Managerial Finance

Confining value from Neural Networks: A sectoral study

prediction of takeover targets in the U.S Technology Sector
M
Journal: Managerial Finance

Manuscript ID MF-12-2017-0523.R1
an
Manuscript Type: Original Article

Keywords: Neural Networks, Takeover targets, Prediction, Technology

ag
er
ial
Fin
an
ce
Page 1 of 31 Managerial Finance

1
2
3 Confining value from Neural Networks: A sectoral study prediction of takeover targets in the
4 U.S Technology Sector
5
6 ABSTRACT
7
8
9
10 Published studies in the area of predicting M&As have made a relatively limited attempt to
11 use neural network systems (NNs henceforth) in such a decision making process. This paper
12 examines the value of utilising a neural networks approach using M&A data confined in the
13 U.S technology domain. Investors value firms before investing in them to identify their true
14 stock price; yet, technology firms pose a great valuation challenge to investors and analysts
15
16
alike as the latest IT stock price bubble in Silicon Valley and as the recent stratospheric rise of
17 Financial Technology (FinTechs henceforth) companies have also demonstrated. At the same
M
18 time, the technology sector in the US commands approximately 8% of GDP and accounts for
19 around 20% of all M&A deals in our sample period. We utilise US technology firms’ data
20
an
from Bloomberg for the period 2000–2016. Our analysis applies and compares a neural
21 network approach to a linear classifier, logistic regression. Our empirical results show that
22
neural networks are a promising method of evaluating M&A takeover targets in terms of
23
ag

24 their predictive accuracy and adaptability. The higher level of accuracy provided by a neural
25 network approach can provide practitioners with a competitive advantage in pricing merger
26 offers. Trade-offs and limitations of using neural nets as an alternative, general modelling
er

27 tool are also discussed. Our findings emphasise the value alternative methodologies
28 potentially provide in high-technology industries in order to achieve the screening and
29
ial

explorative performance objectives, given the technological complexity, market uncertainty

30
and the divergent managerial skill sets required for breakthrough innovations in these
31
32 sectors. This study provides valuable insights to managers aiming to increase the
Fin

33 effectiveness of their decision-making for diversification, growth portfolios and investments.

34
35
36
an

37
38
39
ce

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
 Managerial Finance Page 2 of 31

1
2
3 1. Introduction
4
5 Shareholder theory posits that management has a mandate to maximise shareholder wealth
6 through decisions that add value to investments and stimulate growth. For any large
7 company, growth through Mergers and Acquisitions (M&As henceforth) is often a key part
8
of corporate growth strategy. Growth largely drives value creation and M&As can offer a
9
10 course to growth when esoteric opportunities are restricted through projected financial,
11 strategic and operational synergies achieved at a fair price. Numerous studies though have
12 shown that M&As more often than not destroy value rather than create it. More than 50%
13 of all M&As lead to a decline in relative total shareholder return after one year. Hence,
14 effective target identification must be built on the foundation of a credible strategy that
15 identifies the most promising market segments for growth, assesses whether organic or
16
17
acquisitive growth is the best way forward, and defines the commercial and financial
M
18 hurdles for potential deals. It is thus crucial for companies’ upper management to utilise
19 credible and proven methodologies and models to ensure that target identification is based
20
an
on sound background research. For example, as early as 20 years ago, researchers (Kaastra
21 and Boyd, 1996; Rojas, 1996) argued for the systematic application of neural networks as a
22 method to deal with the problem of non-linearity in financial transactions. Rojas (1996)
23
ag

argues that where there is abundance of data but less theoretical understanding (for
24
25 example behavioural patterns that are not easily identified through established linear
26 methods or continuously keep changing) neural networks can discover statistical regularities
er

27 and keep adjusting parameters even in a changing environment (p.247).

28
29
ial

When valuing a firm three major sources of valuation inputs are considered: (i) Current
30 Financial Statements; (ii) Firm’s Past History; and (iii) Peer Group comparisons. While for
31
most firms such crucial information is ready-made, for technology firms such vital sources
32
might be absent. Their financial statements don’t include much information about growth
Fin

33
34 prospects either. Most technology firms have limited or no past history. They also possess
35 unique businesses and/or products therefore leading to no directly visible peers or
36 competitors (Damodaran , 2001): “As more and more technology firms get listed on financial
an

37 markets, often at very early stages in their life cycles, traditional valuation methods and
38 metrics often seem ill suited to them.” (p. 19). Daniel, et. al (1998) have demonstrated that
39
investors tend to be overconfident when examining unclear information and that mispricing
ce

40
41 is stronger for stocks whose value is closely tied to their growth. The very fact that
42 technology firms more often than not exhibit unconventional growth patterns makes them
43 difficult to evaluate and can lead to their stocks being massively mis-valued (most of the
44 time over-valued) and therefore increasing M&A activity (Rhodes-Kropf & Viswanathan,
45 2004; Jovanovic & Rousseau, 2001). While there are idiosyncratic motives for undertaking
46 M&A-led growth strategies, there are also substantial economy-wide factors which cause
47
48
waves of global M&A activity such as responses to globalization forces and increases in
49 competition, de-regulation and the associated economic reforms and liberalization,
50 block/regional economic integration (i.e. the EU). As such, target firm identification, has
51 become a great research interest area both to the business world and academia alike. The
52 three latest M&A waves (namely, M&As waves 5, 6, and 7) make the case in point:
53
54
55
56
57
58
59
60
Page 3 of 31 Managerial Finance

1
2
3 - Fifth Wave: 1993 – 2000
4 A wave known for its large transactions and overvaluation of firms. Transactions were
5 mostly friendly and financed by equity (Andrade, et al., 2001). It was empowered by cross-
6
border transactions due to the strong economic conditions in the U.S, Europe and Emerging
7
8
Markets. This wave ended with the burst of the internet bubble causing the market to crash
9 (Dieudonne, et al., 2014).
10 - Sixth Wave: 2003 – 2008
11 A wave known for producing less overvalued transactions, with the size of both acquirer and
12 target getting smaller. During this wave firms enjoyed more cash, with the excess liquidity
13 causing this wave (Alexandridis, et al., 2012), and having 75 percent of the transactions paid
14
by cash (Gregoriou & Neuhauser, 2007). This wave ended with the 2008 credit crisis.
15
16 - Seventh Wave: 2010 – Present
17 The wave which we are currently experience started gradually in 2010, and it coincided with
M
18 the emergence of FinTechs. Since 2015, it has reached an all-time high of 2.9 trillion dollars
19 in value (Institute for Mergers, Acquisitions and Alliances, 2016). It is aso owed to the
20
an
system-wide steps taken by central banks after the 2008 credit crisis, such as keeping near-
21 zero interest rate and the quantitative easing procedures which supplied equity and bonds
22
markets with enough liquidity.
23
ag

24
25
With an increasing amount of M&As in the technology sector, it is crucial to identify targets
26 before announcement date, as this can be significantly beneficial for investors, target and
er

27 acquiring firms. From that comes the motivation of a reliable takeover predication model.
28 Standard models have so far been relatively indeterminate in the past, and as a result have
29
ial

not had highly reliable estimations regarding the scale of an outcome or a conclusion on the
30 directional relationships of the variables (Betton et al., 2008; Routledge et al., 2013; Eckbo,
31
2014;). Hence, we switch our approach to an altered empirical exploration model where this
32
is also tested. We elaborate further on our motivation below where by implication we
Fin

33
34 discuss our reasons for utilising NNs as opposed to the traditional regression techniques.
35
36 Having introduced our study topic, we discuss the relevant empirical evidence on the
an

37 characteristics of M&A deals in the US, valuation challenges and the technology sector in
38 section 2 that follows. In section 3 we discuss methodological issues where determinant
39
variables and neural networks are compared with the traditional statistical techniques of
ce

40
41 discriminant analysis and logistic regression with regards to the identification of potential
42 takeover targets. Section 4 discusses our methodology and section 5 presents the results of
43 our analysis. The conclusions of the study are presented in section 6.
44
45 1.1 Motivation Summary
46
47 Financial time series have some characteristics that make them hard to reliably forecast,
48 especially when a traditional statistical method is employed. Such characteristics are as
49
follows (Motiwalla and Wahab 2000; Thawornwong and Enke 2004; Versace et al. 2004):
50
51
52 1. Non-stationarity of data, where due to different business and economic cycles, the
53 statistical properties of financial data change randomly over time, which also introduces:
54 2. Non-linearity of data, where the relationship between the financial and economical
55 independent variables and the desired dependent variable may not be linear. Intensified by:
56
3. Noisiness through daily variations in financial time series.
57
58
59
60
 Managerial Finance Page 4 of 31

1
2
3 On the other hand, NNs are more flexible and adaptable computing methods that provide
4 the ability to potentially capture the patterns among variables more effectively. Hence, the
5 use of NNs to forecast financial time series as an alternative is justified by some of the in-
6
build qualities they posses. Such characteristics make them reasonably well suited for use in
7
8
the financial forecasting domain (Hussain et al. 2007; Lin et al. 2006; Lam, 2004; Eakins and
9 Stansell, 2003):
10 1. Their nonlinearity. NNs can capture nonlinear relations between element (input or
11 independent variables) and response (output or dependent variables).
12 2. Their data driven nature. No prior explicit relational assumptions on the model are made
13 or modelled between inputs and outputs.
14
3. Their generalizability. Once trained, NNs can produce relatable results even when the
15
16 data structure has changed or when they are faced new input patterns.
17 4. Their assumption neutrality. Dissimilar to traditional statistical techniques, NNs do not
M
18 employ pre-constructed assumptions on the input data distribution.
19 Yet, as with any forecasting tool, the robustness of a NN application outcome can equally be
20
an
questioned. This is addressed in our results and discussion sections at the end of this
21 exposition.
22
23
ag

