overfitting or suitable or not may be dependent on the size of the tree (Kruppa et al. 2012). Therefore, a trade-off needs to be made, as a small tree may not accurately reflect complicated structures, but a tree that is too large may be subject to over fitting (Kruppa et al. 2013). Moreover, a RF has many other benefits in comparison to other methods. For example, as RF is more randomized, and it is a better choice than bagging as there will be less correlation (Booth et al. 2014).
The use of random decision forests has not been very predominant within bankruptcy prediction studies, whilst other machine learning methods such as SVM and ANN are more prevalent within literature. Nevertheless, in a review study of bankruptcy prediction models by Kumar and Ravi (2007) it was stressed that it may be worthwhile to study the effects of RF as opposed to ‘older’ methods, such as LR, MDA and ANN. Hence, in a comparative study conducted by Basso et al. (2017), random forest was found to outperform MDA, LR and ANN. Additionally, another example of the use of random forest was in Barboza et al. (2017), in which it was found that random forest outperformed LR, MDA and ANN in predicting bankruptcy. Moreover, in Kruppa et al. (2013) random forest was also used as a way to predict the consumer credit risk, and random forest was found to outperform a LR. Whilst different to bankruptcy, this could indicate that random forest may be a useful tool in predicting financial risks, such as bankruptcy. Therefore, there is good reason to assume that random forest may be suitable for the prediction of bankruptcy.
3.3 Support Vector Machines
There are many benefits to using SVM in comparison to other machine learning methods. First of all, SVM is less vulnerable to overfitting than other methods, which is due to the fact that it uses the principle of structural risk minimization as opposed to the principle of
empirical risk (Min and Lee. 2005). However, SVM is not often regularly used, as it requires a high amount of fine tuning and adjustment, which means that it is not possible to use it ‘out-of-the-box’ (Milenova et al. 2005). Moreover, the interpretation of the outcome of a SVM algorithm may be difficult to interpret, as it does not provide easily interpretable statistics, such as a p-value (Doumpos et al. 2005). However, SVM does ease the burden on part of the researcher by removing the need to outliers (Doran et al. 2007). Additionally, in case the sample is small, SVM can still obtain good results in comparison to other
classification methods (Shin et al. 2005). Hence, it could be concluded that, while not always easy to use and interpret, SVM is a highly suitable classification and regression method. Hence, it is not surprising that SVM has been found to be highly suitable for bankruptcy prediction. For example, in Kim (2011) it was found that SVM had the highest accuracy rate in comparison to LR, MDA and ANN when predicting bankruptcy within the hotel industry. Moreover, this is similar to the findings by Basso et al. (2017) and Min and Lee (2005) for SVM, which also found that SVM was a method that outperformed LR, MDA and ANN in terms of accuracy when predicting bankruptcy. Therefore, it could be concluded that SVM is an efficient method for predicting bankruptcy, which is why the SVM method will be used in this study. Hence, in this study, SVM will be used as well in addition to boosting and random forest in order to predict the occurrence of bankruptcy.
3.4 Artificial Neural Networks.
Artificial Neural Networks (ANN) are, as the name implies, based on biological processes within the human brain. To explain the methodology behind the ANN algorithm goes beyond the scope of this study, however Basheer and Hamseer (2000) wrote a highly recommendable paper on the fundamentals and application of ANN. However, it is interesting to note that ANN is often categorized together with the logistic regression in terms of classification, as they are both based on the maximum likelihood estimation (Dreiseitl and Ohno-Machado. 2002), which is common in the field of marketing.
their ability to handle nonlinear relationships, as ell as being able to deal well with noisy data (Basheer and Hasjmeer. 2000).
4. The effect of Marketing and R&D on Bankruptcy Prediction.
The role of marketing expenses as impacting financial rewards and risks has long been ignored within finance (Day and Fahey. 1988). This is due to the fact that marketing costs have long beeen considered expenses rather than investments (Rust et al. 2004).
Nevertheless, there is a high desire among senior management to shed more light on the effects of marketing and R&D expenditure on financial volatility (Ambler. 2003). This is because there are clear indicators that marketing and R&D can have an effect on predicting risk within firms, and as such may be able to aid in predicting bankruptcy as well.
Moreover, as within the field of accounting most marketing activity related expenses are listed as advertising expenditures, throughout this study marketing activity and advertising expenditure will be considered to be interchangeable. Marketing investments in advertising expenditure will lead to market based assets, such as brands and customer satisfaction.
Hence, the advertising expenditure variable will reflect the direct investment into advertising, but also such market based assets, as these all will affect the likelihood of bankruptcy. These different effects for marketing will be further elaborated on in the following sections, and are illustated in the conceptual model below as well.
Effect of Advertising and R&D expenditure on Bankruptcy.
4. 1.1 The effect of Advertising on Bankruptcy Prediction
First of all, a large share of research has already shown the positive effects of increased advertising expenditure on profitability and earnings. For example, in Luo and Jong (2012) it was shown that there is a direct, significant relationship between advertising expenditure and profitability. The increase in earnings as a result of advertising expenditure should
significantly reduce the risk of bankruptcy. However, advertising expenditure may not only significanly affect profitability once, but can increase earnings for a number of years. For example, in Graham and Frankenberg (2000), it was found that on average the asset value of advertising expenditure has a life of three years.
As a result advertising expenditure can also significantly affect the volatility of future cash flows, as advertising will affect future profitability. For example, in Srivastana et al. (1998) it was found that an increase in advertisiment will significantly reduce cash flow variability. Moreover, investors will generally see an increase in advertising expenditure as a positive sign of increased profitability and reduced cash flow variability (Chauvin and
Herschey. 1993) . The reduced cash flow variability is associated with a decreased likelihood of bankruptcy. Nevertheless, this effect may not always hold true, as for example Merino et al. (2006) found that increasing the advertising budget actually increased the volatility of future cash flows. Therefore, there remains some ambiguity surrounding the effect of advertising on cash flow variance.
