Portal de Periódicos da CAPES

Conventional and neural network target‐matching methods dynamics: The information technology mergers and acquisitions market in the USA

2021; Wiley; Volume: 28; Issue: 2 Linguagem: Inglês

10.1002/isaf.1492

1099-1174

Ioannis Anagnostopoulos, Anas Rizeq,

Financial Markets and Investment Strategies

In an era of a continuous quest for business growth and sustainability it has been shown that synergies and growth-driven mergers and acquisitions (M&As) are an integral part of institutional strategy. In order to endure in the face of fierce competition M&As have become a very important channel of obtaining technology, increasing competitiveness and market share (Carbone & Stone, 2005; Christensen et al., 2011). During the post-2000 era, this is also a point where more than half of the said available growth and synergies in M&As are strongly related to information technology (IT) and its disruptive synergistic potential, as the first decade of the 2000s has shown (Sarrazin & West, 2011). Such business growth materializes at the intersection of internalizing, integrating, and applying the latest data management technology with M&As where there are vast opportunities to develop (a) new technologies, (b) new target screening and valuation methodologies, (c) new products, (d) new services, and (e) new business models (Hacklin et al., 2013; Lee & Lee, 2017). However, while technology and its disruptive capabilities have received considerable attention from the business community in general, studies regarding the examination of technology convergence, integration dynamics, and success of M&As from a market screening and valuation perspective are relatively scarce (Lee & Cho, 2015; Song et al., 2017). Furthermore, little attention has been devoted to investigating the evolutionary path of technology-assisted, target screening methods and understanding their potential for effective target acquisition in the future (Aaldering et al., 2019). We contribute to this by examining the application of neural network (NN) methodology in successful target screening in the US M&As IT sector. In addition, while there are recognized idiosyncratic motivations for pursuing M&A-centered strategies for growth, there are also considerable system-wide issues that introduce waves of global M&A deals. Examples include reactions to globalization dynamics, changes in competition, tax reforms (such as the recent US tax reform indicating tax benefits for investors), deregulation, economic reforms and liberalization, block or regional economic integration (i.e., the Gulf Cooperation Council and the EU). Hence, effective target-firm identification is an important research topic to business leaders and academics from both management and economic perspectives. Technology firms in particular often exhibit unconventional growth patterns, and this also makes firm valuation problematic as it can drive their stocks being hugely misvalued (i.e., overvalued) therefore increasing M&A activity (Rhodes-Kropf & Viswanathan, 2004). Betton et al. (2008) claimed that predicting targets with any degree of accuracy has proven difficult in their general conclusion regarding M&A research aimed at target prediction. Hennessy (2017) argued that such firms add an extra and significant layer of complexity in terms of appropriate target identification and valuation. This is so because such firms either have limited/no past history or go through an invariable and dynamically high volume of transformations throughout their lives; they also enjoy unique businesses and/or products and, consequently, there are not directly visible and comparable peers or competitors. It is extremely challenging to identify and match appropriate targets that could recondition a company, evaluate how much to invest, and how to assimilate them (Christensen, 2016). The problem stems from the fact that upper management (with the exception of motive distortion), either through unintended hubris or myopia, erroneously match targets to the strategic intent of a deal. In this way, managers fail to discriminate between targets that can improve the company's growth prospects from those that could dramatically dampen performance. Most of the research so far concentrates on linearly attempting to predict the probability of targets been acquired through regression analysis and the related distress signals (i.e., bankruptcy probabilities) using publicly available data sets on firms and then applying the aforementioned models. In reality, though, and in most cases, social researchers do not work with data from well-planned laboratory experiments where this is possible. One cannot safely assume that past data, interaction data, and obviously human performance-related data are characterized by an ideal distribution (Sparrowe et al., 2001). Social science studies and the associated data have a tendency for nonlinearity, clustering around certain sections, and being skewed with respect to particular variables (Very et al., 2012). More