2. Literature Review: Mergers and Acquisitions in the U.S Technology Sector

24
25
26 The U.S is well-known as the most preferred international investment destination measured
er

27 by Foreign Direct Investment (FDI) flows. In 1989 the FDI position in the U.S (FDIUS)
28 exceeded $400 billion (Harris & Ravenscraft, 1991), while in 2014 for example it totalled $2.4
29
ial

trillion with an average annual growth of 8.9 percent (Organization for International
30 Investment, 2016; Bureau of Economic Analysis, 2017). It is ranked as the world’s top market
31
for 5 years consecutively (Laudicina & Peterson, 2017). Figure 1 illustrates M&As as the most
32
exercised type of investment in the U.S in volume vs. expansions and new establishments.
Fin

33
34
35 Figure 1: FDI in the United States by type 1994 – 2016, US Bureau of Economic Analysis
36
an

37
38
39
ce

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55 *Due to government budget cuts the Bureau of Economic Analysis was not able to fully conduct a concluding survey from
2009 to 2013.
56
57
58
59
60
Page 5 of 31 Managerial Finance

1
2
3 Rossi & Volpin (2003) suggested the role of the legal system as a factor affecting cross-
4 borders M&A volume. The rationale being that countries with mature legal systems are
5 better able to cope with economic changes, absorb shocks and provide shareholder
6
protection thus improving the liquidity of the market as a whole (Eden & Dobson, 2005;
7
8
Beck, et al., 2003). Harris & Ravenscraft (1991) claim that FDIUS increases when the dollar is
9 weaker compared to the investor’s home currency. Servaes & Zenner (1994) also affirmed
10 that tax regulations have an impact on FDIUS indicating tax benefits for the investors. During
11 the period 2000-2016 the United States occupied the biggest share of worldwide M&A
12 activities. The highest percentage (50 percent) was taken by technology firms in 2000 due to
13 the tech bubble with an average of 37 percent throughout the same 17-year period
14
(WilmerHale , 2017) as shown in figure 2 below; US firms represent on average 20% of
15
16 global M&A as acquirers and 23% as targets by value (Ernst & Young, 2015).
17
M
18 Figure 2: M&A Activity; Worldwide vs. US
19
20
an

21
22
23
ag

24
25
26
er

27
28
29
ial

30
31
32
Fin

33
34
35
36
an

37 The technology sector experienced the largest M&A activity in the U.S, holding the highest
38
number of transactions in the period between 2000 and 2016 which represented 19.9% of
39
all U.S M&A transactions (Institute for Mergers, Acquisitions and Alliances, 2016). The high
ce

40
41 volume of M&A transactions in this sector is attributed to three main causes: (i) Financial
42 Strength: technology firms enjoying large amounts of cash with high stock prices enabled
43 them to make large acquisitions using cash or stocks. In 2016 technology firms held over
44 $773 billion in cash, accounting for 46% of total cash held by U.S non-financial firms of $1.68
45 trillion (Moody’s, 2016); (ii) Industry Trends: i.e. location-based services, digital
46
entertainment, robotics and artificial intelligence, virtual reality, 3-D printing and blockchain.
47
48 These areas have various applications, they are used by millions of users, and at the same
49 time they are rapidly evolving. Acquisitions are favoured by 41% of technology firms as the
50 path to growth and market share capture on one or more tech areas (Ernst & Young, 2016)1.
51 Hagedoorn and Duysters (2002) posit that technology firms in their quest for growth through
52 innovation prefer acquisitions instead of other alternatives such as strategic alliances;
53
54
55 1
Based on a survey conducted by Ernst&Young in October 2016, including 255 respondents from technology
56
firms, of which 51 percent were CEOs, CFOs and other C-level executives.
57
58
59
60
 Managerial Finance Page 6 of 31

1
2
3 (iii) Talent Acquisition: since 2011 the Big Five2 technology firms have added more than
4 418,000 jobs to the market with 76% of technology firms scouting and acquiring other firms
5 in order to secure talent amidst other industries that have lost jobs (Acker, et. al, 2017; Ernst
6
& Young, 2016). The sector accounts for more than 8% of the total US economy. It represents
7
8
about $1.3 trillion of value (CompTIA, 2017) and it employs more than 4% of all US
9 workforce (U.S. Bureau of Labor Statistics, 2016). The dot com bubble (1999 – 2000) was the
10 peak with 371 and 261 tech IPOs respectively, and in 1996 with 274 IPOs (Ritter, 2017). The
11 growth of IPOs during the 1990s was fuelled by venture capitalists excessively funding start-
12 ups as funding rose from $3 billion in 1990 to $60 billion in 1999 (Lowenstein, 2004);
13 furthermore, 57% of these tech firms going public were less than five years old and in some
14
cases even less than two years old (Westenberg, 2009); institutional investors bought stocks
15
16 with thin fundamentals as they purchased more than 63.6% of technology stocks between
17 1997 and 2000 (Griffin, et al., 2011). This was coupled with media coverage and narratives
M
18 from investment bankers, analysts and journalists encouraging individual investors to further
19 invest in the technology sector (Teeter and Sandberg, 2016) making them hold the
20
an
remaining 36.4% of technology stocks and continue to buy them while institutional investors
21 were rapidly selling them (Griffin, et al., 2011). As a result, the market was extremely
22
overvalued when NASDAQ reached its highest level3 on March 2000. It lost more than 50%
23
ag

24 by value in October 2000 (Westenberg, 2000). Its growth has been substantial since the
25 1990s measured by the number of technology IPOs as indicated by figure 3 below.
26
er

27 Figure 3: US Tech IPOs 1980 – 2016 (Ritter, 2017)

28
29
ial

30
31
32
Fin

33
34
35
36
an

37
38
39
ce

40
41
42
43
44
45
46
47 2.1 Valuation of Technology Firms
48
49 In an M&A transaction it is as crucial for the acquirer to determine a fair value of synergies
50 for the target, as it is for the target to come to a value for itself. It is also important for the
51 shareholders of both firms to justify the acquisition price (Petitt & Ferris, 2013). The
52
dynamics of the technology sector are characterized by rapidly evolving firms which operate
53
54 2
55 Big Five: Microsoft, Alphabet, Amazon, Apple and Facebook.
3
NASDAQ Composite Index (IXIC) level of 5,132 was the highest at the time of the tech bubble, it crossed the
56
5,132 level in 2015 (NASDAQ, 2017).
57
58
59
60
Page 7 of 31 Managerial Finance

1
2
3 under high levels of uncertainty and risk (Lev & Zarowin, 1999). This, combined with the lack
4 of positive cash flows (Aydin , 2015) makes their valuation very challenging as also
5 demonstrated by Bakshi & Chenb (2005), where they demonstrate the potential for
6
significant mispricing and departures from fair values. The complexity of valuing technology
7
8
firms can be attributed to reasons such as:
9
10 (i) Tech firms are often young ones, very dependent on innovation and require huge
11 amounts of upfront investments in intangible assets. Chandra et al. (2011) state:
12 “...this arises from the uncertain nature of long-run industry prospects as well as
13 competition among firms for market share through first-mover advantages,
14
creation of entry barriers and establishment of property rights in new
15
16 technology”(p.8) which leads to the second point;
17 (ii) The value of many firms in the technology sector usually comes from intangible
M
18 assets. These assets however don’t always appear on the firm’s financial
19 statements due to the lack of accounting standards to accommodate such
20
an
intangibles, such as innovation, customer satisfaction and human capital,
21 resulting in complexities when it comes to perform an equity valuation (Chan et.
22
al, 2001);
23
ag

24 (iii) Tech firm value is directly dependent on growth; consequently most of the value
25 will originate from future customers or products not from current operations.
26 That makes it challenging for investors to measure firm’s beta (risk);
er

27 (iv) The value of a technology is only known after it is commercialized to the market
28 (Park & Park, 2004).
29
ial

30
There are various methods to value firms; they are however categorized into three
31
32 mainstream methods (Hodges, 2007). Table 1 below shows the pros and cons as listed by
Fin

33 Anadol et al. (2014).

34
35 Table 1: Advantages and Disadvantages of Valuation Methods
36 Method Advantage Disadvantage
an

37
38 Discounted Estimating future cash flows is difficult at
39 Has firm theoretical basis; Easy to
Cash Flow best; Estimating interest rates in the future is
ce

40 compare competing opportunities

(DCF) uncertain
41
42
43 Best when a highly comparable group The whole sector may be over/under valued;
44 Comparable
is available; Units are close in both There are too few comparable examples;
45 Firms (CF)
size and business type Insufficient recent transactions
46
47
48 Looks at all the underlying values in More relevant if the assets can be liquidated
49 the firm's assets; Conservative, not readily; Does not work for initial IPOs; Small
50 Asset-
likely to be criticised; Traditional firms are disadvantaged; Service firms are
51 Based
52 method, people are comfortable with difficult to value this way; Growth rates in
53 it high-tech firms not included
54
55
56
57
58
59
60
 Managerial Finance Page 8 of 31

1
2
3 DCFs as a method, boasts a huge limitation in that firms in the technology sector more often
4 than not either do not pay dividends (even in cases they pay dividends these are often very
5 volatile) or instead choose stock buybacks therefore using this method can undervalue the
6
firm (Palepu, 2003). Any valuation method can be misleading as it does not for example
7
8
incorporate intangibles, yet typically, acquisition premiums achieve more than 50 percent
9 above market value (DeAngelo, 1990). Also, multiple bidding offers can be significantly
10 different in terms of prices (Bradley, 1980). Even hedge funds in the U.S hire technology
11 consultants to provide expert insights about tech firms as they are hard to value from a
12 financial standalone perspective (Benou and Madura, 2005).
13
14
3. Takeover Prediction Techniques
15
16
17 Various researchers have studied the possibility of predicting acquisition targets through
M
18 statistical aggregation and the associated distress signals (i.e. bankruptcy) using publicly
19 available information of firms and then applying different statistical models on them. It is
20
an
important to mention that the methodologies used to predict bankruptcy and predict
21 takeover targets are very similar, (discriminant analysis (DA) and logistic regression)
22 therefore we shall consider both broad approaches below.
23
ag

