Finance, Markets and Valuation
DOI:
10.46503/NLUF8557
Corresponding author
Francisco Guijarro
Received: 25 Mar 2020
Revised: 26 Apr 2020
Accepted: 12 May 2020
Finance, Markets and
Valuation
ISSN 2530-3163.
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
Forecasting stock market trend: a comparison of
machine learning algorithms
La predicción de la tendencia en los mercados bursátiles:
una comparativa entre algoritmos de machine learning
Roberto Cervelló-Royo
ID
1
, Francisco Guijarro
ID
2
1
Economics and Social Sciences Department, Universitat Politècnica de València. Valencia,
España. Email: rocerro@esp.upv.es
2
Research Institute for Pure and Applied Mathematics, Universitat Politècnica de València.
Valencia, España. Email: fraguima@upvnet.upv.es
JEL: G11; G12; G17
Abstract
Forecasting the direction of stocks markets has become a popular research topic in recent years. Dierent
approaches have been applied by researchers to address the prediction of market trends by considering
technical indicators and chart patterns from technical analysis. This paper compares the performance
of four machine learning algorithms to validate the forecasting ability of popular technical indicators in
the technological NASDAQ index. Since the mathematical formulas used in the calculation of technical
indicators comprise historical prices they will be related to the past trend of the market. We assume that
forecasting performance increases when the trend is computed on a longer time horizon. Our results
suggest that the random forest outperforms the other machine learning algorithms considered in our
research, being able to forecast the 10-days ahead market trend, with an average accuracy of 80%.
Keywords: Trend forecasting; Stock markets; Random Forest; Deep Learning.
Resumen
La predicción de la tendencia en los mercados financieros se ha convertido en un área de investigación
popular durante los últimos años. Los investigadores han aplicado diferentes enfoques metodológicos
para abordar la predicción de la tendencia a través de indicadores técnicos y patrones chartistas del
análisis técnico. Este trabajo compara el rendimiento de cuatro algoritmos de machine learning para
validar la habilidad predictiva sobre la tendencia del índice tecnológico NASDAQ de algunos indicadores
técnicos habitualmente utilizados entre los traders. Debido a que los indicadores técnicos se obtienen
a partir de fórmulas matemáticas aplicadas sobre los precios históricos, asumimos que su valor está
ligado a la tendencia pasada de los mercados. Nuestra hipótesis es que la capacidad predictiva de estos
modelos aumenta conforme se amplía el horizonte temporal del análisis.Nuestros resultados sugieren
que el algoritmo de random forest supera al resto de algoritmos analizados en nuestro trabajo, siendo
capaz de predecir la tendencia de los próximos 10 días con una fiabilidad promedio del 80 %.
Keywords: Predicción de la tendencia; Mercados bursátiles; Random Forest; Deep Learning.
How to cite: Cervelló-Royo, R. and Guijarro, F. (2020) Forecasting stock market trend: a comparison
of machine learning algorithms. Finance, Markets and Valuation 6(1), pp. 37–49.
37
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
1 Introduction
The use of machine learning techniques allows exploring new opportunities and develop-
ments for investment in the field of financial markets (Efendi, Arbaiy, y Deris, 2018; Krauss, Do, y
Huck, 2017). In this way, technological developments provide private investors access to any
financial market, so that they can execute their buy and sell decisions in real time by using
tools for price analysis and future trend prediction (Bekiros, 2010; Kara, Boyacioglu, y Ömer
Kaan Baykan, 2011; Kim, 2003; Kim y Han, 2000; Leung, Daouk, y Chen, 2000).
