Available online at www.sciencedirect.

com

Borsa _Istanbul Review

_
Borsa Istanbul Review 20-1 (2020) 13e23
http://www.elsevier.com/journals/borsa-istanbul-review/2214-8450

Full Length Article

Predicting IPO initial returns using random forest

Boubekeur Baba a,*, Güven Sevil b
a
Graduate School of Social Sciences, Anadolu University, Eskis‚ehir, Tepebas‚ı, 26470, Turkey
b
Open Education Faculty, Anadolu University, Eskis‚ehir, Tepebas‚ı, 26470, Turkey

Received 7 May 2019; revised 15 August 2019; accepted 15 August 2019

Available online 21 August 2019

Abstract

Empirical analyses of IPO initial returns are heavily dependent on linear regression models. However, these models can be inefficient due to
its sensitivity to outliers which are common in IPO data. In this study, the machine learning method random forest is introduced to deal with the
issues the linear regression cannot solve. The random forest is used to predict initial returns of IPOs issued on Borsa Istanbul. The prediction
accuracy of the random forest is then tested against methods of robust regression. The prediction results show that random forest has by far
outperformed other methods in every category of the comparison. The variable importance measure shows that the IPO proceeds and IPO
volume are the most important predictors of IPO initial returns. The results also show that the variables that act as potential proxies for ex-ante
uncertainty are more important than variables that are proxies for information asymmetry.
_
Copyright © 2019, Borsa Istanbul Anonim Şirketi. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-
ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

JEL classification: G12; G30; G39

Keywords: Random forest; Initial public offerings; Initial returns; Underpricing; Prediction

1. Introduction during the internet bubble. Consistent with the global evidence,
the IPOs underpricing has been found highly significant in the
IPOs initial returns, often referred to as IPO underpricing in emerging markets (Huang, Lee, Pan, & Nguyen Thi, 2016;
the literature, is one of the most renowned market anomalies Alanazi & Al-Zoubi, 2015; Chang, Chen, Chi, & Young, 2008;
and has been documented in many markets. As early as the Kiymaz, 2000). Empirical evidence, however, shows that the
1970s, researchers such as Ibbotson (1975) observed that the level of IPO underpricing differs considerably among countries.
initial performance of IPOs was exceptionally high, Stoll and A data compiled by Ritter in 2015 shows that the emerging
Curley (1970) as well noticed a remarkable price appreciation markets have much higher IPO underpricing ratios compared to
of equity offerings between the initial offering date and the first developed markets.1 Engelen and Van Essen (2010) examined
market date. In the following years, IPOs underpricing has been IPO data of 21 countries and found a variation of 10% in the
taken seriously and widely discussed in the literature of finance. level of underpricing between countries.
Loughran, Ritter, and Rydqvist (1994) document the occur- Explaining this anomaly has been a prominent focus of
rence of this phenomenon in 25 markets around the globe. academic researchers. Although the notion of IPO under-
Similarly, Ritter and Welch (2002) find that the offer prices of pricing may seem straightforward, practically the process of
IPOs issued by US companies were underpriced by an average IPO is characterized by the complexity of determining the
of 16 percent, this figure then jumped to an extreme level offer price. This complexity arises from the potential conflict

* Corresponding author. 1
Updated global IPO underpricing information can be found on Jay Ritter's
E-mail address: [email protected] (B. Baba). website:https://site.warrington.ufl.edu/ritter/files/2015/05/Initial-Public-
_
Peer review under responsibility of Borsa Istanbul Anonim Şirketi. Offerings-International-Insights-2015-05-21.pdf.

https://doi.org/10.1016/j.bir.2019.08.001
_
2214-8450/Copyright © 2019, Borsa Istanbul Anonim Şirketi. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
14 _
B. Baba, G. Sevil / Borsa Istanbul Review 20-1 (2020) 13e23

of interests among the participants in the IPO process. For this A review by Tkac and Verner (2016) shows that artificial
reason, there has been little consensus regarding whether the neural network applications have made significant inroads in
IPO underpricing is a desirable or undesirable outcome of the finance over the last two decades. According to this review, the
IPO process. For instance, Dalton, Certo, and Daily (2003) majority of these applications are found in studies related to
find that in most of the cases the underwriters are not acting predicting financial distress and bankruptcy and forecasting
in the best of their clients i.e. the IPO firms, but rather favoring shares and bonds performance. As for predicting IPO initial
the recipients of the IPO shares whom often end up receiving returns, the linear regression methods are still the dominant
most of the IPO proceeds. On the other hand, Beatty and Ritter approach. However, there have been significant efforts to
(1986) state that excessive underpricing of the IPO by the analyze IPO returns using a variety of machine learning and
underwriters would be appealing to the uninformed investors, computational intelligence techniques. To name a few, Luque,
but it would not be so to the IPO firms. On the contrary, a Quintana, and Isasi (2012) focus mainly on the offering
higher offer price would benefit the IPO firms but discourages characteristics to predict IPO returns using a genetic algo-
the uninformed investors from buying the IPO. Therefore, the rithm. Huang et al. (2012) also apply a genetic-based algo-
underwriters will seek some optimal level of underpricing that rithm on the IPO's fundamental variables to select the
satisfies both sides. Loughran and Ritter (2002) in efforts to potentially high-growth stocks. The artificial neural networks
understand why the issuers are satisfied leaving a large amount (ANN) and support vector machine (SVM) have been used to
of money on the table by letting the underwriters set a low predict IPO initial returns in studies such as Mitsdorffer and
offer price, they examined the covariance between the issuers' Diederich (2008) and Bastı, Kuzey, and Delen (2015). The
capital sacrifice and their overall wealth after listing. Loughran fuzzy neural network (FNN) an advanced intelligence system
and Ritter found that the issuers attain larger wealth gain on of ANN has been applied recently by Wang et al. (2018) to
the retained shares from a price jump. Another aspect of IPO predict the underpricing of a large sample of U.S. IPOs.
underpricing complexity lies in the difficulty to determine the Robertson, Golden, Runger, and Wasil (1998) construct an
factors that lead to underpricing. In this respect, a considerable OLS regression and neural network models to predict the first
amount of theoretical explanations has been developed to day returns of IPOs, the empirical findings of their study show
rationalize the anomaly of IPO underpricing. Ritter and Welch that the predictions produced by neural network models were
(2002) categorize the theories of IPO underpricing into better than predictions produced by OLS regression. The same
asymmetric and symmetric information. The explanations approach was followed by Reber, Berry, and Toms (2005) to
based on asymmetric information theories have been widely predict IPO initial returns. Their results, however, showed a
supported and followed in the literature of IPO underpricing. slight advantage of neural network models over OLS regres-
Symmetric information theories, on the other hand, have not sion. More recently, the random forest, a powerful and well-
been widely accepted as the primary determinant of under- known machine learning algorithm, has been used by
pricing. One explanation that falls under this category is the Quintana, Saez, and Isasi (2017) to predict the IPO
Tinic (1988) argument which suggests that the IPO is inten- underpricing.
tionally underpriced by the issuers to reduce their legal lia-
bility. Welch and Ritter add another category to IPO 2. Literature review
underpricing theories which focuses on the allocation bias of
IPO shares among the investors. Most of these theories have Ritter and Welch (2002) state that the theories explaining
been subjected to rigorous empirical testing using firm-specific the anomaly of IPO underpricing follow two lines, the first line
and market specific-factors. The empirical evidence presented is based on the information asymmetry problem among the
in the literature is notably in favor of the asymmetric infor- participants in the IPO transactions. The winner's curse model
mation theory. of Rock (1986), the first model to underline the asymmetric
In this study, a machine learning algorithm is employed to information, contends that some investors are better informed
predict IPOs initial returns. Machine learning is a subset of about the market than other investors. The informed investors
data science that learn when exposed to a dataset, but the have the edge over other investors as they only subscribe to
dataset has to be large enough to sustain the learning process, shares of an attractive IPO while the uninformed investors
financial data though is considered small to medium compared subscribe to every IPO. This means that attractive shares will
to other fields where machine learning models are applied to be oversubscribed and dominated by informed investors. The
larger data. However, financial data is also featured with noise uninformed investors, on the other hand, will receive a small
and might be heavily skewed in some cases, for such issues the proportion of the attractive IPO shares and the full supply of
nonlinear techniques are the most appropriate. Fortunately, unattractive IPOs. As a result, they end up receiving expected
machine learning algorithms fit perfectly for this purpose. In returns below the average underpricing or even negative
addition, some of the machine learning algorithms have returns. The uninformed investors then would react by ceasing
already been used in many financial applications, especially in bidding for any IPO shares in the future. Therefore, Rock
risk management and forecasting the future stock returns, and (1986) assumes that the underpricing is pre-market deter-
have outperformed the classical financial methods (see e.g., mined by the issuers in order to increase the demand and
Desai & Bharati, 1998; Hanias, Thalassinos, & Curtis, 2012; reduce the effect of the information asymmetry between the
Maciel & Ballini, 2010; Thawornwong, Enke, & Dagli, 2003). informed and the uninformed investors. Beatty and Ritter
 _
B. Baba, G. Sevil / Borsa Istanbul Review 20-1 (2020) 13e23 15