24
25 3.1 Traditional Analytical Techniques
26
er

27 Regression models: Ohlson (1980) utilised logistic regression analysis in order to examine
28 the relationship between binary or ordinal response probability and explanatory variables.
29
ial

He was the first to point out weaknesses in Altman’s (1968) model and highlighted the
30
importance of using data from firms’ financial statements directly as they will indicate
31
32
whether the firm filed for bankruptcy before or after releasing them which will help the
Fin

33 researcher avoid the “back-casting” issue (i.e. applying the model to firm’s data after being
34 bankrupt). This model produced an accuracy prediction rate of 96 percent with a cut-off
35 point of 0.5.The binary logistic regression, a nonlinear model, is one of the predictions’
36 techniques where the dependent variable is a binary or dummy variable. Very few
an

37 assumptions are required in such model in comparison to other similar dependence

38
techniques such as discriminant analysis. Harris et. al (1982) used a probit model where for
39
example, the dependent variable can take only two values (acquired or not-acquired), in
ce

40
41 order to produce a probability of a firm to be acquired or not as well as what are the
42 characteristics that affected this probability. Dietrich & Sorensen (1984) used logistic
43 regression model to predict acquisition likelihood. Palepu (1986) used a binomial logit
44 probability model with 9 independent variables; his model suggested a good fit of success in
45 predicting a high number of targets. It however, predicted a high number of non-targets as
46
targets, therefore, it was not sufficient to use this model to gain abnormal returns. Barnes
47
48 (1990) used multiple discriminant models with 5 chosen industry-related ratios to increase
49 the predictability of his model. While the previous studies, as above, have shown prediction
50 power between 60 to 90 percent Palepu (1986) argued however that these findings are
51 overstated and suffer a biased estimate due to two main flaws in such methodologies: (i)
52 state-based sampling for model estimation and prediction testing; (ii) using predetermined,
53 arbitrary, optimal cut-off probability. Furthermore, Powell (1997) argued that the
54
characteristics of hostile and friendly takeovers differ therefore using binomial models
55
56 (treating hostile and friendly takeovers in the same group) will cause misleading results.
57
58
59
60
Page 9 of 31 Managerial Finance

1
2
3 Cudd & Duggal (2000) in their study used Palepu’s factors (1986) but they added an industry
4 dispersion factor to account for different industries which improved the accuracy of the said
5 model. In addition, they also found that the dummy variable “industry disturbance” to be
6
significant therefore indicating that a takeover in the same industry in the past 12 months
7
8
will increase the probability of takeover for the remaining firms in that industry.
9
10 Discriminant analysis (DA): allows the researcher to pair two or more firms (or groups of
11 firms) and compare their differences with respect to several variables simultaneously.
12 Depending on how variables behave (i.e. jointly or independently of one another) DA
13 models can be further applied into two sub-categories namely univariate or multivariate
14
models; multivariate models (MDA) consider simultaneously an entire portfolio of
15
16 characteristics common to the firms and their interaction; univariate models are limited to
17 only one characteristic at a time. As a technique, DA does very well provided that the
M
18 variables in every group follow a multivariate normal distribution and the covariance
19 matrices for every group are equal. As early as 1971, Simkowitz and Monroe suggested that
20
an
target firms tend to be usually smaller, with lower P/Es and dividend payout ratio and lower
21 equity growth. Most importantly, they further observed that non-financial characteristics
22
appeared to be as important as financial. Their multivariate discriminant analysis (MDA) in-
23
ag

24 sample results predict 83% of the targets and 72% of the non-targets, while the holdout
25 results are slightly worse predicting 64% of the targets and 61% of the non-target.
26
er

27 3.2 Machine Learning Techniques

28
29
ial

A differentiated methodological approach used by researchers is the use of Neural

30
Networks (NN), Machine Learning (ML) and Data Mining to predict bankruptcy or takeover
31
32 targets. Sharda & Odom (1990) compared the use of both neural networks and multivariate
Fin

33 discriminant models (MDA) in bankruptcy predictions. In their study, the researchers utilised
34 the same ratios used by Altman (1968) and after executing both models their findings
35 suggest that neural networks seem to outperform MDA based on different holdout samples
36 with an accuracy level ranging from 77.78 to 81.48 percent. In another study, Tsai and Wu
an

37 (2008) studied the effect of including multiple neural network classifiers in bankruptcy
38
39
prediction and credit scoring where it was found that single neural network classifiers
outperformed multiple neural network classifiers in both credit scoring and bankruptcy
ce

40
41 prediction. Hongjiu et. al, (2007) used self-organized mapping with Hopfield neural network
42 to cluster data and their model showed accuracy predictions of 80.69 percent for targets
43 and 63.11 percent for non-targets. Their paper suggests also the importance of including
44 non-financial factors to improve the predictability power. Iturriaga and Sanz (2015) used
45
multilayer perceptron (MLP) to predict bankruptcy of U.S banks with a 96 percent success
46
47 rate.
48
49 The evidence regarding method and model fit is far from conclusive though. Coats and Fant
50 (1993) for example, confirmed that NN outperformed Multiple Discriminant Analyses (MDA)
51 in their sample 80 percent of the time. Numerous other studies have supported the use of
52 Neural networks (NN) in outperforming logistic regression (LR) in predicting bankruptcy (see
53
for example, Tam and Kiang, 1992; Jo and Han, 1996; Maher and Sen, 1997; Fan &
54
55 Palaniswami, 2000; Tseng & Hu, 2010). Branch et. al, (2008) utilised both NN and LR to
56 predict whether a takeover attempt will succeed or not with the authors concluding that
57
58
59
60
 Managerial Finance Page 10 of 31

1
2
3 ‘...neural network model outperforms logistic regression in predicting failed takeover
4 attempts and performs as well as logistic regression in predicting successful takeover
5 attempts’ (p. 1186). Salchenberger et. al, (1992) compared NN with LR to test healthy and
6
failed thrift institutions and concluded that NN achieved higher accuracy. In all of the above
7
8
studies a common results’ attribution emerges: NNs seem to possess a higher flexibility and
9 ability to address non-linearities. This echoes Zhang’s et. al (1999) statement that neural
10 networks can potentially be robust and can provide more reliable estimations when applied
11 on different samples only once the optimal architecture is found.
12
13 On the other hand, Altman, et al. (1994) reported that both MDA and NN performed almost
14
the same when trying to predict Italian firms suggesting that contextual and structural
15
16 considerations as well as firm-characteristics’ variables are also important. Equally, Olson et
17 al. (2012) used Logistic Regression, Neural Networks, Support Vector Machines and Decision
M
18 Trees to predict bankruptcy. They demonstrated that different data with different models
19 present different results. There is trade-off between model accuracy and transparency and
20
an
transportability. In a sense, in order to increase model transportability (i.e. applying it to
21 new datasets and observations) the accuracy level will decrease. Table 2 below, by Barnes
22
(1998), summarizes other research showing the prediction rates of various methods for
23
ag

24 North America and the UK.

25
26 Table 2. Previous stuies on the characterisitcs of Target firms.
er

27
28
29
ial

30
31
32
Fin

33
34
35
36
an

37
38
39
ce

40
41
42
43
44
45
46
47
48
3.3 Takeover Determinant Variables
49
50
51 The main takeover relevant metrics/ratios that have been introduced by the financial
52 literature to identify a takeover target are discussed below.
53
54
55
56
57
58
59
60
Page 11 of 31 Managerial Finance

1
2
3 - Inefficient Management
4
5 This hypothesis states that managers who fail to maximize their shareholders’ wealth and
6
firm’s value shall be replaced in accordance with the market for corporate control theory.
7
8
Therefore, incompetent management increases the probability of their firms to be taken
9 over (Jensen, 1986). Investors will seek to replace the management by purchasing a
10 controlling stake in the firm due to the share prices being below their true value, and target
11 managers will typically get replaced if the bid succeeds (Agrawal and Walkling, 1994). This
12 hypothesis can be measured by EBITDA margin ROE, ROCE, ROA and/or asset turnover.
13
14
- Undervalued Firms
15
16
17 This hypothesis suggests that firms with low market value compared to book value are
M
18 targets since they represent a ‘cheap buy’ (Powell, 1997; Palepu, 1986). It utilises market to
19 book and price to earnings ratios where a bidder will bid for an overvalued firm if it was still
20
an
less overvalued than the bidder (Dong, et. al, 2006).
21
22
- Firm Size
23
ag

24 Firm size plays a significant role in takeover probability, the bigger the size the lower the
25 probability of it being taken over, (Palepu, 1986), which explains why usually bigger firms
26 acquire smaller ones (Levine and Aaronovitch, 1981). It has been shown that size is a
er

27 significant factor (Powell, 1997) as measured by market capitalization and total assets.
28
29
ial

- Leverage, Liquidity and Growth

30
Powell and Yawson (2007) debated that many takeovers occur as a way to rescue the target
31
32 firm from a certain bankruptcy due to high debt and poor performance. Therefore firms
Fin

33 with low growth and high leverage are more likely to be classified as targets and measured
34 by debt to equity, current ratio and growth in revenues. While low liquidity does not single-
35 handedly affect the takeover likelihood, when coupled with growth and leverage it can have
36 a significant effect. (Palepu, 1986; Cremers et. al, 2008b)
an

37
38
Table 3: Takeover Determinant Variables and Their Ratios
39
Takeover Hypothesis Ratios
ce