In the last decade, the way of investing has changed at least from two perspectives. First,
investors have access to financialmarkets in realtime. Theuse and proliferation of new technolo-
gies, the speed of access to data and its quantitative treatment, allow them to make investment
decisions quickly and execute them through cross-platform trading soware. Second, artificial
intelligence, supported precisely by technological advances, has led to a new way of investing
where decisions can be carried out by handling huge amounts of information, and quantifying
the risks associated with them much more precisely (Arévalo, García, Guijarro, y Peris, 2017;
Cervelló-Royo, Guijarro, y Michniuk, 2015; Chen y Chen, 2016; Lee, 2009). The increasing global
access to large databases with prices and technical indicators obtained from it, allows investors
to make their decisions regardless of the emotional component, focusing on a more rational
framework based on the analysis of the available quantitative information, and where other
sources of information, like market sentiment, can also be added to price information (Nguyen,
Shirai, y Velcin, 2015).
In this context, machine learning techniques can make an outstanding contribution, facilitat-
ing rational decision making and enabling new opportunities for investors and entrepreneurs
(Abbas, 2018).
Markets and individual stocks move in trends, and history and economic cycles tend to
repeat themselves (Kirkpatrick II y Dahlquist, 2010). There is abundant literature where re-
searchers use historical prices in order to predict the stock price behavior (Ballings, den Poel,
Hespeels, y Gryp, 2015; García, Guijarro, Oliver, y Tamoši
¯
unien
˙
e, 2018) and discover patterns in
data by means of machine learning techniques (Cervelló-Royo y cols., 2015; Göçken, Özçalıcı,
Boru, y Dosdoğru, 2016; Li, Li, y Li, 2017; Patel, Shah, Thakkar, y Kotecha, 2015a, 2015b).
Most current trading platforms incorporate tools for stock market analysis of traded fi-
nancial assets, from individual stocks to indices, financial futures, commodities, currencies
(Forex market) or, more recently, cryptocurrencies. New tools include indicators and technical
oscillators, which are mathematical calculations obtained from the historical price and / or
volume of the financial asset and that are used by investors to try to predict the future behaviour
of stock prices. The number of indicators is large, and in many cases the signals oered by
some may contradict those oered by others. Analysis carried out should help identify invest-
ment opportunities, and investors should be able to handle a large amount of information in
real time (García, González-Bueno, Guijarro, y Oliver, 2020). In this context, machine learning
methods can provide solutions that the individual investor cannot handle eiciently, by si-
multaneously analyzing dierent indicators and their interactions, and thus predicting market
behavior (Ballings y cols., 2015; Huang, Yang, y Chuang, 2008; Ni, Ni, y Gao, 2011; Weng, Lu, Wang,
Megahed, y Martinez, 2018; Zhang, Cui, Xu, Li, y Li, 2018) from the historical analysis of a large
database in minimum time. Machine learning techniques can be applied both to forecast stock
prices and trends (Bekiros, 2010; García y cols., 2018; Hu, Tang, Zhang, y Wang, 2018; Kara y cols.,
2011; Kim, 2003; Kim y Han, 2000; Leung y cols., 2000). This latter one will be the object of this
study. That is, we will not forecast future closing prices but the market trend for dierent time
Roberto Cervelló-Royo and Francisco Guijarro 38
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
horizons. For this purpose, we will focus on stocks from the NASDAQ 100 which lists the largest
technological companies (Industrial, Technology, Retail, Telecommunications, Biotechnology,
Healthcare, Transportation, Media and Service companies). Furthermore, companies listed on
this index like Amazon, Apple or Facebook, are strong references for entrepreneurial ventures.
The paper is structured as follows: Section 2 describes the theoretical background of the
applied learning machine methods and our forecasting models. Section 3 reports the empirical
results from the application of the machine learning methods introduced in the previous section
over a sample of 89 technological companies from NASDAQ-100 Index during the period from
30th March 2009 to 27th December 2019. Finally, our main conclusions are given in Section 4.
2 Machine learning methods for stock market trend forecasting
This Section briefly describes the methods used in our research for the classification fore-
casting of stock market trend. Our interest is focused on two possible states: bullish trend vs.
bearish trend.