recently, Lipton and Lipton (2013)—and earlier Tam and Kiang (1992)—also supported this notion by arguing that the variety and multiplicity of exogenous factors is so persistent at any given time that they affect mergers and make it impossible to predict the level of future merger activity. For all the aforementioned reasons, the assumptions ingrained in such methods constitute an unrealistic prospect for many informational sets utilized by scholars. Prediction performance can potentially be enhanced by utilizing more data in an effort to construct helpful models; yet, Cremers et al. (2009) argued that, equally, a predefined set of data in order to obtain a regression model is an additional issue. However, in both preceding statements, the inability to deal effectively with nonlinearity is a critical drawback of multiple linear regression (Detienne et al., 2003). This de facto observation in the domain of management science renders linear regressions questionable in some instances. If established knowledge of nonlinearity exists, then it is much easier to treat; however, in many instances, researchers may be unaware of the nonlinearity among the chosen variables.Consequently, companies employ less well-suited methods, they pay the wrong (usually higher) price, they fail to realize the synergistic potential, and they significantly increase their costs after integrating the target in the wrong way (Ma & Liu, 2017). Furthermore, both the choice of target screening methodology and the associated performance measures have been long-standing issues facing researchers and managers alike. Past research has shown that, ex ante capital market reactions to an acquisition announcement exhibit little relation to corporate managers' ex post assessment fundamentals (Schoenberg, 2000). M&A announcements archetypally encompass a big premium over going market prices (between 30% and 40% on average) and result in a large and swift change in market prices suggesting the announcement is news to the market (Eckbo, 2014). Approximately half of all M&As fail, with a very large proportion of those considered successful producing negligible gains for their shareholders. At the same time, the upper management of targets departs within 3–5 years following the acquisition completion (Agrawal & Jaffe, 2000; Krug & Aguilera, 2005). Hence, the appropriateness of target screening and valuation methodologies in important economic sectors, such the IT services sector, remain overlooked. It is thus of both business and academic importance to investigate how the underlying trajectories interact in order to seize resulting opportunities. The evolution of artificial intelligence, big data, and modern screening methods is steadily leading managers to reject a one-size-fits-all valuation in favor of more individualized, reconfigurable, and understandable accounting data—data that they can customize and fit into their own structures to meet their own decision-making needs. Consequently, managers increasingly investigate methods and digital data that are less static, less summarized, more concise, and more agile. In that line of thinking, O'Leary (2019) argued that emerging model designs enhance database capabilities and that the parallel development of multiple and different learning models will revolutionize the use of accounting data. In this paper, we explore, apply, and compare two types of target screening classification techniques—the NN and the traditional logistic regression (LR) M&A forecasting techniques—in terms of successful target prediction in the IT M&A market for the USA. We provide for a demonstration of the growing prospects of the use of an NN to systematize feature engineering from raw time series, in a more methodical way as a result of the strategic change in the types of digital commodities that decision-makers demand. In that respect, and within the context of M&As, predicting which companies will become takeover targets and the ability to discriminate between high‑ and low-quality targets is very important for managers and financiers, as well as for regulators and competition market committees. Our findings provide valuable insights to guide managers in financial and other organizations to improve their performance through suitable target (or nontarget) screening methods. Having provided the introduction and our motivation for our study in Section 2 we explore the literature, revolving around three major threads: namely, first examining the relevant empirical evidence on M&A deals in the USA; second, uncovering the valuation challenges; and third, exposing the two models discussed in Section 3. We do so with the aim of aiding the understanding of the evolution of such complex dynamics. In Section 3 we discuss our methodologies, where determinant variables and NNs are also compared vis-à-vis LR. We do so with regard to aiding an understanding of the tensions in model dynamics. Section 4 presents the results of our analysis, and our last two sections provide the conclusions, discussion, and