(1986) conceptually extend the winner's curse model with a unprecedented scale of privatization. The IPO market, in turn,
different division between informed and uninformed investors, experienced a significant uptick in terms of the realized pro-
this model assumes that the underwriters are better informed ceeds. It may be noted that the period 2004e2007 was the
about the market than the issuing corporations. This, in turn, most successful period for IPOs thanks to the large scale of
allows them to set the offer price more accurately. The privatization sales, a total of 44 firms went public in this
signaling model developed by Welch (1989), Allen and period raising around 6.4$ billions of revenues. The period of
Faulhaber (1989) and Grinblatt and Hwang (1989) also com- global recession showed a stark reduction of the number of
plements the argument of information asymmetry by assuming IPOs with only three listings on the BIST in the years 2008
that the issuing corporations are better informed about the and 2009 combined. This reduction in IPO activity was largely
intrinsic value of their companies than the outside investors. driven by the rapid decline in prices on BIST following the
The other line of IPO underpricing theories discusses the breakout of the global recession in the third quarter of 2008.
allocation bias in the shares of IPO. It argues that the un- The IPO activities bounced back in 2010 and maintained its
derwriters may use their discretionary power to purposively positive trend up until 2012, albeit with a reduced total amount
underprice and diffusely allocate the underpriced IPO shares of capital raised. The IPO market in Turkey begun to slow
to favored clients. The underwriters, in turn, can boost their down again in 2012 and has not been so active in the following
profits through the trading commissions they acquire from the years. Throughout the period of the study, we find that 127
clients. offerings out of the total offerings were underpriced, the rest
Despite the myriad of theories put forth by the researchers of the offerings which account for around 41.5% of the total
to explain the anomaly of IPO underpricing, the underlying offerings were either overpriced or correctly priced (Table 2).
causes of this phenomenon are still debated. As pointed out by The IPO data used in the empirical analysis is obtained
Ljungqvist (2007), the asymmetric information theories are from Borsa Istanbul. The data of the companies' financial
the most supported by empirical evidence, but according to operating history prior to IPO was extracted from the financial
Ritter and Welch (2002), it is unlikely to be the primary statements archive of the publicly traded companies provided
determinant of fluctuations in IPO activity and underpricing. by Borsa Istanbul and the public disclosure platform
In addition, most of the theories hold well in the developed (kap.org.tr). Information about the foundation dates was
markets, yet they are less successful in explaining extremely retrieved from the companies' yearbooks provided by Borsa
high underpricing of IPOs in emerging markets. In China, for Istanbul. A summary of this data is provided in Table 4.
example, Tian (2011) and Gao (2010) find no significant
relationship between the IPOs initial returns and the proxies of 3.2. Model selection
asymmetric information. The studies on the Indian IPOs find
mixed evidence. Using IPO data for the period 2005e2012, To predict the IPO first-day returns, we first select the most
Chhabra, Kiran and Sah (2017a) find the variables that signal commonly used measure of underpricing proposed by
information are highly significant and companies with high Ibbotson (1975) and Ibbotson, Sindelar, and Ritter (1988,
information disclosure experience less underpricing. On the 1994), which is expressed as
other hand, Chhabra, Kiran, Sah and Sharma (2017b) find the
informational variables less effective in explaining the IPO IRit ¼ ðCPit  OPit Þ = OPit
underpricing. More examples of studies on IPO from advanced
where CPit is the closing price of the first trading day; OPit is
and emerging markets are summarized in Table 1.
the firm's offer price i at time t; IRit is the IPO's initial return of
the firm i at time t.
3. Data and model selection Secondly, the IPO initial return is regressed over a set of
variables that have been theoretically and empirically linked to
3.1. Data
underpricing in the preceding literature. The variables used in
the empirical model can be classified into company charac-
The data set used in this study consists of first-day trading
teristics, offering characteristics, and market sentiment in-
returns of 127 public offerings listed on Borsa Istanbul (BIST)
dicators. The selected variables are summarized in Table 3.
during the period 1998e2018. In this period, a total of 217
Company characteristics: This category consists of the
firms went public raising around 17.8$ billions of capital. As it
variables firm age, firm size, net income and returns on assets
is shown in Table 2, IPO activities peaked in 2000 with 35 (ROA) to proxy for the firm-specific risk factors. The firm age
deals then dramatically dropped to one deal in the next year,
and firm size are commonly used in the literature as a proxy to
this dramatic decline was due to the liquidity crisis in 2001
measure uncertainty and information asymmetry as well.
and the depreciation of the Turkish Lira by 50% against the
Firms established long before listing should have more in-
dollar in a short period of time. In the following years, espe-
formation available to the public which reduces the informa-
cially since 2004, the global favorable monetary conditions
tion asymmetry among the issuer, the underwriters, and the
spurred economic growth and sectoral developments in many
investors. Ritter (1984) argues that the level of underpricing is
emerging markets. In this period, Turkey saw an
Table 1