2.1 Random Forest
Random Forest is a decision tree algorithm, and it works as an ensemble algorithm in which
more than one model is made (Breiman, 2001). It is eective on a wide range of data sets, since
it needs relatively few parameters. Random forest allows to avoid overfitting and to have lots
of trees. New data can be given to each of those trees and ask each and one of them for their
prediction. When dealing with a classification problem the most popular answer can be chosen,
and when focusing on a regression problem the mean of each tree’s answer can be taken. It is
random because during the training process all the training data is not given to each tree; some
rows and / or columns are randomly hold back. All this makes each individual tree a bit less
eective, but when their results are averaged together the whole is more accurate than any
single one.
2.2 Deep Learning
Unliketheartificial neuralnetworks ofthepast, modern Deep Learning(DL)providestraining
stability, generalization, and scalability with big data. According to Candel, Parmar, LeDell, y
Arora (2016), “since it performs quite well in a number of diverse problems, Deep Learning is
quickly becoming the algorithm of choice for the highest predictive accuracy”.
Many layers of interconnected neuron units constitute the multi-layer-feed-forward neural
networks. First, there will be an input layer to match the feature space, followed by multiple
layers of nonlinearity, and ending with a linear regression or classification layer which should
match the output space. Each non-output layer of the network will include bias units. The
output of the entire network will be determined by the weights obtained by linking neurons and
biases with other neurons (Ballings y cols., 2015). Thus, these weights are adapted to minimize
the error on the labelled training data linking neurons and biases with other neurons and fully
determining the output of the entire network; then and when these weights are adapted to
minimize the error on the labelled training data, learning will occur. The main objective is to
minimize a loss function, for each and one training example. Deep Learning architectures are
models of hierarchical feature extraction, which usually involve multiple levels of nonlinearity.
These Deep Learning models are able to learn useful representations of raw data and have
shown high performance on complex data.
Roberto Cervelló-Royo and Francisco Guijarro 39
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
2.3 Gradient Boosting Machines
Gradient Boosting Machines (GBM) is a powerful machine learning algorithm that is part of
the boosting family algorithms. GBM models have proved to have strong predictive performance
and high flexibility in a wide range of data driven applications (Oppel y cols., 2012; Sayegh, Tate,
y Ropkins, 2016; Touzani, Granderson, y Fernandes, 2018). Boosting algorithms were originally
introduced for classification problems and like the random forest, it is a decision tree algorithm.
The main goal is to combine iteratively several simple models, called “weak learners”, to obtain a
“strong learner” which allows improving the prediction accuracy. Friedman, Hastie, y Tibshirani
(2000) extended the boosting to the regression by introducing the gradient boosting machines
method. The GBM method constitutes a numerical optimization algorithm that looks for an
additive model that minimizes the loss function.
Thus, using a training set
{(
x
1
, y
1
)
, . . . ,
(
x
n
, y
n
)}
of known values of
x
and corresponding
values of
y
, the goal is to find an approximation
ˆ
F (x)
to a function
F (x)
that minimizes the
expected value of this specified loss function L
(
y, F (x)
)
:
ˆ
F = argmin
F
Å
x,y
[
L
(
y, F (x)
)]
(1)
The gradient boosting method assumes a real-valued
y
and seeks an approximation
ˆ
F (x)
in the form of a weighted sum of functions
hi (x )
from some class
H
, the so called base (or
weak) learners:
ˆ
F (x) =
M
Õ
i =1
γ
i
h
i
(x ) + α (2)
where
α
represents a constant. Thus, the GBM algorithm iteratively adds at each step a new
decision tree which best reduces the loss function. In other words, the method tries to find an
approximation
ˆ
F (x)
that minimizes the average value of the loss function on the training set. It
does so by starting with a model, consisting of a constant function
F
0
(x )
), and incrementally
expands it:
F
0
(x ) = argmin
γ
n
Õ
i =1
L(y
i
, γ) (3)
In other words, when applying regression, the algorithm starts by initializing the model by a
first guess, which is usually a decision tree that maximally reduces the loss function, at each
step a new decision tree is fitted to the current residual and added to the previous model to
update the residual.