recommendations. Advancements in technological disruption are fueling a winner-takes-all environment, and firms with the most effective target asset matching will create more distance and differentiation between the largest, most successful firms and the rest of the market (Hennessy & Hege, 2010). Any breakthrough in the capacity to more reliably forecast firms and spotting which companies are likely to translate first-mover advantages into market power that will engage in a merger deal would be very lucrative for an investor in the financial markets. Over the past two decades, approximately more than one-third of worldwide deals have involved US-based companies. This percentage stood above 50% in 2000 due to the tech bubble. In the first part of this section we perform a brief empirical exploration in order to (a) demonstrate how successful M&As of tech-driven, digital business models (e.g., fintechs) have become the instrument of choice to acquire needed technologies, capabilities, and scalabilities in order to close innovation gaps, deflect competition, and (b) as a path to growth through the power of 'ball-point statistics', to demonstrate to business leaders and academics the importance of M&As in the IT sector as well as help our exposition of utilizing agile and reliable methods for successful target acquisition. For the period 2000–2017 the USA has had the biggest stake of global M&A activity. Since then, this trend has been in decline, yet US firms command an average of 37% during the 17 years, as shown in Figure 1 (Bureau of Economic Analysis, 2017). By value, US firms represent 20% of global M&As as acquirers and 23% as targets by value (Ernst & Young, 2015). With a considerable percentage of M&As failing, screening and valuation is a subject of high importance in this field. In an M&A transaction it is as important and crucial for both the acquirer and the target to justify the synergies through matching and come to a fair valuation. It is equally important for the shareholders of both firms to determine a fair value for the deal and justify the acquisition price (Petitt & Ferris, 2013). Failing to realize opportunities on expected synergies during integration is one of the most common risks in any deal. Financial due diligence is steered through imperfect information, where both sides negotiate their best to achieve advantageous terms in a short time frame, and such an effort is typically focused on hypothetical values instead of potential values. This is suitable in terms of containing the risk of overpaying, but, overall, it is not the optimal way to manage a venture for maximizing its potential. This is so because, in order to drive the cost-saving speed of deal completion and efficacy, managers either have an overly unyielding attitude to integration (i.e., the equivalent of disaster myopia)—which fails to recognize the red flags, the unique attributes and requirements of different deal types—or they are equally totally unstructured, ignoring established deal identification and valuation processes, often relying solely on top executives to see their projects through (i.e., the hubris with anchoring bias; Kahneman et al., 2011). As such, there is seldom a screening structure in place in order to review value-creating targets with other executives, board members, and stakeholders. It has been argued that those that can effectively overcome the first puzzle of identifying appropriate targets and then build the capability and acumen required to attain their full capacity for growth can enjoy an enduring competitive advantage (O'Reilly & Pfeffer, 2000; Eisenhardt & Sull, 2001; Rheaume & Bhabra, 2008; Straub et al., 2012). Hence the need of a reliable takeover prediction model.In 1996 there were 274 initial public offerings (IPOs) of IT firms, with the dot.com crash (1999–2000) that fueled bubble investments and IT-related M&As being the peak with 371 and 261 tech IPOs respectively (Ritter, 2017). This 1990s tech-euphoria was the result of: As a result of poor valuation and target screening fundamentals, the market was extremely overvalued when NASDAQ reached its highest level1 in March 2000, where it lost more than 50% by value in October 2000. Since then the number of public technology firms has also decreased by 50% according to Figure 3 as a result of (a) firms being delisted, acquired, or going bankrupt and (b) less IPO activities, where firms choose to remain private (Mauboussin et al., 2017). Figure 4 shows that while the number of deals has shrunk from its peak levels, the value of deals has comparatively increased in the US technology sector. The overall M&A market has grown significantly over the past 5 years, and the value of deals involving a tech target has been rising even faster. There is a concern from the investing community that valuations remain unjustifiably high (Albuquerque et al., 2016; Wong, 2018). As the pace of technology-driven change accelerates, a key question has become how one does identify, value, and