16
Studies of IPO underpricing from developed and emerging markets.
Authors Dataset Variables The main findings
Pande and Vaidyanathan (2009) IPOs issued on NSE in the period between March Dummy variable for market demand, listing delay, Market demand and market percentage change are
2004 and October 2006 marketing expenditures, issue size, market change the main drivers of IPO underpricing.
on the day of listing.
Wadhwa (2014) The analysis was performed on the dataset of 92 Offer price, listing delay, a dummy variable for Underpricing was found to be positively related to
IPOs issued between 2009 and 2011 on NSE market demand, the reputation of the leading the offer price and listing delay.
underwriter, issue size, IPO grade, firm age, internal
risk factors, equity retained, a dummy variable for
private and government issues
Lin and Hsu (2008) The newly listed firms on Taiwan stock exchange Market adjusted returns, Allotment ratio, Debt The evidence does not support the ex-ante
and Hong stock exchange for the period between Ratio, IPO proceeds, trading volume, market cap uncertainty hypothesis in both markets, asymmetric
January 1999 and June 2004 dummy, sectors' dummy information measures are the most consistent
determinant of underpricing in both stock markets

B. Baba, G. Sevil / Borsa Istanbul

Ekkayokkaya and Pengniti (2012) A sample consists of 463 IPOs made during the Market co-movement, lagged underpricing, pre-IPO The study documents a significant reduction in IPO
period 1990e2007, and subsamples for the pre- market returns, pre-IPO market volatility, stock underpricing following major governance reforms.
reform, transitional and post-reform periods market cap. to GDP, aftermarket volatility, issue size The listing activities show improvement after
to industry market cap., elapsed time, dummy governance reforms and the frequency of use of
variable for governance reforms, use of proceeds proceeds disclosures have reduced the ex-ante
disclosure, ownership control retention. uncertainty.
Satta (2017) IPOs issued on international stock exchanges by Firm age, firm size, core business, number of The findings support the validity of the timing and
firms operating in the port industry in the 2001 underwriters, the reputation of the lead underwriter, signaling hypothesis, the age of the issuer is found
e2015 percentage of equity offered, variable represents the to moderate IPO underpricing, IPOs issued by port

_
historical background of the port industry. companies suffer higher levels of underpricing.
Peng and Wang (2007) A sample consists of 647 IPOs issued in Taiwan Proxy for the flotation method, the return of the Ex-ante uncertainty and asymmetric information
stock exchange in the period 1996e2003. market index, the standard deviation of market play a significant role in IPO underpricing, the

Review 20-1 (2020) 13e23

index returns, the ratio shares offered divided by the auction flotation method reduces IPO underpricing.
number of subscriptions, the offer size, sale revenue,
index of earnings management, underwriter's
reputation, the auditor reputation, dummy variable
for electronic firms, dummy variable for firms listed
in the OTC.
Tian (2011) Dataset consists of 1377 IPOs listed on the Shanghai The pricing cap, issuing size, allocation rate, total Financial regulations account for more than half of
and Shenzhen Stock Exchanges between 1992 and assets, Firm age, listing delay, size of the the severe underpricing, investment risks also
2004. government shareholding, size of block contribute to severe underpricing, asymmetric
shareholding, size of employee shareholding, size of information is far from being the major determinant
managerial shareholding. of underpricing.
Jewartowski and Lizinska (2012) IPOs issued on the Warsaw Stock Exchange in the Firm size, privatization, return on equity, the The study documents a strong effect of early
period 1998e2008. average of underpricing over a ninety-day window aftermarket volatility, issuer's size, growth
prior to the IPO date, market condition, market-to- opportunities, and profitability before the offering
book value, early return volatility. on IPO initial returns.
Marshal (2004) A total of 532 IPOs listed on US stock markets in The rank of the lead underwriter, the auditor Financially risky firms with few alternative financial
the period 1993e1995 reputation, percentage of retained shares, percentage sources have higher underpricing, the large supply
of proceeds allocated for usages outside general of IPOs in a particular industry leads to higher
corporate or working capital, offer size, percentage initial-day returns.
of venture capital or corporate funding, debt
capacity, IPO volume at the industry level, measures
of firm's financial risk
 _
B. Baba, G. Sevil / Borsa Istanbul Review 20-1 (2020) 13e23 17

positively related to ex-ante uncertainty about the value of the

beside ex-ante uncertainty improves the explanatory

Rock (1986) found to be significant under specific

The reputation of underwriters helps to reduce the
underpricing, the informed demand hypothesis of
firm. The level of underpricing, therefore, is negatively asso-
Including proxies for ex-post value uncertainty

ciated with the company's age prior to listing. This relationship

has been empirically confirmed in many studies (Ang & Brau,
2002; Loughran & Ritter, 2004; Chahine, 2008). Similarly,
larger firms, as compared to small firms, are perceived as less
risky because it displays less uncertainty about its value. The
large companies also have more public disclosure which de-
creases information asymmetry, winner's curse theory of Rock
power of the model.

(1986) and information asymmetry theory of Beatty and Ritter

(1986) both suggest that greater information asymmetry is
always associated with bigger underpricing. Net income and
conditions.

ROA are also included to proxy for information asymmetry

and to evaluate the firm's quality and performance prior to
IPO, Lowry and Shu (2002) state that firms that experience
stronger operating performance prior to the IPO are subject to
days after listing, the offer size, proxy for hot issues,
The standard deviation of return over the first twenty

The subscription rate, the offer size, firm size, the

first ten days after listing, the market share of the

offer price, standard deviation of returns over the
firm age, a dummy variable for tech and internet
companies, the rank of the lead underwriter, first

less uncertainty.
day's trading volume, proxies for ex-post value

Offering characteristics: The variables selected under this

Note: Except for the study of Peng and Wang (2007) which uses stochastic frontier model, all the other reported studies use linear regression.

category are the IPO volume, the IPO rate, the offer price, and
IPO proceeds.2 The offer price, IPO proceeds, and volume can
be regarded as an indicator of uncertainty. Daily, Certo,
Dalton, and Roengpitya (2003) suggest that the highly-
priced IPOs are characterized by lower uncertainty regarding
underwriter, IPO volume.

the future performance of the firm. In contrast, Ibbotson et al.