Then for
k = 1
to
K
(number of iterations) the pseudo-residuals should be computed and
following one-dimensional optimization problem the model can be updated to:
F
k
(x ) = F
k −1
(x ) + γ
k
h
k
(x ) (4)
Then, the user provides a maximum number of iterations, and the algorithm continues to
iterate until this maximum number of iterations is reached. An important part of the gradient
boosting method is the regularization by means of shrinkage which consists in modifying the
Roberto Cervelló-Royo and Francisco Guijarro 40
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
update rule in the following way:
F
k
(x ) = F
k −1
(x ) + νγ
k
h
k
(x ), 0 < ν ≤ 1 (5)
The learning rate
ν
is a shrinkage parameter that allows the GBM algorithm to have better
results if at each iterative step the contribution of the added decision tree is shrunk using this
learning rate. The learning rate ranges from 0 to 1 and the smaller it is, the more accurate
the model is. Nevertheless, a stronger shrinkage (smaller learning rate) implies a higher num-
ber of iterations to achieve convergence, because the value of the learning rate is inversely
proportional to the number of iterations.
A representation of the algorithm can be simply represented by the following pseudo-code:
1.
First, we should select the depth of the decision trees
d
, the number of iterations
K
, the
learning rate ν, and the subsample fraction η.
2.
Initialization: set the residual
r
0
= y
and
f = 0
. The mean value of
y
has also been suggested
as an initial guess of f (Liaw y Wiener, 2002).
3. For k = 1, 2, . . . , K , do the following:
a)
Randomly choose a subsample
{
y
i
, x
i
}
N
0
from the full training dataset, with
N
0
is the
number of data points corresponding to the fraction η
b) Using
{
y
i
, x
i
}
N
0
fit a decision tree f
k
of depth d to the residual r
k −1
c) Update f by adding the decision tree to the model f (x) ← f (x) + νf
k
(x )
d) Update the residuals r
k
← r
k −1
− νf
k
(x )
2.4 Generalized Linear Models
Generalized Linear Models (GLM) will allow estimating regression models for outcomes
following exponential distributions. In addition to the Gaussian distribution, these models
include Poisson, Binomial, Quasibinomial, Gamma, Ordinal, Negative Binomial and Tweedle
distributions and/or regressions, as well as Binomial and Multinomial classifications. Each one
serves for a dierent purpose, and depending on the kind, can be used either for prediction
and/or classification, i.e. when considering as outcome a binary variable we are talking about
logistic regression.
Thus, for binary classification problems where the response is a categorical variable with
two levels, Logistic regression will be used. Given the data, Logistic regression will provide the
probability of an observation belonging to an output category:
P
i
= P
r
(
y = 1|x
)
(6)
The probability distribution function can be written as:
P
i
= P
r
(
y = 1|x
1
, x
2
, . . . , x
k
)
= F
(
α + β X
i
)
= F (Z
i
) (7)
Where
P
i
is the probability of the dependent variable being 1 subject to the explanatory
variables. There are two basic types of binary choice models: the Logit model and the Probit
model. In the Logit model, the functional form of
F (.)
is the cumulative distribution function
of the logarithmic distribution, while in the Probit model, the functional form of
F (.)
is the
Roberto Cervelló-Royo and Francisco Guijarro 41
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
cumulative distribution function of a standard normal distribution. The Logit function is the
most common function for the binomial family. Its inverse is the logistic function, which takes
any real number and projects it onto the
[0, 1]
range as desired to model the probability of
belonging to a class. We will use the Logit model, which is defined as follows:
P
i
= P
r
(
y = 1|x
1
, x
2
, . . . , x
k
)
= F (Z
i
) =
1
1 + ex p (−Z
i
)
(8)
and it can also be converted to
l og
P
i
1 − P
i
= Z
i
= α + β X
i
(9)
3 Results
This Section presents the application of the machine learning methods introduced in the
previous Section to a large database composed by technological firms from the NASDAQ-100
Index. We have compiled the open, high, low and close prices (OHLC) on a daily basis along
with volume for those stocks listed in the NASDAQ-100 Index during the period 30th March 2009
to 27th December 2019. We excluded those stocks with missing values during the analysed
period; hence, the database is eventually composed by 89 stocks. We have computed the trend
indicator
I
t ,i
by comparing closing prices at times
t
(
C
t
) and
t − i
(
C
t −i
). Equation 3 establishes
a value of 1 in case the trend is positive, and 0 otherwise:
I
t ,i
=
(
1 C
t
> C
t −1
0 C
t
≤ C
t −1
(10)
This way, a value of 1 in the indicator denotes a bullish trend, while a value of 0 translates
into a bearish trend. To extend the analysis, we have computed the trend by considering
dierent distances between closing prices. The shortest version of the trend indicator used in
this research is for
i = 1
, which establishes the dierence between two consecutive closing
prices. We have computed the trend indicators for dierent lagged values of the closing price:
i ∈ {1, 5, 10}.
