Content uploaded by Mohammed Fourmou
Author content
All content in this area was uploaded by Mohammed Fourmou on Apr 21, 2024
Content may be subject to copyright.
Algorithmic high-frequency trading: A
systematic literature review
Abstract. The increase of computing power combined with
the advances in information technologies open the door to
new ways of trading, and created the opportunities for
developing new strategies. Transformations have been
observed through multiple layers, from investment decisions
to order execution. These new techniques relying on ultra-
low-latency reshaped the traditional market, and old
concepts have been taken to another level due to the
proliferation of algorithmic trading and electronic liquidity
provision strategies. Also, The application of AI to financial
investment is a research area that has attracted
extensive research attention since the 1990s, in portfolio
optimization, stock market prediction using AI, and financial
sentiment analysis. The relatively recent phenomenon of
high-frequency trading has had a profound impact on the
micro-structure of financial markets. Several authors hailed
it as a provider of liquidity and a mechanism for controlling
volatility, two highly welcome features, especially
beneficial to retail traders, whereas other authors view
the situation generated by algorithmic trading as damaging
for both small and institutional traders, and the orderly
functioning of the markets.
Keywords: algorithmic trading, high-frequency trading,
electronic market making, evolutionary computation, stock
trading rule, rule discovery, Classification framework,
computational finance, algotradings, artificial intelligence,
finance, Corporate insider trading, Forex Trading, Non-
Trader, Moving Averages, Error Detection, Time Series,
Volatility, Transaction costs, Liquidity, Bid-ask spread,
Financial crises.
By: Mohammed Fourmou
I. Introduction
Organised market system such as regulated markets and
MTFs using continuous double auction as a price
discovery mechanism allow participants to trade with each
other by placing their trade offer and trading demand as
market or limit orders, and where incoming orders are
continuously matched against an order book formed of two
queues of passive limit orders - one for buy (bid) and one
for sell orders (ask) - sorted by price and time priority, so
the trading process, and the collateral price evolution, can
be seen as an outcome of the interplay between order flow
and persistent order book liquidity.
According to Harris (2002), market participants such as
dealers and value traders are always passive liquidity
suppliers. A passive execution has the advantage of lower
cost of trading opposed to sending market orders due to the
bid-ask spread cost. Informed traders are strategically more
aggressive because of their need to fulfill large orders
before the order execution walks up the book. On the other
side, the uninformed dealers accumulate portfolio risk which
might lead to large losses in case the price move against
them, or being counterpart to informed traders due to the
unbalanced order flow - adverse selection or asymmetric
information risk - Dealers can protect themselves from
these risks only by quoting prices where the order flow is
two sided, this is consistent with updating the quotes in the
market direction. This leads to a conflict between
uninformed liquidity providers, which try to quickly detect
market and liquidity shifts, and informed traders which try to
hide their intentions during large order execution (“stealth
trading”).
For the latter reasons, and using the advances of
technology both sides upgraded their strategy
implementation from relying on human operators to
computer-based automated algorithms. The uninformed
liquidity providers have implemented ELPs, on the other
side, the informed traders rely on AT strategies.
Additionally, order anticipators have incorporated high-
frequency technology.
CBT strategies, and their impact on market quality -
captured along three distinct dimensions, i.e. liquidity,
price efficiency and systemic risk -.
Vuorenmaa (2012) reviews media and academic literature,
and discusses the pros and cons of HFT and AT. Media
focused only on the cons side, allocated most of its
publishing space to the Flash Crash of May 6, 2010 and
HFT portrayed as a predatory strategy taking advantage
of slower market participants or lure other traders into
taking toxic positions, as well as manipulative techniques,
such as quote stuffing, smoking, and spoofing. The
negative academic research results point to the
contribution to a higher market systemic risks and worse
contagion effects, and the positive one state that due to the
fast, predictive and accurate reactions of HFT strategies
they were able to decrease bid-ask spreads and
transaction costs, decrease volatility, increase liquidity,
and contribute more to the efficient price discovery.
II. Trading
1. Forex
Foreign exchange is a fast-paced, finance driven market
where trillions of dollars are traded per day and large
amounts of money can be made or lost. Just like getting
your holiday money, foreign exchange simply works by
buying one currency with another currency, but on a much
larger, global scale. Indeed, similar to getting your holiday
money, traders will wait to get the best deal. As prices are
changing every second of every day, traders use a variety
of mathematical tools, and a sprinkling of knowledge to
keep up with the fluctuations of these currencies. In doing
so, they hope that they will spot the exact opportunity to
make the most profit.
The London Stock Exchange, New York Stock Exchange,
Dow Jones et al all deal with companies that have ‘floated’.
To ‘float’ a company means that it changes from privately
owned to being publicly owned. Companies float on a stock
market as a way of raising capital. This floatation is in the
form of shares that people can purchase. Some people buy
these to become shareholders in the company; others,
traders, may use these shares to make a financial profit at a
later stage. Commodities are assets like gold and oil that,
like a company, can be traded on; prices move based on
supply and demand for each commodity, along with how
much is available. The Forex markets work on the same
principle except people trade in world currencies. Forex is a
global market of institutions trading the predominant
currencies of the world. The Market is open from Monday
morning (starting with Asian markets) and closes on
Friday evening (closing with the American markets). Trades
occur 24/7 throughout this time. Traders range from large,
multinational companies to individual investors. The
average daily turnover of trades in the market is averaged
at $5.1 trillion per day (for International Settlements, 2016).
In much the same way that positive/negative information
about a company will alter the price of their share value,
Forex prices are affected by factors such as a nation’s
economy, politics, weather or unforeseen events such as
COVID-19. Currencies are traded in pairs, the value given
with the pair is the cost to buy the base currency. For
example, GBP/USD 1.3505 (Pound Sterling/US Dollar)
indicates £1 would buy $1.3505. If the price is expected to
rise, this would be known as “going long”. To “go short”
would be to sell the Pound with the option to buy it back at a
lower price. The rise and fall of currencies in all markets is
undetermined and completely random. However,
experienced traders are able to spot repeatable price action
patterns and are able to place a trade with some degree of
certainty that the trade will end profitably. Not all traders are
the same, generally, there are four types: Position Trader,
Swing Trader, Day Trader, Scalper.
Most, if not all, methods used in Forex are time series
analysis. These are written to predict the future prices of
currency pairings. One of the earliest was SMA, a simple
averaging technique; this evolved to Weighted Moving
Average (WMA). Exponential Moving Average (EMA) is a
type of WMA that was merged to create Weighted
Exponential Moving Average (WEMA) (Hansun, 2013,
2014). Furthermore, Double Exponential Smoothing
became B-DES and was later merged by Hansun (2016) to
become B-WEMA. Along with the various Moving Average
methods that can be utilised, there is also a means of
comparing the predictions. Hansun (2017) details using
both Mean Square Error (MSE) and Mean Absolute
Percentage Error (MAPE) as a means of error detection
(ED) to test that WEMA and B-WEMA produce the same
accuracy. Mean Absolute Deviation (MAD) and Mean Root
Square Error (MRSE) (or RMSE) are other examples of
error detection. In more detail, error detection is a way
for the user to gauge the accuracy of moving average data
versus the actual data. Mathematically, the goal for error
detection equations is to be as close to zero as possible.
EUR/GBP is presented to 5 decimal places, therefore an
error detection of 0.00000 is the desired result for every
outcome. MAD (Mean Absolute Deviation) is derived from
the positive difference between the data point and the MA
(based on whichever is used); the results are then totalled
and divided by the number of calculations performed.
The performance indicator, when using MAD, is how close
the number is to zero (Chunhua et al., 2011).
● Machine learning and Forex
One type of algorithm used in Forex is SVM which uses
Statistical Learning Theory. This is a supervised learning
model of machine learning that contains a small sample to
research learning rules but also helps to reduce, or
eliminate, overfitting (Hui and Wu, 2012). In their research
(Hui and Wu, 2012), they tested SVM alongside SMA and
resistance/support, Kestner & trading range break filters.
This research was stock exchange based and an interesting
point in the paper was the volume of trades (28.6% in May
2011) that used some sort of computer-based program to
complete a transaction (Hui and Wu, 2012). The paper
highlights the need for “software and auxiliary function to be
used whilst trading”(Hui and Wu, 2012). A buy and sell (or
sell and buy) action is generally referred to as a trade.
Indeed “34% of the trading volume in the second quarter of
2002 comprises some form of computer-aided trading”
(Ellis and Parbery, 2005). In fact, SMA can be used to
support other classifications. Incorporating a neural network
to map news items where positive sentiment had a value of
1 and negative sentiment had a value of -2 (Lauren and
Harlili, 2014) were then able to use SMA 33 to prove that
positive news items had a positive growth on the JKSE
(Jakarta Stock Exchange). Much like Lauren and Harlili
(2014) converting news into a positive or negative number,
Baasher and Fakhr (2011) used a binary method, coupled
with SVM, RBF & MLP machine learning techniques, to
predict the direction of the high rate. (Each currency pairing
has four daily rates: Open, Closed, High & Low). The aim
was not to predict the price, it was to predict the correct
direction of the pairing. One type of Machine Learning is
Deep Learning, and within this field are Neural Networks.
These too can be used as a tool to predict Forex. For
example, Sespajayadi et al. (2015) combine genetic
algorithms with neural networks, with Root Mean Square
Error (RMSE) for error detection, to predict the EUR/USD.
Whilst many chose to predict just the close value,
Sespajayadi et al. (2015) decided upon predicting all
four movements, open, close, high & low. All gave an
RMSE of greater than 0.001.
2. Corporate insider trading
Corporate insider trading relates to the investment
behaviour of corporate employees in the securities of their
own company.
Manne (1966) initiated the debate when he suggested that
the agency problems facing managers and shareholders
would be mitigated if corporate insiders were allowed to
trade and benefit from their activities. This would lead to
improved corporate decision making, resulting in an
overall increase in the value of the firm [Jensen and
Meckling (1976)]. In addition, it has also been argued that
insider trading increases the informational effi efficiency
of markets by contributing to the existing information set
held by investors [Ross (1978); John and Mishra (1990);
John and Lang (1991); Zhang (2001), and Chau and
Vayanos (2008)]. Corporate insiders use their trades as
signals to confirm or contradict the information in public
corporate announcements. Investors view the dual
signals as complementary and act accordingly.
Another reason for allowing insider trading is provided by
Carlton and Fischel (1983) who contend that if insider
trading reduces the value of a firm, investors would demand
more stringent regulation than is currently imposed. They
also argue that there is no relationship between the value of
a firm and the level of corporate insider trading activity in
the firm. The trading of corporate insiders, it is argued, thus
enhances the informational efficiency of the market.
Empirical research is strongly supportive of this notion
with consensus findings of significant price changes
subsequent to insider trading, and in particular insider
buying activity.
An opposing view is the belief that corporate insider trading
harms investor confidence, which leads to a fall in liquidity
trading, thus resulting in a decrease in market efficiency
[Fishman and Hagerty (1992)]. The fall in investor trading
activity can be due to a number of factors. Insider trading
deters outside investors from paying to acquire
information from research and trading. Investors will hold
their orders until they are sure they will not be picked off by
other traders who are more informed than them. They may
also place too much weight on the trades undertaken by
corporate insiders. Even though the level of information is
greater in aggregate, the overall quality of trading is lower
because the presence of informed traders deters others
from participating. Another contributing factor to a loss of
market efficiency comes from the distribution of
information among traders. Corporate insider trading
means that information will be concentrated around
several few individuals, giving an informational advantage
to some traders at the expense of others. Bhattacharya and
Nicodano (2001) show that asymmetric information, by
which this concept is known, has a detrimental effect on the
enthusiasm of less informed investors, which leads to a
drop in their trading.
Scott (1980) and Manove (1989) have argued that insider
trading can also discourage corporate investment when
self-serving managers are allowed to profi t from a firm’s
changing fortunes. Corporate insiders are more likely to
choose riskier investments for their firm so as to benefit
from both increases and decreases in share value,
whichever state occurs. Shareholders who are cognisant
of this would be less supportive of capital expenditure,
causing firm level direct investment to fall below its
economically optimal level. The final argument relates to the
effect of an insider trade on the counterparty, and the
personal loss that is involved [Leland (1992)]. Even as
share prices may change to reflect the information
contained in an insider trade, the buyer of an insider sell
decision and the seller to an insider buy decision would
personally be affected by the trade. As the proportion of
insiders increase in the market, the number of outside
investors losing out will also increase, resulting in a drop in
overall investor confi dence. The aggregate effi ciency of
the market would fall because of the accumulation of
individual sentiments resulting from bad experiences in the
past. Corporate insider trading may also reduce the
importance of annual financial statements. Financial
statements should be the prime source of reliable financial
information regarding a company’s business activities.
While the information resulting from corporate insider
trading will not affect the reliability of financial statements, it
will reduce their relevance and timeliness [Cho and
Shaub (1991)].
The empirical evidence relating to the superior ability of
corporate insiders to detect mispricing in their company’s
shares strongly suggests that there are information
asymmetries in the market and that corporate insiders
benefit from their privileged position vis-à-vis the
information flow. Unfortunately, it is difficult to generalise
these findings to argue that the market is indeed less
efficient as a result of insider trading, and accordingly, there
is sparse empirical literature directly testing this hypothesis.
As yet, there is no consensus on whether corporate insider
trading is benefi cial to fi nancial markets or whether it
harms their viability. In practical terms, the two opposing
views take positions founded on economic theory (i.e.,
information quality) or legal theory (i.e., the equity and
fairness of markets). Clearly, this will infl uence the way in
which the argument develops regarding the benefits and
disadvantages of corporate insider trading.
● The performance of corporate insider trading
The initial research into corporate insider trading focused on
the US and investigated the existence of abnormal trading
performance. The first wave (Lorie and Neiderhoffer, 1968;
Pratt and DeVere, 1970; Jaffe, 1974a; Finnerty, 1976a)
simply measured the total return or risk-adjusted return on
insider trades for a period of several months. However, as
a consequence of using different insider trade definitions,
the results were mixed. Although the methodologies
employed in earlier research are unsophisticated in
comparison to similar, more recent work, much initial insight
was provided into the performance of insider trading.
Results from this earlier period strongly suggested that
insiders can detect and exploit mispricing in their own
company’s securities. Furthermore, the pioneering work of
early researchers into insider trading provided a
springboard for later research in the 80s to re-examine the
issue with more comprehensive data sets, better
methodologies and enhanced computing power. Empirical
research has also reported signifi can't abnormal returns
from the performance of insider trading in other
countries. In the UK, King and Roell (1989) examined a
very limited sample of insider trades and found signifi cant
abnormal returns from buying. Similarly, with a larger
sample over the same period, Pope et al. (1990) report
positive, but much smaller, abnormal returns from buying,
and negative abnormal returns from selling. Corporate
insiders in Hong Kong are able to earn abnormal returns
from both buying and selling (Cheuk et al., 2005). In Spain,
while insiders have earned abnormal returns, outsiders
who pursue a mimicking trading strategy are unable to
capture any of the benefits accrued by insiders (Del Brio et
al., 2002).
The second generation of empirical research into the
performance of insider trading is characterised by much
larger amounts of data and more focussed testing. Seyhun
(1986a) examined the performance of different
categories of insiders and found that directors were more
informed about a firm’s prospects than other insiders. In
addition, his results indicated, for the fi rst time, that fi rm
size and insider returns were negatively related. Jenter
(2005) showed that top managers have contrarian views
on firm value and their personal trading and corporate
decision-making jointly refl ect this perspective. Potentially
under-valued firms experience net insider buying and have
capital structures that refl ect such undervaluation. Rozeff
and Zaman (1988) decomposed the performance of
insider trading into information held by corporate insiders
and the characteristics of securities in which they traded.
After controlling for the size effect and earnings yield, they
found that, once transaction costs were taken into account,
insiders were unable to earn abnormal profits. Rozeff and
Zaman (1988) confirmed the propensity of small firm
insiders to buy and large firm insiders to sell. Taking into
account the concentration of insider trading performance in
smaller firms, the abnormal returns relative to a size-
adjusted benchmark disappear (Gregory et al., 1994).
Gregory et al. (1997) revisited this issue with a consistent
defi nition of an insider trading signal and a more
homogenous sample and found that firm size has a
substantial impact in UK insider trading. Consistent with
Barclay and Warner’s (1993) stealth trading hypothesis,
Friederich et al. (2002) report that medium size insider
trades are more informative over a short term. Garfinkel
and Nimalendran (2003) extend Barclay and Warner’s
(1993) stealth trading analysis to examine the impact of
trader anonymity. Insider trading activity is more transparent
on the NYSE specialist system compared to the NASDAQ
dealer system. They found that abnormal returns after
insider trading are positively related to bid-ask spreads,
which suggests that insider profits may be absorbed by
lower liquidity and transaction costs (Lin and Howe, 1990;
Fishe and Robe, 2004; Cheng et al., 2006). Staying with the
market microstructure theme, Chung and Charoenwong
(1998) report that market makers typically set bigger
spreads for large insider transactions. However, Dolgopolov
(2004) argued that the empirical relationship between bid-
ask spread and the risk of insider trading is inconsistent and
unreliable. Taking corporate governance structures into
account, Fidrmuc et al. (2006) found that the level of insider
shareholdings has a significant impact on insider
abnormal returns. Finally, Zhang et al. (2005) report a
negative relationship between insider trading and pay-
performance sensitivity. Most research in corporate insider
trading has centred on equity trading. When other routes to
trade exist, such as options and futures, insiders may
exploit this avenue instead. In derivative markets, Acharya
and Johnson (2007) find that insider trading in credit default
swaps does not affect prices or liquidity in either equity or
credit markets. Bettis et al. (2001) claimed that insiders
generally use derivatives (collar and swaps in their case) as
hedging instruments for risk reduction, resulting in an
improved alignment of incentives between managers and
shareholders.
● The timing of insider trading
A major focus of research concerns insider trading activity
around merger announcements. Keown and Pinkerton
(1981) examined insider trading prior to merger
announcements and found significant information leakage
in the run-up to the event. Elliot et al. (1984) extended the
work of Keown and Pinkerton (1981) and found that insiders
bought more shares and sold less, twelve months prior to a
merger. Seyhun (1990) also confirmed that insiders in
bidder firms traded more prior to the announcement of a
takeover bid. Furthermore, insiders’ trades were found to be
significantly correlated with the acquisition effect on bidding
firm value. In a case study analysis of the Nestle, S.A.
takeover of Carnation, Chakravarty and McConnell
(1997) examined the individual trades of Ivan Boesky in
Carnation equity. They reported a significant relationship
between Ivan Boesky’s transactions in Carnation and
changes in its share price. Their work has been extended
by Meulbroek and Hart (1997) who report that takeover
premiums on target securities are on average 10
percentage points higher when illegal insider trading has
taken place. It thus appears that illegal insider trading has a
significant impact on the prices of securities around
mergers and acquisitions; however this does not seem to
be the case where legal corporate insider trading occurs.
One reason for this is that corporate directors wish to avoid
possible investigation by the exchange authorities and
therefore trade in periods when there are no important
corporate events. Clearly, illegal insider trading ignores
this threat. Insider trading activity around earnings
announcements have also attracted a significant amount
of academic interest. Penman (1982) found that insiders
purchased shares before good management earnings
forecasts and sold prior to bad forecasts. However, Elliot et
al. (1984), Givoly and Palmon (1985) and Park et al. (1995),
report that insiders do not time their trades around
earnings announcements. In contrast, Allen and Ramanan
(1990), Sivakumar and Waymire (1994), Lustgarten and
Mande (1995) Udpa (1996), Ke et al. (2003), Huddart et al.
(2007) and Cheng and Leung (2008) find that insider
trading activity and earnings surprises are strongly related.
John and Lang (1991) noted the relationship between
insider trading and dividend changes, rather than earning
announcements. Dividend increases accompanied by
unusual insider buying signal good news, resulting in
positive abnormal returns. The converse is also true. Fuller
(2003) extends John and Lang (1991) to consider the
impact of various investors’ trading behaviour. Informed
trading moves share prices closer to their intrinsic value.
Therefore, the price reaction to a dividend increase is lower
when more informed trading takes place. Del Brio and
Miguel (2008) employed Hillier and Marshall’s (2002b)
methodology to classify informed and uninformed Spanish
insider trading. The information effect of the dividend
announcement is found to be explained by the signal from
insider trading activity around the announcement, rather
than the cash flow signal. Recently, Betzer and Theissen
(2007) analysed both the profitability of insider trading and
trading patterns around German earnings
announcements. They report that trading prior to earning
announcements have a larger impact on prices. Further,
ownership structure and accounting standards appear to
impact upon the magnitude of insider profits. In the
accounting literature, a number of studies consider the
relationship between accounting policy changes and insider
trading. The cases in point are Larcker et al. (1983) for
FASB n.o 19; King and O’Keefe (1986) for lobbying with
FASB; and Odaiyappa and Nainar (1992) for SFAS No.33.
Net insider selling is prevalent in the period preceding the
pronouncement of changing accounting policies. In
contrast to these studies, Cheng and Lo (2006)
demonstrate evidence that insiders exploit voluntary
disclosure opportunities for their personal gains in case of
insider purchases. They find that managers increase the
number of bad news forecasts when they plan to purchase
shares. A number of studies also consider insider trading
activity before accounting frauds were revealed.
Accounting frauds are usually accompanied by large falls
in stock prices, top-management turnover, and bankruptcy
filings (Agrawal and Cooper, 2007 and 2008; Desai et al.,
2006; Agrawal and Chadha, 2005; Palmrose et al., 2004).
There is also evidence of significant abnormal insider
selling during the period prior to frauds being exposed.