integrate technology acquisition so that it can be beneficially positioned in a highly disruptive ecosystem. The technology sector, though, is also characterized by rapidly evolving firms that operate under different dynamics compared with other sectors, such as high levels of uncertainty and risk as well as lack of positive cash flows (Anagnostopoulou, 2008; Aydin, 2015; Core et al., 2003; Lev & Zarowin, 1999). This makes their valuation very challenging, especially in IT, where the value of a technology is only known after it is commercialized to the market and, hence, the potential for significant mistargeting and mispricing departures from fair values is a real implication. Consequently, one of the major challenges with technology companies is also valuation, where ideally an accurate measure of value should be utilized that is also hard for management to manipulate. This is the polemic of the next section. As early as the 1990s DeAngelo (1990) argued that, with respect to these types of firms, any conventional valuation method has the potential to be distorted as it does not incorporate intangibles, yet acquisition premiums can typically be more than 50% above market value. Benou and Madura (2005) stated that even hedge funds in the USA hire technology consultants in order to provide expert insights about tech firms as they are hard to value from a standalone financial perspective. A technology firm's value is directly derived and dependent on patents and growth; as such, most of the time, value originates from intangible future capabilities, future customers, and scalabilities, as well as future services/products, not as much from current operations (Chandra et al., 2011). Chan et al. (2001) argued that the core value of technology firms originates from intangible assets. These assets, however, do not clearly emerge (or indeed necessarily materialize) on a firm's financial statements due to the lack of clarity in accounting standards to include such intangibles, such as innovation, customer satisfaction, and human capital, resulting in complexities when it comes to performing an equity valuation. Valuation multiples have also been questioned as appropriate valuation methods. Daniel et al. (1998) have long argued that investors, potential shareholders, and managers tend to be overconfident when probing ambiguous evidence and information. Furthermore, they support the view that mispricing is more persistent for stocks whose value is strictly tied to their growth potential, such as IT stocks. Considering the asset-specificity characteristics of the technology market, it is quite often very difficult to value, predict, and choose suitable target firms. Standard regression-based modeling studies on M&As takeover prediction incur several limitations, such as the exclusion of technological variables, the omission of variables, and misrepresentation when data are located within noisy environments (Wei et al., 2008). Other researchers (Routledge et al., 2013) corroborated this by stating that the explanatory power of models such as LRs is comparatively lower. They advocate the use of machine learning (ML) and textual analysis as potential alternatives capturing nonlinear aspects that conventional linear models fail to consider. The traditional multiple linear regression model provides no direct indicator as to whether the data are best presented linearly, with the choice of predictor variables being another issue in order to enhance the equation prediction power in the traditional regression method (Ansari & Riasi, 2016). This problem can be further exacerbated by the presence of multicollinearity, where correlations among hypothetically independent variables are a major deficiency in regression tradition when they go unidentified. As such, if variables are collinear then using more data is not superior, and in extreme cases of collinearity the regression becomes unstable and minor changes in the inputs can influence output results wildly. In this case, the tendency to regress too many variables at the same time increases error (Liang & Yuan, 2013). Lastly, multiple regression necessitates information on the fundamental distribution of the data in order to specify a model, and a major requirement is that data are normally distributed. At the other end of the spectrum, NN-based modeling is, by its nature, nonparametric and nonlinear, where prior model specification among inputs and outputs, by design, is not required. Such methods can deal with data sets that typically exhibit significant uncertainty and experience uncharacterizable nonlinearity. ML-based models can transform and correspond to multifaceted nonlinear associations more efficiently than regression techniques can. In the M&A field there is research evidence that with the continuous development of globalization, increasing correlations among economies, and increase in uncertainty, the valuation and forecast accuracy of M&A performance can be drastically enhanced through the application of artificial NNs (An