40
41 - EBITDA Margin
42 - ROE
Inefficient Management
43 - ROCE
44 - ROA
45 - Asset Turnover
46 Undervalued Firms - Market to Book
47 - Price to Earnings
48 Firm Size - Market Capitalization
49 - Total Assets
50 Leverage - Debt to Equity
51 - Equity Multiplier
52 Current Ratio
Liquidity -
53
- Net Working Capital
54
Growth
55 - Growth in Annual Sales
56
57
58
59
60
 Managerial Finance Page 12 of 31

1
2
3 4. Sample, Methodology and Data
4
5 Our study required three generic groups of data. M&A transactions records, number of
6 public firms in the technology sector from the year 2000 to 2016, and the relevant financial
7
ratios for the same period. All sample data were gathered from Bloomberg. We define a
8
9 technology firm as a type of business entity that focuses primarily on the manufacturing and
10 development of technology. This also includes the dissemination of information via high
11 tech companies. It also includes information technology (IT) companies as subsets of
12 technology companies as provided by the NAICS coding system where we placed several
13 restrictions and criteria for selecting our sample. We observe that the number of public
14 technology firms in the United States has been declining over the last 17 years as shown in
15
figure 4. It is important also to mention that the decrease in public firms is not only affecting
16
17 the technology sector, as it is affecting the whole U.S stock market. Since 1996 the number
M
18 of public firms in the U.S has decreased by 50 percent, as a result of: (i) firms being delisted,
19 acquired or bankrupt; (ii) less Initial Public Offering (IPO) activities, where firms remain
20
an
private due to available capital provided by Venture Capital and Private Equity firms
21 (Mauboussin, et al., 2017).
22
23
ag

Figure 4: The Decrease in Number of Public Technology Firms in the U.S from 2000 to 2016
24
25
26
er

27
28
29
ial

30
31
32
Fin

33
34
35
36
an

37
38
39
ce

40
41 We pose certain sample restriction criteria for the purposes of our study. First, our study
42 period covers the last 17 years where M&A transactions announced between the year 2000
43 and 2016 are included; second, we eliminate private firms where the target is a publicly
44 traded company and having its domicile in the United States; third, we screen only
45 technology firm targets and exclude firms operating in irrelevant sectors where the target is
46
classified as a technology company by their NAIC code; fourth, we exclude investments,
47
48 joint ventures, spinoffs and buybacks; fifth, we include only M&As transactions that are pure
49 mergers or acquisitions where the acquirer owns more than 50% of the targets’ shares. The
50 total number of M&A transactions based on such criteria reached 966 transactions. Figure 5
51 illustrates the acquirers’ industries by number of deals. More than 80% of the M&As were
52 completed. Ninety-three percent were classified as friendly takeovers with 3% representing
53 hostile takeovers. The rest are classified as unsolicited/unsolicited-to-friendly. Technology
54
firms were 53% of the acquirers’ transactions. Financial firms came in second at 19%.
55
56
57
58
59
60
Page 13 of 31 Managerial Finance

1
2
3 Figure 5: Acquirer Industry by Number of Deals
4
5
6
7
8
9
10
11
12
13
14
15
16
17
M
18
19
20
an

21 Figure 6 below, shows the number and value of deals in the technology sector in the U.S.
22 The total dollar value of these transactions for our period of study reached $1.025 trillion.
23
ag
Most acquirers in our sample came from the United States with 88% and the remaining
24 came from Europe with 10%.
25
26
er

Figure 6: M&A Activity in the Technology Sector in United States (2000 – 2016)
27
28
29
ial

30
31
32
Fin

33
34
35
36
an

37
38
39
ce

40
41
42
43
44
45
4.1 Datasets
46
47
48 Our study sample consists of two datasets, targets and non-targets. The target-group
49 dataset includes firms which got acquired or received a bid to be acquired within our study
50 period. The non-target group dataset includes firms which did not get acquired or received a
51 bid to be acquired during the same period. The number of firms in our target dataset
52 reached 846. Due to data pre-processing and omitted values this number was brought down
53
to 415. We followed Palepu (1986) in choosing pre-determined ratios for the purposes of
54
55 consistency and comparability but also in order to avoid the statistical overfitting issue (see
56 also further support in section 4.3.1 below).
57
58
59
60
 Managerial Finance Page 14 of 31

1
2
3 From this sample, 102 firms (24.5%) did not provide for a meaningful P/E ratio, and a further
4 87 firms (21%) did not have information on liquidity ratios. This further resulted in 189 firms
5 been dropped from the sample producing a final 226 usable observations. The non-target
6
dataset reached 2,340 firms.
7
8
9 4.2 Modelling
10
11 We apply two distinct methods in order to account for the different predictive accuracy of
12 the two categories (target and non-target). A traditional statistical technique as well as a
13 machine learning, predictive analytics technique, the MLP, has been used to model M&A
14
activity at the developed capital markets and to predict potential targets.
15
16
17 4.2.1 Model 1: Multilayer Perceptron Model (MLP) Analysis Method
M
18
19 Over the last decade, a renewed growing interest in neural networks as a tool for data
20
an
analysis has been observed. To a certain extent, the attractiveness of artificial neural
21 networks vis-a-vis other statistical methods may have also been partially caused by human
22 issues that merit some mention: often there is a shortcoming of statisticians to clearly
23
ag

24
communicate their methodologies and algorithms to non-statisticians. A large amount of
25 the extant statistical knowledge raises a hurdle for potential investors of their methods.
26 Neural networks on the other hand, are in a mid-embryonic phase, meaning that the
er

27 current knowledge is thinner compared to statistical techniques. Artificial Neural Networks

28 (ANN) originate from the biological human brain neurons. It is a network of nodes
29
ial

connected with each other through a weighted connection (Roiger, 2016), which can be
30
greatly beneficial for complex non-linear relationships between variables (Hyndman &
31
32
Athanasopoulos, 2013). ANN has been used in many industries such as telecommunications,
Fin

33 industrials, banking, airlines and healthcare, and has been successfully showcased by
34 (Widrow et. al, 1994). An example/representation of this model is shown below in figure 7.
35
36 Figure 7: Multilayer Perceptron Model
an

37
38
39
ce

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54 The nodes in the input layers are passive nodes as they only pass the data from the input
55 layer to the hidden layer. In the hidden layer, a weight (Wn) will be generated for each input
56 node.
57
58
59
60
Page 15 of 31 Managerial Finance

1
2
3 For the first iteration it is a randomly generated number based on Gaussian distribution.
4 Then each input (Xn) will be multiplied by its weight (Wn) to produce a weighted input (XWn).
5 The summation of these weighted inputs goes into the activation function to produce an
6
output between (0, 1). The output number gets transferred to the output layer, there they
7
8
get multiplied again with another set of randomly generated weights to produce the final
9 output number between (0, 1). The model then compares the output number with the
10 target number and calculates the difference. It then adjusts the weights in order to decrease
11 the sum of margin errors (i.e. the cost function).
12
13 Input Layer Variables
14
15
16 This layer consists of pre-determined, industry related financial ratios derived and in line
17 with established financial literature. Table 4 shows the financial ratios used in this study
M
18 which have been used by a number of influential research papers (Ohlson, 1980; Palepu,
19 1986; Powell & Yawson, 2007). EBITDA and ROA for most technology firms in our sample
20
an
were non-existent hence had to be dropped as candidate variables as they would limit the
21 sample to less than 100 firms. The number of nodes in this layer simply equals the number
22
of independent variables, in our case 6 for each instance.
23
ag

24
25 Table 4: Financial Ratios Used in our Study
26 Takeover Hypothesis Ratio
er

27 Inefficient Management Return on Equity

28 Undervalued Firms Price to Earnings Ratio
29
ial

Firm Size Market Capitalization

30 Leverage Debt to Equity Ratio
31 Liquidity Current Ratio
32
Growth Rate of change of Annual Revenues
Fin

33
34
35 Hidden layer
36
an

37 This layer will receive the nodes sent from the input layer. It will generate a weight for each
38 connection between any node in input layer and any node in the hidden layer. Then it will
39 multiply each node with its weight as shown in equation 1 below:
ce

40
41
42 Equation 1: Summation of Weighted Inputs
43
44
45
46 b : Bias node
47
48
X1 : Financial ratio (Ex: ROE for the first instance)
49 W1 : Weight associated with X1 (Randomly generated number between 0 and 1)
50
51 The net input function z, will go into a non-linear activation function (sigmoid function). It
52
will act as a smooth thresholding function to determine the relationship between inputs and
53
54 outputs. Our sigmoid function performs better for negative variables and classifiers (Zhang,
55 et al., 1998) based on equation 2 below.
56
57
58
59
60
 Managerial Finance Page 16 of 31

1
2
3 Equation 2: Sigmoid Activation Function
4
5
6
7
8
9
10
With differentiation φ’ (z) = φ (z) (1- φ (z)), in updating the curve. The cost function used in
11 the study was sum of squared errors using an optimisation Gradient Decent method with
12 the following parameters: Initial Learning Rate = 0.4 and Momentum = 0.9. The Cost Function
13 C, given as:
14
15
16
C = 1/n ∑ni=1(zi – φ (xi))2, with φ(xi) the output for xi.
17
M
18 Output layer
19
20
an
The output value will then be multiplied with its connection weights again and the final
21 value will go into the Output layer. Hidden layers adjust the weightings on those inputs until
22
they reach the optimization stage that is, the error of the neural network is minimized. An
23
ag

24 interpretation of this is that the hidden layers extract salient features in the input data
25 which have predictive power with respect to the outputs. This is the discussed feature
26 extraction function and it is parallel to the function of statistical techniques such as principal
er

27 component analysis. This layer consists of a binary node4; it will receive the value from the
28 hidden layer, indicating the dataset which the firm is predicted to be in. One indicates a
29
ial

target, zero indicates a non-target. The final output value will be compared with the
30
desirable target value. This whole process is called Standard Forward Propagation. The final
31
32 model architecture is shown in table 8 below:
Fin

33
34 Figure 8: Model I Final Architecture
35
36
an

37
38
39
ce

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56 4
The end-outcome will be either output ‘target’ or ‘non-target’; i.e. 1 or 0
57
58
59
60
Page 17 of 31 Managerial Finance

1
2
3
4
5 Validation
6
7
8
We used the cross-validation method that involves dividing the data records into three sets:
9 (i) Training data set: data records that are used to train the model; (ii) Testing data set:
10 records that are used to observe the error rate while training in order to further tweak the
11 model; (iii) Holdout data set: this set of records is used to assess the model’s final error rate
12 and performance. Validation is used to measure the performance and the generalization
13 ability of this model (Kaastra & Boyd, 1996). While there is no standardized rate of division
14
in the literature some researchers (Hammerstrom, 1993) recommend using the 70/30 ratio.
15
16 In our study, our data is randomly divided into three groups as follows: 70% for training;
17 20% for testing; and 10% as a holdout. We first clustered the data into years (Sample 1).
M
18 Then it was clustered into target and non-target firms. Next, the records were randomly
19 sorted, and the analysis was performed on 3 sets:
20
an

21 • Sample 1: All records of all years, randomly sorted.