These results suggest that managers wished to offload their
stock at inflated prices (Agrawal and Cooper, 2008). More
recent insider trading and accounting information papers
have examined the relationship between corporate
decisions and insider trading behaviour. Darrough and
Rangan (2005) find that reductions in R&D expenditures
around IPOs are negatively related to insider selling. The
reduction in R&D expenditure increases current earnings
and so insiders sell to take advantage of the temporary
inflation of current earnings. Within the literature, insider
buys and sells are found to have differential information
signals. Insider sales have been consistently shown to have
little information about future firm performance for large
firms (Lakonishok and Lee, 2001; Jeng et al., 2003 and
Jenter, 2005). Lakonishok and Lee (2001) explained that
insider selling, motivated by private information, is
dominated by personal portfolio rebalancing for
diversification purposes. Cheng et al. (2007) argue that the
non-informative sale transactions are a result of poor
corporate governance. Insiders in the US can report private
transactions between executives and their firms after the
end of the fi scal year via a route called Form 5. Insider
sales reported through Form 5 were found to signal
negative future returns and lower future earnings relative
to analyst forecasts. Marin and Olivier (2008) meanwhile
argue that selling by insiders prior to market crashes occurs
heavily prior to the crash in the far past, and that the level of
insider selling close to the crash is much lower. Given the
possibility that insiders trade with superior information,
regulators have imposed restrictions on insiders to
prevent them from trading during firm specific
announcement periods. A number of studies examine the
effect of trading bans on insiders’ strategic trading
behaviour. Garfinkel (1997) examines the impact of the
Insider Trading and Securities Fraud Enforcement Act
(ITSFEA) of 1988. The timing of insider trades was affected
by the imposition of the new law. Insiders in the US now
tend to postpone their trades as a result of the new
regulation until after earnings announcements to avoid
investigation by the SEC. Insider trading is also suppressed
by company prohibitions in the same ways as by the SEC.
Bettis et al. (2000) found that insider trading corporate
policies are widespread in the US, with 78 percent of their
sample having explicit blackout periods. Roulston (2003)
found that those firms which restrict insider trading offer
their insiders more incentive based compensation. In the
UK, Hillier and Marshall (2002b) examined the effect of the
London Stock Exchange Model Code, which bans directors
from trading two months prior to their firm’s earnings
announcement. Insiders were found to decide their level of
stock trading based on the degree of mispricing in the
period immediately after the end of the trading ban.
Specifically, insiders buy after abnormally bad earnings
news and sell after good news. The remainder of the
literature into the timing of insider trading is mixed in its
conclusions. For example, with IPO lockup periods,
research considers the level of insider sales subsequent
to the IPO. Firms signal their quality to the market through
lockup agreements in order to achieve a higher IPO or
seasoned offering price. Brav and Gompers (2003) found
that the lockup agreement does not signal the true value of
the firm and insiders do not use the length of lockup to
signal higher firm quality. In contrast, Cao et al. (2004)
document high levels of insider trading after lockup
expiration. With seasoned equity offerings, Karpoff and Lee
(1991), Eyssell and Reburn (1993), Kahle (2000) and
Clarke et al. (2001) reported abnormal increases in
insider sales prior to the announcement. Insiders may
therefore exploit this window of opportunity to issue
overvalued equity. Furthermore, the long-run performance
of seasoned offerings is significantly related to the level of
abnormal insider trading after seasoned equity offerings. In
contrast, Lee (1997) found long-run stock returns for
seasoned equity offerings where insiders sell shares prior to
the announcement are not significantly different from
those where insiders buy shares Hauser et al. (2003)
hypothesized that the insider trading pattern around
seasoned equity offerings depends on the firm’s ownership
structure. Insiders with a concentrated stake take a long-
term view and buy shares prior to the announcement to
preserve their control over the firm. Finally, evidence
relating to insider trading and bankruptcy is mixed.
Gosnell et al. (1992) found that insiders sell prior to
bankruptcy while Loderer and Sheehan (1989) did not find
any evidence of selling. The discrepancies between these
results are indicative of sample differences that may be
related to institutional or time specific factors.
● Insider trading regulation
Regulators have implemented country specific codes of
best practice on insider trading through mandatory public
regulations or voluntary corporate policies. Insider trading
regulations and laws have been enacted to control or
reduce the possibility of insiders trading on price sensitive
private information. With this, the definition of an insider has
become clearer whereby insiders may be corporate
employees or any person who has privileged access to a
firm’s non-public information. This latter group may include
large shareholders, financial consultants, bankers,
auditors, lawyers and other related parties. A number of
researchers (Hebner and Kato, 1997; Goshen and
Parchomovsky, 2001; Roulstone, 2003) suggest that
security analysts should be allowed to trade on inside
information, since such activity will inevitably create
more liquidity and competition in securities markets.
However, numerous papers advocate the implementation of
insider trading regulations because ordinary investors are
disadvantaged when trading with market professionals (See
Haddock and Macey, 1987 and Bushman et al., 2005). In
the US, the Securities and Exchange Commission (SEC)
has prohibited fraud and market manipulation as well as
insider trading since 1934. According to the SEC Act
section 10(b), insiders are corporate insiders, or anyone
who obtains material, non-public information from a
corporate insider or the issuer, or who steals the information
from another source. The main responsibility of the SEC is
to enforce insider trading regulations. Against a defendant,
the SEC may bring civil charges, refer the case to the
Justice Department for criminal prosecution, or suspend
the professional license. Subsequent Acts have made
the penalties for insider trading stronger. The 1984 Insider
Trading Sanctions Act (ITSA) covers derivatives trading and
allows for both civil and criminal charges; the 1988 Insider
Trading and Securities Fraud Enforcement Act (ITSFEA)
increased the criminal fines and the maximum jail term to
both the firm and its employees; the 1990 Securities
Enforcement Remedies Act and the 2000 Financial
Disclosure Regulation Act ban selective disclosure of
corporate information to large shareholders and analysts.
Many other countries have similar regulations to that of
the US. However, different governments and regulators
apply their own definition of insider trading. One clear
distinction that occurs is the definition of what constitutes
inside information. For most countries, the relevant
definition of insider information is materiality and the
potential impact of the information on security prices. In the
UK, for example, insider trading regulation defi nes
materiality in that if a specific piece of information is made
public, it would be likely to have a signifi cant effect on the
price of a security. German law meanwhile defines inside
information as, «the knowledge of a fact not publicly known
relating to one or more issuers of insider securities and
which fact is capable of substantially influencing the price of
the insider securities in the event of it becoming publicly
known». Further, the definition of an insider is broader and
more general in many countries. Non-US countries do not
follow the US fiduciary relation to define illegal insider
trading and stock tipping. UK statute specifies that an
insider may be anyone who possesses non-public
information from any source, whereas in the US the insider
must have some formal or informal connection to the
company. As a result, the UK definition is argued to be
superior in form to that defined by US law (Tridimas,
1991; Watson, 1995). Nevertheless, the vast majority of
all empirical research in the UK and the US has dealt with
corporate insiders who are employed by a firm and who
may or may not have fiduciary responsibilities to the
shareholders of the firm. Prior to the introduction of the
Insider Trading Directive in 2003 by the European
Community, insider trading was regulated by country
specific mandate by each member state. Insider dealing
(that is, trading on private, specific, and precise information
that is likely to have a material impact on prices) is illegal
throughout European countries. The European Community
posits that insider dealing undermines investor
confidence, leading to suboptimal security markets.
Supervisory authorities in European countries have also
strengthened their existing insider trading regulations (e.g.
Denmark, Greece, France, Luxembourg, Portugal, the
Netherlands, Spain and the United Kingdom) or initiated
new insider trading rules (e.g. Belgium, Germany, Ireland
and Italy) in order to protect investors from insider
dealing and to help with the internationalisation of their
securities markets. In the European Union, supervisory
authorities coordinate their efforts between national
regulators as a result of the European Economic
Community Directive (EECD) on Insider Trading and the
Council of Europe’s Convention on Insider Trading. The
Directive, based on the Single European Act Article 100a,
was designed on the French and English insider trading
regulations. The EECD imposed the first community
prohibition of insider trading, through the 1989 Directive
154, which has subsequently been replaced by Directive
155 in 2003. Directive 155 requires all member states to
develop and enforce national laws on insider trading. In
addition to Directive 155, the EECD also imposed the
definition of and the prohibition from market
manipulation in Directive 156 and Directive 157,
respectively. Moreover, inside information is clearly
specified in both the 1989 and 2003 Directives, unlike in the
US regulations. In defining insider dealing, the EECD
Directives applies the possession of non-public
information, whereas the US law is concerned with
breaches in fiduciary duty. In the same way as US law,
Directive 161-Directive 167 justifies the rationale for the
prohibition of insider trading to ensure investor confidence
and to maintain market efficiency. However, unlike the US,
all Member States did not classify insider dealing as a
criminal offence at the time of adoption of Directive.
Directive 168 prohibits insider dealing more widely to
include not only persons who acquired inside information
due to their position as a director, manager, employee or
majority shareholder, but also those who acquired the
information illegally.
As markets become ever more integrated, insider trading
can cross national borders and international cooperation
is necessary for successful investigations and prosecutions
to take place. To this end, the US has agreements (the
Mutual Legal Assistance Treaties in Criminal Matters
and Memoranda of Understanding) with other countries to
share information and cooperate in the investigation and
prosecution of security law violations. Likewise, the EU
Directive 156 requires members to cooperate whenever
necessary for the purpose of implementing the Directive. In
general, the definitions of insider trading are much broader
in the EU Directive than in the US laws where the focus is
on the narrower breach of fiduciary duty definition.
However, unlike US law, the Directive does not propose
specific penalties for violating insider trading laws.
Governments and regulators in each country decide on the
appropriate penalties for violating the Directive based on
their own legal regime. Penalties and enforcement are
therefore dependent, not only on the will of regulators, but
also upon country specific factors such as culture and the
legal environment. This approach leads to very different
legal consequences for insider trading. In the UK for
example, The Financial Services Authority (FSA) states that
there are unlimited civil fines for insider trading, whereas
the Dutch regulatory authority, Autoriteit Financiële Markten
(AFM), has only imposed criminal penalties for insider
dealing. The findings of empirical studies on the
effectiveness of insider dealing prohibition and
enforcement are mixed. In most countries, researchers
have found that regulation is generally weak in preventing
insider dealing (Bhattacharya and Daouk, 2002; Bris, 2005).
Moreover, the stricter enforcement of insider trading
regulations does not directly lead to more participants in
capital markets. As such this casts doubt on the premise
that fairer markets will have higher market participation and
also raises questions about the value of insider trading law.
In contrast, a number of studies find evidence of a positive
relation between capital markets and insider trading law
enforcement. Bushman et al. (2004) find a positive
relationship between insider trading enforcement,
corporate disclosure, institutional investment, and media
penetration. In a later study Bushman et al. (2005) also
documents increased levels of analyst coverage after the
initial enforcement of insider trading law.
3. Algorithmic trading
In their definitions of AT, regulators underline the
automated and computer-based decision process – with no
human intervention – of determining the individual order
trading parameters regarding timing, pricing and quantity
setting, as well as the managing of orders after their
submission. As expressed in Kissell and Malamut (2005), a
trader faces the trade-off/dilemma of trading too quickly
(aggressively) and trading too slowly (passively). At one
end, an order can be executed at once - using market order
- with a high trading cost, at the other end, it can be equally
split and scheduled at a constant execution rate over the
entire trading period, the optimal scheduling lies in-between
this range bounded by the minimum variance strategy, at
one side, and the minimum impact strategy, at the other
side.
A formalization of the order scheduling problem is provided
in Almgren and Chriss (2001).
Johnson (2010) provides a classification of AT based on
its target objectives into impact-driven (TWAP, VWAP,
POV) strategies, cost-centric (Implementation Shortfall),
and opportunistic algorithms (Price Inline).
Fabozzi, Focardi, Kolm et al. (2010) identify IS and VWAP
as the two most popular execution strategies.
1. VWAP
VWAP is an ideal option for passive traders with no alpha
and no need for urgency. The actual implementation
consists in splitting the initial order over the entire trading
period based on a model of the historical fractional daily
volume pattern. The VWAP can be improved by replacing
the standard historical pattern with other models of intraday
volume dynamics. E.g., Bialkowski, Darolles and Le Fol
(2008) suggest a dynamical volume model, which
decomposes the traded volume in two parts: one reflecting
the common and seasonal market evolution – modeled by
an extension of CAPM with factors estimated by principal
component analysis – and a second one capturing the
intraday specific volume dynamics by means of an
ARMA(1,1) or a SETAR model.
2. Implementation Shortfall
According to Fabozzi et al. (2010), this strategy is especially
appropriate for market participants who know their risk-
aversion profile and who have a strong belief about the
future returns. IS is benchmarked to the arrival price and
is optimized towards minimizing the overall potential risk-
adjusted costs with respect to a predefined coefficient of
risk aversion. A model for the trading schedule x(t), similar
to the one proposed by Almgren and Lorenz (2006), can be
assumed to be the sum of two distinct trajectories: (i) a
linear execution (neutral/ TWAP) and (ii) a deviation – a
quadratic function with the roots at t=0 and t=T, which
corresponds to a given “speed of execution” κ optimized
by the proprietary trading strategy. Under certain
assumptions, e.g. Glosten-Milgrom-Harris framework
(Glosten and Milgrom (1985), Glosten and Harris (1988))
and a model for the temporary market impact, an
analytically derivation of the optimal function κ’ is possible.
3. Application of Evolutionary Computation for Rule
Discovery
Generally, AT refers to the use of sophisticated
computer algorithms to automatically make certain trading
decisions in the trading cycle, including pre trade analysis
(data analysis), trading signal generation (buying and
selling recommendations), and trade execution (order
management). One of the advantages of AT is the
effectiveness and efficiency of machine learning techniques
in financial big data analysis. However, some AT learning
models are considered as “black boxes” because they
involve difficulty in providing easy-to-understand
explanations on the interactions between the model inputs
and the outputs. Trading with black boxes makes
investors uncomfortable and elicits mistrust in the model. To
address this issue, an increasing number of researchers
have investigated rule discovery techniques for finding
explicit trading rules that can provide explicit knowledge
to guide trading. Thus, investors can justify system
decisions using their domain knowledge, and potential
investment risks can be reduced, facilitating the discovery
of new knowledge and integrating new and old
knowledge, reducing errors derived from noise, feature
subset selection, or inaccurate parameter settings.
Evolutionary computation (EC) has been widely employed
in rule discovery. EC is generally defined as a computing
tool to solve realistic problems by simulating the
evolutionary mechanisms of nature. It is mainly based on a
population, uses probabilistic transition rules, and directly
applies the objectives from the user as “fitness”.
Over the past few years, several literature reviews have
been conducted on stock prediction models. Atsalakis and
Valavanis reviewed the application of neural and neural-
fuzzy techniques in stock prediction. Guresen et al.
presented a comparative survey of different neural network
models in NASDAQ Stock Exchange index prediction.
Bahrammirzaee provided a comparative analysis among
artificial neural networks, expert systems and hybrid
intelligent systems in credit evaluation, portfolio
management and financial prediction and planning.
● Classification of analysis method
All stock analysis approaches can be classified into
fundamental, technical, or blending analysis.
Fundamental analysis. This type of analysis method
is based on the assumption that the internal value of each
stock is determined by its potential profitability. It is mainly
based on three essential aspects, which are described as
follows:
i. Macroeconomic analysis, which analyzes the effect of the
macroeconomic environment on the future profit of a
company. The popular indicators include GDP, CPI, M1B,
etc.
ii. Industry analysis, which estimates the value of the
company based on industry status and prospect, such as by
analyzing the billings (or revenues) of upper stream entities
in an industry.
iii. Company analysis, which analyzes the current operation
status of a company to evaluate its internal value, mostly
by examining the company financial reports.
Technical analysis. This analysis method considers
the movement of stock price and volume as reflections of all
the related information about the stock market, in
addition, previous market behavior patterns repeat in the
future, thus by analyzing the previous behavior patterns of
price and volume, trading rules can be generated. Unlike
fundamental analysis, technical analysis lacks a consistent
taxonomy. Therefore, by referencing Bodie et al., CFA
Institute, Colby, Fidelity Mutual Fund, Goldman Sachs,
NASDAQ, Pring, Wikipedia, Yahoo Finance, and Market
Technicians Association, we found that domains of
technical analysis can be grouped into rational classification
scheme as follows:
i. Sentiment, which mainly represents the behaviors of
various market participants. The analysis of these indicators
is grounded on the hypothesis that different types of
investors show different behaviors at the main market
turning points. Indicators such as expert-public ratio,
consulting services sentiment indicator, short interest ratio
and put options to call options.
ii. Flow-of-funds, which is a type of indicator used to
investigate the financial status of various investors to pre-
evaluate their strength in terms of buying and selling
stocks, then, corresponding strategies, such as short
squeeze, can be adopted. The analyzable data for this
strategy contains the capital of mutual funds or large
investors, and the events like new and additional issues.
iii. Raw data, which include stock price series as well as
price patterns such as K-line diagrams and bar charts. The
former is commonly used for time series analysis or trend
judgment combined with other indicators, the latter
generally suggests price patterns can reflect the changes of
market sentiment which affect short movements of stock.
iv. Trend, which is a type of price-based indicator for tracing
the stock price trends. The corresponding strategy, called
trend following strategy, proposes that economic and
political events usually change market prices through
changing market trends rather than by instantly returning to
the most rational point. Thus, investors can gain profit by
tracing the occurrence of price trends. Common trend
indicators include SMA and EMA.
v. Momentum, which is also a kind of price-based
indicator but is used to evaluate the velocity of price change
and judge whether a trend reversal in stock price is about to
occur. The momentum indicators are analyzed based on
the hypothesis that stock prices undergo a nominal cycle,
and the cycle will manifest as price rebound and callback
trend. Such indicators include RSI, MACD, and ROC etc.
vi. Volume. Volume-based indicators reflect the enthusiasm
for investing of both buyers and sellers, which is also a
basis for predicting stock price movements. Strategy and
applying volume indicators is grounded on the hypothesis
that price movement is determined by the enthusiasm of
buyers and sellers. Transactions are commonly suggested
to expand during rising tendencies and shrink during
downward tendencies. Popular indicators include volume,
volume ratio, and OBV.
vii. Cycle. According to cycle theories, stock prices vary
periodically, a long cycle may last more than ten years and
contain various short cycles that can be as short as a few
days or weeks. Strategies that use cycle indicators aim to
analyze the position of the current stock price in the cycle.
For example, Elliott wave and seasonal patterns are often
used to examine the cyclical variations of stock prices.
viii. Volatility, which is commonly used to investigate the
fluctuation range of stock prices. It can be used to evaluate
the risk and identify the level of the support and
resistance. Stock prices are generally recognized to
fluctuate between the level of support and resistance, but
continue to rise (fall) once they break through the level of
resistance (support). Common volatility indicators include
average true range and Bollinger band.
Blending analysis (fundamental and technical analysis). The
analysis combining fundamental analysis and technical
analysis will be regarded as the blending method.
● Classification of EC technique
Evolutionary algorithm (EA). EAs were proposed in the late
1950s and the early 1960s. Initially, the representative
algorithms were EP, ES and GA. EA represents
optimization algorithms that search the space by simulating
the genetic evolutionary process of Darwin's theory
including selection, mutation, recombination and
reproduction, it also uses fitness functions as performance
measures to drive the evolution towards better regions of
the search space. In present, the main subfields of EAs
include EP, ES, GA, GP, LCS etc.
Swarm intelligence (SI). SI, first proposed by Beni and
Wang, primitively describes a paradigm about Cellular
Robotic Systems. Afterwards, Bonabeau et al. expanded
this definition to “algorithms or distributed problem solving
devices inspired by the collective behavior of insect
colonies and other animal societies.” The latter definition is
widely accepted. In present, the main subfields of SI include
PSO and ACO.
Hybrid EC techniques (evolutionary algorithm and swarm
intelligence). Algorithms that combine EA and SI will be
regarded as hybrid EC techniques.
III. High-frequency trading
High-frequency trading is not a strategy per se, but a
technology which allows for the automation of a wide
spectrum of trading strategies, propelled by the ongoing
advances in computer technology. Researchers (see, e.g.,
Aldridge (2009), Brogaard (2010)) as well as regulators
across the world have tried to specify what are the
attributes of HFT-based strategies. Their findings could
be summarised as follows:
1. Computer-based non-discretionary, automated
strategies (no human intervention in order
initiation, generation, routing or execution)
2. Proprietary trading (as opposed to agency
activity)
3. use of low latency HFT technologies (e.g., co-
location services, proximity hosting, direct market
access, individual data feeds offered by exchanges)
4. real-time tick-by-tick data processing
5. high amount of intraday messages (orders,
quotes or cancellations) and trades
6. small margins per trade (“scalping”)
7. high capital turnover
8. flat overnight positions (no positions or fully
hedged)
According to EUREX, the main HFT-based strategies
are liquidity provision, (statistical) arbitrage, short term
momentum and liquidity detection.
● Volatility
Volatility is one of the most critical market parameters,
especially when HFT is involved, as suspicions have not yet
vanished about the role and responsibility of ultra-fast
traders in the May 2010 Flash Crash and in the supposedly
numerous mini-flash crashes.
By modelling a continuous double auction, a widely used
structure shown by nearly all exchanges, an agent-based
simulation produced by Myers and Gerig (2014) shows a
drop in volatility. Using the same approach, Leal et al.