et al., 2006; James et al., 1998). Such network-based techniques are free of the assumption that variables are autonomous and uncorrelated with each other (Ragothaman et al., 2003). Thus, by comparison, NNs do not require the assumptions of regression; they are an extremely flexible tool; and they are less sensitive to the conundrum of dimensionality that hinders forecast testing. For example, Hou et al. (2015) recently presented a cascading, multidimensional, dynamic-merger agent-based computational model driven by competition pressure diffusion to describe the forming process of industry merger waves. Their results support the argument that NNs have many desirable features that can synchronize with modern economy characteristics. Such methods have the aptitude to absorb knowledge from experience, use it to allow adaptive adjustments to the predictive model as new examples become available, and generalize the results in complex scenarios. As new data are fed into the model, past data are not ignored either; they are used but their weight is rather reduced gradually as new data sets are fed into the network. Such features are potentially even more prevalent today, where the modern economy moves exponentially to big-data analytics. The latter argument also provides scope for an opportune and clear example regarding the role, suitability, and value of NNs in the modern economy. A very real and highly topical illustration is the competence of real-time artificial applications being highly adaptive to the data in hand and increasingly being applied in today's modern financial and nonfinancial companies absorbing and integrating IT; examples include firms like fintechs, regtechs, and insurtechs and companies such as Amazon, Uber, and Airbnb. All these are a subset of a very dynamic setting characterized by rapidly changing agents, markets, and valuations. In turn, all call for the use of agile, flexible, and real-time tools that promote effective decision-making, where (a) artificial intelligence and NNs offer nearly limitless online processing and evaluation capabilities and (b) materially deliver more effective decision-making. The World Economic Forum (2015), in their recent and globally highly influential report, projected that the rise and continuous development of artificial intelligence is going to rather be an unceasing force for innovation that will impact both investor behavior and the long-term structures of the services sector. Hence, an additional important managerial aspect for consideration and research is whether NNs seem to indeed be managerially superior tools for forecasting and valuation purposes, particularly in situations loaded with high data velocity and complexity. As already discussed herein, "typical" prediction models have so far been relatively indeterminate in the past and, as a result, have not had highly reliable estimations regarding the scale of an outcome or a conclusion on the directional relationships of the variables. As such, modeling issues become crucial, and these are discussed in the section that follows. Researchers have also suggested that valuation and acquisitions' studies should consider employing multiple methodologies and multiple performance measures in order to gain a holistic view of the target outcome, while, in the longer term, opportunities remain to identify and refine improved metrics (Schoenberg, 2006). Starting with binary LR, this is a nonlinear model, where the dependent variable is a binary or dummy variable. This model requires very few assumptions compared with other similar statistical dependence techniques, such as discriminant analysis. A classic example is Ohlson's (1980) study that utilized LR analysis in order to examine the relationship between binary/ordinal response probability and descriptive accounting variables. This research emphasized the significance of utilizing data directly from a firm's financial statements—as they indicate whether a firm filed for bankruptcy before or after announcing them—which assists with avoiding the "back-casting" problem (i.e., applying the model to a firm's fundamentals after going through bankruptcy). Ohlson's model produced an accuracy rate of prediction of 96% with a cut-off point of 0.5. The use of probit models posits that the dependent variable can take only one of two values (i.e., 1 for acquired or 0 for not-acquired), in order to arrive at a probability of a firm being acquired or not as well as uncovering what are the characteristics that affected such a probability (Harris et al., 1982). LR was also used by Dietrich and Sorensen (1984) to predict acquisition likelihood. Palepu's (1986) use of a binomial logit probability model with nine independent variables suggested a good fit of success in predicting a high number of targets. The results, though, were mixed, in that the model also predicted a high number of nontargets as targets; therefore, the model suffered from the investment/managerial limitation that it was not enough as a construct for gaining