Hence, as advertising is likely to increase earnings, profitability and reduce cash flow variability, this will also increase the value of the firm. A reduction in cash flow variability will also increase the market-to-book ratio of a firm, as research has shown there is a negative relationship between the market value of a firm and cash flow volatility (Barnes. 2002). Moreover, it was found in McAlister et al. (2007) that when correcting for accounting variables, advertising spending had a negative effect on a firm’s systematic risk and a positive effect on firm value. Similarly, in Ho et al. (2005), it was found that advertising expenditure will significantly affect equity for nonmanufacturing firms, although a similar result could not be found for manufacturing firms. Moreover, while advertising expenditure will significantly increase a firm’s own valuation, it may simultaneously decrease the
valuation of its competitors (Joshi and Hanssens. 2010). Hence, advertising expenditure may not only decrease the likelihood of a firm’s bankruptcy, but simultaneously increase the likelihood of others.
it was shown that negative marketing events, such as deceptive marketing, can significantly, negatively affect the value of a firm. Moreover, in Kim et al. (2014) it was found that in a sample of restaurants an increase in advertising expenditure significantly increased the total and unsystematic risk, although this was probably sample specific. Hence, it could be that bad or negative advertising may also occur as an effect of advertising, which would significantly decrease the market-to-book-ratio of a firm, as well as decrease earnings. In such cases an increase in advertising expenditure would increase the likelihood of bankruptcy rather than decrease. Hence, in this study the variable of interest is marketing expenditure as share of sales in order to measure marketing effectiveness.
In addition to the direct effect of advertising expenditure on the likelihood of bankruptcy, the investment in advertising and marketing activity also yields marketing assets, such as brands and customer satisfaction. For example, there has been found to be a strong link between brand loyalty and advertising expenditure (Ha et al. 2011), as well as relationship between advertising expenditure and brand value and profitability (Eng and Keh. 2007). Hence, the effect of brands on the likelihood of bankruptcy prediction will be further elaborated upon in section 4.1.2. Moreover, it has also been found that advertising
expenditure will significantly affect customer satisfaction (Ha and Muthaly. 2008), and that a decrease in advertising expenditure will significantly decrease customer satisfaction (Malshe and Agarwal. 2015). Therefore, the effect of customer satisfaction on the likelihood of bankruptcy will also be subject to discussion in section 4.1.3.
4.1.2 The Risk Reducing Effect of Brands
Brands will be able to increase the returns on investments, but also lower the risk associated with these returns (Fornell et al. 2006). Hence, through the building of brands, marketing will be able to reduce the risk associated with firms. There has been strong, significant support showing the positive relationship between brand (quality) and firm value. For example, in Kerin and Sehturaman (1998) it was found that there is a positive relationship between brand quality and the market-to-book ratio of a firm. Such findings were also supported by Rego et al. (2009), which found that consumer-based brand equity (CBEC) could significantly predict (unsystematic) risk.
expenditure is expected to directly, significantly affect brand loyalty (Ha et al. 2011). Hence, a higher brand loyalty obtained through advertising expenditure is expected to reduce the variability of cash flows. As cash flow variability decreases, so should the likelihood of bankruptcy.
Additionally, a firm should experience greater profitability when it has more consumers with a high degree of consumer-equity (Rego et al. 2009). This is in part because consumers with higher consumer-based equity are more likely to engage in repeating purchases (Keller. 2003). Hence, while advertising expenditure may directly increase profitability and earnings, through the creation of brands this effect could be explained or even strenghtened.
Furhermore, the marketing of brands can have a significant effect on the risk of
competitors entering the marketing (Kim et al. 2012). This would be in consensus with the study conducted in Hsu and Jang (2008), which found that advertising can be an effective tool in establishing brand equity within a market. Moreover, in Kapferer (2012) it is suggested that a brand portfolio can be useful in lowering firm risk when entering a new market, as a firm can use a brand as leverage. Hence, branding and advertising can
significantly reduce risks of competitors entering, but also significantly reduce risks when entering a new market.
Moreover, the reputation of a brand can also significantly contribute to reducing the risks associated with firms. For example, in Alhuwalia et al. (2000) it was found that positive associations with a brand minimize the impact of negative information on a firm to consumers and investors. This is also reflected in the study of Rego et al. (2009), which suggested that there may be a ‘corporate reputation effect’ of positive brand associations with investors, who consider familiar brands to be less risky. Additionally, in Rego et al. (2009) it was found that consumer-based brand equity (CBEC) could significantly predict
unsystematic and systematic risk. Hence, as a result investors prefer to have the stocks of brand they are more familiar with as they perceive these to be less risky (McAlister. 2007). Hence, it appears that a brand’s reputation can signal to consumers, investors and others that a firm is facing less risk which increases the market value of a firm, which in turn reduces the likelihood of bankruptcy
firm risk (Morgan et al. 2009). So, while brands are hypothesized to negatively affect the likelihood of bankruptcy, this is dependent on a firm’s marketing strategy.
To summarize, brands can significantly affect both the equity and profitability of a firm, and reduce (un)systematic firm risk. This is mainly due to the signalling effect brands have towards shareholders, and because brands lead to increased consumer loyalty, which reduces risks and increases profitability. Hence, the strength of a firm’s brand can positively reduce the risk of bankruptcy.