(2014) instead identify higher volatility and a fundamental
HFT role in market flash crashes, mostly driven by a
high-order cancellation rate, although the authors also
admit a role in the fast market recovery that usually follows
such extreme events. The test of algorithmic impact on
volatility in a modelled market made up of informed,
momentum and noise traders, leads Gsell (2008) to
conclude that lower latency yields a statistically
significant reduction in volatility. Linton and
Mahmoodzadeh (2018), who investigate whether it is HFT
that causes volatility or rather is the latter attracting the
former with the promise of ripe profits, reach no conclusive
result, whereas Verousis et al. (2018) give more relevance
to the tick size, rather than HFT, as a volatility-driving factor.
Aldridge (2014) reaches a similarly doubtful conclusion, by
high-lighting that not all phenomena of high volatility give
rise to a crash and that distinction between a crash and
strong noise is sometimes nuanced. The financial firm
Nanex produces data and analyzes widely used by
academics and practitioners to investigate the impact of
HFT activities on markets, and the potential risk they create.
Among the researchers who made use of Nanex data,
Zervoudakis et al. (2012) analyze whether the practice of
high-frequency quoting lowers volatility, finding “no direct
and unambiguous evidence of causality between HFT and
increased volatility” (ibid. p. 7). They also argue that “[i]f
HFT contributed to volatility, then HFT diffusion should
have increased the intra-day-to-overnight volatility ratio;
but this correlation is not evident”. A note by Credit Suisse
(Chaparro 2017a) attributes to HFT the phenomenon of
‘flickering quotes’, rapid price bounces mainly experienced
by large capitalization stocks, especially around the end of
the day, when most HF traders rush to close off their
positions. Another study (Bollen and Whaley 2015) based
on electronic trade-by-trade data provided by the
InterContinental Exchange, Eurex, NYSE Euronext, and
the CME Group, yields similar results for the futures market.
Kelejian and Mukerji (2016) use a sample of end-of-day
prices from the 200 most traded S&P 500 stocks for the
period 1 May 1985 through 31 March 2012, finding that
“HFT has had a significant effect on daily asset price
volatilities [...and] the relationship between stock
fundamentals and volatility has weakened since the advent
of HFT”. The findings of Jarrow and Protter (2012), who
investigate the exploitation of arbitrage opportunities by
ultra-fast traders, lean towards increased volatility and in
general a dysfunctional role of HFT, as HF activity rises.
Zhang (2010) uses exogenous shocks of NYSE autoquote
to HFT and, after taking into account the firm's
fundamentals and other volatility drivers, finds a positive
correlation between HFT and volatility, especially for
large capitalization stocks. An institutional research
(Caivano 2015) finds that an increase by one standard
deviation in HFT activity raises volatility between 0.5 and
0.8 standard deviations, a higher impact than any other
control variable affecting volatility. Aldridge and Krawciw
(2015) find a correlation between aggressive HFT activity
and volatility, although there is uncertainty about the causal
relationship: “[i]t is not immediately clear [...] whether
aggressive HF traders seek out high volatility, whether
aggressive HFT participation induces higher volatility in
stocks, or both”.
Another issue raised in the literature is whether trading
strategies similar to each other may lead to crowding
effects, with the consequence of exacerbating price
movements. Search for speed leads to simplification of
algorithms and therefore to follow simple, common
strategies. This phenomenon is investigated by Brogaard
(2010), who interprets the results found as a confirmation
that HF traders engage in less diverse strategies than non-
HF traders. The paper recognizes HFT activity as a
potential volatility factor because of the herding effect. The
underlying logic is that several algorithms following similar
strategies cumulate in pushing prices abnormally, and a
significant correlation between HFT strategies is confirmed
by Chaboud et al. (2014). Also according to Kirilenko et al.
(2017), HF traders seem to have common strategies: they
are found to follow the trend in the 4s after a stock price
showing a clear direction and going contrarian after 10s.
Friederich and Payne (2012) acknowledge that strategy
crowding does not seem an unusual occurrence, whereas
Farmer and Skouras (2013) report wide perception by
market participants of ‘crowdedness of computer trading’,
although presenting no evidence.
One important point about HFT is whether it behaves the
same under different market conditions. Hasbrouck and
Saar (2013) observe that more low-latency activity implies
lower short-term volatility, and this effect is noticeable under
quiet and stressed market conditions. Analyzing the 30
most traded stocks on the Swedish index OMXS30 on the
NASDAQ Stockholm exchange Hagströmer and Nordén
(2013) suggest that HFT, whether following market making
or opportunistic strategies, mitigates intraday price volatility.
This result is consistent in both highly volatile months (as
shown by August 2011 data) and much less volatile ones
(February 2012 data). As a counter-proof, the same
research investigates the opposite hypothesis, verifying that
“a decrease in trading activities of the opportunistic HF
trader causes an increase in stock return volatility”.
Brogaard (2010) only finds “not strong” evidence of HFT
having an impact on reducing volatility. The same paper
also finds little change in HFT activity under extreme
conditions. The same weak or lack of evidence is
empirically found by Groth (2011) about HFT withdrawal
during periods of high volatility on the Xetra platform at
the Frankfurt Stock Exchange. Zingrand et al. (2012) also
fail to find direct evidence of a positive impact of HFT on
volatility, but they argue that, under some circumstances,
computer-based trading can trigger self-reinforcing
feedback loops, potentially increasing volatility. This topic is
further developed by Abrol et al. (2016), who admit that
positive feedback loops can amplify shocks and pose
systemic risk, since events occur well beyond human
reaction time. HFT as a source of systemic risk is the thesis
of Jain et al. (2016) who, by measuring order flow,
correlation and CoVAR, evaluate trading risk as volatility
shock propagation. An agent-based model described in
Virgilio (2016) shows that, in the presence of both HF and
LF traders, markets are more prone to exacerbate volatility
when it is already high than when it is low. The simulation
creates a fast market in which quotes are posted and trades
are executed as soon as they arrive at the exchange server.
However, LF orders, whether passive or aggressive, by
definition experience a delay with respect to HF orders,
before reaching the exchange. This causes quotes and
trades being served at prices potentially different from the
one originally intended. In times of quiet markets, this effect
is scarcely noticeable, but in the presence of even
moderate volatility, the effect gets amplified. A similar
conclusion is reached by a research at Bundesbank which,
in its October 2016 Monthly Report (Bundesbank 2016),
explains that, during periods of heightened volatility, fast
market takers augment their activities whereas passive HF
traders delete their limit orders to avoid being picked off, as
price uncertainty rises. The combination of these two
factors further increased volatility and risk of market
turmoil.
● Transaction costs
The following sections address the transaction costs in the
form of monitoring costs, cost-minimizing sequence of
trades, fees and commissions, implementation shortfall,
systemic risk, front-running and withdrawal in troubled
times. Because of its peculiarities, the share part of
transaction costs represented by bid-ask spread, will be
addressed in a subsection of its own.
As far as transaction costs are concerned, many
researchers claim findings which tend to put HFT in a rather
favourable light, albeit with some notable differences.
Foucault et al. (2013) develop a theoretical model
describing the interaction between market makers and
market takers and find that algorithmic trading directly
increases the trading rate via a reduction in monitoring
costs. Kirilenko and Lo (2013) use stochastic dynamic
programming evaluating the expected cost-minimizing
sequence of trades to dispel any doubt that algorithmic
trading yields “tremendous cost savings, operating
efficiency and scalability in every financial market it
touches”, a result confirmed by Harris (2013), who states a
substantial decrease in transaction costs for both
institutional and retail investors. Conrad et al. (2015)
analyze 2009 Trade And Quote (TAQ) data stamped at
second granularity, and 2010–2011 TAQ National Best Bid
and Offer (NBBO) data stamped at millisecond, finding that,
on average, high-frequency activity significantly lowers
trading costs. A similar conclusion is reached by
Anagnostidis and Fontaine (2018), who analyze order flow
and the liquidity process in the Paris CAC 40 market and
notice how transaction costs diminish. Baron et al. (2017)
provide experimental evidence that HFT activities reduce
indirect costs for LFTs. More cautious but still pointing to
the same direction is Menkveld (2013). Investigating the
impact on trading fees by new high-tech entrant markets
specifically suited for HFT, like BATS in the US and Chi-X in
Europe, the study finds large cost reductions, even though
competition among exchanges seems to play a crucial role
in this case. Another study by the same author, Menkveld
(2016), confirms the general opinion, by bringing new
evidence and comparing both explicit and implicit
transaction costs as reported by various researchers.
Stocks listed at NASDAQ and NYSE in 2001 (presumably
with low or no HFT activity) and 2011 (certainly including
a lot of HFT) show at least a 50% decline in all of the four
following aspects analyzed: effective spread, commissions
to retail investors, to institutional investors and
implementation shortfall. Also positive about trading cost
reduction is Harris (2013), and Brogaard (2010) goes as far
as to support the much criticized habit of HFT to post and
quickly cancel quotes by stating evidence of “net
economically significant benefits by reducing the non-
execution cost that would otherwise occur”.
Bid-ask spread is a primary component of transaction costs,
and therefore, it deserves to be discussed in a separate
subsection. Indeed, a major focus of market quality is bid-
ask spread, that is, the cost traders incur in buying at the
ask and selling at the bid (and correspondingly the profit
market makers earn). By focusing on the effective bid-ask
spread costs in the Ancerno data set, Brogaard et al.
(2014a, b) find no measurable effect by the increase in HFT
activity on execution costs for institutional investors. The
impact of HFT on market quality is the topic investigated by
Brogaard (2010), who tests whether HF traders flee in
volatile markets by analyzing their activity as volatility
increases and during varying degrees of 15-min period
price changes. The paper reports that HF traders often
book the best bid and ask, reducing the spread, especially
for larger firm size. Jarrow and Protter (2012) adopt a
theoretical model by using two equations, one representing
the stock price process in the absence of HFT and the other
one in the presence of it. The research somehow
cautiously states ‘preliminary evidence’, suggesting that
HFTs narrow spreads, even if it affirms no final verdict on
this matter. A similar opinion is displayed by Menkveld
(2013), according to which stocks affected by HFT activity
experienced a bid-ask spread reduction around 30% in a
year with respect to other stocks. A clear downward trend in
bid-ask spread is the result of Friederich and Payne
(2011), who plot London Stock Exchange (LSE) best
spreads and book depth for the FTSE-100 stocks between
January 2009 and April 2011. The conclusion is that the
authors “do not see which forces other than the growth of
CBT could explain these trends”, where CBT stands for
computer-based trading, a prerequisite of HFT. A critical
aspect is investigated by Menkveld and Zoican (2013), who
develop a model with some HF traders providing liquidity
and some consuming it. They find that spreads fall when
limit orders execute at higher speed than market orders.
Formal statistical hypotheses are tested by the authors,
in particular whether latency reduction leads to an
increase in bid-ask spread. The paper regresses the
adverse selection component of the spread aggregated
across stocks and finds that a drop in market latency has a
positive significant effect of about 7%. Zervoudakis et al.
(2012) follow the main stream of thought by arguing that
“HFT systems reduce [...] spreads by allocating liquidity”.
By studying cross-sectional determinants of HFT
participation in LSE and Euronext Paris between 2001 and
2011, Aitken et al. (2012) find that a higher level of HFT
activity tends to reduce bid-ask spreads. Hasbrouck and
Saar (2013) use ordinary NASDAQ book order data for Q4
2007 and June 2008 to estimate regression coefficients and
conclude that “higher low-latency activity implies lower
posted and effective spreads”. So, most researchers
tend to align themselves on the opinion of HFT improving
market quality, via spread reduction. However, a different
result is reached by Hendershott and Mouton (2011) who
analyze data from NYSE and CBOE in the year around
hybrid activation running from 1 June 2006, through 31 May
2007. The research finds that the change, which reduced
execution time from about 10 s to less than one, increased
the bid-ask spread because of the increase in adverse
selection. This result is confirmed by Brogaard (2010),
who looked at trading statistics of a HFT sample on
NASDAQ and NYSE and found that HFT tends to trade
aggressively in securities with lower spreads, so widening it.
Although not unanimous, academic research shows a clear
propensity towards a positive HFT impact on reducing bid-
ask spreads.
Ding et al. (2014) are rather negative about the impact of
HFT on general transaction costs. They find several price
dislocations between different NBBO data feeds every
second and such dislocations last no more than 2 ms,
causing extra costs for slow traders frequently active on the
market, like institutional investors. A similar opinion is
expressed by Hoffmann (2014), who notices how slow
traders submit limit orders which have lower execution
probability because of the presence of fast traders, and
“because speed is a source of fast traders to extract rents
from other market participants” (ibid. p. 156). Hirschey
(2018) positively tests the hypothesis that HF traders
anticipate trading of other investors and states that such
time advantage may be detrimental to slower players as
far as trading costs are concerned. An increase in
transaction costs is mentioned by several authors, and
some of them refer to ‘legalized front-running’, the topic
described in ‘Flash Boys’ by Michael Lewis. A merit of
Kirchner (2016) is referring to the book yet noticing that it
fails to present any evidence of the alleged ‘rigged’ market.
Indeed, most authors referring to Lewis (2014)
incomprehensibly fail to cite its sharpest critic, Kovac
(2014), who rejects nearly every single statement made by
Lewis in his highly acclaimed book. Yet, the highly reputed
trading analysis firm Nanex took the opportunity offered by
a trade execution report which stated a very similar
problem to ‘legalized front-running’ (Nanex 2014): an order
for 20,000 Ford Motors Co. shares failed to fill completely
despite an advertised liquidity on eight different venues of
nearly 25,000 shares. As the order was launched, part of
the available liquidity suddenly disappeared although no
news about the company was released shortly beforehand.
In just 5ms activity jumped from near zero to the equivalent
of 80,000 quotes per second. The conclusion by Nanex is
“based on the other 4 examples, we are sure that no trades
would have occurred during these few milliseconds of time
if it wasn’t for this trader’s order”, so indirectly confirming
higher trading costs for traditional market participants, as
stated by Lewis (2014). Indeed, sudden disappearance of
liquidity leads to filling the remaining part of the trader’s
order at a worse price and possibly to the benefit of those
predators that swept away the available liquidity. Several
papers (Brogaard et al. 2014b; Biais et al. 2014; among
others) similarly argue that HF traders impose higher
transaction costs to slower traders to their own benefit. But
this is not the end of the story. The advantage HFTs have in
arbitraging to the detriment of slow traders is a topic widely
discussed in the literature. Differently from academia, there
is widespread opinion among traditional practitioners that
HFT is a malicious innovation, disrupting consolidated
market habits and casting well-founded doubt on financial
stability. Some theoretical models, by using low-latency as
the main characteristic of HFT, conclude that they make
profit at the expenses of slower traders. Jarrow and Protter
(2012) find the differential speed advantage of HFT causing
inequity and compare this to insider trading, as in the HFT
case “the ‘inside’ information is not based on the
‘fundamental’ price process but the ‘order flow’ process” . A
similar view is reported by van Kervel (2015), who
recognizes that HFT increases adverse selection costs for
slower traders. Their large latency causes slow
participants to often trade against stale quotes in an
environment in which prices change at very high speed.
Experimental evidence is reported by De Luca et al.
(2011) to corroborate the intuitive fact that
“outperformance of the algorithmic trading systems over
humans are primarily speed-related”. Studying the impact
of HFT on market making leads Weaver (2012) to state that
increasing market fragmentation and HFT activity will
erode profits of LF market makers, warning that additional
compensation may be necessary to ensure that market
making activity continues. A theoretical model
considering interaction between slow and fast traders
developed by Biais et al. (2014) shows that HFT obtains
advantageous access to valuable information, creating
adverse selection for non-HF traders. Bringing this
statement to its logical consequences, Cvitanić and
Kirilenko (2010) conclude that the more orders humans
submit, the more money HFTs make. In other words, either
a firm reaches a critical speed at which it can compete, or it
is condemned to lose money. Instead, evidence of a neutral
role of HFT is found by Brogaard (2011) in case of
macroeconomic announcements. An intuitive reason for
HFT’s profit to have an upper limit is that, as the number of
HF trades increases more than linearly with the number of
HF traders present in the market, the fast ones will
eventually start to mainly trade with each other. In this zero-
sum scenario, profits of one HF trader will imply losses of
another one. Rather counter-intuitively Brogaard (2010)
finds that HF traders tend to trade with other HF
counterparts less than expected and with non-HF traders
more than expected. The reason provided by the author is
that HF traders have a less diverse variety of strategies
than non-HF traders, so when one HF trader decides to sell
(buy) it will more likely find non-HF traders willing to buy
(sell) than another HF trader. The logical conclusions would
be an expansion of the available pool for HF traders’ profits,
more competition among HF traders to pick up the
profitable opportunities first, and a more pronounced
crowding effect in case of extreme unidirectional price
movements, possibly leading to a self-reinforcing feedback
loop.
The reduction in transaction costs, according to Foucault
and Menkveld (2008), is mostly due to RegNMS, which
paved the way to proliferation of venues and triggered fee-
based competition among incumbent market makers and
other limit order quoting agents. Also dubious are Sornette
and von der Becke (2011), who dismiss generalized
welfare gains as minimal in the short-term and negative in
the longer-term. Also Verousis et al. (2018) recognize the
cost reduction as mostly due to reduction in tick size, as it
was the case in the USA since 2001, with the migration of
minimum price change from fractions down to one cent of a
dollar. Other negative contributions to the market at large
are indicated by the same authors as far as arbitrageurs,
and not ordinary investors, benefit most by the transaction
cost reduction, and by Zhang (2017), who notices how, in
the presence of a shock, HF traders are the fastest to
withdraw their limit orders, therefore suffering less in terms
of adverse selection costs. In summary, it seems that
different researchers often have different opinions about the
supposed beneficial effect of HFT on trading costs.
● HFT‐led liquidity provision and consumption
Market making versus taking has a paramount impact on
market liquidity. However, whether on one side HFT
introduced massive order cancellation as a major liquidity
factor, it also assigned new roles to market making and
market taking.
In the aftermath of 6 May 2010, HFT became the target for
allegations of excessive liquidity consumption, in its turn,
leading, under severe stress conditions, to stub quotes
being displayed and executed. A simulation by Myers and
Gerig (2014) displays a continuous double auction agent-
based model (ABM) in the presence of HFT and finds better
liquidity as a result. It also reports a higher probability of
transactions to happen, which the study interpreted as
another sign of higher liquidity directly linked to HFT.
Friederich and Payne (2011) analyze London Stock
Exchange data finding an upward trend of liquidity since
HFT entrance, whereas Hagströmer and Nordén (2013)
subdivide the traders into HFT and non-HFT, and their
model shows the HFTs group supplying more liquidity than
they consume. Yet, according to the paper, although these
results are all statistically significant, no evidence can be
drawn about cause–effect relationship. Chaboud et al.
(2014) analyze the impact of algorithmic trading in the
ForEx market and find it beneficial for liquidity. A similar
result is reached by Brogaard (2011), who points out how
HF traders supply liquidity in both high and low-volatility
days. Even stronger is the result found by Brogaard et al.
(2014a, b), who report 18% higher HFT activity during jump
intervals compared to other periods, whereas Brogaard
(2010) observes that even in times of abnormally high
volatility, HFT does not stop providing liquidity, on both bid
and ask side. A Bank of England working paper (Benos and
Sagade 2016) analyzes transaction data about four stocks
picked up from the FTSE-100 data set at one-second
granularity and finds that overall HF traders tend to supply
somewhat more liquidity that they consume. Abundant
liquidity is always welcome by the exchanges, as it allows
any agent, whether HF or other, to trade with minimal or no
impact on the price, to the benefit of market stability.
Yet, the beneficial effect of HFT on liquidity is such only for
those who can exploit it. A common practice shown by ultra-
fast market makers is quoting the same liquidity onto
several different exchanges, giving the impression of
depth in all of them, but ready to cancel all the non-filled
limit orders as soon as the first one has been picked up
(AFM 2016). This way, markets experience the so-called
ghost liquidity, which falsely shows prices that do not exist
for real. HFT activities vary a lot over time and across
countries. In the USA, HFT share of trades was around
20% in 2005 to reach between 55% (Miller and Shorter
2016) and 60% (Kaya 2016) just 4 years later, only to
decline afterwards. The percentages of HFT posting quotes,
as opposed to executing trades, are even higher. Indeed,
an important feature of HFT is quoting competitive
prices only for an extremely short time and then canceling
limit orders to avoid being picked off by more informed
investors. Then, ultra-fast market makers re-analyze the
information arrived in the meantime, calculate a new
competitive price and submit a new ‘flash’ limit order,
starting the whole cycle all over again. Whereas on one
side this way of operating dramatically reduces the risk of
passive trading, it also provides markets with large amounts
of liquidity that can never be executed and that is
essentially non-existent. Arnoldi (2016) recognizes that
order-to-trade ratios that were floating around the 1:1 ratio
before HFT time, suddenly jumped to 50:1, and Kirchner
(2016) sets the hi-water mark at occasional 1000:1.
Menkveld (2016) agrees that the main price discovery
exercise is carried out through quote updates rather than
trade execution, reaching the conclusion that HFT is
‘intimately related’ to order flow fragmentation.
Baron et al. (2012) use transaction-level data for the E-mini
S&P 500 futures for the period August 2010 through August
2012 and find that HF traders earn substantially higher
profits when consuming liquidity than by providing it. This
seems a strong motivation for HF traders to diminish the
available liquidity instead of increasing it. Of the same
opinion are Cvitanić and Kirilenko (2010), who develop a
mathematical model showing that during crisis HF traders
provide less liquidity than in normal times. However, in the
literature there seems to be some consensus on HFT being
beneficial for liquidity supply, although diverse viewpoints
are not infrequent. This fact is summarized by Aitken et al.