(1988) finds firms that offer a low price have a high level of
uncertainty, and that their offered equities can be subjected to
speculative trading. The size of the IPO often measured by the
uncertainty.

proceeds is supposed to have a strong impact on underpricing,

Clarkson (1994) find the IPO size to be an effective proxy for
ex-ante uncertainty, whereas How, Izan, and Monroe (1995)
report a significant negative relationship between the size
and underpricing. As for the IPO rate, which represents the
IPOs listed on the AMEX, NASDAQ, or NYSE

All the new offerings listed on the Hong Kong

percentage of shares offered, the signaling theory suggests that

this factor conveys information about the quality of the firm
(Leland & Pyle, 1977). Consequently, the higher percentage of
Exchange over the period 1994e2005

shares offered, the lower underpricing. In addition to these

variables, a dummy variable is included to indicate for the
during the period 1993e1998

foreign investors' participation in the IPO process.

Market sentiments: The 30 days and 60 days of market
performance prior to IPO date are commonly used as an in-
dicator for market sentiments. The cyclical behavior hypoth-
esis argues that IPOs realized during the hot market are
heavily underpriced compared to those realized in periods of
cold market. Ritter (1984) documented the existence of these
type of behaviors in US markets during the 1980s. In addition,
Hanley (1993) reports a positive relationship between IPO's
Falconieri, Murphy, and Weaver (2009)

initial return and market index returns prior to IPO. Due to the
existence of a short-term momentum effect in Borsa Istanbul
as evidenced by Ejaz and Polak (2015), the two measures are
considered in the empirical analysis. Therefore, two different
Vong and Trigueiros (2010)

empirical models are developed to study the effect of each

measure separately.

2
To remove the inflation effect that may distort the results, the IPO proceeds
specifically have been taken as US dollars in most of the studies on Turkish
companies' IPOs such as Bastı et al. (2015), Durukan (2006), Yüksel and
Yüksel (2006), Durukan (2002) and Kiymaz (2000).
18 _
B. Baba, G. Sevil / Borsa Istanbul Review 20-1 (2020) 13e23

Table 2 Quintana et al. (2017), the random forest has some unique
IPO activities on Borsa Istanbul during the period 1998e2018. features over other tree-based techniques which make it
Year Number of IPO deals IPO proceeds (millions of USD) Percent (%) potentially suitable for predicting IPO returns. Predicting IPO
1998 20 383.35 2.15 initial return has been a challenging task due to the involve-
1999 10 90.72 0.51 ment of a large number of determinants with very different
2000 35 2806.22 15.76 explanatory power and the presence of outliers. In this regard,
2001 1 0.24 0.00
2002 4 56.47 0.32
the random forest with its ability to combine weak and strong
2003 2 11.25 0.06 variables and its robustness to outliers can be a very useful tool
2004 12 482.58 2.71 for this task. In general, the machine learning algorithms and
2005 9 1743.96 9.80 particularly the random forest work effectively on large data.
2006 15 930.50 5.23 Therefore, sectoral segmentation of the IPOs which is a
2007 9 3298.31 18.53
2008 2 1876.92 10.54
common practice in the IPO literature would significantly
2009 1 6.91 0.04 shrink the data sample and ultimately leads to poor results. On
2010 22 2104.02 11.82 account of this, sectoral segmentation of the IPOs is avoided to
2011 25 826.49 4.64 ensure better results. The main purpose of this study is to
2012 16 297.08 1.67 extend the work of Quintana et al. (2017) to other markets and
2013 8 721.65 4.05
2014 9 308.87 1.73
provide further supporting evidence to the advantage of using
2015 3 23.21 0.13 the random forest in predicting IPO initial returns.
2016 2 117.14 0.66
2017 3 351.76 1.98 4.1. Random forest
2018 9 1366.44 7.67
Total 217 17804.1 100
Random forest, developed by Breiman (2001), is an
Source: Borsa Istanbul (BIST). ensemble learning method in which multiple decision trees are
constructed and merged together to get a more accurate and
4. Research methodology stable prediction or classification. The trees in the random
forest are drawn from the original sample using bootstrap
The empirical analyses carried out to explain and predict resampling, and each tree is grown using a randomized subset
the IPO initial returns are often based on linear regression of features. The procedures to produce the random forest of
models. There are also cases where non-linear models such as regression trees are explained below.
logistic regression are implemented. The use of machine- Let's assume we have the dataset D ¼ fðx1 ; y1 Þ……ðxn ; yn Þg
based methods has been on the rise recently. In this study, and the aim is to find the function f : X/Y where X is the inputs
the random forest -one of the most popular machine methods and Y is the produced outputs. Furthermore, let M be the number
in both classification and regression-is used to predict the of inputs.
returns of IPOs issued on Borsa Istanbul in the period between
1998 and 2018. In addition, since the random forest is a novel 1 Random forest randomly selects n observations from D
technique to the IPO literature, its predictive accuracy is with replacement to form a bootstrap sample.
compared to some of the well-known robust regression 2 Each tree is grown using a subset of m features from the
methods. To the best of the authors' knowledge, there is only overall M features. For regression, it is recommended to
one study conducted by Quintana et al. (2017) which uses the set the subset of features at M=3.Then at each node, m
random forest to predict the IPO returns. According to features are selected at random and the best performing
split among the m features is selected according to the
impurity measure (Gini impurity).
Table 3
3 The trees are grown to a maximum depth without pruning.
Variables definitions.
Dependent variable IR Fist-day IPO return
Growing trees without pruning and selecting the best fea-
Firm characteristics FA Firm age prior to the listing date tures split at each node allow the random forest to maintain
FS Firm size (assets)
prediction strength. In addition, the random selection of fea-
ROA Return on assets
NI Net income tures reduces the correlation between the trees. Unlike other
Offering characteristics IPOV Number of shares offered tree-based techniques, the random forest is immune to over-
IPOR IPO rate (the proportion fitting as more trees are added to the forest. The overall pre-
of shares offered) dictions are produced by taking the average of the predictions
OP The offer price
of the individual trees in the forest. Random forests do not
Pd The IPO proceeds (in USD)
D20 Dummy variable equals 1 if the only generalize the predictions of trees over the entire sample,
foreign investors' purchase but also provide variable importance measure using the out-of-
of IPO exceeds 20%, and 0 otherwise. bag sample. The main idea is eliminating the dependence of
Market sentiments MP30 30 days market return prior to IPO the predictor with the response variable by permuting its
MP60 60 days market return prior to IPO
values across all trees, then the loss of prediction accuracy of
 _
B. Baba, G. Sevil / Borsa Istanbul Review 20-1 (2020) 13e23 19

the forest is estimated, high loss implies high importance of should at least be more than half of the number of observa-
the predictor and vice versa. It should be noted that random tions. Put differently, the trimming parameter is the threshold
forest is frequently implemented with K-folds cross-validation that separates the outlier points from the rest of the observa-
when accurate assessment against other machine methods is tions. Therefore, LTS would be less efficient if the trimming
required. However, such a procedure may not be necessary parameter is small. However, it should be noted that LMS and
since the unbiased estimate of error can be estimated internally LTS have a high percentage of breakdown points, which
in the random forest. means that these two methods are insensitive to outliers.