4.1.3 The Risk Reducing Effects of Customer Satisfaction
In addition to the risk reducing effects of brands, it was also found that customer satisfaction can significantly affect the future cash flows of a firm (Gruca and Rego. 2005). Customer satisfaction is an important component within marketing and marketing activity can
significantly influence customer satisfaction (Andreasan and Kotler. 2003). A firm with high customer satisfaction should experience higher profitability and lower costs (Anderson et al. 1994). All in all, customer satisfaction can be considered to be an economic asset, which could significantly affect the profitability of a firm (Fornell. 1992). Therefore, a firm’s liabilities could be lower whilst its earnings could be higher when a firm is experiencing high customer satisfaction, which in turn should reduce the likelihood of bankruptcy.
Moreover, a higher customer satisfaction will make customers more loyal (Anderson and Sullivan. 1993), which will make customers more likely to defect lower priced competitor products (Fornell et al. 1993). As a result, customer satisfaction was found to significantly reduce the variability in cash flows (Gruce and Rego. 2005). Hence, customer satisfaction is a highly advantageous asset to firms, as it both reduces the variability of cash flows, as well as increase the cash flow level (Tarasi et al. 2013) Hence, as customer satisfaction reduces the variability in cash flows, this will positively affect the market-to-book ratio of a firm, and reduce the likelihood of bankruptcy.
Nevertheless, as a high customer satisfaction can reduce the variability and risks
associated with a firm, a low customer satisfaction can have an opposite result. For example, it was found in Luo (2007) that a low customer satisfaction and negative customer
concluded that while customer satisfaction can be risk-reducing, a lack of customer satisfaction could also be risk-inducing.
To summarize, customer satisfaction can significantly reduce the risks in firms, and as such reduce the risk of bankruptcy. Through increased loyalty, customer satisfaction can reduce the variability of future cash flows and increase profitability. As marketing activity has a positive, significant effect on customer satisfaction, these activities can in turn reduce the risk of bankruptcy.
4.1.4 Hypothesis development
Therefore, it could be concluded that marketing activities, such as branding, advertising and optimizing customer satisfaction can significantly increase a firm’s profitability. Moreover, higher customer satisfaction and corporate branding can significantly, positively impact the market-to-book ratio of a firm. Hence, as firms with effective marketing activities experience higher profitability and higher market value, it could be hypothesized that marketing activity can significantly reduce the risks of bankruptcy. Therefore, the following hypothesis is made:
H1: Advertising effectiveness of a firm will significantly affect the likelihood of
bankruptcy.
4.2 The effect of R&D on Firm Risk and Bankruptcy Prediction
Although there are a few exceptions (e.g. educational service electrical equipment, chemicals etc) a majority of firms will not employ both high advertising and R&D, but usually one or the other (Chauvin and Hirschey. 1993). Moreover, increasingly so marketing and R&D have become more integrated with one another, and are significantly influenced by one another (Griffin and Hauser. 1996). Therefore, it would also be interesting to research the effect of R&D expenditure on the risk of bankruptcy, as R&D can be considered to be closely connected to marketing.
within that specific sample R&D also lead to an increase in profitability (Jaisinghani. 2015). Therefore, investments in R&D could significantly increase a firm’s earnings, which would reduce the likelihood of bankruptcy.
Hence, it is not surprising that behavioural studies have found that managers are more likely to invest in R&D when profitability is low as a means of attempting to increase
profitability (Kotlar et al. 2014), and thus avoid bankruptcy. However, this may not always be the case, as the investment in R&D is one that is highly dependent on risk and uncertainty with regards to profitability (Amoroso et al. 2016). For example, in Cooper (1984) it was found that on average firms had a success rate of 67% and a kill rate of 16% for new, developed products.. Nevertheless, it could be argued that such risky decisions may be required in order to obtain profitability (Amoroso et al. 2016). However, the risk associated with product innovations and R&D spending could significantly be reduced by pursuing the right strategies (Cooper. 1984). Hence, while R&D spending could significantly increase a firm’s earnings, there is no guarantee of such a return on the investment. Therefore, there is a high amount of uncertainty associated with R&D expenditure and its effect on earnings, which makes its effect on the likelihood of bankruptcy ambiguous.
Additionally, R&D investments may affect the variability of cash flows, which could affect bankruptcy. For example, it was found in Mizik and Jacobson (2003) that R&D expenditure may stabilize cash flows through the value added when R&D leads to superior products and returns. A similar result was found in Srivastava et al. (1998), which found that R&D expenditure can increase future cash flows on the long term, as well as decrease the variability of such cash flows. Therefore, it may be that an increase in R&D expenditure will decrease the cash flow variability of firms, which in turn could significantly decrease the likelihood of bankruptcy.
Nevertheless, generally literature is in consensus there may be some uncertainty regards the effects of R&D expenditures on future earnings, which increases the risk associated with bankruptcy. For example, in Kotharia et al. (2002) it was found that the future earnings yielded from R&D investments are considerably less certain than other capital
investments. Similarly, in Shi (2003) it is also shown that R&D expenditure leads to an increase in the variance of a firm’s future cash flow. Hence, the effect of R&D expenditure on the variability of cash flows remains unclear, as academic literature has found support for the fact it may decrease cash flow variability, as well as increase variability. Hence, if R&D expenditure leads to an increase in variability, this would increase the likelihood of
Additionally, as it has been stated that there is a high amount of certainty associated with the cash flows, this is also reflected in the valuation of the firm. For example, it has been founds that R&D expenditure increases volatility of stock returns (Chan et al. 2001) and systematic risk (Ho et al. 2004). Such findings were also confirmed by Erickson and Jacobson (1992) who found that R&D does not have any large effect on the firm value of a firm than any other type of expense or investment. However, some firms investors may consider R&D expenditure to be a predicament of an increase in future cash flows, which would positively affect the market value of the firm (Chauvin and Hirschey. 1993). Therefore, there is a high amount of uncertainty with regards to the effect of R&D expenditure on the market-to-book ratio of firms, which could significantly reduce or increase the likelihood of bankruptcy.