(2012) who observe that “the majority of papers that deal
with HFT empirically find a predominantly positive overall
impact” on liquidity. Barker and Pomeranets (2011)
summarize a common view as “HFT appears to be having a
profound impact on market liquidity, and its rise has
coincided with an increase in trading volumes”.
As said earlier, many sources have indicated HFT as either
the originator or the main culprit of the Flash Crash and in
particular of the dramatic liquidity scarcity experienced on
that day. However, Menkveld (2013) carries out an in-depth
study, based on TAQ data on Dutch local index stocks for
both Chi-X and Euronext in the period 1 January 2007
through 17 June 2008, and finds that “in both markets the
vast majority of HFT trades are passive: 78.1% in Euronext
and 78.0% in Chi-X”. A compatible result is found by
Hagströmer and Nordén (2013), who compute order-to-
trade ratio (number of limit order divided by the number of
executions), leading to the conclusion of HFTs acting most
often as market makers. Moreover, they observe 63–72% of
HFT trading volume being made of market making and 81–
86% of limit order traffic. The discrepancy between the two
ranges is due to the presence of fleeting limit orders, whose
features make them alike to market orders. Hasbrouck and
Saar (2009) call ‘fleeting limit orders’ those limit orders
cancelled within 2 s of their submission. They are close
substitutes of market orders as their purpose is seeking
immediate execution without running the risk of being
picked off by a fast moving adverse market trend. One of
the goals of fleeting orders is to pick up hidden liquidity, that
is, orders not publicly shown on dark pools. Hasbrouck and
Saar (2009) analyze data from the Island ECN, an
electronic communication network organized as limit order
book, and find that about 14% of the trades occur against
hidden liquidity. They argue that traders seeking immediate
execution are not restricted to posting market orders, and
limit orders are no longer the tool used by passive traders
patiently waiting for their order being hit whenever an
aggressive trader deems it profitable. This situation (which
may well show different figures after more than a decade
later), has the potential to make the trading scenario less
clear-cut. The old days when the mark of a chalk on a
blackboard used to say everything about the marketplace
are gone forever. As gone is the time when a limit order
was posted with the genuine purpose to be executed. The
same study by Hasbrouck and Saar (2009) reports 83% of
all incoming orders in the Island ECN being limit orders but
only 18.4% getting fully or partially executed. Many limit
orders are cancelled within an extremely short time and
27.7% are fleeting orders. HFT seems to have changed the
scenario completely. Kirilenko and Lo (2013) translate these
figures into hard money by arguing that traders who enjoy
best access to customer order flow earn top rewards.
Another advantage of HFT market making is underlined by
van Kervel (2015), who argues that “high-frequency traders
who operate as market makers can strongly benefit from
the increased execution probability. That is, placing
duplicate limit orders increases their trading rate and
expected profits”. Obviously, a slow trader could not afford
the luxury of placing multiple limit orders on different
markets for the same stock in the hope that only one gets
through execution. In such a case, the risk of going off
balance with multiple (unintended) executions before being
able to cancel the limit orders other than the first one, would
be far too high. Leland (2011) develops a formal model
showing that the ability of HF traders to process information
faster than other traders grants them an advantage. This
sounds intuitive as the fast market maker will have good
chances to cancel its limit order as soon as it receives
information, suggesting that such order is mispriced
according to the very latest news. In fact, Jovanovic and
Menkveld (2016) assume HFTs are both faster and more
informed than their counterparts. Based on this assumption,
their findings on market efficiency are mixed. Another
hypothesis found in several studies is the different behavior
of HFT in times of crisis with respect to normal times. This
is the position held by Easley et al. (2011), who investigate
the E-mini S&P 500 futures market and find that some
liquidity providers turned into liquidity consumers during the
Flash Crash, somehow exacerbating the price downfall.
Quite boldly, Kirilenko and Lo (2013) state that “[i]n contrast
to a number of public claims, high-frequency traders do not
as a rule engage in the provision of liquidity”. The same
concept is restated in a slightly different manner in Baron et
al. (2012), where the authors notice how liquidity-taking
HFT is especially profitable and its aggressiveness is
consistent across days.
● Discovering the right price
The research carried out by Brogaard et al. (2014a, b)
mentioned above also states that HFT shows the positive
effect of increasing price efficiency as HF participants tend
to trade in the direction of permanent price change, so
speeding up price discovery, and in the opposite direction
of temporary mispricing, both on days showing average and
above average volatility. Foresight (2012) acknowledges
“evidence suggesting that price efficiency has generally
improved with the growth of CBT”. In addition, Brogaard
(2010) shows that HFT activities on the market contribute
more to price discovery than traditional traders do.
Manahov and Hudson (2014) confirm efficiency in the price
discovery process using 1-min high-frequency data from
the six most often traded currency pairs (USD/EUR,
USD/JPY, USD/GBP, USD/AUD, USD/CHF and
USD/CAD). Similar results are also reported by
Hendershott (2011): the study analyzes price efficiency
trends over a 4-year period (2006–2009) at NYSE and
NASDAQ and argues that, based on intraday variance
ratios, growing HFT activity leads to higher market
efficiency, which includes quicker price discovery. On the
other hand, Anagnostidis and Fontaine (2018) report that
market fragmentation allows HF traders to collect
information quickly from multiple sources and this
increases their information advantage (and right pricing) to
the detriment of other traders. Ordinary traders usually rely
on officially distributed price information that suffer from
higher latency than direct feeds sold by trading venues to
those fast traders who can exploit it. Therefore, whereas HF
traders enjoy the best possible view of the markets, LFTs
must contend with lagged prices that, in case of rapidly
moving markets, may penalize them. Frequent price
cancellations are also a valuable source of information for
price discovery purposes, according to Blocher et al.
(2016), who base their research on 5.78 terabytes of data
on all the S&P stocks for 2012. “The HFTs process the
information so quickly that price discovery comes from the
cancellations rather than from executions. This is the more
effective method, since no dollars need change hands”.
Aitken et al. (2012) study cross-sectional determinants of
HFT participation over long time series on the Euronext
Paris and LSE exchanges, finding that “the increase in the
level of HFT activity has increased [price] efficiency without
harming the integrity of the market”. Using data provided by
NASDAQ to academics under non-disclosure agreement
and by NYSE, Brogaard et al. (2014a, b) show that HFT
contribution to price discovery is statistically significant and
that HFT activity is negatively correlated with pricing errors.
Smaller pricing errors reduce market inefficiency to the
benefit of long-term investors. However, given the time-
frame HFTs operate at, and the minuscule price
fluctuations observed over such ultra-short periods on the
average trading days, the very same concept of pricing
error seems far from clear. This matches intuition as, in
order to exploit small price movements, it is necessary that
prices show some degree of inaccuracy, a certain level of
volatility in both directions within short time frames, and
narrow spreads to allow minimization of costs and risk of
speed-based strategies. Brogaard (2011) finds HFT playing
an important role in the price discovery process, suggesting
that “HFT’s trades provide more private information than do
non HFT’s trades”. Rather critical of HFT’s merits seems
to be another Nobel laureate, Paul Krugman who, in an
article on the New York Times, 2 August 2009, mentioned
in Linton and O’Hara (2012), states it being “hard to imagine
a better illustration (of social uselessness) than high-
frequency trading. The stock market is supposed to allocate
capital to its most productive uses, for example by helping
companies with good ideas raise money. But it’s hard to
see how traders who place their orders one thirtieth of a
second faster than anyone else do anything to improve that
social function”. On the same side sit Hasbrouck and
Saar (2013), who cast doubt over the social benefits of low-
latency trading. Eliminating transient price disturbances has
its value “but such an argument at the millisecond
environment is a bit tenuous”. Less straight but substantially
aligned to this opinion are Benos and Sagade (2016), who
find a direct relationship between price discovery occurring
as a response to information about fundamentals (not the
core business of HFT). Indeed, excessive emphasis on
short-term information by short-term speculators is
recognized by Brogaard (2010), to the detriment of
fundamentals, a concept usually associated with sound
financial information.
● High‐frequency trading and flash crashes
The 6 May 2010 Flash Crash spread great worry among
practitioners and regulators alike and attracted a lot of
interest from the academic world. Some argue that, leaving
aside worries for the abnormal market behavior, May 6 was
just another troubled day on the financial market as many
were observed in the aftermath of the 2008 subprime crisis.
Other researchers think differently: that event was by no
means ‘business as usual’. Indeed, several sources talk
about frequent mini-flash crashes, rapid events occurring at
sub-second scale, affecting only one or very few securities,
much less noticeable than the May 6 event. Still, in the view
of Vuorenmaa and Wang (2014), they are relevant for
market stability and investors’ confidence. If a mini-flash
crash affects only one security, as suggested by Golub et
al. (2012), the event is much harder to detect than one
displaying a world-wide multi-market contagion. Yet, mini
events may cast doubt on the stability of the financial
markets, exactly as the major Flash Crash did. Recent
reports highlighted that this phenomenon is in no way
restricted to the US market but it seems to affect the whole
financial world. Because of the extreme speed at which
things happen, lacking a thorough understanding of the
causes leading to minor or major flash crashes, real-time
human supervision would not be able to help at all, leaving
the financial markets at bay of events whose
unpredictability only relates to ‘when’ rather than ‘if’. A
self-reassuring position is taken by Gomber et al. (2011),
who fingerpoint the US market structure. The report
commissioned by the Deutsche Börse states that, since
Europe enjoys a more flexible ‘best execution’ regime
thanks to the Markets in Financial Instruments Directive
(MiFID) and a circuit breaker regime based on individual
securities, “no market quality problems related to HFT have
been documented so far”. Another European report,
Foresight (2012) commissioned by the UK Government
Office for Science, keeps a more balanced position.
Maybe because it was published about 1 year after
Gomber et al. (2011) and more evidence was by then
available, Foresight (2012) recognizes that “[t]here has
been a variety of other, smaller illiquid events in the markets
since the Flash Crash”. The report lists the 8.1% natural
gas drop on 8 June 2011, which bounced back in a few
seconds; the eightfold spike in volatility of oil futures on 2
February 2011; the 98% fall in Morningstar ETFs in March
2011; the very rapid changes in BT Group, Hays, Next,
Northumbrian Water Group, and United Utilities Group, all
listed at the London Stock Exchange. In all these cases, no
significant news seems to have caused the swings, which in
some cases affected European market as well. But the list
does not terminate here. Zervoudakis et al. (2012) mention
the Dow Jones Industrial Average flash crash on 29
September 2008, the cocoa futures mini-flash crash on 1
March 2011 and the dollar–yen sell-off 15 days later. On 2
May 2011, it was the turn of gold to drop by $20, just to
quickly recover more than $15; silver followed suit on the
next day and in July of the same year crude oil futures
showed large swings (Cliff 2011). More recent years have
witnessed other similar events. Sornette and von der Becke
(2011), referring to a few other studies, report rather
frequent occurrences of mini-flash crashes in single
stocks and although the definitive proof of HFT
involvement is missing, some crashes seem to have been
accompanied by an increase in quoting frequency.
According to Nanex, the last few years experienced
thousands of mini-flash crashes. Investigation of the
number of both crash and spike ‘black swans’ (Taleb 2007)
in 100-ms windows in Johnson et al. (2013) (that confirms
and expands Johnson and Zhao 2012) shows that such
number increases as observation time shortens, with near-
perfect superposition of the curves representing crashes
and spikes. In particular, within the 100–200-ms window the
number of events was about ten times greater than the
number in the 900–1000-ms window. The authors also find
18,520 black swan events (defined as ten down- or up-ticks
in a row with more than 0.8% price change) that lasted less
than 1.5 s on multiple exchanges between 2006 and 2011,
that is, nearly 10 per every working day. Foresight (2012)
suggests that a high number of mini-flash crashes may be
caused by the feedback loop generated solely by computer
algorithms. A possible reason for such short events to be
neglected is that they tend to cure themselves nearly as
quickly as they arise, in a matter of milliseconds, and are
often assimilated to market noise. As seen before, several
studies (Kirilenko et al. 2017, among others) find that HFT
tends to improve liquidity in normal times (something that
prevents crashes) but recognize a different behavior and a
different effect under stress. This seems confirmed by
Golub et al. (2012), who analyze mini-flash crashes in the
US equity markets during the four most volatile months in
the period 2006–2011. Their findings confirm the adverse
impact of HFT on liquidity during the mini-flash crashes,
which may have exacerbated those events. Moreover,
analyzing quoted liquidity during the same crashes, they
note a stronger reduction on the bid side by HF traders,
resulting in sell-side pressure, something that matches
both common intuition and practitioners’ perception.
● Profits of high‐frequency traders
A critical issue about HFT is whether this practice ensures
profitability. If HF traders make money out of speed, then it
is plausible to investigate the impact of ultra-fast trading on
the regular functioning of the markets and the nuances of
the high-speed algorithmic strategies. If not, that is, if HF
traders do sometimes make money, and at some other
times lose money, the conclusion would be to consider
them as ordinary traders, subject to the usual random walk
of the market. Once again, scholars display a wide range of
opinions. Menkveld (2013) shows that all HF traders’
earnings arise from passive orders and that in general they
lose money on their positions but make money on the bid-
ask spread and this looks consistent across the stock
universe analyzed. Moreover, the profit component of the
inventory seems to be restricted to trades closed within 5s,
whereas virtually all those lasting longer than 1-min yield
negative results and those in the middle (5 s to 1 min) are
mixed. The rationale behind this seems to be the adverse
selection of participants posting limit orders. Opposite
findings are reached via the analysis of NASDAQ and
NYSE in 2008–2009 by Brogaard et al. (2014a, b),
according to which HFTs lose on limit orders and gain on
market orders. Quite surprisingly, the same paper finds that,
at 90% confidence interval, HFT incurs losses during price
jumps while traditional traders seem to profit from that. A
question would come natural at this point: why should HF
traders engage in liquidity supplying activities if they lose
money? The answer resides in the market fee structure.
Foucault et al. (2013) report trading fees for five US
trading platforms (NYSE Arca, Nasdaq, BATS, EDGX,
LavaFlow) ranging from − $0.30 to − $0.20 for make fees
(i.e. liquidity suppliers earn a few cents) per 100 shares
traded, and + $0.25 to + $0.30 for take fees. This supports
the view of HFTs losing money on limit orders as they get
compensated by the incentive. However, free market is as
various as you can think and indeed Foucault (2012)
reports make and take fees for a set of 10 US markets,
showing that some exchanges offer no rebate for limit
orders and there is even one exchange that offers − $0.18
take fee (a rebate) against a + $0.14 make fee, both per
100 shares. According to Brogaard (referred to by Weaver
2012), HFTs in the US earn $0.075 to $0.09 per $100
traded; this accounts for a tiny one-seventh of traditional
market makers’ profit. Brogaard (2011) provides statistics
reporting the after-fee revenue per stock and day. The
research shows that HFT is profitable overall and in
particular when demanding liquidity. Arnoldi (2016) reaches
a similar result about arbitraging capabilities of ultra-fast
algorithms over the same security on different markets or,
even more often, minuscule differences between
securities that are statistically correlated. More clear-cut is
the conclusion by Kirilenko et al. (2017), who state that
“High Frequency Traders are consistently profitable
although they never accumulate a large net position”. One
of the main findings of Baron et al. (2012) is that, in the E-
mini futures market they analyze, HF traders are always
profitable, especially when they consume liquidity. They find
HF traders to earn on average $0.25 per each contract
traded and in August 2010 aggressive HFTs earned a daily
mean of $45,267 (against $19,466 and $2460 for mixed
and passive HF traders, respectively). Yet, the standard
deviation for aggressive HF traders was $167,411, which
means they are also running serious risk to incur heavy
losses. This result is opposite to Kirilenko and Lo (2013),
who believe HFT profits are large and persistent whereas
the risk they take is very little. The profit calculation on the
E-mini S&P 500 futures contract for August 2010 at CME
made by Baron et al. (2012) leads to an aggregate of $23.6
million in gross profit, corresponding to an annualized profit
of over $280 million, whereas Brogaard (2010), using
different source data, from NASDAQ and CBOE, compute
an annualized gross profit of nearly $3 billion. Foresight
(2012) reports “indications that the profitability of HFT is
reaching its limits and in the next 10 years may come under
further pressure”, a forecast that makes sense as no
innovation may reasonably be successful over a long period
of time (during which it will be copied by competitors)
without being innovated further and in this business the
speed of light is the upper limit set by the laws of physics.
Indeed, the time frame is the big question mark. Baron et
al. (2012) also report speculation that “profitability of HFT
has decreased over time, perhaps due to increased
competition”. Among the possible reasons suggested
there are new entrants decreasing total profits for the
average firm, or competition driving down profits for each
HFT firm. Yet, performing a hypothesis testing exercise on
both average firm and per contract profit, they find
aggressive and passive HF traders having actually
increased their profits between August 2010 and August
2012, and only mixed HF traders experiencing no significant
change during the same period. In the theoretical model
developed by Cvitanić and Kirilenko (2010) with infinitely
many slow traders, the profit available to HFTs is bounded,
with the consequence that, as the number of machines
increases the profit for each one, it decreases on average.
The topic of declining HFT profits has been hotly debated
over recent years. One interesting finding by Baron et al.
(2017) investigating the Swedish equity index OMX S30 is
that HFT uses speed in ways that partially offset each other,
although the same paper recognizes that all HF traders
make profits, with the fastest ones earning top money. Also
Brogaard et al. (2017) find evidence of HFTs being able to
exploit extreme price movements to earn profits. However,
the business question is whether the non-top HFT firms are
still capable of making up for the extremely expensive arms
race in technological equipment and network connections.
The low-volatility era which world financial markets seem
to have recently entered is identified by Chaparro (2017b)
as the main cause, together with operational costs, of
squeezing HFT profits, with officially reported revenues for
the sector dropping from more than $7bn in 2009 down to
just over one billion in 2016, according to the TABB Group.
A similar view is shared by Kaya (2016), who finds
increased infrastructural costs and competition as the
probable cause of the profit slowdown, highlighting that co-
location costs have risen two or three times whereas
simultaneously per-share profit has halved from $0.001
down to 1/20th of a penny. Linton and Mahmoodzadeh
(2018) also point out aggressive competition as the reason
for reduced profits. Maybe business journalists are
exaggerating their emphasis of the trend because of the
well-known search for headlines and perhaps titles like
“High-speed traders have hit the wall” (Massa and Chilton
2017), “high frequency trading is done, it’s over” (Worstall
2017) or “the bonanza has now ended” (Meyer et al. 2018)
are overstating the matter but it seems a widely recognized
fact that, at least among practitioners, there is some
commonality of views about shrinking HFT profits. This
does not seem to be the case for recent rigorous
academic studies. While Verousis et al. (2018)
acknowledge market making (a major field of business for
HFTs) as having become largely unprofitable not because
of decreased volatility but because of reduced tick size, Aït-
Sahalia and Saglam (2013) argue that having HFT
‘plateaued’ on the markets it was traditionally present, it is
now addressing new and mostly unregulated markets
like cryptocurrencies where, until recently, P&L percentage
did still show double figures. Table 7 summarizes academic
opinions about whether HFT is gaining abnormal profits,
and on the expected trend over the next future.
1. Passive HFT è liquidity traders
Passive HFTs apply strategies similar to traditional
market making ones, which are also known as electronic
liquidity provision (ELP), seeking to capture both the bid-ask
spread and the rebates paid by the trading venues as
incentives for posting liquidity. The strategies employed
by passive HFTs are sometimes referred to as quasi
market making, because HFTs can suspend their activity
whenever the market state would lead to their
unprofitability, there is a general critique that the liquidity
provided by HFT is often illusory. In Government Office For
Science (2012), this phenomenon is named periodic
illiquidity and is also explained by the opportunistic style
and tight risk management of HFT market makers.
Moreover, a particular type of ELP, known as rebate
arbitrage, tries to profit only from collecting liquidity
rebates paid by the trade venue by opening and closing
positions at the same price on different exchanges (no
spread capture), without really contributing to liquidity.
Menkveld (2011) states that HFT “cream skimming”
strategies, requiring little capital, have driven out traditional
market makers with large capital and inventories, leading to
a drastic change in the structure of liquidity provision
sources, with large potential implications for the market
dynamics in times of stress. On the other hand, the Eurex
Exchange finds evidence against this critique, by analyzing
the behavior of HFTs on the 25th of August, 2011 – a highly
volatile day for the DAX Futures (FDAX). The report
concludes that HFTs actually contribute to liquidity and
prevent fast price movements during periods of high
volatility or strong directional trading.
Aldridge (2013) identifies two main types of automated
market making models: inventory models concerned
only with the effective management of inventory, without
any opinion on the drift or any autocorrelation structure e.g.,
Avellaneda and Stoikov (2008) and information-based
models tries to extract the information which other market
participants may possess, by analysing the order flow
(buying and selling pressure) and/or the shape of the order
book. The theoretical model proposed in Glosten and
Milgrom (1985) relies on Bayesian learning to combine new
information into the market maker’s prior beliefs about the
true market value of the traded asset. Das (2005) also
proposes a basic profit-making strategy, achieved by
widening the spread around the break-even bid-ask spread
with a fixed amount. Das (2008) suggests placing bids and
asks just inside the current spread as long as they are
associated with a non-negative expected profit, i.e. the
Glosten and Milgrom condition. Another market making
model based on the Bayesian framework is introduced in
Lin (2006). This time, the true value belief is updated by
means of a discrete Kalman filter, where the
measurement is given by the net order flow observed over
some period of time. A different framework used to derive
market making algorithms is reinforcement learning. Chan
and Shelton (2001) train a market-maker strategy which
observes three (discrete) state variables: own inventory,
order flow imbalance, bid-ask spread.