22
23
ag

24 However, once this approach was finalised we discovered that this would potentially
25 create a considerable over-training issue because then the requirement would be to
26 repeat the steps above 17 times (17 years) with the same companies appearing on all
er

27 data sets. We then took a 2nd sample approach.

28
29
ial

• Sample 2:
30
o 50/50: 50% of the training data consisted of non-target firms, and 50% target.
31
32
Fin

33 Data itemization in our study could potentially suffer from unreliability owed to sample
34 limitations where the data available were not enough to train different networks on
35 different subsets of the data. Consistent with Srivastava et al. (2014), Dekel et al. (2010) and
36 Hinton & Salakhutdinov (2006), at this stage we only performed the training once in order
an

37 not to fall into the over fitting and overtraining where the network would just memorise the
38
outcome and not learn thus making it only usable in our specific data set. The data was fed
39
to the network in at once but it used the data 10 times (learning epochs = 10) to update the
ce

40
41 weights. Following the above authors’ prior work on data size and data diversity
42 considerations we performed the experiment based on 3 trials and then took the average of
43 these trials as shown in the analysis.
44
45 4.2.2 Model 2: Logistic Regression Model (LR)
46
47
48 Logistic regression method was used as the nature of our study is to forecast takeover
49 targets. Therefore, the output is always binary (i.e. target, non-target) so it is important to
50 use a technique that can classify a data instance into two classes by predicting the
51 probability of an input being in a certain class. We convert the values of our independent
52 variable from a string format to a numerical format assignining the following codes: 0 = Non-
53 Target, 1 = Target. The logistic regression model starts with no predictive variables and only
54
includes the intercept (constant) and measures the prediction power of this model using -2
55
56 Log Likelihood.
57
58
59
60
 Managerial Finance Page 18 of 31

1
2
3 It then adds one predictive variable per step and calculates -2 log likelihoods again to
4 measure if the new variable improved the prediction accuracy for all predictive variables.
5 The model allows us to calculate the odds of an input (firm) to be acquired or not using
6 ( ) where, a is the intercept (constant), b is the predictive variable added in
7
8 step i, and X is the independent variable (Acquisition Status, 0 or 1).
9
10 Next, we convert the odds to probabilities using ( ). Based on the
11
12 probabilities result for each input the model classifies them into target or non-target based
13 on a threshold (0.5), any input with a probability equals 0.5 or more will be classified as
14 target, anything less than that will be classified as non-target. Based on this classification,
15 the model produces a classification showing the number of cases correctly classified versus
16 incorrect classifications in order to produce an overall prediction accuracy.
17
M
18
19 Predictive Variables
20
an

21 The same inputs from Model 1 are utilised, therefore maintaining consistency in our
22 predictive variables (independent variables) namely, Return on Equity, Price to Earnings,
23
ag
Market Capitalization, Debt to Equity, Current Ratio, Rate of change of Annual Revenues.
24
Our model is based on 3 traditional empirical formulae as proposed by Swaminathan and
25
26
Rogers (1990) formulation of the logistic regression procedures where:
er

27
28
29 Equation 3: Odds Function
ial

30
31
32
Fin

33
34 Equation 4: Probability Using Odds
35
36
an

37
38
39 Equation 5: Logistic Regression Equation
ce

40
41
42
43
44
45 Odds : Ratio of probability occurring divided by the probability of it not occurring
46 Pi : The probability of firm i being taken over
47
β0 : The intercept
48
49 Zi : The weighted sum of the predictive variables
50 βn : The coefficients for the financial ratio Xn
51
52
53
54
55
56
57
58
59
60
Page 19 of 31 Managerial Finance

1
2
3 5. Results and Findings
4
5 Model I: MLP
6
7
8
The results of our sample include a total number 226 technology firms, 50 percent target
9 firms and 50 percent non-target firms. Table 5 shows the prediction percentages for
10 training, testing and holdout datasets based on three trials. As the table shows 70% of the
11 data is reserved for training, 20% for testing and the final 10% for our final holdout sample.
12
13 Table 5: 50/50 Sample Cases Summary – Model I
14
Case Processing Summary
15
16
17 N Percent
M
18
19 Sample Training 157 69.45%
20
an

21 Testing 46 20.35%
22
23
ag
Holdout 23 10.20%
24
25 Valid 226 100.0%
26
er

27 Excluded 0
28
29
ial

Total 226
30
31
32
Fin

33 We apply a standard feedforward propagation neural network with a single hidden layer in
34 our sample in order to identify potential takeover targets (and hence for example, the
35 possibility to yield positive abnormal returns from investing in these targets stocks).
36
an

37 Below, table 6 summarizes the results of our analysis where the predictive ability of the
38
model is tested. The variables applied to the neural network models are the return on
39
equity, price to earnings ratio, market capitalization, debt-to-equity, current ratio, and rate
ce

40
41 of change of annual revenues and by default, industry. Overall the results are promising
42 compared to the standard binary regression technique. Our sample had an out-of-sample
43 overall average prediction accuracy of 71.4 percent, with an average of 28.6 percent of
44 incorrect predictions.
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
 Managerial Finance Page 20 of 31

1
2
3 Table 6: Model 1, 50/50 Sample Cases Results
4
5 Trial 1 Trial 2 Trial 3
6
7 Training
8
9 Non-Target 86.1 71.6 58.3
10
11 Target 40.5 49.4 61.9
12
13 Overall 61.5 61.1 60.1
14
15
Testing
16
17
M
18
Non-Target 72.0 52.9 60.9
19
20 Target 28.6 60.9 52.4
an

21
22 Overall 52.2 57.5 56.8
23
ag

24 Holdout
25
26 Non-Target 87.5 87.5 50.0
er

27
28 Target 25.0 54.5 87.5
29
ial

30 Overall 66.7 68.4 71.4

31
32
Fin

33
The attempt rate (i.e. trials) at 3 trials showed improvement in our holdout sample for
34
35 correctly identifying the target companies with a 87.5% accuracy prediction rate. Yet it has
36 to be recognised that it also correctly identified non-targets only 50% of the time giving an
an

37 overall prediction accuracy rate of 71.4% at trial three. Our neural network model attempts
38 to provide a tool that can adaptively sift through noise and identify patterns in complicated
39 financial relationships where non-linearity might pose problems. Using 6 inputs considered
ce

40 to be the most relevant, and having only 4 hidden nodes our sample gets around the issue
41
of having a relatively small dataset. Adaptability also lies in the recognition of not adding too
42
43 many nodes which could lead to mode overfitting.
44
45 Our results support that such an approach can potentially provide meaningful explanation
46 regarding dependent and independent variables compared to a traditional regression
47 model. We turn to this below in tables 7 and 8.
48
49
50
51
52
53
54
55
56
57
58
59
60
Page 21 of 31 Managerial Finance

1
2
3 Table 7: Model 2: Regression: 50/50 Sample Cases Summary
4 Unweighted Cases N Percent
5
6 Selected Cases: Included in Analysis 226 100
7
8 Missing Cases 0 .0
9
10 Total 226 100
11
12 Unselected Cases 0 .0
13
14 Total 226 100
15
16
17 Table 8: 3-Step Classification Tablea - Model II
M
18
19 Predicted
20
an

21 Observed Acquisition Status

22 Percentage
23 Non-Target Target Correct
ag

24
25 Step 1 Acquisition Status Non-Target 59 54 52.2
26
er

27 Target 38 75 66.4
28
29
ial

Overall Percentage 59.3

30
31 Step 2 Acquisition Status Non-Target 57 56 50.4
32
Fin

33 Target 35 78 69.0
34
35
Overall Percentage 59.7
36
an

37
38 Step 3 Acquisition Status Non-Target 65 48 57.5
39
Target 38 75 66.4
ce

40
41
42 Overall Percentage 61.9
43
44 a. The cut value is .500
45
46 Table 8 above, illustrates the three steps taken by our regression model when adding new
47 predictive variables to the model and the accuracy of correct predictions on each step. The
48 model was able to increase the accuracy with each step, albeit marginally, reaching an
49 overall accuracy of 61.9%. This model correctly identified the target companies with a 66.4%
50 accuracy prediction rate and it also correctly identified non-targets only 57.5% of the time.
51
Comparatively, the first model achieves a higher accuracy overall over model 2 providing
52
53 some support for the utilisation of NNs.
54
55
56
57
58
59
60
 Managerial Finance Page 22 of 31