Hence, there remains a lot of ambiguity about the effects of R&D on cash flow variability and firm value. Some research, such as Ahmed et al. (2001) may even detect none or neutral effects of R&D on profitability and a firm’s market-to-book ratio. Due to this ambiguity the effect of R&D on the risk associated with bankruptcy remains unclear as well. Hence, this study wishes to further examine the effect of R&D on the risk within firms, and specifically the effect of R&D on the risk of bankruptcy. Therefore, the following hypothesis has been made:
H2: Increased R&D effectiveness of a firm will significantly, affect the likelihood of
5. The Moderating effect of Market Share on Advertising and R&D.
However, whilst it is argued that there will be an effect of R&D expenditure and advertising expenditure on the likelihood of bankruptcy, there may be some variables which influence this effect. For example, throughout this study it will be assumed that the market share of a firm will significantly moderate the effect of R&D as well as advertising on the likelihood of bankruptcy.
Effect of Advertising and R&D expenditure on Bankruptcy.
5.1 Conceptualization of effects of Marketing and R&D on Bankruptcy Prediction moderated by market share.
Similarly, in Buzzell et al. (1975) the strategic importance of market share is stressed with regards to marketing costs, as marketing costs as percentage of sales will be significantly higher for smaller firms. In order to finance these costs, smaller firms will be more
dependent on external capital (Titman and Wessels. 1988). Consequentially, the liabilities of a firm could significantly increase and market equity value decrease, which could increase the likelihood of bankruptcy.
Moreover, advertising may become a necessity in order to maintain or gain market share (Kessides. 1986), the losing of which could significantly affect a firm’s earnings. For large firms advertising becomes a necessity in order to maintain their market share and prices, whereas for smaller firms it becomes a necessity in order to gain market share (Nagle.1981). By investing in advertising as well as R&D a firm creates a barrier to entry for other firms, which would reduce the competition and increase future earnings (Srivastava et al. 1998).. Therefore, it could be expected that the effect of advertising and R&D expenditue becomes more important as market share increases, because a decrease in either could mean a decrease in market share and a reduction of barriers of entry, which would significantly affect
earnings. Hence, market share would positively moderate the effect of advertising and R&D expenditure on the likelihood of bankruptcy.
Furthermore, as previously mentioned the strategy maintained by a firm will contribute to the success of marketing activity reducing the risk of bankruptcy. However, particularly for smaller firms this will be less likely the case, as smaller firms are more likely to lack the appropriate resources and knowledge to invest successfully (Carson. 1990). For example, Banerjee and Bandyopadhyay (2003) found that advertising was not an efficient method for competition for small firms. Hence, for smaller firms it could be considered to be
increasingly risky to increase advertising spending as opposed to larger firms. Such risk is also found for smaller firms with regards to R&D spending. Similarly, Chakavarty and Grewal (2010) found that as market share decreases, managers will be more likely to treat marketing budgets in a more short-sighted manner, which would be less efficient. Hence, it could be concluded that as market share decreases, the likelihood of marketing activity being successful in increasing sales or earnings decreases due to more inefficient management, which makes the likelihood of bankruptcy more prevalent. Therefore, marketing could be less likely to reduce the risks associated with bankruptcy when the firm sizer is smaller.
To summarize, marketing and R&D expenditures are significantly less efficient for
important as market share increases, in order to maintain this market share and the associated earnings. Hence, it could be hypothesized that the market share could significantly moderate the effect of marketing effectiveness on bankruptcy.
H3: The market share of a firm will significantly moderate the effect of marketing effectiveness on bankruptcy.
Additionally, as larger firms typically have a larger amount of capital available to buffer against failled R&D innovations, larger firms are exposed to significantly less risk than smaller firms when enaging in R&D (Cooper. 1984). Therefore, larger firms are more capable of dealing with the risks associated by R&D investments. However, larger firms also receive a higher return on this risk in comparison to their smaller counterpants (Cooper. 1984). Such findings have also been confirmed by Chauvin and Hirschey (1993) and Connolly and Hirschey (2005), which found that the returns on risk are significantly higher for larger firms than their smaller competition. Hence, it could be concluded that the risk of R&D investment is significantly decreased when a firm has a higher market share, and that it also yields a higher return on these investments. As such, market share could significantly , negatively moderate the effect of R&D on the likelihood of bankruptcy.
Schumpeterian theory holds true, then increased market share would lead to a significant reduction in costs as well, which would reduce liabilities and increase working capital, and thus decrease the likelihood of bankruptcy. However, if it does not, the opposite effect applies. Hence, there are two conflicting theories on whether or not a firm with a relatively large market share has incentive for cost-reducticing innovation . Nevertheless, while the theory of Arrow (1959) may hold true sometimes, in other times the Schumpter (1934) theory might hold true, as was found by Garella (2012). Therefore, it is possible for both effects to occur, and this remains ambiguous effect.
To summarize, a larger market share will increase productivity, efficiency and returns of R&D investments, although these may or may not become more cost reducing. Hence, it could be concluded that there may be a moderating effect of market share on the effect of R&D expenditure and efficiency, although it remains somewhat ambiguous whether this effect will be positive or negative. Hence, the following hypothesis will be subject to testing in this thesis:
6. Data.
First of all, throughout the following section the data used throughout this study will be discussed. The data was acquired through the thesis supervisor dr. Bhattacharya and contains data of various corporations in different industry across different years, and whether or not the firm had gone bankrupt.