2. Active HFT - order anticipators
Active HFTs act as liquidity takers, by trading with
aggressive, liquidity-consuming market-orders. There are
two main lines of active HFT development: the first one
seeks to predict market momentum and incoming order-flow
and makes profits from short-term market shifts, while
the second one tries to exploit the technological structure of
the trading network system and superior speed advantage
in order to detect and “front run” distributed executions of
large orders. According to Harris (2002), order anticipators
try to profit from information about third parties’ trading
intentions, rather than from their own fundamental
information regarding the traded asset. This type of profit-
motivated speculators, also described as “parasitic
traders” or “predatory”, can be further subdivided
between front runners, sentiment-oriented technical
traders and squeezers. While front-running and squeeze is
considered to be illegal strategies, sentiment-oriented
technical traders (technical analysts or chartists) use
publicly available information in order to predict future price
returns. End-of-day strategies process aggregated
information related to the daily price (open, high, low, close,
average, median) or to the daily turnover in order to
generate profitable trading signals. On the other side,
intraday trading relies on additional information available
at the market microstructure level, such as order book
liquidity, order flow, trade and tick history. The main idea
is to identify persistent relationships between various
market indicators and future short-term market moves or
incoming customer order flow, upon which profitable trading
strategies can be built.
In a high-frequency setting, the analysed time-series are
usually in a raw format (tick-by-tick). The actual trading
rules can either be specified by a human specialist or can
be “learned” from past data through various quantitative
and computational methods. A related general concern,
confirmed by Brogaard (2010), is that HFT-based strategies
are not very diversified or, equivalently, their trading
signals are highly correlated, leading to an exacerbation of
market movements. One simple HFT strategy proposed by
Aldridge (2009) is based on order flow short-term
autocorrelation, which consists in opening position in the
direction of the order flow imbalance, i.e. the difference
between the cumulative number/volume of buy and sell
orders, or buyer- and seller-initiated trades. Another
strategy relies on mimicking aggressive trading, as a proxy
for informed traders’ expectations. An indicator measuring
the orders’ aggressiveness can be computed as the
percentage of the market as opposed to limit orders.
A second class of active HFTs tries to identify patterns of
execution algorithms (processing large institutional orders)
and trade ahead the remaining execution program
(“electronic front-running”). On the execution side, smart
routing strategies which are able to minimize the speed
advantage of ultra-HFTs have been developed. E.g., Royal
Bank of Canada’s THOR R routing technology
precomputes the routing latencies and sends the different
slices of the total order in such a way so they reach the
targeted trading venues at the same time and therefore
eliminating any information leakage.
IV. Artificial Intelligence Applied to Stock Market
Trading
Beginning in the 1990s with the introduction of
computational methods in finance, much research has
focused on applying Artificial Intelligence (AI) to financial
investments in the stock market. The main advantages of
using computational approaches to automate the
financial investment process include the elimination of
“momentary irrationality” or decisions made based on
emotions, ability to recognize and explore patterns that are
looked over by humans, and immediate consumption of
information in real-time. This area of knowledge has
become known as Computational Finance. More recently,
within computational finance, there is increasing use of
and research on AI techniques applied in financial
investments. Although a computer conducts the vast
majority of hedge fund trades in an automated way, 90% of
these operations are still performed by a hardcoded
procedure. Thus, the ever-increasing application of artificial
intelligence still has great potential for development.
1. Portfolio optimization
Portfolio Optimization, or Portfolio Selection, is a problem
that consists of determining a set of financial assets that
best suits a particular investor, usually aiming at
maximizing profits. The Modern Portfolio Theory (MPT),
created by Markowitz, was the first contribution to portfolio
optimization models. Markowitz introduced two metrics for
evaluating a portfolio’s performance: the expected return
and the risk. The expected return expresses the idea that
an asset that has performed well in the recent past tends
to maintain such performance in the future. As a forecast,
the risk is the proposed metric to model the return’s
uncertainty. Markowitz’s theory has become very
widespread and several changes have been made to its
original proposal. The use of portfolio variance as a risk
measure, for example, has been widely criticized since the
variance takes into account both negative and positive
deviations. Then, downside risk measures emerged,
taking into account only the worst historical returns of the
portfolios. Conditional Value at Risk (CVaR) is a downside
risk measure widely used because it is a coherent measure.
Numerous works have improved these models, creating
more risk measures and proposing restrictions that bring
them closer to the practical aspects of stock market trading.
Several exact, heuristic and hybrid optimization methods
have been proposed to solve these portfolio optimization
models, which have become increasingly more complex.
Moreover, a portfolio optimization Mean-Absolute
Deviation (MAD) model was proposed with the objective
to minimize the absolute deviation (risk measure).
T-J.Chang, N. Meade, considered a Mean-Variance model,
in which variance is the risk measure to be minimized, and
the mean return is limited to a lower bound value. The
portfolio cardinality is constrained to a single value. The
paper proposed three heuristic methods based on the
Genetic Algorithm, Tabu Search, and Simulated Annealing.
Y. Crama and M. Schyns also considers a Mean-
Variance model, but its model considers a variable
cardinality constraint, which allows the portfolio cardinality
to assume a range of values. This work proposes a
simulated Annealing algorithm to solve the model.
M.Ehrgott, K.Klamroth considered a mono-objective
Mean-Variance model that combines the portfolio variance
minimization with the mean return maximization in a single
objective function using complementary weighted factors for
each objective. Genetic Algorithm, Tabu Search, and
simulated Annealing based methods are proposed to solve
the portfolio optimization model.
R.Subbu proposed a model with three objectives: mean
return maximization, variance minimization, and Value-at-
Risk (VaR) minimization. The proposed heuristic
initializes the population with the Randomized Linear
Programming (RLP), generates an interim Pareto front with
Pareto Sorting Evolutionary Algorithm (PSEA) and Target
Objective Genetic Algorithm (TOGA), completes gaps in the
Pareto front with TOGA, and stores the result in a
repository.
Moral-Escudero performed hybrid methods by combining
proposed heuristics and Genetic Algorithm to solve the
mono-objective Mean-Variance model. A. Fernández and S.
Gómez presents a combination of Neural Artificial Networks
with heuristics to solve the weighted factors monobjective
Mean-Variance model. G. Hassan and C. D. Clack
proposed a multiobjective evolutionary algorithm based on
SPES2 to solve the bi-objetctive Mean-Variance model. T-
J.Chang, S-C.Yang compared different weighted factors of
monobjective models composed by risk minimization and
mean return maximization. The models differ by the risk
measures applied: variance, semivariance, and absolute
deviation. Genetic Algorithms with specific operators and
repair methods for each model were performed. With the
increase in the complexity of the models and considering
larger numbers of assets and a larger historical series, the
works started to use heuristics instead of exact methods,
since these methods solve complex models in polynomial
time. Pindoriya proposed a portfolio optimization model with
three objectives: maximize return (mean), minimize risk
(variance), and maximize historical returns skewness. For
this model, a Multiobjective Particle Swarm Optimization
(MOPSO) algorithm was performed.
H.Zhu, Y.Wang performed Particle Swarm Optimization
(PSO) algorithm with specific operators and repair
operators for three different mono-objective optimization
models: Mean-Variance model with minimum return
constraint, Mean-Variance utility function model using
weight factors, and Sharpe-ratio maximization portfolio
optimization model.
W-G.Zhang, Y-J.Liu applied Genetic Algorithm and
Simulated Annealing to a proposed portfolio optimization
monobjective model in which the objective is a utility
function formed by variance minimization and portfolio
diversity maximization, and the minimum return constraint
defines a lower bound for the mean return.
Ponsich and Jaimes surveyed two biobjective portfolio
optimization models, both considering mean return
maximization and different risk measures: variance and
VaR minimization. The work considered cardinality,
transaction lots, and turnover constraints, and found that
these methods can be solved by several Multiobjective
Evolutionary Algorithms (MOEAs). Finally, the paper
compares the models mentioned above to a mono-objective
Sharpe-ratio model solved by a Genetic Algorithm.
R.Mansini, W.Ogryczak, and M.G.Speranza, reviewed
mono-objectives and bi-objectives portfolio optimization
methods, which aim to minimize risk, subject to a minimum
mean return value, and to minimize risk and maximize
mean return, respectively. The considered risk measures
include absolute deviation, minimum return, Gini’s Mean
Difference (GMD), and Conditional Value-at-Risk (CVaR).
The paper analyzed transaction costs, cardinality, and
transaction lots constraints. The methods reviewed can
be divided into heuristics: Non-dominated Sorting
Genetic Algorithm II (NSGA-II), Pareto Envelope-based
Selection Algorithm (PESA) and Strength Pareto
Evolutionary Algorithm 2 (SPEA2) for multiobjective
models, and Genetic Algorithm (GA) and Threshold
Accepting (TA) for the monobjective ones; and exacts:
branch-and-bound and branch-and-cut based methods; and
hybrids: methods in which heuristics find a good relatively
small subset of assets before exact algorithms find optimum
solutions for these subsets. Another trend that was
subsequently observed is the use of multiobjective models,
since the choice of prioritizing return or risk depends on
the profile of a particular investor. In a multiobjective
model, a set of non-dominated portfolios are provided so
that the best portfolio according to the investor profile can
be subsequently chosen. Chen proposed a utility function
monobjective model representing historical portfolio returns
as a random fuzzy variable. The utility function combines
the random fuzzy variable variance (risk measure)
minimization and its mean (expected return) maximization.
The model considers cardinality, transaction costs, and
turnover constraints. A modified Artificial Bee Colony (ABC)
algorithm was performed to solve the model.
Seyedhosseini proposed a hybrid harmony search and
artificial bee colony algorithm to solve the mono-objective
mean-semi variance portfolio optimization model with
minimum return and cardinality constraints. Kalayci
presented specific repair operators for an Artificial Bee
Colony algorithm performed to solve the monobjective
mean-variance model using a utility function with
weighted factors, limited to a single portfolio cardinality
value.
Ertenlice surveyed Swarm Intelligence algorithms applied
to portfolio optimization models. The analyzed algorithms
are PSO (Particle Swarm Optimization), BPO (Business
Process Optimization), ACO (Ant Colony Optimization),
ABC (Artificial Bee Colony), CSO (Cat Swarm
Optimization), FA (Firefly Algorithm), IWO (Improved
invasive weed optimization), BA (Bat Algorithm) and FWA
(Fireworks Algorithm). These models aim to minimize risk,
subject to minimum return, cardinality, transaction costs,
and transaction lots constraints. The considered risk
measures are variance, variance with skewness,
semivariance, mean absolute deviation (MAD), Value-at-
Risk (VaR), minimum return, and Conditional Value-at-
Risk (CVaR).
Kizys proposed an Iterated Local Search (ILS) heuristic in
which the local searches are performed using a quadratic
programming algorithm to solve a monobjective 30904
mean-variance portfolio optimization model subject to
minimum return and cardinality constraints.
Recently, C.Chen and Y.Wei introduced a robust
multiobjective optimization model based on the mean-
variance model and elaborates a multiobjective Particle
Swarm Optimization (PSO) algorithm for the specific
problem.
Y.L.T.V. Silva, also uses a multiobjective PSO algorithm,
but proposes a PSO with ranks to solve the medium-
variance model with variable cardinality constraints. M.
Kaucic, M. Moradi, and M. Mirzazadeh developed NSGA-II
and SPEA2 algorithms with specific operators for three
different multiobjective models, which intend to maximize
the return and minimize the risk, differing in relation to the
risk measure considered: semivariance, CVaR, and a
combination of both. G.H.M. Mendonça proposed a
biobjective mean-CVaR model with lots, variable cardinality,
and turnover constraints, in addition to an evolutionary
algorithm based on the NSGA-II to solve it. Finally, the
paper suggested three different decision-making methods
for selecting a single portfolio on the Pareto-optimal border,
based on a given investor’s profile. When both are used.
There is still the possibility of using Machine Learning (ML)
algorithms to assist in the optimization process. Multi-
Attribute Utility Theory (MAUT) a posteriori methods are
often used for multiobjective problems when the selection of
only one solution is required.
Portfolio optimization models are becoming increasingly
complex, presenting more restrictions and, in some
cases, several objectives. This way, there is a tendency to
use heuristics to solve them since exact methods cannot
solve some more complex models in polynomial time. A
more significant number of objectives stems from the
growing number of metrics proposed to represent the return
and, principally, the risk of a financial portfolio.
2. Stock market prediction using AI
Stock market prediction or forecasting using historical time
series has become a technique widely used by researchers
and investors to obtain financial profits in stock trading.
These predictions, initially carried out by statistical methods,
have been increasingly performed by Artificial Intelligence
algorithms. Therefore, AI applied to investments
constitutes a recent research area that has already
achieved a large amount of publications.
Since 1965, many researchers have defended the
hypothesis of an efficient market A.W.Lo al, which states
that the market incorporates all the information that all
market participants have and their expectations, so that the
price changes are completely random and unpredictable.
In contrast to the efficient market hypothesis, other
researchers believe that the market prices fluctuate with a
trend. Considering this hypothesis, two schools of market
analysis can be regarded as: 1) technical analysis, which
defends trends in stock price movements and tries to
predict them through historical asset prices, and 2)
fundamental analysis, which argues that the socioeconomic
context of a company interferes with its future stock price
and, therefore, provides information that can be used for
forecasting future asset prices J.J.Murphy al.
Golan proposes a Rough Set Theory method that generates
rules to assist in Stock Market trading actions, which
includes buying, selling, and keeping an asset. This
method uses fundamentalists indicators as input data.
Schierholt and Dagli proposed a Probabilistic Neural
Network to predict financial prices movement trends using
historical asset price series. Three different classes of
trends were considered, each one indicating a different
action: buying, keeping, and selling a given asset.
Kim and Chun presented an Artificial Probabilistic
Network (APN) that considers historical prices and
fundamentalist indicators as input variables and performs
trends classification considering six classes of return
levels. Z.Zhanggui, H.Yan, and A.M.Fu employed Support
vector machine (SVM), k-Nearest Neighbor Classifier,
Probabilistic Neural Network (PNN), Classification and
Regression Tree (CART), boosting (Adaboost), and
bagging algorithms aiming to perform binary classification of
financial assets. Historical prices of the assets were used
as an input variable, and the paper’s results indicate the
better performance of the Boosting algorithm.
Kaboudan used historical prices of assets as input data and
predicted price and return with a proposed genetic
programming. The paper concludes that asset prices are
more predictable than returns. Kuo proposed a Genetic
Algorithm (GA) integrated with a Fuzzy Neural Network
(FNN) model to predict financial trends of asset price
movements (considering three different classes of trends)
using technical indices as input variables. Reference
Y.Wang introduced a Fuzzy grey prediction system that
uses historical prices, volume, and fundamentalist index
data of assets to predict future prices. Y.Wang proposed a
fuzzy rough set system that predicts financial assets future
prices using their historical prices and volume data as input.
J-Y.Potvin and P.Soriano predicted financial rules that
indicate buying and selling signals, performing a proposed
genetic programming feed by historical prices and volume
data of financial assets. W.Huang, Y.Nakamori collected
fundamentalist indices from financial assets to predict their
future price movement trend (in a binary classification)
using a Support Vector Machine (SVM) algorithm.
Kim and Min proposed a Genetic Algorithm (GA) that
selects the best weights for several classifiers predictions
and combines them into a single prediction. The
classifiers use technical indicators as input variables and
predict financial price movement trends considering four
different classes: Bear, Edged-Down, Edged-Up, and Bull.
Roh presented three hybrid Artificial Neural Network time
series models: NN-EWMA, NN-GARCH, and NN-
EGARCH, which uses EWMA, GARCH, and EGARCH to
determine input variables before applying the ANN to
predict financial assets volatility. The model employs
prices, volume, and fundamentalist indicators as input
variables, and the results indicate that NN-EGARCH
performs best. Since the first works on prediction of asset
trends, Machine Learning algorithms were already being
used. It was also common to use heuristics for optimization
as Genetic Algorithms, especially for the combination of
different classification algorithms.
For instance, C.Huang, D.Yang, and Y.Chuang combined
Support Vector Machine (SVM), Kth Nearest Neighbor
(KNN), Back-propagation neural network, decision tree, and
logistic regression using a voting committee after
performing a wrapper feature selection method. Using
prices, volume, and technical indices data to perform a
binary assets classification, the paper concludes that
voting performs better than single classifiers.
C-L.Huang and C-Y.Tsai, first performed a filter-based
feature selection for historical prices and technical
indicators data. A proposed Self-Organizing Feature Map
(SOFM) combined with Support Vector Regression (SVR)
was developed to predict future asset prices. SOFM
divides training data into several clusters before
different SVR models are applied to each cluster. Test
data are predicted using the SVR model trained with the
most similar cluster.
E.Hadavandi, H.Shavandi, and A.Ghanbari collected
assets historical prices series, which were selected in a
stepwise regression analysis (SRA). A Self-organization
Map (SOM) neural network was used to divide the training
data into clusters, and a fuzzy genetic system was applied
to predict assets future prices.
Hsu proposed a Self-Organizing Map (SOM) combined
with Genetic Programming (GP) method. SOM divides
training data into several clusters, where each cluster is
composed of training, validation, and test data. Validation
data select the best GP model for that cluster, and test
data evaluate the prediction performance. Prices,
volume, and technical indicators are used as input data,
and the model predicts the future prices of assets.
S.Asadi proposed the Preprocessed Evolutionary LM
Neural Networks (PELMNN). Stepwise Regression
Analysis (SRA) for variable selection. A genetic algorithm
was used as a global search method to evolve artificial
neural networks initial weights, and a Levenberg–
Marquardt Back Propagation (LMBP) neural network was
trained and used to predict future prices. The input data are
technical indices and trading volume.
Hsu proposed a price forecasting method to perform an
iterative feature selection procedure using a
backpropagation neural network invalidation data
composed of prices, volume, and technical indicators.
Finally, a backpropagation neural network was performed
on test data to predict the future prices of assets. Booth and
Gerding selected prices and technical indicators data in a
backward elimination method using a Random Forest
algorithm. Multiple Random Forest algorithms were used to
predict asset’s future prices. Final prediction used an
average of the predictions of all the Random Forest
predictors weighted by their training error. Patel and Shah
compared ANN, SVM, random forest, and Naive-Bayes
algorithms in performing binary classification of asset prices
trends. The classifiers are fed by technical indicators data,
and a validation step optimizes the classifiers’ hyper-
parameters.
Cavalcante surveyed Stock Market forecasting Machine
Learning models. The paper first surveys the
preprocessing techniques: normalization, outliers
exclusion, clustering, and feature selection. The
considered forecasting models are the Artificial Neural
Network (ANN) and Support Vector Machine and
ensembles methods that combine both and integrate the
models mentioned above with heuristics to predict assets
future prices or price trends.
Chong, Han, and Park compare three different feature
selection and transformation methods: Principal
Components Analysis (PCA), AutoEncoder (AE), and the
Restricted Boltzmann Machine (RBM). Then, a Machine
Learning algorithm is performed to predict future asset
return. For this, log returns data are collected every five
minutes. Fischer and Krauss use a Deep Learning Long
Short Term Memory (LSTM) Neural Network for a binary
asset prices trend classification, analyzing a historical
series of asset returns. Long, Lu, and Cui propose a Deep
learning model with convolutional and recurrent neuron
layers which classifies future asset price trends in three
different classes, using prices and volume historical data as
input. A trend that has been increasingly explored is the use
of complex techniques for preprocessing the input data,
which facilitates the execution of Machine Learning
algorithms and increases its accuracy, since the noisy data
tend to be eliminated, leaving only the most relevant data.
Recently, there has been an increasing tendency to apply
deep neural networks for stock market forecasting. For
example, Zhong and Enke developed a Deep Neural
Network (DNN) to classify future trends of asset prices
(considering two price directions of the next price). The data
set consists of 60 attributes (including returns and technical
indicators) of assets belonging to the SPDR S&P 500 ETF
between June 2003 and May 2013, with daily frequency.
Compared to an Artificial Neural Network (ANN), results
show that, although the DNN presents higher accuracy, the
ANN provides greater returns and lower risks (variance) in a
Stock Market trading simulation. Vignesh collected open,
close, low, and high prices of Yahoo and Microsoft assets
from January 2011 to December 2015 and computed five
technical indicators as features: momentum, volatility, index
momentum, index volatility, stock momentum, stock price
volatility. The paper further compares the SVM and LSTM
models for the problem of binary classification of stock
trends, and the results indicate the better accuracy of the
LSTM algorithm. Nabipour computed ten technical
indicators from opening, close, and low. The high prices of
assets during November 2009 to November 2019 were
used to predict prices by applying Decision Tree, Bagging,
Random Forest, Adaboost, Gradient Boosting, XGBoost,
Artificial Neural Network (ANN), Recurrent Neural Network
(RNN) and Long Short Term Memory (LSTM) algorithms,
concluding that the LSTM algorithm performs better than
the others.
J.M.T.Wu, Z.Li, and G.Srivastava proposes a new two-
dimensional CNN, which uses a matrix composed of
futures, options, opening, closing, high and low prices, and
the transaction volume of each asset in a time series of 120
days of data. For that purpose, 5 Taiwanese assets and 5
United States assets are used. The proposed CNN is used
to predict trends in the movement of financial asset prices.
It considers three different classes: class 1, for days when
the return exceeds 1% (upward trend), −1, for days when
the return is less than −1% (downward trend) and 0,
otherwise (lateral movement). The accuracy of the novel
classifier is compared to that of SVM, Neural Network and
one-dimensional CNN and the results show that the novel
CNN surpasses all other classifiers considering each of the
10 assets. For predicting asset prices and their trends, the
vast majority of methods found in the literature employed
Machine Learning techniques. Regression techniques
are commonly used to price prediction, while
classification techniques are frequently used to predict
trends in asset price movements. An increasing amount of
data used as inputs or features was identified, which implies
the need to use methods to select and preprocess this input
data to filter only the most essential information. Deep
Learning methods are another trend showing promising
results in recent works, despite its great computational
complexity, which makes selecting and preprocessing data
even more important.