Price and volume information was extended with the technical indicators showed in Table
??
. The parameter values indicated in the table are the default values included in the TTR library
of R soware, as these are the commonly used in most trading platforms.
Table 1. Technical indicators and parameters used in the empirical research
Name of indicators and
parameters
Parameter values
Aroon indicator Number of periods to use in the calculation: 20
Average True Range (ATR) Number of periods for moving average: 14
Bollinger Bands (BB)
Number of periods for moving average: 20
Number of standard deviations used (sd): 2
Chaikin Volatility Number of periods for moving average: 10
Roberto Cervelló-Royo and Francisco Guijarro 42
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
Table 1. Technical indicators and parameters used in the empirical research
Name of indicators and
parameters
Parameter values
Close-to-close volatility Number of periods for the volatility estimate: 10
(volat_close) Number of periods per year: 260
Commodity Channel Index (CCI) Number of periods for moving average: 20
Directional Movement Index (ADX) Number of periods to use for direction calculation: 14
Exponential Moving Average (EMA)
Number of periods for moving average. we have used
two versions of this technical indicator: 10 and 25
Money Flow Index (MFI) Number of periods to use in volume calculations: 14
Moving Average Convergence /
Divergence (MACD)
Number of periods for fast moving average: 12
Number of periods for slow moving average: 26
Number of periods for signal moving average: 9
On Balance Volume (OBV) No parameters are needed, just price and volume
Parkinson formula for historical Number of periods for the volatility estimate: 10
volatility (volat_parkinson) Number of periods per year: 260
Relative Strength Index (RSI) Number of periods for moving average: 14
Stochastic Momentum Index (SMI)
Number of periods for initial smoothing: 2
Number of periods for double smoothing: 25
Number of periods for signal line: 9
Figure 1 shows the correlation matrix between all variables considered in the prediction
of the market trend. We can observe that OHLC prices are highly correlated with technical
indicators EMA10, EMA25, MACD and ATR. However, most of them show a low correlation co-
eicient among them. This is of strong interest for the purpose of our research; thus, we can
conclude that indicators can contribute in a dierent and specific way to the explanation of
trend behaviour.
The methods introduced in the previous section have been used to estimate the trend
indicator,
ˆ
I
t
, by considering two possible scenarios. The first one only considers OHLC prices
and volume from period
t − i
in the prediction of trend indicator of period
t
; i.e., five variables in
the forecasting of the trend. The second one extends these input variables with the 14 technical
indicators referred in Table 1. According to previous literature (Kara y cols., 2011; Patel y cols.,
2015b), we assume that technical indicators do have incremental value in the prediction of
the market trend. OHLC prices and volume from t − i just consider information preceding the
closing price from period
t − i
. However, technical indicators are calculated by considering a
larger period or prices and volumes, hence its potential ability to summarize the past behaviour
of stocks. The dierent values considered for computing the lagged closing prices in Equation 3
allow to analyse the forecasting ability of technical indicators on the trend indicator. We assume
that the larger the value of
i
, the more significant forecasting ability of technical indicators, as
its calculation involves the use of historical prices and volumes.