4.2. Robust regression methods 5. Empirical results

The ordinary least squares method is known to be sensitive The ability of the random forest to produce accurate pre-
to outlier points. In robust regression, the influence of the dictions of IPO initial returns is the main focus of empirical
outliers on the fitted regression line is reduced using weight analysis. To this end, the predictive accuracy of random forest
function. This has the additional advantage of making outliers is compared to that of robust regressions in term of mean
stand out more strongly against the line. There are many square errors (MSE) and root mean square errors (RMSE). In
weighting functions proposed in the literature. In this study, addition, the comparison also includes the measures of
three robust regression methods are used namely the itera- descriptive statistics of the predictions produced by each
tively reweighted least-squares (IRLS) algorithm, the least method and the actual initial returns. The second part of the
median squares (LMS), and the least trimmed squares (LTS). analysis discussion is devoted to studying the relative
In the IRLS algorithm, the outlier points are weighted using explanatory power of the independent variables.
Huber psi and Tukey's bisquare psi functions.
5.1. Predictions results
4.2.1. The iteratively reweighted least-squares (IRLS)
The method of iteratively reweighted least squares consists To obtain the best predictions from the random forest, the
of an underlying weighted least squares fit that is placed inside initial value of features split to be used at each node was set at
an iteration loop. When a loop is passed at each iteration, a five, and all the other parameters of the trees to be grown were
least-squares fit is carried out using a set of weights, weight is let at default. Then, the value of features split was decreased
assigned to each observation. Moreover, the weights are gradually at each experiment. The best predictions were pro-
derived from the current residual and are updated from itera- duced at value three, meaning that all the trees inside the forest
tion to iteration. Since the new weights are derived from the were constructed using three random variables. In addition, the
residuals, the iteration process goes on as the residuals keep same value of features split was given when automated search
changing, then the process terminates when the residuals for the optimal value was implemented, the automated search
remain unchanged over two passes. The IRLS heavily depends was executed with the value of features split initially set at
on the weighting functions (see Heiberger and Becker. (1992) two. During the course of the experiments, we noticed that the
for more details about the most commonly used weighting change in the number of grown trees did not have a major
functions). impact on the results, but the change of features split value was
crucial in obtaining the best predictions.
4.2.2. The least median squares (LMS) and least trimmed Table 5 summarizes the prediction errors measured by
squares (LTS) mean square and root mean square of errors. MSE and RMSE
Linear least squares estimator minimizes the sum of are measures of the absolute fit of the regression model pre-
squared residuals to find the parameters that best fit a set of dictions to the observed values, which also refers to the
data points. The least median squares estimator replaces the proximity of predicted values to the observed values and in the
sum of squared residual with the median of squared residuals.
As stated by Rousseeuw (1984), the creator of the technique,
the resulting estimator from this process can resist the effect of Table 4
nearly 50% of contamination in the data. Rousseeuw later Descriptive statistics.
introduced the least trimmed squares (LTS) to improve the Min. Median Mean Max. Std.dev
asymptotic efficiency of LMS. This method consists of finding IR 0.0032 0.12 0.1314 0.38 0.08024737
a subset of data points whose deletion from the data set would FA 0.2849 12.7288 15.4671 72.211 13.89858
lead to the regression with the smallest residual sum of FS (in millions) 0.1769 37.55 585 34480 3176.11346
squares. The idea is to mitigate the influence of outlier points IPOV (in millions) 0.05 3.6 21.520896 625 71.038799
ROA 0.10143 0.03719 0.07669 0.77646 0.1149453
by minimizing the sum of the smallest squared residuals rather
NI (in millions) 62.976 0.880541 12.160637 864.259 79.532473
than the complete sum of squares. This is done by ordering the IPOR 0.00345 0.2518 0.30292 0.9907 0.1773995
squared residuals from smallest to largest, then the number of OP 1 4 10.06 100 14.33584
the smallest squared residuals is determined by specified Pd (in millions $) 0.2427 13.97 89.81 1837 264.401905
trimming parameter which also leaves out the percentage of MP30 24.0255 0.4077 1.5025 47.3977 11.83767
MP60 34.049 2.777 2.893 59.782 14.46375
outliers among all the observations, the trimming parameter
20 _
B. Baba, G. Sevil / Borsa Istanbul Review 20-1 (2020) 13e23

Table 5 the other models performed better when the 30 days pre-IPO
Prediction errors. market performance was replaced with the 60 days of pre-
Method First model Second model IPO market performance. LMS and LTS methods, which
MSE RMSE MSE RMSE have been established to be insensitive to outliers, have per-
Random forest 0.001156 0.033998 0.001133 0.033669 formed poorly in both models and they even produced less
LMS 0.050222 0.224102 0.045886 0.214211 accurate predictions than the IRLS methods. This was not
LTS 0.036997 0.192346 0.013282 0.115249 expected considering the high breakdown point percentage of
IRLS-T 0.006011 0.077536 0.005949 0.077135 LMS and LTS, but it should be noted that LMS and LTS do not
IRLS-H 0.006012 0.077541 0.006022 0.077603
weigh down the outlier points but rather ignore them. There-
Note: IRLS-T and IRLS-H refer to iteratively reweighted least squares fore, it is highly likely that the strong presence of outlier
implemented with Tukey's bisquare psi function and Huber psi function
respectively.
points in the data has far exceeded the resistance level of the
methods. The IRLS methods, on the other hand, were able to
perform better because it weighs down each observation.
Table 6 In term of the descriptive statistics of the predictions, as it
Descriptive statistics of the predicted values.
is shown Table 6, the predictive performance of the random
Method of prediction Median Mean Std.dev forest again has been exceptionally strong, the median and
First model Random forest 0.126 0.1321 0.050041 mean predictions by random forest fall close to the mean and
LMS 0.0736 0.1035 0.171793 the median of the observed values, even the standard deviation
LTS 0.1942 0.165 0.178523
IRLS-T 0.1303 0.1312 0.020103
of random forest predictions is slightly different than the
IRLS-H 0.1299 0.1306 0.019427 standard deviation of the observed values. The IRLS pre-
Second model Random forest 0.1319 0.1321 0.050117 dictions in term of the mean and median are acceptable, but
LMS 0.1852 0.184 0.192766 the standard deviations of their predictions are far less than the
LTS 0.1811 0.1621 0.090621 standard deviation of the observed values. The MLS and LTS
IRLS-T 0.1307 0.1311 0.021117
IRLS-H 0.1291 0.1297 0.019409
offered unreliable predictions in each of the descriptive sta-
tistics measures. Overall, the descriptive statistics measures
make it apparent that the distribution of random forest pre-
same time it indicates for the unexplained variance in the re- dictions is relatively identical to the distribution of the
siduals. Therefore, lower values of MSE and RMSE indicate observed values.
better fit. The random forest, as shown in Table 5, is able to The IPO underpricing in Turkey has been explored in a
produce far better predictions compared to the other methods, number of studies (see e.g., Kiymaz, 2000; Durukan, 2002;
this can be seen more clearly in the plot of the random forest Atas‚ et al., 2003; Yüksel & Yüksel, 2006; Bildik & Yılmaz,
predictions against the actual values in Fig. 1. The errors 2008; and Ozdemir & Kizildag, 2017). However, all these
found in the random forest predictions are extremely small in studies were basically concerned with demonstrating the sig-
both models, but the predictions of the second model which nificance of factors affecting the initial returns rather than
accounts for the 60 days pre-IPO market performance are making predictions of IPO returns. Moreover, some of these
slightly better than the predictions of the first model, this studies focus on a certain aspect of IPO underpricing. For
means that the market short-term momentum effect is an instance, Ozdemir and Kizildag (2017) study the relationship
important factor for the initial performance of IPOs. In fact, all between the franchising activities of IPO candidates and the