3. Financial sentiment analysis
Sentiment analysis is the field of study that analyzes
people’s feelings and moods towards an entity, such as a
product to evaluate whether the opinions are negative or
positive and how negative or positive they are. To perform
sentiment analysis, it is first necessary to collect a large
number of texts, such as those extracted from social media
or news websites, written in a natural language. Thus, the
application of natural language processing methods is also
necessary B.Liu al. Sentiment analysis has been
increasingly applied in several areas of knowledge, such
as computational finance.
In finance, several researchers test the effect of news and
opinions on future asset prices, putting in check the
hypothesis of the efficient market, based on the idea that
news and opinions guide investors, creating trends in
market prices and, thus, providing possibilities for an
investor to obtain profits in stock market trading.
Mittermayer collected press related to stocks in NYSE or
NASDAQ-AMEX, considering only press releases of
companies with a turnover higher than US$5,000,000 per
day. Text reprocessing removes stopwords, numbers, and
less meaningful terms according to the TFxIDF measure.
These press releases are classified as “Good News” “Bad
News” and “No Movers” (for neutral news). Good news
provokes a rise of 3% on the stock price within at least 60
minutes after the press release and increases the price
average by at least 1% during this interval. The other news
is classified as “No Movers.” An SVM model was applied to
predict stock prices using the news sentiment and historical
price series. Results show a recall of approximately 60%
and a better cumulative return and average return per trade
than the random trader in a stock market simulation.
Takahashi analyzed 77,256 analyst reports from Thomson
Financial Web service related to the companies in the
Tokyo Stock Exchange from January 1, 2001 to March 31,
2003. 12 keywords were extracted from the obtained
reports’ title and classified into three classes: Good, Bad,
and Neutral News. Change of monthly consensus earnings
estimate for the next fiscal year (CESFY1) is used as a
numerical indicator for future stock performance. A Machine
Learning model uses a binary classification label
according to monthly earning forecast changes: upward or
downward revision to predict stock movements. The
proposed model performance indicates that the sentiment
is related to stock prices 20 days before and 20 days after
the press.
Takahashi extracted specific news that has a high
possibility to affect the stock prices from news data offered
from JIJI Press about companies in the Tokyo stock
exchange market between August 10, 2006 and November
24, 2006. News is classified into three classes: “Good
News” “Bad News” and “Neutral News” using Naive Bayes,
which presents 78% accuracy. Results show that the
average return for 30 days prior and 30 days after the news
varied for these different categories, and the average daily
return was lower before and after Bad News and greater
before and after Good News. Schumaker and Chen
proposed four stock price prediction models with different
features. The first model applies linear regression using 60-
minute stock quotations before given news, the second one
consists of an SVM that uses only extracted article terms
for its prediction, the third one is an SVM that uses
extracted article terms and the stock price at the time the
article was released, and the fourth one is an SVM that
uses extracted terms and a regressed estimate provided by
the linear regression model of the stock price 20 minutes
after the news. The paper concludes that the third model
performs the best.
Schumaker and Chen compared their methodology to the
top 10 quant funds operating for a full year at the time of the
study. The paper concludes that their methodology
provides the fourth-best return among the top 10 and the
best return when compared with only the quant funds for
companies in S&P 500. In relation to the works that perform
analysis of feelings, it is observed that the oldest studies
used news about a certain asset to assign a mood to it. A
disadvantage of using news is that it tends to be more
neutral, which makes it difficult to differentiate between
good and bad ones, in addition to the smaller amount, in
relation to comments on social media.
Bollen employed lexicon dictionaries to perform sentiment
analysis to public tweets recorded from February 28 to
December 19, 2008 from stocks in DJIA. After stopwords
and punctuation removal, a one-dimension public mood
time series was generated by OpinionFinder, which
classifies text moods as positive or negative. Besides,
seven dimensions of the public mood time series were
generated by GPOMS, each representing a different aspect
of the public’s mood on a given day: Calm, Alert, Sure,
Vital, Kind, and Happy. Bivariate Granger causality analysis
showed that only the GPOMS’s Calm dimension correlates
(linear) with the stock prices series. Self-Organizing Fuzzy
Neural Network was proposed to predict stock movements
using three previous historical daily prices and permutations
of the seven mood indicators as features. Self-Organizing
Fuzzy Neural Network predictions match Granger analysis
results, showing that the calm indicator increases the
prediction accuracy, but the other indicators do not improve
the prediction using only the stock price series.
B.Wang, H.Huang, and X.Wang collected six security
companies’ quarterly and annual reports and their ROEs
time series and selected terms that occurred three times or
more. The paper performed future asset prices forecasting
using ARIMA and SVR, which uses historical return series,
and historical term series from reports, respectively. Results
reveal that the hybrid model, combining ARIMA and SVR,
presents the best forecasting accuracy.
Smailović used a data set of 1,600,000 (800,000 positive
and 800,000 negative) tweets collected and labeled by
Stanford University. The preprocessing step removed
usernames and links, identified occurrences of more than
two letters in a word and changed it to only one letter, drew
explicit negation words, exclamation, and question marks,
besides performing text tokenization, removal of
stopwords, stemming, N-gram construction (size 2), and
removed words that occurred only once. Two sentiment
analysis models were applied. The first classifies the
documents into two classes: negative and positive, while
the second classifies them into three categories: negative,
neutral, and positive. Sentiment analysis presents an
accuracy close to 80%. The study then analyzed a
correlation between 152,572 tweets discussing stock
relevant information about eight companies in nine
months in 2011 and the stock closing prices of these eight
companies for the same period. A statistical hypothesis test
for stationary time series was performed to determine
the linear correlation level between the sentiment and stock
closing price and whether one contains predictive
information about the other (Granger causality analysis).
Results indicate that tweets’ sentiment can predict stock
price movements for several assets, and the introduced
neutral class can improve the correlation between the
opinionated tweets and the stock closing price in certain
situations.
X.Li, H.Xie, L.Chen, J.Wang, and X.Deng examined a
news archive from FINET, containing both company-
specific and market-related news VOLUME 9, 2021 from
January 2003 to March 2008. The news was classified by
Harvard IV-4 sentiment dictionary (HVD), considering 15
sentiment dimensions, and Loughran–McDonald financial
sentiment dictionary (LMD), considering six sentiment
dimensions. An SVM model was developed to predict stock
movement using historic daily open-to-close price return as
a feature, in addition to the news sentiment moods values.
For the SVM classification, the stocks were labeled into
three classes according to return value: positive, neutral,
and negative. The paper concludes that sentiment
analysis helps to improve prediction accuracy.
T.H.Nguyen, K.Shirai, and J.Velcin proposed different
methods to classify untagged messages in five classes:
Strong Buy, Buy, Hold, Sell, and Strong Sell. These
messages were taken from 18 message boards of the 18
stocks from Yahoo Finance Message Board for one year
(July 23, 2012 to July 19, 2013). Historical daily adjusted
close prices extracted from Yahoo Finance for 18 stocks
were also used as features for a stock movement
prediction performed by an SVM algorithm, in which its
labels represent actual price movement (up or down).
Results indicate that the incorporation of the sentiment
analysis improves the prediction accuracy.
N.Oliveira, P.Cortez, and N.Areal employed all tweets
containing cashtags of all stocks traded in US stock
markets from December 22, 2012 to March 27, 2015,
before the Stanford CoreNLP tool was applied to execute
common natural language processing methods, including
tokenization, Part Of Speech (POS) tagging, and
lemmatization. Sentiment analysis classification considers
three sentiment classes: “bullish”, “bearish”, and “neutral”.
The work measured the correlation between Twitter
sentiment indicators and two popular survey sentiment
indicators: the American Association of Individual
Investors (AAII) and Investors Intelligence (II). A strong
correlation between them indicates that the microblogging
sentiment indicator can provide important information.
Pagolu collected a total of 250,000 tweets about Microsoft
from August 31, 2015 to August 25, 2016, extracted from
Twitter API, in addition to stock opening and closing prices
of Microsoft in the same period obtained from Yahoo!
Finance. Tweet preprocessing included tokenization,
stopwords removal, and regex matching for removing
special characters. Sentiment analysis considered three
different classes for each tweet: positive, neutral, and
negative. A total of 3,216 tweets were examined,
labeled, then used to train a Random Forest classification
model. The trained model predicts other tweets’
sentiments. After performing sentiment analysis, a
classification model used positive, neutral, and negative
tweets in 3 days as features to perform a binary stock
movement classification. Results show around 70%
accuracy for sentiment classification and around 70%
accuracy for stock movement prediction.
Batra and Daudpota retrieved around 300,000 tweets about
Apple from StockTwits during 2010-2017, each tweet
composed by its content, date, and user sentiment. Tweet
preprocessing included tokenization and removal of
stopwords and Twitter symbols. An SVM model was trained
for sentiment prediction, in which daily sentiment is
positive if there are more positive tweets than negative
ones and is negative if otherwise. Besides, Apple’s
historical price data were extracted from Yahoo Finance
from 2010 to 2017 and used for an SVM classifier that
predicts binary class stock movement (up or down). For
sentiment prediction, the achieved test accuracy was
63.5%, 75.3% recall, and 76.8% precision, and for stock
movement prediction, the completed test accuracy was
76.68%, 100% recall, and 69.5% precision.
Allen, McAleer, and Singh assigned daily sentiment
scores for the DJIA market by accumulating high-
frequency sentiment scores of the DJIA’s constituents
obtained from the TRNA dataset from January 2006 to
October 2012. DJIA’s constituents were tagged according
to three classes: positive, neutral, or negative (and a
probability associated with each one). Then, the daily
sentiment was computed as the average sentiment
prediction weighted by probabilities of each of these
predictions to be corrected. The paper applied linear and
quantile regression to these daily scores, analyzing its
correlation with the Thomson Tick History database’s
stock prices. The regressions considered daily sentiments
up to 5 days before a given price and that price. Results
demonstrate that daily financial news sentiment can predict
prices, as they show a significant correlation.
Mohan considered news articles for the S&P 500
companies from February 2013 to March 2017 from
international daily news websites, comprising a total of
265463 articles, in addition to daily closing stock prices of
these assets for the same interval. The paper developed an
ARIMA model for price regression, considering only the
stock price data. The Facebook Prophet algorithm was
used to predict future stock prices using historical closing
prices. Also proposed three Recurrent Neural Network Long
Short Term Memory (RNN LSTM) methods for stock price
prediction. The first method uses only the price data as
features; the second uses historical closing prices and the
textual polarity (negative or positive) for each asset
provided by the natural language toolkit (NLTK); and the
third uses the prices and textual data as features. The
paper concludes that there is a strong relationship between
stock prices and financial news articles, as the RNN LSTM
models that use textual data or sentiment information
perform better than models that use only the prices data.
Zadrozny first captured the semantic information of stock-
related tweets texts by applying LSTM models that produce
a textual representation for each set of texts about a
particular asset on a given day. Subsequently, another
LSTM model followed by a regression MLP was used to
predict future asset prices. Reference [41] proposed using
the Convolutional Neural Network (CNN) algorithm to
extract features from the set of words formed by the text of
each tweet. The paper also presented the optimization of
the generated features using the Cuckoo Search (CS)
Algorithm and, finally, the classification of the texts’ polarity
(positive or negative) using a Neural Network algorithm.
Results showed that this proposed approach surpasses
previous approaches found in literature, considering the
accuracy measure. Xing, Cambria, and Zhang proposed
a Sentiment-aware volatility forecasting (SAVING) method.
The sentiment analysis considered social media
messages on StockTwits for 10 US stocks from August 14,
2017 to August 22, 2018. The work assessed the polarity of
feeling for a given asset taking into account the intensity
and quantity of messages about the asset. This polarity
was coupled to the volatility prediction models as
variables or features. For this, the paper used the Recurrent
Neural Network (RNN) algorithm. The proposed SAVING
model was compared with volatility prediction models that
do not use sentiment analysis: GARCH, EGARCH, TARCH,
GJR, GP-vol, VRNN, NSVM and LSTM (Long Short Term
Memory) models. A t-test compared SAVING with each
other method and results indicate that SAVING statistically
outperforms all other methods, except EGARCH, TARCH
and GJR, in addition to not being dominated by any other
method.
References
Aldridge, I. (2009). High-frequency trading: a practical
guide to algorithmic strategies and trading systems, Vol.
459, John Wiley & Sons, Hoboken, New Jersey.
Aldridge, I. (2013). High-frequency trading: a practical
guide to algorithmic strategies and trading systems, Second
Edition, John Wiley & Sons, Hoboken, New Jersey.
Almgren, R. and Chriss, N. (2001). Optimal execution of
portfolio transactions, Journal of Risk 3: 5–40.
Almgren, R. and Lorenz, J. (2006). Bayesian adaptive
trading with a daily cycle, The Journal of Trading 1(4): 38–
46.
Avellaneda, M. and Stoikov, S. (2008). High-frequency
trading in a limit order book, Quantitative Finance 8(3):
217–224.
Bialkowski, J., Darolles, S. and Le Fol, G. (2008). Improving
vwap strategies: a dynamic volume approach, Journal of
Banking & Finance 32(9): 1709–1722.
Brogaard, J. (2010). High frequency trading and its impact
on market quality, Northwestern University Kellogg School
of Management Working Paper 66.
Chan, N. T. and Shelton, C. (2001). An electronic market-
maker, Technical Report 195, Massachusetts Institute of
Technology – Artificial Intelligence Laboratory.
Das, S. (2005). A learning market-maker in the glosten–
milgrom model, Quantitative Finance 5(2): 169–180.
Das, S. (2008). The effects of market-making on price
dynamics, Proceedings of the 7th international joint
conference on Autonomous agents and multiagent
systems-Volume 2, International Foundation for
Autonomous Agents and Multiagent Systems, pp. 887–
894.
Fabozzi, F. J., Focardi, S. M., Kolm, P. N. et al. (2010).
Quantitative equity investing: Techniques and strategies,
John Wiley & Sons.
Glosten, L. R. and Harris, L. E. (1988). Estimating the
components of the bid/ask spread, Journal of financial
Economics 21(1): 123–142.
Glosten, L. R. and Milgrom, P. R. (1985). Bid, ask and
transaction prices in a specialist market with
heterogeneously informed traders, Journal of financial
economics 14(1): 71–100.
Government Office For Science (2012). Foresight: The
future of computer trading in financial markets, Technical
report.
Harris, L. (2002). Trading and exchanges: Market
microstructure for practitioners, Oxford University Press,
USA.
Johnson, B. (2010). Algorithmic Trading & DMA: An
introduction to direct access trading strategies, 4Myeloma
Press, London.
Lin, H. (2006). Essays in market microstructure, PhD thesis,
University of Warwick.
Menkveld, A. (2011). Electronic trading and market
structure, UK Government Office for Science, Foresight
Driver Review - The Future of Computer Trading in
Financial Markets.
Vuorenmaa, T. (2012). The good, the bad, and the ugly of
automated high-frequency trading, Technical report, Valo
Research and Trading.
Yong Hu Kang Liu Xiangzhou Zhang Lijun Su E.W.T. Ngai
Mei Liu (2015) Application of Evolutionary Computation
for Rule Discovery in Stock Algorithmic Trading: A
Literature Review
FERNANDO G. D. C. FERREIRA, AMIR H. GANDOMI
AND RODRIGO T. N. CARDOSO: Artificial Intelligence
Applied to Stock Market Trading: A Review
Mandes, Alexandru (2016) : Algorithmic and high-
frequency trading strategies: A literature review, MAGKS
Joint Discussion Paper Series in Economics, No. 25-2016,
Philipps-University Marburg, School of Business and
Economics, Marburg
Atsalakis, G.S. and K.P. Valavanis, Surveying stock market
forecasting techniques–Part II: Soft computing methods.
Expert Systems with Applications, 2009. 36(3): p. 5932-
5941.
Guresen, E., G. Kayakutlu and T.U. Daim, Using artificial
neural network models in stock market index prediction.
Expert Systems with Applications, 2011. 38(8): p. 10389-
10397.
Bahrammirzaee, A., A comparative survey of artificial
intelligence applications in finance: artificial neural
networks, expert system and hybrid intelligent systems.
Neural Computing and Applications, 2010. 19(8): p. 1165-
1195.
Bodie, Z., A. Kane and A. Marcus, Investments. 6th ed.
2005, New York: Mcgraw-Hill.
CFA Institute, Ethical and Professional Standards and
Quantitative Methods. 1st ed. CFA curriculum 2013 level1.
2013, London: Pearson.
Wikipedia, Fundamental analysis. 2013.
Colby, R.W., The Encyclopedia of Technical Market
Indicators. 2nd ed. 2002, New York: McGraw-Hill. Fidelity
Mutual Fund,
Fidelity Mutual Fund & Investment Management. 2013.
Goldman Sachs. 2013.
NASDAQ, NASDAQ Stock Market-Stock Quotes-Stock
Exchange News. 2013.
Wikipedia, Technical analysis. 2013.
Yahoo Finance, Yahoo Finance Charts User Guide. 2013.
Market Technicians Association, The MTA Knowledge
Base. 2014.
Beni, G. and J. Wang, Swarm intelligence in cellular robotic
systems, in Robots and Biological Systems: Towards a New
Bionics? P. Dario, G. Sandini and P. Aebischer, P. Dario,
G. Sandini and P. Aebischer^Editors. 1993, Springer:
Berlin Heidelberg. p. 703-712.
Bonabeau, E., M. Dorigo and G. Theraulaz, Swarm
intelligence: from natural to artificial systems. 1st ed. 1999,
New York: Oxford university press.
M. Chang. (2018). How A.I. Traders Will Dominate Hedge
Fund Industry. [Online]. Available:
https://www.youtube.com/watch?v=lzaBbQKUtAA
H. Markowitz, “Portfolio selection”, J. Finance, vol. 7, no. 1,
pp. 77–91, 1952. [Online]. Available:
https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1540-
6261.1952.tb01525.x
R. T. Rockafellar and S. Uryasev, “Conditional value-at-risk
for general loss distributions”, J. Banking Finance, vol. 26,
no. 7, pp. 1443–1471, Jul. 2002. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S03784266
02002716
M. G. Speranza, “A heuristic algorithm for a portfolio
optimization model applied to the milan stock market”,
Comput. Oper. Res., vol. 23, no. 5, pp. 433–441, May 1996.
T-J.Chang, N.Meade, J.E.Beasley, and Y. M. Sharaiha,
“Heuristics for cardinality constrained portfolio optimisation,”
Comput. Oper. Res., vol.27, no.13, pp. 1271–1302, Nov.
2000. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S03050548
9900074X
Y.Crama and M.Schyns, “Simulated annealing for
complex portfolio selection problems”, Eur. J.Oper. Res.,
vol.150, no.3, pp.546–571, Nov. 2003. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S03772217
02007841
M. Ehrgott, K. Klamroth, and C. Schwehm, “An MCDM
approach to portfolio optimization,” Eur. J. Oper. Res.,
vol.155, no.3, pp. 752–770, Jun. 2004.
R. Subbu, P. Bonissone, N. Eklund, S. Bollapragada, and
K. Chalermkraivuth, “Multiobjective financial portfolio
design: A hybrid evolutionary approach,”
R. Moral-Escudero, R. Ruiz-Torrubiano, and A. Suárez,
“Selection of optimal investment portfolios with cardinality
constraints,” in Proc. IEEE Int. Conf. Evol. Comput., Jul.
2006, pp. 2382–2388.
A. Fernández and S. Gómez, ‘‘Portfolio selection using
neural networks,’’ Comput. Oper. Res., vol. 34, no. 4, pp.
1177–1191, Apr. 2007. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S03050548
05002042
G.Hassan and C.D.Clack, “Multiobjective robustness for
portfolio optimization in volatile environments,” in Proc. 10th
Annu. Conf. Genetic Evol. Comput., 2008, pp. 1507–1514.
T-J.Chang, S-C.Yang, and K-J.Chang, “Portfolio
optimization problems in different risk measures using
genetic algorithm,” Expert Syst. Appl., vol. 36, no. 7, pp.
10529–10537, Sep. 2009.
N.M.Pindoriya, S.N.Singh, and S.K.Singh, “Multi-objective
mean–variance–skewness model for generation portfolio
allocation in electricity markets”, Electr. Power Syst. Res.,
vol. 80, no. 10, pp. 1314–1321, Oct. 2010.
H.Zhu, Y.Wang, K.Wang, and C.Y.Chen, “Particle Swarm
Optimization (PSO) for the constrained portfolio
optimization problem,” Expert Syst. Appl. vol. 38, no. 8, pp.
10161–10169, 2011.
W-G.Zhang, Y-J.Liu, and W-J.Xu, “A possibilistic mean-
semivariance-entropy model for multi-period portfolio
selection with transaction costs”, Eur. J. Oper. Res., vol.
222, no. 2, pp. 341–349, Oct. 2012.
A.Ponsich, A.L.Jaimes, and C.A.C.Coello, “A survey on
multiobjective evolutionary algorithms for the solution of
the portfolio optimization problem and other finance and
economics applications,’’ IEEE Trans. Evol. Comput., vol.
17, no. 3, pp. 321–344, Jun. 2013.
R.Mansini, W.Ogryczak, and M.G.Speranza, “Twenty years
of linear programming based portfolio optimization”, Eur.