1
2
3 It has to be said though that, the Training, Testing, and Holdout results differ from each
4 other in each of the 3 trials for the 50/50 samples. We suggest that it is due to the random
5 number generator where the network starts with a random initial numbers to start with and
6
then keeps updating the weights accordingly; this is important in order to create a global
7
8
optimum solution. In the first instance we actually had an average 6% change from one step
9 to the next but a variability of 16.5% in-between the steps. Compared to the second model
10 the step difference is 1% with a variability of 14.5% in-between the steps. The observations
11 drawn are: (i) the regression model is static throughout the sample and trials whereas the
12 NN model shows evolution and adaptability, (ii) there are large swings in variable values
13 where for example, the RoE, D/E and liquidity swing wildly – deep in negative and high up in
14
positive territory - from year to year, and (iii) the number of observations is relatively
15
16 limited where the holdout sample is strictly anecdotal data since it covers only a limited
17 number of observations, hence the expressive power of the network is potentially not
M
18 enough to capture the target function. One alternative would be to add more layers or more
19 hidden units in fully connected layers. So while it’s helpful to test different methods, and
20
an
provide for greater accuracy, it does not by itself, conclusively determine which method is
21 best owed to data limitations. In addition, an examination of the variables also provides
22
some interesting insights.
23
ag

24
25 Table 9 below shows the importance of each variable fed into the model in terms of
26 characterising its output. While the variable importance analysis below shows the input
er

27 effects on the output it can be also clearly seen that the 3 variables mentioned above
28 account for over 80% of the effects on output.
29
ial

30
Table 9. Variable Importance Analysis
31
32
Fin

33
34
35
36
an

37
38
39
ce

40
41
42
43
44
45
46
47
48
49
50
51 The variable importance analysis showed a great importance for ROE, D/E ratio and
52 liquidity. These are consistent with the inefficient management, leverage and liquidity
53 takeover hypotheses but the direction of the relationship between the independent and
54
55
dependent variables is not clear. It is also these 3 variables that showed the greatest
56 volatility throughout our sample period.
57
58
59
60
Page 23 of 31 Managerial Finance

1
2
3 6. Conclusions and limitations
4
5 This paper examines the use of a neural network method for pricing mergers. With over
6
50% of mergers failing, it is critical for acquiring firms to identify the characteristics of a
7
8
target prior to a merger that will provide synergies once the merger is complete. The neural
9 network model presented in this paper simplistic as it may be at this stage, shows overall
10 improvements on the accuracy of predicting merger targets over linear regression results.
11 ANN has outperformed logistic models in both senses of discrimination and calibration,
12 although from the arbitrary standpoint of accuracy (cutoff point 0.5), logistic models can be
13 superior to ANN models. The fact is that in some applications neural networks fit better
14
than other models such as linear regression and this usually occurs when there are
15
16 nonlinearities involved though it is important to evaluate other aspects. For example: a
17 linear regression model will have less parameters to estimate compared to a NN for the
M
18 same set of input variables. Hence, a NN will require a larger dataset for its calibration and
19 subsequent optimization in order to get the required benefit of generalization, applicability
20
an
and nonlinear mapping. In the absence of critically enough data, despite existing
21 nonlinearities involved, a linear regression model may indeed be better calibrated.
22
23
ag

24 Improvements as per the accuracy of target prediction can translate into significant savings
25 in offering prices for target companies. Reliable predictions can improve the quality of
26 decisions and business strategy in target determination and fair price decisions. Neural
er

27 network methods permit the use of an expanding number of prospective venture

28 opportunities with the added benefit that as market changes are introduced and more
29
ial

dynamic analysis is eventually involved new and more inputs can be loaded onto the model
30
with less resource devotion. Having said that, it is also important to identify that neural
31
32 network methods do not provide for a fuller analysis of significance for each of the
Fin

33 autonomous variables in the model as traditional regression methods do.

34
35 Using a different activation function and a ‘deeper’ network with more hidden layers could
36 potentially account for how each successive layer uses the output from the previous layer as
an

37 input. It could also further show how the algorithm self learns from multiple levels of
38
39
representations that correspond to different levels of abstraction (i.e. the levels form the
hierarchy of concepts above). The quantity of data at our disposal though is relatively limited
ce

40
41 for more hidden layers to be involved in terms of describing potential causal connections
42 between input and output. The transfer function is the calculated derivative sigmoid
43 function utilized; we see this as important when calculating the weight updates in the
44 network based on the amount of data and the computational load of our simulation. Lastly,
45
while the extra layers could potentially help in learning features indeed the authors felt that
46
47 with such a sample introducing LeRu we may also run the risk of naively training a ‘deeper’
48 neural network. As argued above, the possibility of added layers of abstraction could also
49 show rare dependencies modelling in the training data.
50
51 It could arguably have been also interesting to investigate how model performance is
52 influenced by using different activation functions (for example utilising the so called ReLU
53
method – the Rectified Linear Unit - instead of Sigmoid,) or also involve in the analysis a
54
55 higher number of hidden nodes. This is another area for research where traditionally,
56 machine learning evaluation works best in producing an extrapolative model. The trade-off
57
58
59
60
 Managerial Finance Page 24 of 31

1
2
3 though, of creating a flexible, nonparametric predictive model on the other hand, is that
4 causal interpretations can potentially be lost. Equally, linear regression is a relatively
5 inflexible approach yet it is less complicated in its interpretation. Flexible constructs avoid
6
the assumptions of a particular functional form for a model, but they also require a larger
7
8
number of observations and are more complicated and challenging to interpret. In addition,
9 it can also be supported that NNs with different initializations produce different signals for a
10 certain feature. As seen above, our NN with a certain initialization produced better signals in
11 some cases and incorrect signals in some other.
12
13 Our results are also consistent with 20 years of research and some seminal papers that date
14
as back as the 90s until today (see for example, Sen and Gibbs, 1994; Sinha and Richardson,
15
16 1998; Fescioglu-Unver and Tanyeri, 2013; Spangler et al., 2015; Tkáč, and Verner, 2016).
17 Such studies indicate that although neural networks map the data satisfactorily, it is still
M
18 questionable whether they predict merger targets significantly better than logistic
19 regression. This strongly suggests that the financial models used to predict mergers are
20
an
relatively inadequate. Firms should approach the development of merger prediction models
21 cautiously and identify other factors that are more likely to predict mergers. Neural
22
networks give the best overall results for the largest multiple classification cases. There is
23
ag

24 substantial room for improvement in overall performance for all techniques. The results
25 indicate that data mining methods and data proportions and characteristics have a
26 significant impact on classification accuracy. Zhu et al. (2001) for example, state that within
er

27 data mining methods, rough sets provide better accuracy, followed by neural networks and
28 inductive learning.
29
ial

30
The generalization breadth of this study is limited within a specific sector (technology) in a
31
32
specific country (United States) covering a specific period (2000–2016). One of the most
Fin

33 important limitations was data collection, as we had to omit approximately 50 percent of the
34 initial sample due to unavailable data on firms and their financial ratios. The takeover
35 determinants were chosen from previous studies done by other researchers that showed
36 statistical significance; this may affect the results of this analysis as the sample size, sector
an

37 and study period are different. Further research can be done to extend this model and
38
maybe improve the accuracy of it by including for example: (i) Technology Firms’ Specific
39
Ratios, this will allow the model to study technology firms not just from a financial but from
ce

40
41 an operational perspective too; (ii) Social Profiling, Social Media and softer Social variables
42 not captured or modelled by standardised techniques, where these can be leveraged in
43 order to discover opportunities or create maps for those interested audiences (Beese, 2015).
44 Monitoring social media impression of the firm or its management might give an indication
45 of its takeover probability.5
46
47
48 Old may be, but this echoes also Kuo and Reitsch’s (1995) early research in the managerial
49 forecasting problem; many managers value the ‘softer’ features of neural nets, particularly
50 when standard regression models tend to emphasize the causal interpretations (more the
51 why) of the problem and not the solution.
52
53
54
55 5
Similar research has been done in this field by Xiang et al. (2012). A Supervised Approach to Predict Company
56
Acquisition with Factual and Topic Features Using Profiles and News Articles on TechCrunch. in ICWSM.
57
58
59
60
Page 25 of 31 Managerial Finance

1
2
3 References
4
5 Acker, O., Hagen, H. and Hajj, J., (2017) 2017 Technology Trends. [Online] available at:
6
https://www.strategyand.pwc.com/trend/2017-technology-trends, [Accessed 16 May 2017].
7
8
Agrawal, A. and Walkling, R. A., (1994) Executive Careers and Compensation Surrounding Takeover
9
10 Bids, The Journal of Finance, 49(3): 985 - 1014.
11
12 Alexandridis, G., Mavrovitis, C. F. and Travlos, N. G., (2012) How Have M&As Changed? Evidence
13 from the Sixth Merger Wave, The European Journal of Finance, 18(8): 663-688
14
15 Altman, E. L., Marco, G. and Varetto, F., (1994) Corporate distress diagnosis: comparisons using
16 linear discriminant analysis and neural networks (the Italian experience), Journal of Banking and
17 Finance, 18(3): 505 - 529.
M
18
19 Altman, E. L., (1968) Financial Ratios, Discriminant Analysis And The Prediction Of Corporate
20
an
Bankruptcy, The Journal of Finance, 23(4): 589 – 609.
21
22 Anadol, B., Paradi, J. C., Simak, P. and Yang, X., (2014) Valuing Private Companies: A DEA Approach,
23
ag
International Journal of Business and Management, 9(2): 16 - 28.
24
25 Andrade, G., Mitchell, M. & Stafford, E., (2001) New Evidence and Perspectives on Mergers, Journal
26
er

of Economic Perspectives, 15(2): 103 – 120.