Panel A: Predictions of the fist model with 60 Panel B: Predictions of the second mdel with 30
days pre-IPO market performace. days pre-IPO market performance
0.4 0.4
0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
1
9
17
25
33
41
49
57
65
73
81
89
97
105
113
121
1
9
17
25
33
41
49
57
65
73
81
89
97
105
113
121

Predicted values Actual iniƟal returns Predicted values Actual iniƟal returns

Fig. 1. Random forest predictions vs observed values.

 _
B. Baba, G. Sevil / Borsa Istanbul Review 20-1 (2020) 13e23 21

price of their initial offerings, whereas Yüksel and Yüksel received an average score in the second model, this variable
(2006) focus on the effect of the trading volume on under- has been found to be significant in Ertuna et al. (2003) while
pricing. In addition, the empirical investigations of these Kiymaz (2000) reports an insignificant relationship between
studies were carried out using linear regression methods. the IPO rate and IPO returns. The market sentiment is rela-
Meaning that predicting IPO initial returns was beyond the tively important in both models, Kiymaz (2000) finds that the
scope of these studies. Therefore, the lack of research on this market trend between IPO date and first trading day of IPO has
matter urges the need to explore the subject of underpricing a significant impact on IPO returns. Similarly, Yüksel and
with methods other than classical linear models. In this regard, Yüksel (2006) find the 40 days change in market index prior
the random forest can deliver better results on both ends. As it to IPO to be significant. For company characteristics, firm size
is already shown, the random forest is able to produce accurate and return on assets are the most important for predicting IPO
predictions for IPO returns. On top of this, variable importance initial returns but not as important as the offering character-
measure, which is statistically equivalent to variable signifi- istics. The firm's size and age have been reported to be sig-
cance in linear regression, can be carried out with the random nificant in most of the aforementioned studies except in
forest. Kiymaz (2000). The variables that act as potential proxies
for ex-ante uncertainty such as proceeds, IPO volume, and
5.2. Variable importance firm size are highly important in predicting IPO returns, while
proxies for information asymmetry such as firm age, return on
As previously mentioned, the variable importance is assets and net income has been less important than proxies for
measured by the loss of the model's prediction accuracy when ex-ante uncertainty.
the variable of interest is disassociated from the response The use of machine learning methods to study IPOs of
variable. The results of this process are represented by Fig. 2. Turkish companies has been seen previously in a study con-
In both models, the variable of IPO proceeds is ranked as the ducted by Bastı et al. (2015), this study investigated the short-
most relevant variable to IPO underpricing followed by the term performance of IPOs using decision trees and support
volume of the IPO. Note that introducing the 60 days pre-IPO vector machines. To calculate the variable importance Bastı et
market performance to the model changed some of the out- al. (2015) employed sensitivity analysis, in which the impor-
comes. The pre-IPO market performance and net income tance scores given by each method were combined and aver-
maintained the same rank, but the firm size jumped up two aged on the weight of the model to obtain the final importance
rows and the variable of return on assets fell to the rank five in score. In contrast to our study, their findings suggest that the
the second model. In addition, the importance of the offer proceeds to be the least relevant variable while the market
price decreased while the IPO rate importance increased in the sentiments had the highest score. However, the sample used in
second model. The firm age and the dummy variable signaling the study of Bastı et al. (2015) consists of the IPOs made by all
for foreign investors' share of IPO are reportedly the least the companies except investment trusts. In addition, before the
relevant variables. As can be seen in the figure below, the two execution of empirical analysis the data was screened and the
variables exchange the bottom positions between the two outlier points were cleaned from the sample, such procedure
models. The results also suggest that the offerings character- besides being questionable, it may have significantly affected
istics represented by IPO proceeds and IPO volume are the their results. Their findings though were partially similar to the
main determinants of IPO initial returns, the studies of results of this study in the sense that proxies for information
Durukan (2002), Atas‚ et al. (2003), Ertuna, Ercan, and asymmetry being less important in predicting IPO returns.
Akgiray (2003) and Yüksel and Yüksel (2006) all report sig- Practically, the empirical findings of this study may have
nificant relationship between IPO returns and the IPO pro- further importance in light of the complications involved in the
ceeds. The IPO rate, the supposed measure of firm quality IPO process. These complications, which the bulk of IPO

Panel A: Variable importance of the first model Panel B: Variable importance of the second
with 30 days pre-IPO market performance. model with 60 days pre-IPO market performance

Pd Pd
IPOV IPOV
ROA FS
MP30 MP60
FS ROA
OP IPOR
NI NI
IPOR OP
D20 FA
FA D20
0 5 10 15 20 0 5 10 15 20

Fig. 2. Out-of-bag variable importance measured by the increase in MSE.