J. Oper. Res., vol. 234, no. 2, pp. 518–535, Apr. 2014.
C.Chen and Y.Wei, “Robust multiobjective portfolio
optimization: A set order relations approach”, J. Combinat.
Optim., vol. 38, no. 1, pp. 21–49, Jul. 2019, doi:
10.1007/s10878-018-0364-9.
W.Chen, “Artificial bee colony algorithm for constrained
possibilistic portfolio optimization problem,” Phys. A, Stat.
Mech. Appl., vol. 429, pp. 125–139, Jul. 2015.
S.M.Seyedhosseini, M.J.Esfahani, and M.Ghaffari, “A novel
hybrid algorithm based on a harmony search and artificial
bee colony for solving a portfolio optimization problem
using a mean-semi variance approach”, J. Central South
Univ., vol. 23, no. 1, pp. 181–188, Jan. 2016.
C.B.Kalayci, O.Ertenlice, H.Akyer, and H.Aygoren, “An
artificial bee colony algorithm with feasibility enforcement
and infeasibility toleration procedures for cardinality
constrained portfolio optimization”, Expert Syst. Appl., vol.
85, pp. 61–75, Nov. 2017.
O.Ertenlice and C.B.Kalayci, “A survey of swarm
intelligence for portfolio optimization: Algorithms and
applications”, Swarm Evol. Comput., vol. 39, pp. 36–52,
Apr. 2018.
R.Kizys, A.Juan, B.Sawik, and L.Calvet, “A biased-
randomized iterated local search algorithm for rich portfolio
optimization”, Appl. Sci., vol. 9, no. 17, p. 3509, Aug. 2019.
Y.L.T.V.Silva, A.B.Herthel, and A.Subramanian, “A multi-
objective evolutionary algorithm for a class of mean-
variance portfolio selection problems”, Expert Syst. Appl.,
vol. 133, pp. 225–241, Nov. 2019.
M.Kaucic, M.Moradi, and M.Mirzazadeh, “Portfolio
optimization by improved NSGA-II and SPEA 2 based on
different risk measures”, Financial Innov., vol. 5, no. 1, pp.
1–28, Dec. 2019, doi: 10.1186/s40854- 019-0140-6.
G.H.M.Mendonça, F.G.D.C.Ferreira, R.T.N.Cardoso, and
F.V.C.Martins, “Multi-attribute decision making applied to
financial portfolio optimization problem”, Expert Syst. Appl.,
vol. 158, Nov. 2020, Art. no. 113527. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S09574174
20303511
A.W.Lo, “Efficient markets hypothesis”, in The New
Palgrave Dictionary of Economics. Basingstoke, U.K.:
Palgrave Macmillan, 2007.
J.J.Murphy, Technical Analysis of the Financial Markets: A
Comprehensive Guide to Trading Methods and Applications
(New York Institute of Finance Series). New York, NY, USA:
New York Institute of Finance, 1999. [Online]. Available:
https://books.google.com.br/books?id=5zhXEqdr_IcC
R.H.Golan and W.Ziarko, “A methodology for stock market
analysis utilizing rough set theory,’’ in Proc. Conf. Comput.
Intell. Financial Eng., Apr. 1995, pp. 32–40.
K.Schierholt and C.H.Dagli, “Stock market prediction using
different neural network classification architectures”, in
Proc. IEEE/IAFE Conf. Comput. Intell. Financial Eng., Mar.
1996, pp. 72–78.
S.H.Kim and S.H.Chun, “Graded forecasting using an array
of bipolar predictions: Application of probabilistic neural
networks to a stock market index”, Int. J. Forecasting, vol.
14, no. 3, pp. 323–337, Sep. 1998.
Z.Zhanggui, H.Yan, and A.M.Fu, “A new stock price
prediction method based on pattern classification”, in Proc.
Int. Joint Conf. Neural Netw., vol. 6, 1999, pp. 3866–3870.
M.A.Kaboudan, “Genetic programming prediction of stock
prices”, Comput. Econ., vol. 16, no. 3, pp. 207–236, 2000.
R.J.Kuo, C.H.Chen, and Y.C.Hwang, “An intelligent stock
trading decision support system through integration of
genetic algorithm based fuzzy neural network and artificial
neural network”, Fuzzy Sets Syst., vol. 118, no. 1, pp. 21–
45, Feb. 2001. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S01650114
98003996
Y.Wang, “Predicting stock price using fuzzy grey prediction
system”, Expert Syst. Appl., vol. 22, no. 1, pp. 33–38, Jan.
2002. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S09574174
01000471
Y.Wang, “Mining stock price using fuzzy rough set system”,
Expert Syst. Appl., vol. 24, no. 1, pp. 13–23, Jan. 2003.
J-Y.Potvin, P.Soriano, and M.Vallée, “Generating trading
rules on the stock markets with genetic programming”,
Comput. Oper. Res., vol. 31, no. 7, pp. 1033–1047, Jun.
2004.
W.Huang, Y.Nakamori, and S-Y.Wang, “Forecasting stock
market movement direction with support vector
machine”, Comput. Oper. Res., vol. 32, no. 10, pp. 2513–
2522, Oct. 2005.
M-J.Kim, S-H.Min, and I.Han, “An evolutionary approach
to the combination of multiple classifiers to predict a stock
price index”, Expert Syst. Appl., vol. 31, no. 2, pp. 241–247,
Aug. 2006.
T.H.Roh, “Forecasting the volatility of stock price index”,
Expert Syst. Appl., vol. 33, no. 4, pp. 916–922, Nov. 2007.
C.Huang, D.Yang, and Y.Chuang, “Application of
wrapper approach and composite classifier to the stock
trend prediction”, Expert Syst. Appl., vol. 34, no. 4, pp.
2870–2878, May 2008.
C-L.Huang and C-Y.Tsai, “A hybrid SOFM-SVR with a filter-
based feature selection for stock market forecasting”,
Expert Syst. Appl., vol. 36, no. 2, pp. 1529–1539, Mar.
2009.
E.Hadavandi, H.Shavandi, and A.Ghanbari, “Integration of
genetic fuzzy systems and artificial neural networks for
stock price forecasting”, Knowl-Based Syst., vol. 23, no. 8,
pp. 800–808, Dec. 2010. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S09507051
10000857
C-M.Hsu, “A hybrid procedure for stock price prediction
by integrating self-organizing map and genetic
programming”, Expert Syst. Appl., vol. 38, no. 11, pp.
14026–14036, May 2011.
S.Asadi, E.Hadavandi, F.Mehmanpazir, and
M.M.Nakhostin, “Hybridization of evolutionary Levenberg–
Marquardt neural networks and data pre-processing for
stock market prediction”, Knowl.-Based Syst., vol. 35, pp.
245–258, Nov. 2012.
C-M.Hsu, “A hybrid procedure with feature selection for
resolving stock/futures price forecasting problems”,
Neural Comput. Appl., vol. 22, nos. 3–4, pp. 651–671, Mar.
2013.
A.Booth, E.Gerding, and F.McGroarty, “Automated
trading with performance weighted random forests and
seasonality”, Expert Syst. Appl., vol. 41, no. 8, pp. 3651–
3661, Jun. 2014.
J.Patel, S.Shah, P.Thakkar, and K.Kotecha, “Predicting
stock and stock price index movement using trend
deterministic data preparation and machine learning
techniques”, Expert Syst. Appl., vol. 42, no. 1, pp. 259–268,
Jan. 2015. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S09574174
14004473
R.C.Cavalcante, R.C.Brasileiro, V.L.F.Souza,
J.P.Nobrega, and A.L.I.Oliveira, “Computational
intelligence and financial markets: A survey and future
directions”, Expert Syst. Appl., vol. 55, pp. 194–211, Aug.
2016.
E.Chong, C.Han, and F.C.Park, “Deep learning networks for
stock market analysis and prediction: Methodology, data
representations, and case studies”, Expert Syst. Appl., vol.
83, pp. 187–205, Oct. 2017.
T.Fischer and C.Krauss, “Deep learning with long short-
term memory networks for financial market predictions”,
Eur. J. Oper. Res., vol. 270, no. 2, pp. 654–669, Oct. 2018.
W.Long, Z.Lu, and L.Cui, “Deep learning-based feature
engineering for stock price movement prediction”, Knowl-
Based Syst., vol. 164, pp. 163–173, Jan. 2019.
X.Zhong and D.Enke, “Predicting the daily return direction
of the stock market using hybrid machine learning
algorithms”, Financial Innov., vol. 5, no. 1, pp. 1–20, Dec.
2019. C.K.Vignesh, “Applying machine learning models in
stock market prediction”, EPRA Int. J. Res. Develop., vol. 5,
no. 4, pp. 395–398, Apr. 2020.
M.Nabipour, P.Nayyeri, H.Jabani, A.Mosavi, and
E.Salwana, “Deep learning for stock market prediction”,
Entropy, vol. 22, no. 8, p. 840, Jul. 2020.
J.M.T.Wu, Z.Li,G.Srivastava, M.H.Tasi, and J.C-W.Lin, “A
graph- based convolutional neural network stock price
prediction with leading indicators”, Softw Pract Exper., vol.
2020, pp. 1–17, Oct. 2020.
B.Liu, “Sentiment analysis and opinion mining”, Synth.
Lectures Hum. Lang. Technol., vol. 5, no. 1, pp. 1–167,
2012.
S.Takahashi, M.Takahashi, H.Takahashi, and K.Tsuda,
Analysis of Stock Price Return Using Textual Data and
Numerical Data Through Text Mining (Lecture Notes in
Computer Science: Lecture Notes in Artificial Intelligence
and Lecture Notes in Bioinformatics), vol. 4252. Berlin,
Germany: Springer, 2006, pp. 310–316.
R.P.Schumaker and H.Chen, “Textual analysis of stock
market prediction using breaking financial news: The AZFin
text system”, ACM Trans. Inf. Syst., vol. 27, no. 2, pp. 1–19,
Feb. 2009.
R.P.Schumaker and H.Chen, “A discrete stock price
prediction engine based on financial news”, Computer, vol.
43, no. 1, pp. 51–56, Jan. 2010.
J.Bollen, H.Mao, and X.Zeng, “Twitter mood predicts the
stock market”, J. Comput. Sci., vol. 2, no. 1, pp. 1–8, Mar.
2011. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S18777503
1100007X
B.Wang, H.Huang, and X.Wang, “A novel text mining
approach to financial time series forecasting”,
Neurocomputing, vol. 83, pp. 136–145, Apr. 2012.
J.Smailović, M.Grčar, N.Lavrač, and M.Žnidaršič,
Predictive Sentiment Analysis of Tweets: A Stock Market
Application (Lecture Notes in Computer Science: Lecture
Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics), vol. 7947. Berlin, Germany: Springer, 2013,
pp. 77–88.
X.Li, H.Xie, L.Chen, J.Wang, and X.Deng, “News impact on
stock price return via sentiment analysis”, Knowl-Based
Syst., vol. 69, no. 1, pp. 14–23, Oct. 2014.
T.H.Nguyen, K.Shirai, and J.Velcin, “Sentiment analysis
on social media for stock movement prediction”, Expert
Syst. Appl., vol. 42, no. 24, pp. 9603–9611, Dec. 2015.
N.Oliveira, P.Cortez, and N.Areal, “Stock market sentiment
lexicon acquisition using microblogging data and statistical
measures”, Decis. Support Syst., vol. 85, pp. 62–73, May
2016.
V.Pagolu, K.Reddy, G.Panda, and B.Majhi, “Sentiment
analysis of Twitter data for predicting stock market
movements,” in Proc. Int. Conf. Signal Process., Commun.,
Power Embedded Syst., 2017, pp. 1345–1350.
R.Batra and S.M.Daudpota, “Integrating StockTwits with
sentiment analysis for better prediction of stock price
movement”, in Proc. Int. Conf. Comput., Math. Eng.
Technol., Jan. 2018, pp. 1–5. D.E.Allen, M.McAleer, and
A.K.Singh, “Daily market news sentiment and stock prices,”
Appl. Econ., vol. 51, no. 30, pp. 3212–3235, Jun. 2019.
S.Mohan, S.Mullapudi, S.Sammeta, P.Vijayvergia, and
D.C.Anastasiu, “Stock price prediction using news
sentiment analysis”, in Proc. IEEE 5th Int. Conf. Big Data
Comput. Service Appl. (BigDataService), Apr. 2019, pp.
205–208.
W.Zadrozny, “A comparison of neural network methods for
accurate sentiment analysis of stock market tweets”, in
Proc. ECML PKDD Workshops, MIDAS PAP, vol. 11054,
Dublin, Ireland. Cham, Switzerland: Springer, Sep. 2019, p.
51.
F.Z.Xing, E.Cambria, and Y.Zhang, “Sentiment-aware
volatility forecasting”, Knowl-Based Syst., vol. 176, pp. 68–
76, Jul. 2019.
Iain Clacher, David Hillier & Sunthare Lhaopadchan:
Corporate insider trading: A literature review, 2014.
MANNE, H. 1966. Insider trading and the stock market.
New York: Free Press.
JENSEN, M., and MECKLING, W. 1976. Theory of the firm:
Managerial behaviour, agency costs and ownership
structure. Journal of Financial Economics 3: 305-360.
FASB. 1978. Statement of Financial Accounting
Concepts n.o 1: Objectives of Financial Reporting by
Business Enterprises. Financial Accounting Standards
Board. Connecticut 06856.
JOHN, K., and MISHRA, B. 1990. Information content of
insider trading around corporate announcements: The case
of capital expenditures. Journal of Finance 45 (3): 835-854.
JOHN, K., and LANG, L. 1991. Insider trading around
dividend announcements: Theory and evidence. Journal of
Finance 46: 1361-1390.
ZHANG, G. 2001. Regulated managerial insider trading as
a mechanism to facilitate shareholder control, Journal of
Business Finance & Accounting 28 (1-2): 35-62.
CHAU, M., and VAYANOS, D. 2008. Strong-form efficiency
with monopolistic insiders. The Review of Financial Studies
21 (5), 2275-2306.
FISHMAN, M. J., and HAGERTY, K.M. 1992. Insider trading
and the effi ciency of stock prices. RAND Journal of
Economics 23 (1): 106-122.
BHATTACHARYA, S., and NICODANO, G. 2001. Insider
trading, investment, and liquidity: a welfare analysis. The
Journal of Finance 56 (3): 1141-1156.
LELAND, H. 1992. Insider Trading: Should It be
Prohibited? Journal of Political Economy 859-887.
CHO, J., and SHAUB, M. 1991. The Consequences of
Insider Trading and the Role of Academic Research.
Business and Professional Ethics Journal 10: 83–98.
LORIE, J.H., and NIEDERHOFFER, V. 1968. Predictive
and statistical properties of insider trading. Journal of Law
and Economics 11: 35-53.
PRATT, S.P., and DE VERE, C.W. 1970. Relationship
between insider trading and rates of return for NYSE
common stocks, 1960-66, in LORIE, J., and BREALEY, R.
Eds.: Modern Developments in Investment Management
Praeger, New York.
JAFFE, J.F. 1974a. The effect of regulation changes on
insider trading. Bell Journal of Economics and Management
Science 5: 93-121.
FINNERTY, J.E. 1976a. Insiders’ activity and inside
information: A multivariate analysis. Journal of Financial
and Quantitative Analysis 11: 205-216.
KING, M., and ROELL, A. 1988. Insider trading. Economic
Policy 7: 163-193.
POPE, P.; MORRIS, R.C., and PEEL, D. 1990. Insider
trading: some evidence on market efficiency and directors
share dealings in Great Britain. Journal of Business Finance
and Accounting 17: 359-380.
CHEUK, M.Y.; FAN, D.K., and SO, R.W. 2006. Insider
trading in Hong Kong: some stylized facts. Pacific-Basin
Finance Journal 14: 73-90.
DEL BRIO, E.B.; MIGUEL, A. DE., and PEROTE, J. 2002.
An investigation of insider trading profi ts in the Spanish
stock market. The Quarterly Review of Economics and
Finance 42 (1), 73-94.
SEYHUN, H.N. 1986a. Insiders profits, costs of trading,
and market efficiency. Journal of Financial Economics 16:
189-212.
Jenter, D. 2005. Market timing and managerial portfolio
decisions. The Journal of Finance 60 (4): 1903-1949.
ROZEFF, M.S., and ZAMAN, M.A. 1988. Market efficiency
and insider trading: New evidence. Journal of Business
61(1): 25-44.
GREGORY, A.; MATATKO, J., and TONKS, I. 1997.
Detecting information from directors trades: signal definition
and variable size effects (forthcoming). Journal of Business
Finance and Accounting 24(3): 309-342.
GREGORY, A.; MATATKO, J.; TONKS, I., and PURKIS , R.
1997. UK directors trading: The impact of dealings in
smaller firms. The Economic Journal 104 37-53.
BARCLAY, M.J., and WARNER, J. 1993. Stealth trading
and volatility (which trades move prices? Journal of
Financial Economics 34: 281-305.
FRIEDERICH, S.; GREGORY, A.; MATATKO, J., and
TONKS, I. 2002. Short-run returns around the trades of
corporate insiders on the London Stock Exchange.
European Financial Management 8 (1): 7-30.
GARFINKEL, J.A., and NIMALENDRAN, M. 2003. Market
structure and trader anonymity: an analysis of insider
trading. Journal of Financial and Quantitative Analysis 38
(3): 591-610.
LIN, J., and HOWE, J.S. 1990. Insider trading in the OTC
market. Journal of Finance 45: 1273 1284.
FISHE, R.P.H., and ROBE, M.A. 2004. The impact of illegal
insider trading in dealer and specialist markets: evidence
from a natural experiment. Journal of Financial Economics
71 (4): 461-488.
CHENG, L.; FIRTH, M.; LEUNG, T.Y., and RUI, O. 2006.
The effects of insider trading on liquidity. Pacific-Basin
Finance Journal 14: 467-483.
CHUNG, K.H., and CHAROENWONG, C. 1998. Insider
trading and the bid-ask spread. The Financial Review 33: 1-
20.
DOLGOPOLOV, S. 2004. Insider Trading and the Bid-Ask
Spread: A Critical Evaluation of Adverse Selection in Market
Making. Capital University Law Review 33: 83–180.
FIDRMUC, J.P.; GOERGEN, M., and RENNEBOOG, L.
2006. Insider trading, new releases and ownership
concentration. The Journal of Finance 61 (6): 2931-2973.
ZHANG, W.; CAHAN, S.F., and ALLEN, A.C. 2005. Insider
trading and pay-performance sensitivity: an empirical
analysis. Journal of Business Finance & Accounting 32 (9-
10): 1887-1919.
ACHARYA, V.V., and JOHNSON, T.C. 2007. Insider trading
in credit derivatives. Journal of Financial Economics 84 (2):
110-141.
BETTIS, J.C.; COLES, J.L., and LEMMON, M.L. 2000.
Corporate policies restricting trading by insiders. Journal of
Financial Economics 57 (2): 191-200.
GOSNELL, T.; KEOWN, A.J., and PINKERTON, J.M. 1992.
Bankruptcy and insider trading: differences between
exchange-listed and OTC firms. The Journal of Finance 47:
349-362.
ELLIOT, J.; MORSE, D., and RICHARDSON, G. 1984. The
association between insider trading and information
announcements. Rand Journal of Economics 15 (4): 521-
536.
CHAKRAVARTY, S.; GULEN, H.; MAYHEW, S., and
MCCONNELL, J.J. 1997. An analysis of prices, bid/ask
spreads, and bid and ask depths surrounding Ivan Boesky’s
illegal trading in Carnation’s stock. Financial Management
26 (2): 18-34.
MEULBROEK, L., and HART, C. 1997. The effect of illegal
trading on takeover premia. European Finance Review (l):
51 80.
PENMAN, S.H. 1982. Insider trading and the
dissemination of fi rms forecast information. Journal of
Business 55: 479 504.
GIVOLY, D., and PALMON, D. 1985. Insider trading and the
exploitation of inside information: Some empirical
evidence. Journal of Business 58: 69-87.
SIVAKUMAR, K., and WAYMIRE, G. 1994. Insider trading
following material news events: Evidence form earnings.
Financial Management (Spring): 23-32.
LUSTGARTEN, S., and MANDE, V. 1995. Financial
analysts earnings forecasts and insider trading. Journal of
Accounting and Public Policy (14): 233-261.
UDPA, S. 1996. Insider trading and the information content
of earnings. Journal of Business Finance and Accounting
23 (8): 1069-1095.
KE, B.; HUDDART, S., and PETRONI, K. 2003. What
insiders know about future earnings and how they use it:
evidence from insider trades. Journal of Accounting and
Economics 35 (3): 315-346.
HUDDART, S.; HUGHES, J.S.; LEVINE, C.B.; KE, B., and
SHI , C. 2007. Jeopardy, non-public information, and insider
trading around SEC 10-K and 10-Q filings. Journal of
Accounting and Economics 43 (1): 3-36.
CHENG, L.; FIRTH, M.; LEUNG, T.Y.; RUI, O., and
LEUNG, T.Y. 2008. Is there information content from insider
trading activities preceding earnings and dividend
announcements in Hong Kong? Accounting & Finance 48
(3): 417-437
FULLER, K.P. 2003. The impact of informed trading on
dividend signalling: a theoretical and empirical examination.
Journal of Corporate Finance 9 (4): 385-407.
DEL BRIO, E.B., and MIGUEL, A.DE. 2008. Dividends and
market signalling: an analysis of corporate insider trading,
European Financial Management (October): 1-18.
HILLIER, D., and MARSHALL, A.P. 2002a. The market
evaluation of information in directors trades. Journal of
Business Finance & Accounting 29: 77-110.