27
28
Aydin , N., (2015) A Review of Models for Valuing Young and Innovative Firms, International Journal
29
ial

30
of Liberal Arts and Social Science , 3(9), p. 1 - 8.
31
32 Bakshi, G. and Chenb, Z., (2005) Stock Valuation in Dynamic Economies, Journal of Financial
Fin

33 Markets, May, 8(2): 111 - 151.

34
35 Barnes, P., (1998) Can Takeover Targets be identified by statistical techniques? some UK evidence,
36 The Statistician, 47(4): 573-591
an

37
38 Barnes, P., (1990) The Prediction of Takeover Targets in the U.K. by means of Multiple Discriminant
39 Analysi, Journal of Business Finance & Accounting, 17(1): 73–84.
ce

40
41 Beck, T., Demirgüç-Kunt, A. and Levine, R., (2003) Law and finance: why does legal origin matter?,
42 Journal of Comparative Economics, 41(4): 653 - 675.
43
44 Beese, J., (2015) What Is Social Listening & Why Is It Important?. [Online]
45
Available at: https://sproutsocial.com/insights/social-listening/, [Accessed 21 July 2017].
46
47
Benou, G. and Madura, J. (2005) High-tech acquisitions, firm specific characteristics and the role of
48
49 investment bank advisors, The Journal of High Technology Management Research, 16(1):101-120
50
51 Betton, S. Eckbo, B. E. and Thorburn, K. S. (2008) Corporate takeovers, Handbook of Corporate
52 Finance: Empirical Corporate Finance, edited by B. E. Eckbo, Elsevier, Oxford, 2: 291-416
53
54 Bradley, M., (1980) Interfirm Tender Offers and the Market for Corporate Control, The Journal of
55 Business, 53(4): 345 - 376.
56
57
58
59
60
 Managerial Finance Page 26 of 31

1
2
3 Branch, B., Wang, J. and Yangc, T., (2008) A note on takeover success prediction, International
4 Review of Financial Analysis, 17(5): 186–1193.
5
6 Bureau of Economic Analysis (2017), Available: https://www.bea.gov, [3 September 2017].
7
8 Chandra, U., Wasley, C. E. and Waymire, G. B., (2011) Income Conservatism in the U.S. Technology
9 Sector, Accounting Horizons, 25(2): 285 - 314.
10
11 Chan, L. K. C., Lakonishok, J. and Sougiannis, T., (2001) The Stock Market Valuation of Research and
12 Development Expenditures, The Journal of Finance, 56(6): 2431 – 2456.
13
14
Coats, P. K. and Fant, L. F., (1993) Recognizing Financial Distress Patterns Using a Neural Network
15
Tool, Financial Management, 22(3): 142 - 155.
16
17
M
18 CompTIA (2017), Cyberstates 2017, Research Report, accessed September 2017:
19 http://www.cyberstates.org/pdf/CompTIA%20Cyberstates%202017.pdf
20
an

21 Cremers, M., Driessen, J., Maenhout, P., and Weinbaum, D. (2008b) Individual stock-option prices
22 and credit spreads, Journal of Banking and Finance, 32(12): 2706–2715.
23
ag

24 Cudd, M. and Duggal, R., (2000) Industry Distributional Characteristics of Financial Ratios: An
25 Acquisition Theory Application, The Financial Review, 35(1): 105 – 120.
26
er

27 Damodaran , A., (2001) The Dark Side of Valuation : Valuing Old Tech, New Tech and New Economy
28 Companies. s.l.:Financial Times.
29
ial

30 Daniel, K., Hirshleifer, D. and Subrahmanyam, A., (1998) Investor Psychology and Security Market
31 Under- and Overreactions, The Journal of Finance, 53(6): 1839 – 1885.
32
Fin

33 DeAngelo, L. E., (1990) Equity Valuation and Corporate Control, The Accounting Review, 65(1): 93 -
34 112.
35
36
Dekel, O., Shamir, O. and Xiao, L. (2010) Learning to classify with missing and corrupted features,
an

37
Machine Learning, 81(2):149–178
38
39
Dietrich, K. J. and Sorensen, E., (1984) An application of logit analysis to prediction of merger targets,
ce

40
41 Journal of Business Research, 12(3): 393-402.
42
43 Dieudonne, S., Fabienne, C. and Slimane, B. (2014), M&A Activity: Where are we in the Cycle?
44 September 17
45 Available at SSRN: https://ssrn.com/abstract=2504412 or http://dx.doi.org/10.2139/ssrn.2504412
46
47 Dong, M., Hirshleifer, D., Richardson, S. and Hong Teoh, S., (2006) Does Investor Misvaluation Drive
48 the Takeover Market?, The Journal of Finance, 61(2): 725 - 762.
49
50 Eakins, S.G. and Stansell, S. R. (2003) Can value-based stock selection criteria yield superior risk-
51 adjusted returns: an application of neural networks, International Review of Financial Analysis,
52 12:83–97
53
54 Eckbo, B. E. (2014) Corporate takeovers and economic efficiency, Annual Review of Financial
55 Economics 6(1): 51-74.
56
57
58
59
60
Page 27 of 31 Managerial Finance

1
2
3 Eden, L. and Dobson, W., (2005) Governance, Multinationals, and Growth. s.l.:Edward Elgar
4 Publishing.
5
6 Ernst & Young, (2015) Buying and Selling: Cross-border mergers and acquisitions and the US
7 corporate income tax , s.l.: Business Roundtable.
8
9 Ernst & Young, 2016. Technology: Capital Confidence Barometer.
10 Available at: http://www.ey.com/Publication/vwLUAssets/ey-ccb-15-technology/$FILE/ey-ccb-15-
11 technology.pdf, [Accessed 16 May 2017].
12
13 Fan , A. and Palaniswami , M., (2000) Selecting Bankruptcy Predictors Ssing a Support Vector
14
Machine Approach. Como, IEEE.
15
16
17 Fescioglu-Unver, N. and Tanyeri, B. (2013) A comparison of artificial neural network and multinomial
M
18 logit models in predicting mergers, Journal of Applied Statistics, 40(4): 712-720
19
20
an
Gregoriou, G. N. and Neuhauser, K., (2007) Mergers and Acquisitions: Current Issues. United
21 Kingdom: Palgrave Scholarly.
22
23
ag

Griffin, J. M., Harris , J. H., Shu, T. and Topaloglu, S., (2011) Who Drove and Burst the Tech Bubble?,
24 The Journal of Finance, 66(4): 1251 - 1290.
25
26
er

Hagedoorn, J. and Duysters, G., (2002) External Sources of Innovative Capabilities: The Preferences
27
for Strategic Alliances or Mergers and Acquisitions, Journal of Management Studies, 39(2): 167 –
28
29 188.
ial

30
31 Hammerstrom, D., (1993) Working With Neural Networks. IEEE Spectrum, 30(7): 46 - 53.
32
Fin

33 Harris, R. S. and Ravenscraft, D., (1991) The Role of Acquisitions in Foreign Direct Investment:
34 Evidence from the U.S. Stock Market, The Journal of Finance, 46(3): 825 - 844.
35
36 Harris, R. S., Stewart, J. F. & Guilke, D. K., (1982) Characteristics of Acquired Firms: Fixed and Random
an

37 Coefficients Probit Analyses, Southern Economic Journal, 49(1): 164 - 184.

38
39 Hinton G. and Salakhutdinov R. (2006) Reducing the dimensionality of data with neural networks,
ce

40 Science, 313(5786):504 – 507

41
42 Hodges, M. B., (2007) Three Approaches?, The Appraisal Journal , 75(1): 34 - 44.
43
44 Hongjiu, L., Huimin, C. and Yanrong, H., (2007) ‘Financial characteristics and prediction on targets of
45
M&A based on SOM- Hopfield neural network,’ 2007 IEEE International Conference on Industrial
46
Engineering and Engineering Management, Singapore, 2007, p. 80-84.
47
48
49 Hussain, A.J., Knowles, A., Lisboa, P.J.G. and El-Deredy, W. (2007) Financial time series prediction
50 using polynomial pipelined neural networks, Expert Systems with Applications, 35:1186–1199
51
52 Hyndman, R. J. and Athanasopoulos, G., (2013) Forecasting: Principles and Practice. s.l.:OTexts.
53 Institute of Mergers, Acquisitions and Alliances, (2016) M&A Statistics.
54 Retrieved October 8, 2017, from
55 http://www.imaa-institute.org/recources/statistics-mergers-acquisitions/
56
57
58
59
60
 Managerial Finance Page 28 of 31

1
2
3 Iturriaga, F. J. L. and Sanz, I. P. (2015), Bankruptcy visualization and prediction using neural networks:
4 A study of U.S. commercial banks, Expert Systems with Applications, 42(6): 2857-2869
5
6 Jensen, M. C., (1986) Agency Costs of Free Cash Flow, Corporate Finance, and Takeovers, The
7 American Economic Review, 76(2): 323 - 329.
8
9 Jo, H. and Han, I., (1996) Integration of Case-Based Forecasting, Neural Network, and Discriminant
10 Analysis for Bankruptcy Prediction, Expert Systems with Applications, 11(4): 415 - 422.
11
12 Jovanovic, B. and Rousseau, P. L., (2001) Mergers and Technological Change: 1885-1998, Vanderbilt
13 University, Working Paper No. 01-W16, August
14
15
Kaastra, I. and Boyd, M., (1996) Designing A Neural Network For Forecasting Financial And Economic
16
17
Time Series, Neurocomputing, 10(3): 215 – 236.
M
18
19 Kuo, C. and Reitsch, A. (1995) Neural networks vs. conventional methods of forecasting, Journal of
20 Business Forecasting Methods and Systems, 14(4): 17-22.
an

21
22 Lam, M. (2004) Neural network techniques for financial performance prediction: integrating
23 fundamental and technical analysis, Decision Support Systems, 37:567–581
ag

24
25 Laudicina, P. A. & Peterson, E. R., (2017) The 2017 A.T. Kearney Foreign Direct Investment Confidence
26 Index.[Online] Available at: https://www.atkearney.com/foreign-direct-investment-confidence-index
er

27 [Accessed 3 September 2017].