One of the primary obstacles encountered with the data is the imbalance of the dataset. It has been found that there is significantly more missing data available for companies which went bankrupt than for companies that were not. Moreover, there are significantly more cases of non-bankruptcy than there are of bankruptcy. In total, there were 155 cases of bankruptcy, whereas there are thousands of cases of non-bankruptcy. An imbalanced dataset will significantly, negatively affect the accuracy rate of classification methods, particularly for machine learning methods, such as SVM (Akbani et al. 2004). There are several methods for resolving an imbalanced dataset, such as under sampling, oversampling and feature selection (Kostiantis et al. 2006). Hence, in order to balance the imbalanced dataset, the Synthetic Minority Over-Sampling Technique (SMOTE) method was applied. The SMOTE method randomly applies an over-sampling of the minority class and an under-sampling of the majority class, which would significantly improve classification performance and remove the imbalance from the dataset (Chawla et al. 2002). The SMOTE method is considered one of the state
Table 6.1: Variables and meaning.
Variables Description
AdEff Advertising Expenditure / Sales
RdEff R&D Expenditure / Sales
Leverage Total Debt / Total Assets
Msh Market share.
marketg Market growth.
LRcap Long term marketing capabilities. SRcap Short term marketing capabilities.
fss Firm size.
mturb Market turbulence.
ModAD Market Share * AdEff
ModRD Market Share * RdEff
Returns Revenue / Total Assets
Profitability Retained Earnings / Total Assets
Productivity Earnings before Interest and Taxes / Total Assets Liquidity Working Capital / Total Assets
of the art methods for handling imbalanced datasets, and has been found to significantly improve classification accuracy (Kumar and Sheshradi. 2012; Kostiantis et al. 2006).
Additionally, the missing data proved to be problematic in the implementation of machine learning methods, especially as the missing data was primarily present for the rare cases of bankruptcy. While not ideal, in such circumstances where it is best to use all available data the imputation of missing data becomes an important task (Jerez et al. 2010). Moreover, it had been found that the best results for the imputation of missing data have been obtained through machine learning methods
(Lakshminarayan. 1999). Hence, missing data had been imputed using the package ‘missForest’ in R, which imputes missing data by creating a random forest for every variable where tag = 1. The usage of this package for the imputation of missing data had been found to be highly successful and to outperform other imputation methods (Steekhoven and Buhlmann. 2011). Hence, missing values for cases of bankruptcywere imputed using this package.
The final distribution of the dataset can be found in the table below, which shows the various descriptive statistics for the different variables.
Table 6.2. Descriptive Statistics
Coefficients Min 1st Qua Mean 3rd Qua Max SD
The data will be used in different models, where different variables will be used for different models. First of all, the absolute, predictable power of the Altman (1968) will be measured by only including these variables as proposed in Altman (1968) for predicting bankruptcy, which are leverage, returns, profitability, productivity and liquidity. Additionally, the absolute,
predictable power of the variables as proposed in the preceding literature review. In model III this will be further expanded upon by including control variables. These control variables include market growth and market turbulence, as economic determinants can significantly effect the likelihood of bankruptcy (Aleksanyan and Huiba. 2014). Moreover, the leverage variable as discussed in Altman (1968) was added as a control variable. Lastly, firm size was also taken into account into this model , as firm size was found to significantly effect the likelihood of bankruptcy, as smaller firms are more likely to be more vulnerable (Wijn and Bijnen. 2001).
In model IV the variables SRcap and LRcap have been added, which takes into account the efficiency of marketing expense in the long and short term. This will allow to determine the importance of marketing experience as opposed to marketing expenses.
In the final model, model V, all the Altman (1968) variables and the variables of model IV have been combined. This model is final model, and should give the most accurate
predictions, as it includes all relevant variables.
Table 6.3: Variable Inclusion in Different Models
Variables Model I Model II Model III Model IV Model V
7. Results.
7.1. Performance Metrics.
The inspiration for the format in which the results in this research are presented is largely derived from the work of Barboza et al. (2017). Hence, the performance metrics that will be looked at throughout this study is the Type I error, the type II error, AUC, ACC. This makes for easy comparison with the Barboza study, as well as the Altman (1968) study, which also included the same performance metrics and thus makes for an easy comparison in accuracy. The accuracy rate is the most natural of all predictive performance metrics, which is the sum of the accurately predicted observation as a percentage from all observation, which can be seen in the calculation below.
The Receiver Operating Characteristic (ROC) curve is a plot of the test true-positive rate on the (sensitivity) x-axis against the false-positive rate (1 – specificity) on the y-axis. The area underneath this curve represents the likelihood that a sample is ranked with greater likelihood than if it were ranked on a completely random basis (Hanley and McNeil. 1982). Hence, the area underneath the receiver operating characteristic is known as AUROC, or more
commonly as AUC. The AUC is considered to be a strong and consistent, single-number method for assessing the predictive accuracy of a machine learning model (Ling et al. 2003). In addition to the above-mentioned performance metrics, the predictive positive value (PPV) and negative predictive value (NPV) have been included into the results tables. The PPV reflects the percentage of bankruptcy cases that has been predicted true Particularly the PPV is of interest to this study, because the primary interest of this study is the prediction of true (bankruptcy).
7.2. Results of Model I
As can be seen in table 7.1 below, the model was initially run with exclusively the Altman variables, as also shown in table 6.3. It can be observed that the highest AUC, ACC and PPV was found to be the case for the Random Forest model. As can be seen, this model experienced the lowest Type I and Type II error, which is also why the model was so highly able in forecasting bankruptcy. The least predictive strength was found to be the case for the Logit Model and linear SVM model. Interestingly enough, the SVM linear model scores high on the AUC, just like the Logit model scores high on the ACC, but both the SVM linear and logit model score relatively low on the PPV.
The relevance of the Altman (1968) variables in predicting bankruptcy is also confirmed by the output of the logit model, which can be seen in table 7.2 below. As can be seen, all variables, particularly leverage and productivity significantly affected the likelihood of bankruptcy.