BETZER, A., and THEISSEN, E. 2007. Insider trading and
corporate governance: the case of Germany. European
Financial Management (december): 1-28.
LARCKER, D.F.; REDER, R.E., and SIMON, D.T. 1983.
Trades by insiders as evidence of the existence of
economic consequences of accounting standards. The
Accounting Review (58): 606-620.
KING, R., and O'KEEFE, T. 1986. Lobbying activities and
insider trading. The Accounting Review (61): 76-90.
ODAIYAPPA, R., and NAINAR, S.M.K. 1992. Economic
consequences of SFAS n.o 33 an insider trading
perspective. The Accounting Review 67: 599-609.
CHENG, Q., and LO, K. 2006. Insider trading and voluntary
disclosures. Journal of Accounting Research 44 (5): 815-
848.
DESAI, H.; KRISHNAMURTHY, S., and
VENKATARAMAN , K. 2006. Do short sellers target firms
with poor earnings quality? Evidence from earnings
restatements. Review of Accounting Studies 11: 71-90.
AGRAWAL, A., and CHADHA, S. 2005. Corporate
governance and accounting scandals. Journal of Law and
Economics 48: 371-406.
PALMROSE, Z-V., and SCHOLZ, S. 2004. The
circumstances and legal consequences of non-GAAP
reporting: Evidence from restatements. Contemporary
Accounting Research 21, 139-180.
DARROUGH, M., and RANGAN, S. 2005. Do insiders
manipulate earnings when they sell their shares in an initial
public offering? Journal of Accounting Research 43 (1): 1-
33. LAKONISHOK, J., and LEE, I. 2001. Are insider trades
informative? The Review of Financial Studies 14 (1): 79-
111.
Jeng, L. A., Metrick, A. and Zeckhauser, R. 2003.
Estimating the returns to insider trading: a performance-
evaluation perspective. Review of Economics and
Statistics 85 (2): 453-471.
CHENG, S.; NAGAR, V., and RAJAN, M. V. 2007. Insider
trades and private information: the special case of delayed-
disclosure trades. The Review of Financial Studies 20 (5):
1833-1864.
MARIN, J.M., and OLIVIER, J.P. 2008. The dog that did not
bark: insider trading and crashes. The Journal of Finance
63 (5): 2429-2476.
GARFINKEL, J.A. 1997. New evidence on the effects of
federal regulations on insider trading: The Insider Trading
and Securities Fraud Enforcement Act (ITSFEA). Journal
of Corporate Finance (3): 89-111.
ROULSTONE, D.T. 2003. The relation between insider-
trading restrictions and executive compensation. Journal
of Accounting Research 41 (3): 525-551.
BRAV, A., and GOMPERS, P.A. 2003. The role of lockups
in initial public offerings. The Review of Financial Studies 16
(1), 1-29.
CAO, C.; FIELD; L.C., and HANKA, G. 2004. Does insider
trading impair market liquidity? Evidence from IPO lockup
expirations. Journal of Financial and Quantitative and
Analysis 39 (1): 25-46.
KARPOFF, J.M., and LEE, D. 1991. Insider Trading Before
New Issue Announcements. Financial Management, 20: 18-
26.
EYSSELL, T.H., and REBURN, J. 1993. The effects of the
insider trading sanctions Act of 1984: the case of seasoned
equity offerings. Journal of Financial Research (16): 161-
170.
KAHLE, K.M. 2000. Insider trading and the long-run
performance of new security issues. Journal of Corporate
Finance 6(1): 25-53.
CLARKE, J.; DUNBAR, C., and KAHLE, K.M. 2001. Long-
run performance and insider trading in completed and
cancelled seasoned equity offering. Journal of Financial and
Quantitative and Analysis 36 (4): 415-430.
LEE, I. 1997. Do firms knowingly sell overvalued equity?
The Journal of Finance 52: 1439-1466.
HAUSER, S.; KRAIZBERG, E., and DAHAN, R. 2003. Price
behaviour and insider trading around seasoned equity
offerings: the case of majority-owned firms. Journal of
Corporate Finance 9 (2): 183-199.
GOSNELL, T.; KEOWN, A.J., and PINKERTON, J.M. 1992.
Bankruptcy and insider trading: differences between
exchange-listed and OTC firms. The Journal of Finance 47:
349-362.
LODERER, C., and SHEEHAN, D.P. 1989. Corporate
Bankruptcy and Managers Self-Serving Behavior. The
Journal of Finance 44: 1059-1075.
HEBNER, K.J., and KATO, T. 1997. Insider Trading and
Executive Compensation: Evidence from the U.S. and
Japan. International Review of Economics and Finance 6:
223 37.
GOSHEN, Z., and PARCHOMOVSKY, G. 2001. On Insider
Trading, Markets, and «Negative» Property Rights in
Information. Virginia Law Review 88: 1249-51.
HADDOCK, D.D., and MACEY, J.R. 1987. Regulation on
Demand: A Private Interest Model, with an Application to
Insider Trading Regulation. Journal of Law and Economics
30: 311 352.
BUSHMAN, R.M.; PIOTROSKI, J.D., and SMITH, A.J.
2005. Insider trading restrictions and analysts’ incentives to
follow firms. The Journal of Finance 60 (1): 35-66.
TRIDIMAS, T. 1991. Insider trading: European
harmonisation and national law reform. International and
Comparative Law Quarterly 40: 919-937.
WATSON, A. 1995. From legal transplants to legal
formants. American Journal of Comparative Law 43 (3):
469-476.
BHATTACHARYA, U., and DAOUK, H. 2002. The world
price of insider trading. The Journal of Finance 57 (1): 75-
108.
BRIS, A. 2005. Do insider trading laws work? European
Financial Management 11 (3): 267-312.
Hansun, S., 2013. A new approach of moving average
method in time series analysis. Conference on New Media
Studies doi:10.1109/CoNMedia.2013.6708545.
Hansun, S., 2014. A novel research of new moving average
method in time series analysis. International Journal of New
Media Technology .
Hansun, S., 2016. A new approach of brown’s double
exponential smoothing method in time series analysis.
Balkan Journal of Electrical and Computer Engineering 4,
pp.75–78.
Hansun, S., 2017. Wema versus b-wema methods in forex
forecasting. Publications of the 9th International Conference
on Machine Learning and Computing, pp.268–
271doi:10.1145/3055635.3056565.
Chunhua, S., Huimin, W., Fancheng, Y., Zhengliang, R.,
2011. Arima and neural network prediction of foreign
exchange reserve. Proceedings of 2011 6th International
Forum on Strategic Technology 2, pp.986–989.
doi:10.1109/IFOST.2011.6021186.
Hui, X.f., Wu, Y.j., 2012. Research on simple moving
average trading system based on svm. 2012
International Conference on Management Science
Engineering 19th Annual Conference Proceedings,
pp.1393–1397doi:10.1109/ICMSE.2012.6414356.
Ellis, C.A., Parbery, S.A., 2005. Is smarter better? a
comparison of adaptive, and simple moving average trading
strategies. Research in International Business and
Finance 19, pp.399–411.doi:10.1016/j.ribaf.2004.12.009.
Lauren, S., Harlili, S.D., 2014. Stock trend prediction
using simple moving average supported by news
classification. International Conference of Advanced
Informatics: Concept, Theory and Application
(ICAICTA),pp.135–139doi:10.1109/ICAICTA.2014.7005929.
Baasher, A.A., Fakhr, M.W., 2011. Forex trend
classification using machine learning techniques.
Proceedings of the 11th WSEAS international
conference on Applied computer science (pp. 41-47). World
Scientific and Engineering Academy and Society (WSEAS).
Sespajayadi, A., Indrabayu, Nurtanio, I., 2015. Technical
data analysis for movement prediction of euro to usd using
genetic algorithm-neural network. International Seminar on
Intelligent Technology and Its Applications , pp.23–26.
Myers, B., Gerig, A.: Simulating the synchronizing behavior
of high-frequency trading in multiple markets. In: Bera, A.,
Ivliev, S., Lillo, F. (eds.) Financial Econometrics and
Empirical Market Microstructure, pp. 207–213. Springer,
Berlin (2014)
Leal, S., Napoletano, M., Roventini, A., Fagiolo, G.: Rock
Around the Clock: An Agent-Based Model of Low- and
High-Frequency Trading. Working Paper, February (2014)
Gsell, M.: Assessing the Impact of Algorithmic Trading on
Markets: A Simulation Approach. Goethe Universität,
Frankfurt am Main (2008)
Linton, O., Mahmoodzadeh, S.: Implications of High-
Frequency Trading for Security Markets. Working Paper,
January (2018)
Verousis, T., Perotti, P., Sermpinis, G.: One size fits all?
High frequency trading, tick size changes and the
implications for exchanges: market quality and market
structure considerations. Rev. Quant. Financ. Account. 50,
353–392 (2018). https://doi.org/10.1007/s11156-017-0632-
2
Zervoudakis, F., Lawrence, D., Gontikas, G., Al, M.M.:
Perspectives on High-Frequency Trading. University
College London, London (2012)
Aldridge, I.: High-frequency runs and flash-crash
predictability. J. Portf. Manag. 40(3), 113–123 (2014).
https ://doi.org/10.3905/jpm.2014.40.3.113
Chaparro, F.: Here’s how high-frequency trading has
changed the stock market. Credit Suisse, March (2017a)
Bollen, N., Whaley, R.: Futures market volatility: what has
changed? J. Futures Mark. 35(5), 426–454 (2015).
https://doi.org/10.1002/fut.21666
Kelejian, H.H., Mukerji, P.: Does high frequency algorithmic
trading matter for non-AT investors? Res. Int. Bus.
Finance 37, 78–92 (2016).
https://doi.org/10.1016/j.ribaf .2015.10.014
Jarrow, R., Protter, P.: A dysfunctional role of high
frequency trading in electronic markets. Int. J. Theor. Appl.
Finance 15, 3 (2012). https://doi.org/10.1142/s0219
02491250022
Zhang, F.: High-Frequency Trading, Stock Volatility, and
Price Discovery (2010). Available at SSRN:
ssrn.com/abstract=1691679.
https://doi.org/10.2139/ssrn.1691679. Accessed 31 July
2014
Caivano, V.: The impact of high-frequency trading on
volatility—evidence from the Italian market. Working Paper.
CONSOB No. 80, March (2015)
Aldridge, I., Krawciw, S.: Aggressive high-frequency trading
in equities. Huffington Post Business (2015).
www.huffingtonpost.com/irene-aldridge/aggressive-high
frequency-_1_b_6698982.html?. Accessed 05 Feb 2016
Chaboud, A., Chiquoine, B., Hjalmarsson, E., Vega, C.:
Rise of the machines: algorithmic trading in the foreign
exchange market. J. Finance 69(5), 2045–2084 (2014).
https://doi.org/10.1111/jofi.12186
Baron, M., Brogaard, J., Hagströmer, B., Kirilenko, A.: Risk
and Return in High-Frequency Trading. Working Paper,
November (2017)
Friederich, S., Payne, R.: Computer-Based Trading and
Market Abuse. Foresight Driver Review DR20. UK
Government Office for Science (2012)
Farmer, J.D., Skouras, S.: An ecological perspective on the
future of computer trading. Quant. Finance 13(3), 325–346
(2013). https://doi.org/10.1080/14697 688.2012.75763 6
Hasbrouck, J., Saar, G.: Low-latency trading. J. Financ.
Mark. 16(4), 646–679 (2013)
Hagströmer, B., Nordén, L.: The diversity of high frequency
traders. J. Financ. Mark. 16(4), 741–770 (2013)
Groth, S.: Does Algorithmic Trading Increase Volatility?
Empirical Evidence from the Fully-Electronic Trading
Platform Xetra Wirtschaftsinformatik Proceedings Paper
112. Goethe Universität, Frankfurt am Main (2011)
Zingrand, J.P., Cliff, D., Hendershott, T.: Financial Stability
and Computer Based Trading. Foresight Driver Review
WP2. UK Government Office for Science (2012)
Abrol, S., Chesir, B., Mehta, N.: High frequency trading and
US stock market microstructure: a study of interactions
between complexities, risks and strategies residing in
U.S. equity market microstructure. Financ. Mark. Inst.
Instrum. 25(2), 107–165 (2016)
Jain, P., Jain, P., McInish, T.: Does high-frequency trading
increase systemic risk? Working Paper (2016)
Virgilio, G.: The impact of high-frequency trading on market
volatility. J. Trading 11(2), 55–63 (2016).
https://doi.org/10.3905/jot.2016.11.2.055
Bundesbank: Significance and impact of high-frequency
trading in the German capital market. Deutsche
Bundesbank Monthly. Report October 2016, pp. 37–60
(2016)
Foucault, T., Kadan, O., Kandel, E.: Liquidity cycles and
make/take fees in electronic markets. J. Finance 68(1),
299–341 (2013). https://doi.org/10.1111/j.1540-
6261.2012.01801.x
Kirilenko, A., Lo, A.: Moore’s law versus Murphy’s Law:
algorithmic trading and its discontents. J. Econ. Perspect.
27(2), 51–72 (2013). https://doi.org/10.1257/jep.27.2.51
Harris, L.: What to do about high-frequency trading. Financ.
Anal. J. 69(2), 6–9 (2013).
https://doi.org/10.2469/faj.v69.n2.6
Conrad, J., Wahal, S., Xiang, J.: High-frequency quoting,
trading, and the efficiency of prices. J. Financ. Econ. 116,
271–291 (2015). https://doi.org/10.1016/j.jfine
co.2015.02.008
Anagnostidis, P., Fontaine, P.: Liquidity provision,
commonality and high frequency trading. EUROFIDAI
Working Paper (2018)
Baron, M., Brogaard, J., Hagströmer, B., Kirilenko, A.: Risk
and Return in High-Frequency Trading. Working Paper,
November (2017)
Menkveld, A.: High frequency trading and the new-market
makers. J. Financ. Mark. 16(4), 712–740 (2013)
Menkveld, A.J.: The economics of high-frequency trading:
taking stock. Annu. Rev. Financ. Econ. 8, 1–24 (2016)
Brogaard, J., Hendershott, T., Hunt, S., Latza, T., Pedace,
L., Ysusi, C.: High-frequency trading and the execution
costs of institutional investors. Financ. Rev. 49(2), 345–
369 (2014a). https://doi.org/10.1111/fire.12039
Friederich, S., Payne, R.: Computer Based Trading,
Liquidity and Trading Costs. Foresight Driver Review DR5.
UK Government Office for Science (2011)
Menkveld, A., Zoican, M.: Need for Speed? Exchange
Latency and Liquidity Tinbergen Institute Discussion Paper
14-097/IV/DSF78 (2013). Available at SSRN:
ssrn.com/abstract=2442690.
https://doi.org/10.2139/ssrn.2442690. Accessed 7 May
2016
Aitken, M., de Harris, F.H.B., McInish, T., Aspris, A., Foley,
S.: High frequency trading—assessing the impact on
market efficiency and integrity. Foresight Driver Review
DR28. UK Government Office for Science (2012)
Hendershott, T., Mouton, P.C.: Automation, speed, and
stock market quality: the NYSE’s hybrid. J. Financ. Mark.
14(4), 568–604 (2011). https://doi.org/10.1016/j.finma
r.2011.02.003
Ding, S., Hanna, J., Hendershott, T.: How slow is the
NBBO? A comparison with direct exchange feeds. Financ.
Rev. 49, 313–332 (2014)
Hoffmann, P.: A dynamic limit order market with fast and
slow traders. J. Financ. Econ. 113(1), 156–169 (2014).
https://doi.org/10.1016/j.jfineco.2014.04.002
Hirschey, N.: Do High-Frequency Traders Anticipate Buying
and Selling Pressure? Working Paper. London Business
School, February (2018)
Kirchner, S.: High frequency trading: facts and fiction.
Policy 31(4), 8–20 (2016)
Lewis, M.: Flash Boys. W. W. Norton & Company Inc., New
York (2014)
Kovac, P.: Flash Boys: Not So Fast. Directissima Press,
Cambridge (2014)
Nanex: Nanex~15-Jul-2014~Perfect Pilfering. Nanex blog
(2014). http://www.nanex.net/aqck2/4661.html. Accessed
13 June 2018
Van Kervel, V.: Competition for order flow with fast and slow
traders. Rev. Financ. Stud. 28(7), 2094–2127 (2015).
https://doi.org/10.1093/rfs/hhv023
De Luca, M., Szostek, C., Cartlidge, J., Cliff, D.: Studies of
interactions between human traders and algorithmic trading
systems. Foresight Driver Review DR13. UK Government
Office for Science (2011)
Biais, B., Foucault, T., Moinas, S.: Equilibrium fast trading.
J. Financ. Econ. 116(2), 292–313 (2014).
https://doi.org/10.1016/j.jfineco.2015.03.004
Cvitanić, J., Kirilenko, A.: High Frequency Traders and
Asset Prices. California Institute of Technology, Pasadena
(2010)
Foucault, T., Menkveld, A.J.: Competition for order flow and
smart order routing systems. J. Finance 63, 119–158 (2008)
Sornette, D., von der Becke, S.: Crashes and High
Frequency Trading. Swiss Finance Institute. Research
Paper Series No. 11-63, Zürich (2011)
Zhang, S.: Need for speed: Hard information processing
in a high-frequency world. J. Futures Mark. 38, 3–21
(2017). https://doi.org/10.1002/fut.21861
Benos, E., Sagade, S.: Price discovery and the cross-
section of high-frequency trading. J. Financ. Mark. 30, 54–
77 (2016). https://doi.org/10.1016/j.finmar.2016.03.004
Miller, R., Shorter, G.: High Frequency Trading: Overview
of Recent Developments. Congressional Research
Service. 7-5700, April (2016)
Kaya, O.: High frequency trading: reaching the limits.
Autom. Trader Mag. 41, 23–27 (2016)
Arnoldi, J.: Computer algorithms, market manipulation and
the institutionalization of high frequency trading. Theory
Cult. Soc. 33(1), 29–52 (2016).
https://doi.org/10.1177/0263276414566642
Baron, M., Brogaard, J., Kirilenko, A.: The Trading Profits
of High Frequency Traders. Princeton University,
Princeton (2012)
Barker, W., Pomeranets, A.: The growth of high-
frequency trading: implications for financial stability.
Reports, Bank of Canada, Financial System Review, pp.
47–52 (2011). https://www.bankofcanada.ca/wp-conte
nt/uploads/2011/12/fsr-0611-barke r.pdf
Hasbrouck, J., Saar, G.: Technology and liquidity provision:
the blurring of traditional definitions. J. Financ. Mark.
12(2), 143–172 (2009). https://doi.org/10.1016/j.finma
r.2008.06.002
Leland, H.: Leverage, Forced Asset Sales and Market
Stability: Lessons from Past Market Crises and the Flash
Crash’. Foresight Driver Review DR9. UK Government
Office for Science (2011)
Jovanovic, B., Menkveld, A.: Middlemen in Limit-Order
Markets (2016). Available at SSRN:
ssrn.com/abstract=1624329.
https://doi.org/10.2139/ssrn.16243 29. Accessed 18 Dec
2014
Easley, D., Lopez de Prado, M., O’Hara, M.: The
microstructure of the ‘flash crash’: flow toxicity, liquidity
crashes, and the probability of informed trading. J. Portf.
Manag. 37(2), 118–128 (2011).
https://doi.org/10.3905/jpm.2011.37.2.118
Manahov, V., Hudson, R.: The implications of high
frequency trading on market efficiency and price discovery.
Appl. Econ. Lett. 21(16), 1148–1151 (2014).
https://doi.org/10.1080/13504851.2014.914135
Blocher, J., Cooper, R., Seddon, J., Van Vliet, B.: Phantom
liquidity and high-frequency quoting. J. Trading 11(3), 6–
15 (2016). https://doi.org/10.3905/jot.2016.11.3.006
Linton, O., O’Hara, M.: The Impact of Computer Trading on
Liquidity, Price Efficiency/Discovery and Transaction Costs.
Foresight Driver Review WP2. UK Government Office for
Science (2012)
Golub, A., Keane, J., Poon, S.: High Frequency Trading and
Mini Flash Crashes. University of Manchester, Manchester
(2012). Available at SSRN: ssrn.com/abstract=2182097.
https://doi.org/10.2139/ssrn.2182097. Accessed 28 April
2014
Gomber, P., Arndt, B., Lutat, M., Uhle, T.: High-
Frequency Trading. Goethe Universität, Frankfurt am Main
(2011)
Taleb, N.: The Black Swan. Random House Inc., New York
(2007)
Johnson, N., Zhao, G.: Brave New World: Quantifying the
New Instabilities and Risks Arising in Subsecond
Algorithmic Trading. Foresight Driver Review DR27. UK
Government Office for Science (2012)
Chaparro, F.: The fastest traders on wall street are in
trouble. Business Insider (2017b). http://www.businessin
sider.com. August 2017. Accessed 13 June 2018
Massa, A., Chilton, C.: They’re The World’s Fastest
Traders. Why Aren’t They Thriving? Bloomberg (2017).
https://www.bloomberg.com/news/articles/2017-07-13/they-
re-the-world-s-fastest-traders-why-aren-t-they-thriving.
Accessed 04 June 2018
Worstall, T.: Don’t Worry, Be Happy—High Frequency
Trading is Over, Dead, It’s Done. Forbes (2017). 25 Mar
2017.
Meyer, G., Bullock, N., Rennison, J.: How High-
Frequency Trading Hit a Speed Bump. The Big Read, New
York (2018)
Aït-Sahalia, Y., Saglam, M.: High-Frequency Traders:
Taking Advantage of Speed. National Bureau of Economic
Research. Working Paper 19531, Cambridge, MA (2013)