28
29
ial

Levine, P. and Aaronovitch, S., (1981) The Financial Characteristics of Firms and Theories of Merger
30 Activity, The Journal of Industrial Economics, 30(2): 149 - 172.
31
32 Lin, C., Huang, J., Gen, M. and Tzeng, G. (2006) Recurrent neural network for dynamic portfolio
Fin

33 selection, Applied Mathematics and Computation, 175:1139–1146

34
35 Lowenstein, R., (2004) Origins of the Crash: The Great Bubble and Its Undoing. s.l.:Penguin Books.
36
an

37 Maher, J. J. and Sen, T. K., (1997) Predicting Bond Ratings Using Neural Networks: A Comparison with
38 Logistic Regression, Intelligent Systems in Accounting, Finance and Management, 6(1): 59–72.
39
ce

40 Mauboussin, M. J., Callahan, D. and Majd, D., (2017) The Incredible Shrinking Universe of Stocks: The
41
Causes and Consequences of Fewer U.S. Equities. Global Financial Strategies, 22 March.
42
43
Moody’s Corporation, (2016) US non-financial corporates' cash pile increases to $1.68 trillion, tech
44
45 holding the lead.
46 [Online] Available at: https://www.moodys.com/research/Moodys-US-non-financial-corporates-
47 cash-pile-increases-to-168--PR_349330 [Accessed 16 May 2017].
48
49 Motiwalla, L. and Wahab, M. (2000) Predictable variation and profitable trading of US equities: a
50 trading simulation using neural networks, Computers and Operations Research, 27:1111–1129
51
52 NASDAQ, (2017) NASDAQ Composite Index (COMP).
53 [Online] Available at: http://www.nasdaq.com/markets/nasdaq-composite
54 [Accessed 6 September 2017].
55
56
57
58
59
60
Page 29 of 31 Managerial Finance

1
2
3 Ohlson, J. A., (1980) Financial Ratios and the Probabilistic Prediction of Bankruptcy, Journal of
4 Accounting Research, 18(1): 109 - 131.
5
6 Olson, D. L., Delen, D. and Meng, Y., (2012) Comparative Analysis of Data Mining Methods for
7 Bankruptcy Prediction, Decision Support Systems, 52(2): 464 – 473.
8
9 Organization for International Investment, (2016) Foreign Direct Investment in the United States,
10 Washington DC: www.ofii.org.
11
12 Palepu, K. G., (1986) Predicting Takeover Targets: A Methodological and Empirical Analysis, Journal
13 of Accounting and Economics, 8(1): 3-35.
14
15
Palepu, K. G., (2003) Business Analysis and Valuation Using Financial Statements. Third Edition ed.
16
17
s.l.:South-Western.
M
18
19 Park, Y. and Park, G., (2004) A New Method for Technology Valuation in Monetary Value: Procedure
20 and Application, Technovation, 24(5): 387 - 394.
an

21
22 Petitt, B. S. and Ferris , K. R., (2013) Valuation for Mergers and Acquisitions. s.l.:Pearson Education,
23 Inc.
ag

24
25 Powell, R. and Yawson, A., (2007) Are Corporate Restructuring Events Driven by Common Factors?
26 Implications for Takeover Prediction, Journal of Business Finance and Accounting , 34(7-8): 1169 –
er

27 1192.
28
29
ial

Powell, R. G., (1997) Modelling Takeover Likelihood, Journal of Business Finance and Accounting,
30 24(7 - 8): 1009 – 1030.
31
32 Rhodes-Kropf, M. and Viswanathan, S., (2004) Market Valuation and Merger Waves, The Journal of
Fin

33 Finance, 59(6): 2685 - 2718.

34
35
Ritter, J. R. (2017) Initial Public Offering: updated statistics. Working paper, University of Florida
36
an

37
38 Roiger, R. J., (2016) Data Mining: A Tutorial-Based Primer, Second Edition ed. s.l.:Chapman and
39 Hall/CRC.
ce

40
41 Rojas, R (1996), Neural Networks: A Systematic Introduction, Springer-Verlag, Berlin
42
43 Rossi, S. and Volpin, P. F., (2003) Cross-Country Determinants of Mergers and Acquisitions, Journal of
44 Financial Economics, 74(1): 277 – 304.
45
46 Routledge, B. R., Sacchetto, S. and Smith N. A. (2013) Predicting Merger Targets and
47 Acquirers from Text, working paper, Carnegie Mellon University
48
49 Salchenberger, L. M., Cinar, E. M. and Lash, N. A., (1992) Neural Networks: A New Tool for Predicting
50
Thrift Failures, Decision Sciences, 23(4): 899 – 916.
51
52
Sen, T. K. and Gibbs, A. M. (1994) An Evaluation of the Corporate Takeover Model Using Neural
53
Networks, Intelligent Systems in Accounting, Finance and Management, 3(4): 279-292
54
55
56
57
58
59
60
 Managerial Finance Page 30 of 31

1
2
3 Servaes, H. and Zenner, M., (1994) Taxes and the Returns to Foreign Acquisitions in the United
4 States, Financial Management, 23(4): 42 - 56.
5
6 Sharda, R. and Odom, M. D., (1990) A Neural Network Model for Bankruptcy Prediction. s.l., IEEE
7 Xplore.
8
9 Simkowitz, M. and Monroe, R. J. (1971) A Discriminant Analysis Function for Conglomerate Targets,
10 The Southern Journal of Business, 38: 1-16
11
12 Sinha, A. P. and Richardson, M. A. (1998) A Case-Based Reasoning System for Indirect Bank Lending,
13 Intelligent Systems in Accounting, Finance and Management, 5(4): 229-240
14
15
Spangler, W. E. May J. H. and Vargas L. G. (2015) Choosing Data-Mining Methods for Multiple
16
17
Classification: Representational and Performance Measurement Implications for Decision Support,
M
18 Journal of Management Information Systems, 16(1): 37-62,
19
20 Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. and Salakhutdinov, R. (2014) Dropout: A
an

21 Simple Way to Prevent Neural Networks from Overfitting, Journal of Machine Learning Research, 15:
22 1929-1958
23
ag

24 Swaminathan, H. and Rogers, H. J., (1990) Detecting Differential Item Functioning Using Logistic
25 Regression Procedures, Journal of Educational Measurement, 27(4): 361 - 370.
26
er

27 Tam K. Y. and Kiang, M. Y., (1992) Managerial Applications of Neural Networks: The Case of Bank
28 Failure Predictions, Management Science, 38(7): 926-947.
29
ial

30 Teeter, P. and Sandberg, J., (2016) Cracking the Enigma of Asset Bubbles With Narratives, Strategic
31 Organization , 15(1): 91 - 99.
32
Fin

33 Thawornwong, S. and Enke, D. (2004) The adaptive selection of financial and economic variables for
34 use with artificial neural networks, Neurocomputing, 56:205–232
35
36 The Institute for Mergers, Acquisitions and Alliances, (2016) United States - M&A Statistics.
an

37 Available at: https://imaa-institute.org/m-and-a-us-united-states/, [Accessed 21 February 2017].

38
39 ,
Tkáč M. and Verner, R. (2016) Artificial neural networks in business: Two decades of research,
ce

40
41
Applied Soft Computing, 38: 788-804
42
43 Tsai, C. F. and Wu, J. W. (2008) Using neural network ensembles for bankruptcy prediction and credit
44 scoring, Expert Systems with Applications, 34(4): 2639–2649
45
46 Tseng, F.-M. and Hu, Y.-C., (2010) Comparing four bankruptcy prediction models: Logit, quadratic
47 interval logit, neural and fuzzy neural networks, Expert Systems with Applications, 37(3): 1846 -
48 1853.
49
50 U.S. Bureau of Labor Statistics, (2016) U.S. Bureau of Labor Statistics. [Online]
51 Available at: https://www.bls.gov/home.htm, [Accessed 7 September 2017].
52
53
54 Versace, M., Bhatt, R., Hinds, O. and Shiffer, M. (2004) Predicting the exchange traded fund DIA with
55 a combination of genetic algorithms and neural networks, Expert Systems with Applications, 27:417–
56 425
57
58
59
60
Page 31 of 31 Managerial Finance

1
2
3
4 Westenberg, D. A. (2009) Initial Public Offerings: A Practical Guide to Going Public, Practising Law
5 Institute
6
7 Westenberg, D. A., (2000) The IPO Report, s.l.: Hale and Dorr LLP.
8
9 Widrow, B., Rumelhart, D. E. and Lehr, M. A., 1994. Neural Networks: Applications In Industry,
10 Business and Science, Communications of the ACM, 37(3): 93 - 105 .
11
12 WilmerHale, (2017), 2017 M&A Report, Accessed November 11, 2017:
13 https://www.wilmerhale.com/uploadedFiles/Shared_Content/Editorial/Publications/Documents/20
14
17-WilmerHale-MA-Report.pdf
15
16 Xiang, G., Zheng, Z., Wen, M., Hong, J., & Rose, C. (2012) A Supervised Approach to Predict Company
17
M
Acquisition with Factual and Topic Features Using Profiles and News Articles on TechCrunch,
18
Artificial Intelligence, 607–610.
19
20
an
Zhang, G., Hu, M. Y., Patuwo, B. E. and Indro, D. C., (1999) Artificial Neural Networks In Bankruptcy
21
22 Prediction: General Framework And Cross-Validation Analysis, European Journal of Operational
23 Research, 116(1): 16 - 32.
ag

24
25 Zhang, G., Patuwo, E. B. and Hu, M. Y., (1998) Forecasting With Artificial Neural Networks: The state
26 of The Art, International Journal of Forecasting, 14(1): 35 – 62.
er

27
28 Zhu, D., Premkumar, G., Zhang, X. and Chu, C. H. (2001) Data Mining for Network Intrusion
29
ial

Detection: A Comparison of Alternative Methods, Decision Sciences, 32(4): 635-660

30
31
32
Fin

33
34
35
36
an

37
38
39
ce

40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60