abnormal returns. All these previous studies have shown varying prediction power anywhere between 50% and 90%. All these findings are overstated and suffer a biased estimate due to two main flaws in such methodologies: (a) state-based sampling for model estimation and prediction testing; (b) using predetermined, arbitrary, optimal cut-off probabilities (Palepu, 1986). In addition, Powell (1997) criticized the extensive use of binomial models, arguing that the characteristics of hostile and friendly takeovers differ; therefore, treating hostile and friendly takeovers in the same group will inevitably cause misleading results. Cudd and Duggal (2000) utilized Palepu's (1986) nine factors but added an industry dispersion factor to account for different industries. The authors claimed this improved the accuracy of the said model. In addition, they also found the dummy variable "industry disturbance" to be significant. This indicated that an acquisition in the same industry (in the past 12 months) will increase the probability of takeover for the remaining firms in that industry. Since then, NNs and artificial intelligence using genetic algorithms have been supported by the proponents of what is generally called the ML prediction area. Relevant research so far in the utilization of ML within the business/finance domain generally seems to support the view that alternative prediction techniques can potentially generate a higher forecast accuracy. Furthermore, they highlight the proposition that it is possible to enhance the performance of managerial decision-making via the utilization of complementary techniques as opposed to competing methodologies. This is further examined below. The use of NNs, ML, and data mining are differentiated methodologies utilized by researchers in order to predict bankruptcy and/or takeover targets. Salchenberger et al. (1992) compared NNs with LR to test healthy and failed thrift institutions and concluded that NNs achieved higher accuracy. Other studies have supported the use of NNs in outperforming LR in predicting bankruptcy (e.g., Fan & Palaniswami, 2000; Jo & Han, 1996; Maher & Sen, 1997; Tam & Kiang, 1992; Tseng & Hu, 2010). Liu et al. (2007), using self-learning and organized mapping models with Hopfield NNs to cluster data, showed accuracy predictions of 80.69% for targets and 63.11% for nontargets. In a parallel study, Tsai and Wu (2008) studied the effects of incorporating multiple NN classifiers in credit scoring and the associated bankruptcy prediction. They found that single NN classifiers outperform multiple NN classifiers in both credit scoring and bankruptcy prediction. A contribution of their paper is that it also suggests the inclusion of nonfinancial variables for improving predictability power. Branch et al. (2008) utilized both NNs and LR to predict whether a takeover attempt will succeed or not, with the authors concluding that "...neural network model outperforms logistic regression in predicting failed takeover attempts and performs as well as logistic regression in predicting successful takeover attempts" (p. 1186). In all of the aforementioned studies a common results' attribute emerges: NNs seem to possess a higher flexibility and ability to address nonlinearities. This echoes the statement of Zhang et al. (1999) that NNs can potentially be robust and can provide more reliable estimations when applied on different samples only once the optimal architecture is found. More recently—and over the last decade in genera—there is a growing body of literature (importantly, from a variety of both academic and professional disciplines) supporting the use of alternative, complementary methodologies within the interplay domain of finance and technology. For example, in the context of finance, Lahmiri and Gagnon (2016) used a two-stage algorithm for bankruptcy prediction in financial institutions and governments. They used a regression model to process and select important features that affect business failure prediction and then they adopted a probabilistic NN model to achieve high probability of success for data classification. Following this, Lahmiri (2017) proposed a two-step prediction approach within a bank telemarketing problem, showing it is robust enough to deal with nonlinear and noisy data and it is suitable to make fast and easy predictions for large data sets. Khashman (2010) advocated the use of NNs in evaluating credit scoring in loan application approval. Khandani et al. (2010) applied ML techniques to construct nonlinear, nonparametric forecasting models of consumer credit risk. The authors showed that through constructing out-of-sample forecasts NNs can significantly improve the classification rates of credit-card-holder delinquencies and defaults. These also improve linear regression R2 values of forecasted/realized delinquencies of 85%. The authors concluded that complementary, aggregated consumer credit-risk analytics may also have important applications in forecasting systemic risks. Brown and Mues (2