Table 7.1: Results of Model I
TP TN FN FP Type I Error (%) Type II Error (%) AUC (%) ACC (%) PPV (%) NPV (%) Ada-Boosting 443 493 28 40 7,5% 5,9% 98,17% 93,2% 91,7% 94,6% Random Forest 518 460 11 15 2,8% 2,3% 99,60% 97,4% 97,2% 97,7% SVM - Linear 398 451 73 82 15,4% 15,5% 91,18% 83,4% 82,9% 86,1% SVM – BKF 423 464 48 69 12,9% 10,2% 94,56% 88,3% 86,0% 90,6% Logit Model 392 451 79 82 15,4% 16,8% 83,9% 91,21% 82,7% 85,1% Neural Network 436 495 35 38 7,1% 7,4% 92,7% 96,76% 92,0% 93,4%
Table 3.6: Logit Model of Model I
Coefficients Estimate Std.. Error Z-Value Pr (>|z|)
7.3 Results of Model II, III & IV.
After considering the predictable quality of the Altman variables, the absolute effect of the variables as proposed in the conceptual model was tested. In Model II, depicted in table 7.3 below, this includes only the AdEff, RdEff, msh, ModAd and ModRd variables. As can be seen, the Random Forest and adaptive boosting model still perform with a relatively high accuracy by nearly all measurements. However, the accuracy rates for the remaining methods has significantly decreased for all different performance metrics. If anything, it can be observed that it becomes particularly hard for the remaining models to accurately predict cases of bankruptcy, which is reflected by the relatively high type I error and low PPV.
Nevertheless, in order to further expand upon Model II, there were certain control variables added as discussed in the data section. As can be seen in table below, the accuracy significantly improved across all performance metrics. Nevertheless, the accuracy performance metrics are still all significantly lower in comparison to the Altman variables.
Table 7.3: Results of Model II
TP TN FN FP Type I Error (%) Type II Error (%) AUC (%) ACC (%) PPV (%) NPV (%) Ada-Boosting 492 436 35 41 7,7% 7,4% 96,3% 92,4% 92,3% 92,6% Random Forest 500 439 32 33 6,2% 6,8% 97,8% 93,5% 93,8% 93,3% SVM - Linear 290 382 89 243 45,6% 18,9% 71,0% 66,9% 54,4% 81,1% SVM – BKF 382 379 89 154 28,9% 18,9% 85,0% 75,7% 71,3% 81,0% Logit Model 323 356 115 210 39,4% 24,4% 70,8% 67,7% 60,6% 75,6% Neural Network 353 349 118 184 34,5% 25,1% 75,7% 70,0% 65,7% 74,7%
Table 7.4: Results of Model III
Hence, in order to further improve upon model III, the variables LRcap and SRcap are added to this model, which takes into account the long-term and short-term marketing capabilities. The results are highly interesting in comparison to model III, as the addition of these variables significantly increased the performance metrics of all different machine learning methods. As can be seen in the table below, the AUC of all models is above 94% and in most cases now outperforms model I solely based on the Altman (1968) variables. Hence, the incorporation of long-term and short-term marketing capabilities has had a large effect on the predictability of bankruptcy, which suggests that beyond marketing expense, the capabilities are also important when it comes to the likelihood of bankruptcy.
This effect of the LRcap variable n the accuracy is also reflected in the variable importance table, which is based upon the RF model. As can be seen, the LRcap variable was highly important in the prediction of bankruptcy, as well as the leverage variable. It can also be bserved that SRcap and modRD have played a substantial role in the prediction of bankruptcy.
Table 7.5: Results of Model IV
TP TN FN FP Type I Error (%) Type II Error (%) AUC (%) ACC (%) PPV (%) NPV (%) Ada-Boosting 437 505 34 28 5,3% 7,2% 98,27% 93,8% 94,0% 93,7% Random Forest 457 522 14 30 2,1% 3,0% 99,57% 97,5% 93,8% 97,4% SVM - Linear 400 475 71 58 10,9% 15,1% 94,37% 87,3% 87,3% 87,0% SVM – BKF 429 491 42 42 7,9% 8,9% 96,87% 91,6% 91,1% 92,1% Logit Model 400 477 75 54 10,5% 15,1% 94,44% 84,8% 88,1% 86,4%
Table 7.6: Variable Importance (Random Forest Model IV)
Coefficients Variable Importance Mean
The findings as shown in table 7.6 are also reflected back in the output of the logit model, as shown in table 7.7 above. It can be seen that the LRcap variable is highly significant and negatively affected the likelihood of bankruptcy. Hence, based upon this output it could be argued that stronger, long-term marketing capabilities will significantly reduce the likelihood of bankruptcy. Moreover, the RdEff variable was also found to be significant at the 5% level, which also has a negative effect on the likelihood of bankruptcy. It can be weakly shown that advertising expenditure as percentage of sales will reduce the likelihood of bankruptcy. Moreover, this effect appears to be positively moderated by the market share.
Table 7.7: Logit Model of Model IV
Coefficients Estimate Std.. Error Z-Value Pr (>|z|)
7.3 Results of Model V.
In the final model all the Altman (1968) variables will be added to model IV in order to more accurately measure the effects of the marketing variables relative to the Altman variables, which already proved to be highly valuable on their own. As can be seen in in table 7.8 and 7.9, the addition of these variables enhanced the predictable qualities of the model. As a result, model V had the most optimal rate for all performance metrics, with the exception of the type I error. Hence, the inclusion of marketing variables helped improve the predictive qualities of the different machine learning
methods.
Furthermore, as can be seen on the following page, after the profitability variable, the LRcap is one of the most important variables. This would suggest that the long-term effects of marketing capabilities are highly important in affecting the likelihood of bankruptcy, which is also reflected by the high significance this variable obtained in the logit model.