22 _
B. Baba, G. Sevil / Borsa Istanbul Review 20-1 (2020) 13e23

literature hinge on, stem mainly from the determination of IPO Conflict of interest
price and the post-market IPO performance. In this regard, the
use of the random forest to analyze IPO returns may be of We have no conflicts of interest to disclose.
particular relevance to the IPO's issuers and investors as both
parties are highly concerned with the uncertainty regarding References
IPO price and performance.
Alanazi, A. S., & Al-Zoubi, H. A. (2015). Extreme IPO underpricing and the
6. Conclusion legal environment in wealthy emerging economies. Journal of Multina-
tional Financial Management, 31, 83e103.
Allen, F., & Faulhaber, G. R. (1989). Signaling by underpricing in the IPO
Predicting IPO initial return has been a challenging task market. Journal of Financial Economics, 23(2), 303e323.
due to the involvement of a large number of determinants Ang, J. S., & Brau, J. C. (2002). Firm transparency and the costs of going
with very different explanatory power and the presence of public. Journal of Financial Research, 25(1), 2002.
outliers. Linear regression models have dominated the Atas‚, R., Karan, M. B., & Aydogan, K. (2003). Forecasting short run perfor-
mance of initial public offerings the Istanbul Stock Exchange. The Journal
empirical investigations in this domain. Linear regressions, of Entrepreneurial Finance, 8(1), 69e85.
however, can be inefficient due to its high sensitivity to Bastı, E., Kuzey, C., & Delen, D. (2015). Analyzing initial public offerings
outlier points. In this study, the random forest, a powerful short-term performance using decision trees and SVMs. Decision Support
machine method, is introduced to deal with the issues linear Systems, 73, 15e27.
regression cannot solve. The predictive performance of Beatty, R. P., & Ritter, J. R. (1986). Investment banking, reputation, and the
underpricing of initial public offerings. Journal of Financial Economics,
random forest is tested on a sample of underpriced equity 15(2), 213e232.
offerings issued on Borsa Istanbul in the period between Bildik, R., & Yılmaz, M. K. (2008). The market performance of initial public
1998 and 2018. Then the results of the random forest are _
offerings in the Istanbul stock exchange. BDDK Bankacılık ve Finansal
compared to the prediction accuracy of robust regression Piyasalar, 2(2), 49e75.
methods in term of MSE, RMSE, mean, median and stan- Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5e32.
Chahine, S. (2008). Underpricing versus gross spread: New evidence on the
dard deviation. Robust regressions are by design less sen- effect of sold shares at the time of IPOs. Journal of Multinational
sitive to data contamination. The LMS uses the median of Financial Management, 18(2), 2008.
squared errors instead of the sum while the LTS uses the Chang, E., Chen, C., Chi, J., & Young, M. (2008). IPO underpricing in China:
trimming parameter to exclude the outlier points, in IRLS New evidence from the primary and secondary markets. Emerging Markets
every point in the data is weighted using weighting func- Review, 9(1), 1e16.
Chhabra, S., Kiran, R., & Sah, A. N. (2017a). Information asymmetry leads to
tions. The outcomes of the comparison show that random underpricing: Validation through SEM for Indian IPOs. Program, 51(2),
forest has by far the better performance in every category of 116e131.
the comparison. The random forest predictions for the mean Chhabra, S., Kiran, R., Sah, A. N., & Sharma, V. (2017b). Information and
and the median are almost identical to the observed values, performance optimization: A study of Indian IPOs during 2005e2012.
while the standard deviation of random forest predictions Program, 51(4), 458e471.
Clarkson, P. M. (1994). The underpricing of initial public offerings, ex ante
falls slightly below the standard deviation of the observed uncertainty, and proxy selection. Accounting and Finance, 34(2), 67e78.
values. In fact, random forest predictions are obviously close Daily, C. M., Certo, T. S., Dalton, D. R., & Roengpitya, R. (2003). IPO
to the actual initial returns as demonstrated in Fig. 1. The underpricing: A meta-analysis and research synthesis. Entrepreneurship:
main conclusion to be drawn from this comparison is that Theory and Practice, 27(3), 271e295.
random forest has all the potentials to be an ideal alternative Dalton, D. R., Certo, T. S., & Daily, C. M. (2003). Initial public offerings as
web of conflicts of interests: An empirical assessment. Business Ethics
to linear regression methods which are commonly imple- Quarterly, 13(3), 289e314.
mented to explain IPO returns. In addition, the predictive Desai, V. S., & Bharati, R. (1998). A comparison of linear regression and
accuracy of the random forest can potentially boost and neural network methods for predicting excess returns on large stocks.
facilitate the decision-making process for the IPO partici- Annals of Operations Research, 78(0), 127e163.
pants. In the IPO market where making decisions is probably Durukan, B. M. (2002). The relationship between IPO returns and factors
influencing IPO performance: Case of Istanbul stock exchange. Manage-
highly uncertain compared to other stock market activities, rial Finance, 28(2), 18e38.
IPO issuers, as well as investors, can potentially benefit from Durukan, B. M. (2006). IPO underpricing and ownership structure: Evidence
the use of the random forest algorithm to reduce the un- from Istanbul stock exchange. In G. N. Gregoriou (Ed.), Initial public
certainty regarding the IPO pricing and the post-market IPO offerings: An international perspective (pp. 263e278). Butterworth-Hei-
performance. nemann: Elsevier.
Ejaz, A., & Polak, P. (2015). Existence of short term momentum effect and
The variable importance of independent variables was also stock market of Turkey. Investment Management and Financial In-
studied using out-of-bag mean square error. The results reveal novations, 12(4), 9e15.
that the IPO proceeds and IPO volume to be the most Ekkayokkaya, M., & Pengniti, T. (2012). Governance and IPO underpricing.
important predictors of IPO initial returns. Market sentiments Journal of Corporate Finance, 18(2), 238e253.
return on assets and firm size were shown to have less Engelen, P. J., & Van Essen, M. (2010). Underpricing of IPOs: Firm-, issue-
and country-specific characteristics. Journal of Banking & Finance, 34(8),
explanatory power in predicting IPO initial returns. In general, 1958e1969.
the results show that the variables that act as potential proxies Ertuna, B., Ercan, M., & Akgiray, V. (2003). The effect of the issuer-
for ex-ante uncertainty are more important than variables that underwriter relationship on IPOs: The case of an emerging market. Jour-
are proxies for information asymmetry. nal of Entrepreneurial Finance and Business Ventures, 8(3), 43e55.
 _
B. Baba, G. Sevil / Borsa Istanbul Review 20-1 (2020) 13e23 23