Table 7.8: Results of Model V
TP TN FN FP Type I Error (%) Type II Error (%) AUC (%) ACC (%) PPV (%) NPV (%) Ada-Boosting 457 508 14 25 4,7% 3,0% 99,3% 96,1% 94,8% 97,3% Random Forest 463 521 8 12 2,3% 1,7% 99,9% 98,0% 97,5% 98,5% SVM - Linear 420 457 51 76 14,3% 10,8% 93,8% 87,3% 84,7% 90,0% SVM – BKF 436 489 35 44 8,3% 7,4% 96,7% 92,1% 90,8% 93,3% Logit Model 409 433 62 70 13,1% 13,2% 93,8% 86,8% 85,4% 87,5% Neural Network 436 503 36 30 5,6% 7,4% 97,7% 93,5% 93,6% 93,3%
Table 7.9 :Average Performance Metric Values across model I – V
Table 7.10: Variable Importance Random Forest Model V Coefficients Variable Importance Mean Decrease Accuracy Mean Decrease Gini 0 1 Profitability 31,50 40,88 43,96 45,97 LRcap 31,05 39,04 43,49 40,52 Leverage 20,92 32,15 34,07 29,02 Productivity 19,61 29,61 31,96 23,75 RdEff 16,08 16,52 21,14 12,92 Fsss 15,35 16,05 19,32 9,41 ModAd 12,85 14,20 18,11 8,04 ModRd 11,70 14,92 17,96 8,25 SRcap 11,97 14,78 17,85 8,07 Returns 12,80 14,43 18,53 7,5 Liquidity 10,80 12,75 17,20 7,32 Msh 10,01 12,40 15,35 5,66 mturb 3,94 14,36 13,59 5,11 AdEff 6,52 9,92 11,83 4,5 marketg 3,11 11,79 11,32 4,83
Table 7.11: Logit Model of Model V
Coefficients Estimate Std.. Error Z-Value Pr (>|z|)
8. Discussion.
First of all, it can be observed that 50 years past the Altman (1968) paper, the variables proposed by Altman are still highly accurate and valuable in the prediction of bankruptcy. In fact, for this particular sample the Altman variables on their own where highly capable of predicting bankruptcy with an average AUC and ACC of 93,35% and 91,71%. In this regard the results were quite different from the research of Barboza et al. (2017), which found significantly lower predictive values for some models for exclusively the Altman variables. Hence, Barboza et al. (2017) argued for the importance of the inclusion of additional variables, which in this study was argued to best be marketing related variables. These
arguments and findings are in accordance with other studies on the Altman variables, such as Hillegeist et al (2004), which question the importance still attributed to the Altman variables in predicting bankruptycy. However, according to this study, the Altman variables are still highly valuable.
Unfortunately, the same could not be said for the absolute effect for the marketing and R&D variables argued for in this study. As could be seen, the predictive qualities of these variables was signifiantly lower than those of the Altman variables. One of the primary issues with the selection of these variables was the signiciant increase in the type I error, which almost trippled relatively to the Altman model. Moreover, the PPV was also considerably low. Hence, the relatively low PPV and type I error relative would indicate that the models were not very succesful at predicting if a firm was bankrupt or not. Nevertheless, even though the models were not excellent, according to the traditional academic point system the
accuraccy found could be described as generally fair to poor. While not incredible, this does mean that the marketing variables by themselves outperformed a completely random model. Hence, this does seem to suggest that marketing and R&D variables can make a positive contribution to predicting bankruptcy.
Hence, these results would sugges that a reduction in R&D and marketing effectivness would have a negative effect on the likelihood of bankruptcy.
Hence, it appears that marketing activities, such as branding, advertising and optimizing customer satisfaction can significantly increase a firm’s profitability and reduce the
likelihood of bankruptcy, as was hypothesized in the first hypothesis. Additionally, the moderator effect of market share on the effect of advertising effectiveness on the likelihood of bankruptcy was found to be significantly positive. Hence, this would suggest than an increase in market share would increase the effect of advertising effectiveness on the likelihood of bankruptcy. This would be in accordance with the traditional view on advertising expenditure and how it is an important tool in order to maintain market share and/or gain market share (Nagle. 1981). The significant effects were found to hold true for model IV, as well as model V, which also included all the Altman variables.
Moreover, the effect of R&D expenditure was also found to have a negative, significant in both models. However, the effect of R&D expenditure was only found to be weakly significant, and no significant support could be found for a moderating effect of market shareon the effect of R&D expenditure. However, the slightly significant results do suggest a negative effect of R&D on the likelihood of bankruptcy, which would confirm the hypothesis of a negative effect. This is in accordance with the increase in research which has started to consider R&D investments as an asset facilitating long term growth rather than an
expenditure (Chan et al. 2001). Hence, it seems that R&D effectiveness would have a positive effect on long-term profitability, which in turn has decreased the likelihood of bankruptcy. A possible explanation for the fact no strong moderation effect was found is that innovative R&D projects may create a whole new market, or because R&D projects may affect market share rather than vice versa. However, this would be interesting for further research.
Lastly, most interesting is the finding for the long term marketing capabilities variable. This variablehas been found to be highly significant in the logit model, but also was an important variable in the prediction of bankrutpcy in the RF model. This is particularly interesting, because it suggests that the long term, marketing effectiveness is highly
significant and important in the prediction of bankruptcy. Hence, marketing effectiveness and expenditure is significant and important, but it is also highly relevant that marketing
9. Managerial implications.
Marketing has largely been considered purely a short-term expense within the field of finance (Day and Fahey. 1998), and could thus not aid in predicting a long-term future. Hence, it has become increasingly important for marketing to be connected to financial rewards and risks in order for marketing to get a “seat at the table” and part take in corporate decision making. In this study it was shown that marketing and R&D expenditure can have a significant effect on the likelihood of bankruptcy, which could add significant weight to the importance of the marketing department.
Moreover, this study highlighted the imporrtance of long-term marketing capabilities. The long term marketing capabilities turned out to be one of the most important variables in the prediction of bankruptcy. First of all, the effect of long term marketing capabilities was highly signifcant, but also considerably more significant than the short term marketing capabilities. This would suggest that marketing managers should pay particular attention to the long-term effects of their projects, as they could have a significant effect on the financial prosperity and stability of a firm. This may prove difficult as it can be difficult to distribute marketing resources where they will have the largest long-term effect (Dekimpe and Hanssens. 1999).
Lastly, it appears that literature is correct on the role of market share and advertising. According to this study, the effect of marketing effectiveness was significantly, positively moderated by market share. Therefore, as market share increases manager should pay even more attention to marketing expenditure and the effectiveness of this expenditure in terms of sales, as this would become signficantly more important.
10. Conclusion.
This study aimed to study the effect of marketing and R&D variables on the likelihood of bankruptcy through machine learning methods. It was argued that old, accounting based ratios were no longer sufficient enough to predict bankruptcy in recent datasets. Hence, marketing variables were argued to be a good addition to increase the accuracy rates of bankruptcy prediction models.
significantly, negatively affect the likelihood of bankruptcy. This result was found to be positively moderated by the market share. Additionally, the long term marketing capabilities were found to be highly important.
11. Recommendations & Limitations
11.1 Recommendation: Explore the moderating effects of industry concentration.
Instead of market share as considered in this study, it would also be highly interesting to further research the moderating effects of industry concentration marketing activity becomes increasingly more important as industry concentration increases. Within low concentration industries competition is primarily price-based, whereas in high concentration industries branding and advertising become the primary method of competition (e.g. Coca Cola and Pepsi in the soda industry). Therefore, in high concentration industries abnormal profitability is possible as opposed to low concentration industries which have ‘normal’ profitability levels (Collins and Preston. 1969). Hence, as industry concentration increases, the importance of advertising increases, as in higher concentrated markets it becomes more important to maintain or gain market share, which leads to high profitability. it could also be argued that when industry concentration increases, R&D becomes increasing important in reducing firm risk. Such effects were found to hold true for R&D spending, which would significantly increase when industry concentration is higher (Nootenboom and Vossen. 1995). This is seemingly logical, as firms in high concentrated markets will typically have higher cash reserves and profits to spend on R&D (Vossen. 1999). However, increased R&D spending is significantly more valued within high concentrated industries. For example, according to Doukas and Switzer (1992) it was found that announcing R&D expenditure significantly increased the market-to-book ratio in high concentrated industries, and has the opposite effect in low concentrated industries. This is in accordance with the Schumpeter’s theory on innovation and industry concentration, which suggests that innovation becomes more important for firms in high concentrated industries in order to maintain market power (Schumpeter. 1954).
11.2 Recommendation: Using RF for bankruptcy prediction.
Furthermore, in general it appears that the RF model generally outperformed the other
et al. (2017) and Min and Lee (2005), which argued that the SVM model was the most successful in predicting bankruptcy. This is an interesting observation, because the RF model is still relatively missing in most bankruptcy studies, which is why in Kumar and Ravi (2007) it was stressed that it may be worthwhile to study the effects of the RF model. Hence, it may be interesting to further explore the role of the RF model in further research.
11.3 Limitation: Data manipulation.
One of the major limitations of this study was the manipulation of the data through SMOTE, and the imputation of some missing data through the RF-method. The first problem SMOTE resolved was the issue of the imbalanced dataset. This limited this study in two ways. Firstly, a large amount of (non-bankruptcy) data was omitted. The omission of this true data could have significantly contributed to make a more accurate model.
The second problem was with the dataset, where cases of bankruptcy were in a significant minority and there was also a large amount of missing data for these rare cases of bankruptcy. The imputation of this data is, no matter how good the imputation may be, not ‘true’ data, and thus decreases the validity of the model.
12. References
(Jane) Hsu, L. and (Shawn) Jang, S. (2008). Advertising expenditure, intangible value and risk: A study of restaurant companies. International Journal of Hospitality Management, 27(2), pp.259-267.
Abi.org. (2018). The Paradox of Corporate Bankruptcy in a Robust Economy | ABI. [online] Available at: https://www.abi.org/abi-journal/the-paradox-of-corporate-bankruptcy-in-a-robust-economy [Accessed 11 Jun. 2018].
Agarwal, V. and Taffler, R. (2008). Does Financial Distress Risk Drive the Momentum Anomaly?. Financial Management, 37(3), pp.461-484.
Ahluwalia, R., Burnkrant, R. and Unnava, H. (2000). Consumer Response to Negative Publicity: The Moderating Role of Commitment. Journal of Marketing Research, 37(2), pp.203-214.
Alaka, H., Oyedele, L., Owolabi, H., Kumar, V., Ajayi, S., Akinade, O. and Bilal, M. (2018). Systematic review of bankruptcy prediction models: Towards a framework for tool selection.
Expert Systems with Applications, 94, pp.164-184.
Aleksanyan, L. and Huiban, J. (2016). Economic and financial determinants of firm bankruptcy: evidence from the French food industry. Review of Agricultural, Food and
Environmental Studies, 97(2), pp.89-108.
Alexander, D., Flynn, J. and Linkins, L. (1995). Innovation, R&D productivity, and global market share in the pharmaceutical industry. Review of Industrial Organization, 10(2), pp.197-207.
Al-Hroot, Y. (2016). Bankruptcy Prediction Using Multilayer Perceptron Neural Networks In Jordan. European Scientific Journal, 12(4).
Altman, E. (1968). The Prediction of Corporate Bankruptcy: A Discriminant Analysis. The