Falconieri, S., Murphy, A., & Weaver, D. (2009). Underpricing and ex-post Marshal, B. B. (2004). The effect of firm financial characteristics and the
value uncertainty. Financial Management, 38(2), 285e300. availability of alternative finance on IPO underpricing. Journal of Eco-
Gao, Y. (2010). What comprises IPO initial returns: Evidence from the Chi- nomics and Finance, 28(1), 88e103.
nese market. Pacific-Basin Finance Journal, 18(1), 77e89. Mitsdorffer, R., & Diederich, J. (2008). Prediction of first-day returns of initial
Grinblatt, M., & Hwang, C. Y. (1989). Signaling and the pricing of new issues. public offering in the US stock market using extraction from support
The Journal of Finance, 44(2), 393e420. vector machines. In J. Diederich (Ed.), Rule extraction from support vector
Hanias, M., Thalassinos, E. L., & Curtis, P. (2012). Time series prediction with machines (Vol. 80, pp. 185e203). Berlin: Springer.
neural networks for the Athens Stock Exchange indicator. European Ozdemir, O., & Kizildag, M. (2017). Does franchising matter on IPO per-
Research Studies Journal, 15(2), 23e31. formance? An examination of underpricing post-IPO performance. Inter-
Hanley, K. W. (1993). The underpricing of initial public offerings and the national Journal of Contemporary Hospitality Management, 29(10),
partial adjustment phenomenon. Journal of Financial Economics, 34(2), 2535e2555.
231e250. Pande, A., & Vaidyanathan, R. (2009). Determinants of IPO underpricing in
Heiberger, R. M., & Becker, R. A. (1992). Design of an S function for robust the National Stock Exchange of India. ICFAI Journal of Applied Finance,
regression using iteratively reweighted least squares. Journal of Compu- 25(1), 84e107.
tational & Graphical Statistics, 1(3), 181e196. Peng, Y., & Wang, K. (2007). IPO underpricing and flotation methods in
How, J. C., Izan, H. Y., & Monroe, G. S. (1995). Differential information and Taiwan-a stochastic frontier approach. Applied Economics, 39(21),
the underpricing of initial public offerings: Australian evidence. Ac- 2785e2796.
counting and Finance, 35(1), 87e105. Quintana, D., Saez, Y., & Isasi, P. (2017). Random forest prediction of IPO
Huang, C., Chang, C., Kuo, L., Lin, B., Hsieh, T., & Chang, B. (2012). underpricing. Applied Sciences, 6(7).
A genetic-search model for first-day returns using IPO fundamentals. In Reber, B., Berry, B., & Toms, S. (2005). Predicting mispricing of initial public
International conference on machine learning and cybernetics (pp. offerings. Intelligent Systems in Accounting, Finance and Management,
1662e1667). Xian: IEEE. 13(1), 41e59.
Huang, S. Y., Lee, C.-H., Pan, L.-H., & Nguyen Thi, B. H. (2016). IPO initial excess Ritter, J. R. (1984). The “hot issue” market of 1980. Journal of Business,
return in an emerging market: Evidence from Vietnam's stock exchanges. 57(2), 215e240.
Review of Pacific Basin Financial Markets and Policies, 19(2), 1e23. Ritter, J. R., & Welch, I. (2002). A review of IPO activity, pricing, and allo-
Ibbotson, R. G. (1975). Price performance of common stock new issues. cations. The Journal of Finance, 57(4), 1795e1828.
Journal of Financial Economics, 2(3), 235e272. Robertson, S. J., Golden, B. L., Runger, G. C., & Wasil, E. E. (1998). Neural
Ibbotson, R. G., Sindelar, J. L., & Ritter, J. R. (1988). Initial public offerings. network models for initial public offerings. Neurocomputing, 18(3),
The Journal of Applied Corporate Finance, 1(2), 37e45. 165e182.
Ibbotson, R. G., Sindelar, J. L., & Ritter, J. R. (1994). The market's problems Rock, K. (1986). Why new issues are underpriced. Journal of Financial
with the pricing of initial public offerings. The Journal of Applied Economics, 15(2), 187e212.
Corporate Finance, 7(1), 66e74. Rousseeuw, P. J. (1984). Least median of squares regression. Journal of the
Jewartowski, T., & Lizinska, J. (2012). Short- and long-term performance of American Statistical Association, 79(388), 871e880.
Polish IPOs. Emerging Markets Finance and Trade, 48(2), 59e75. Satta, G. (2017). Initial public offerings in the port industry: Exploring the
Kiymaz, H. (2000). The initial and after-market performance of IPO in an determinants of underpricing. Maritime Policy & Management, 44(8),
emerging market: Evidence from Istanbul stock exchange. Journal of 1012e1033.
Multinational Financial Management, 10(2), 213e227. Stoll, H. R., & Curley, A. J. (1970). Small business and the new issues market
Leland, H. E., & Pyle, D. H. (1977). Informational asymmetries, financial for equities. Journal of Financial and Quantitative Analysis, 5(3),
structure, and financial intermediation. The Journal of Finance, 32(2), 309e322.
371e387. Thawornwong, S., Enke, D., & Dagli, C. (2003). Neural networks as a decision
Lin, C., & Hsu, S. (2008). Determinants of the initial IPO performance: Ev- maker for stock trading: A technical analysis approach. International
idence from Hong Kong and Taiwan. Applied Financial Economics, Journal of Smart Engineering System Design, 5(4), 313e325.
18(12), 955e963. Tian, L. (2011). Regulatory underpricing: Determinants of Chinese extreme
Ljungqvist, A. (2007). IPO underpricing. In E. B. Eckbo (Ed.), Handbook IPO returns. Journal of Empirical Finance, 18(1), 78e90.
of empirical corporate finance (Vol. 1, pp. 375e422). New York: Tiniç, S. M. (1988). Anatomy of initial public offerings of common stock. The
Elsevier. Journal of Finance, 43(4), 789e822.
Loughran, T., & Ritter, J. R. (2002). Why don't issuers get upset about leaving Tkac, M., & Verner, R. (2016). Artificial neural networks in business: Two
money on the table in IPOs? The Review of Financial Studies, 15, decades of research. Applied Soft Computing, 38, 788e804.
413e444. Vong, A. P., & Trigueiros, D. (2010). The short-run price performance of
Loughran, T., & Ritter, J. (2004). Why has IPO underpricing changed over initial public offerings in Hong Kong: New evidence. Global Finance
time? Financial Management, 33(3), 5e37. Journal, 21(3), 253e261.
Loughran, T., Ritter, R. J., & Rydqvist, K. (1994). Initial public offerings: Wadhwa, B. (2014). Insights into the IPO underpricing for listing on National
International insights. Pacific-Basin Finance Journal, 2(3), 165e199. stock exchange. Journal of Business Thought, 5, 38e58.
Lowry, M., & Shu, S. (2002). Litigation risk and IPO underpricing. Journal of Wang, D., Qian, X., Quek, C., Tan, A., Miao, C., Zhang, X., et al. (2018). An
Financial Economics, 65(3), 309e335. interpretable neural fuzzy inference for predictions of underpricing in
Luque, C., Quintana, D., & Isasi, P. (2012). Predicting IPO underpricing with initial public offerings (Vol. 30, pp. 102e117).
genetic algorithms. International Journal of Artificial Intelligence, 8(S12), Welch, I. (1989). Seasoned offerings, limitation cost and the underpricing of
133e146. initial public offerings. The Journal of Finance, 44(2), 421e449.
Maciel, L., & Ballini, R. (2010). Neural networks applied to stock market Yüksel, A., & Yüksel, A. (2006). The link between IPO underpricing and
forecasting: An empirical analysis. Journal of the Brazilian Neural trading volume: Evidence from the Istanbul Stock Exchange. Journal of
Network Society, 8(1), 3e22. Entrepreneurial Finance and Business Ventures, 11(3), 57e78.