We have extended the analysis with two additional scenarios. In the first one we have
applied the machine learning methods to each individual stock. One limitation of this approach
Roberto Cervelló-Royo and Francisco Guijarro 43
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
Figure 1. Correlation matrix for variables considered in the prediction of trend market
is that the sample size is constrained to the period analyzed: 2,516 daily observations per
stock. This would potentially limit the eectiveness of the machine learning methods which are
supposed to provide more accurate results with large datasets, therefore patterns can be more
eectively captured. However, technical analysis assumes that price patterns are globally valid
regardless of the considered stock (Cervelló-Royo y cols., 2015; Tsinaslanidis, 2018). This led us
to establish a second scenario by considering all the stocks as a whole dataset, thus enlarging
the original sample and gathering 223,924 observations in the final sample.
Table 2 summarizes the dierent scenarios considered in this research. The three versions
of the trend indicator according to the lagging of closing prices (
i ∈ {1, 5, 10}
) are used for all
scenarios.
We have used a common framework for the dierent machine learning techniques used in
the research. The training set was randomly selected by considering the 80% of the sample
while the remaining 20% was separated for testing. We carried out 50 experiments per each
scenario and machine learning method to assess the consistency of the results. Table 3 shows
the parameter values used in the dierent machine learning algorithms. These were taken as
the default values in the h2o library from R Soware.
Table 4 shows the average accuracy ratios obtained along the experiments carried out in
our research. We have trained and tested 4 dierent machine learning methods: GBM, RF, DL
Roberto Cervelló-Royo and Francisco Guijarro 44
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
Without technical indicators With technical
indicators
Individual stocks
Scenario A Scenario B
2,516 observations 2,516 observations
5 trend predictors 19 trend predictors
Aggregated sample of stocks
Scenario C Scenario D
223,924 observations
223,924 observa-
tions
5 trend predictors 19 trend predictors
Table 2. Simple size and number of predictors for the dierent scenarios considered in the research
Machine learning algorithms Parameters Value
Deep Learning
Number of folds for crossvalida-
tion
5
Hidden layers (200, 200)
epochs (number of iterations) 100
rho (adaptive learning rate time
decay factor)
0.99
epsilon (adaptive learning rate
smoothing factor)
1E-08
Learning rate 0.005
Random forest
Number of folds for crossvalida-
tion
5
Number of trees 50
Maximum tree depth 20
Gradient Boosting Machine
Number of folds for crossvalida-
tion
5
Number of trees 50
Maximum tree depth 20
Learning rate 0.1
Generalized Lineal Model
Number of folds for crossvalida-
tion
5
Family Binomial
Theta 1E-10
Table 3. Parameter values considered in the machine learning algorithms
Roberto Cervelló-Royo and Francisco Guijarro 45
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
and GLM. For all the considered scenarios, the table reports the average accuracy ratio for
the 50 experiments along with the standard deviation of these values. The three alternatives
of time horizon of the trend indicator are also reported. We can observe that the machine
learning algorithms did not obtain satisfactory results in scenarios A, B and C. Despite the
average accuracy ratios are above 50%, we have to consider that during the period analyzed the
NASDAQ market has experienced a long bullish market. This translates into more bullish than
bearish trading days, and thus the trend indicator
I
t ,i
has a biased mean value, greater than 0.5.
We can observe that 52.3% of days were bullish (last column of Table 4). And this percentage is
55.1% for 5-days periods and 57.0% for 10-days periods. Therefore, a naïve forecasting model
couldobtainthese values by unfailinglypredicting the value1 regardless of the inputs considered
in the algorithm.
However, some results of scenario D are above these mean trend indicator values. In the
case of
i = 1
, the percentage of bullish trading days is the same than the one obtained by all of
the 4 machine learning algorithms: 52.3%. Interestingly, this accuracy ratio improves as the
value of i increases. For
i = 5
, the percentage of bullish trading days is 55.1%, but all algorithms
report a better accuracy ratio: 57.5% (GBM), 72.4% (RF), 56.2% (DL) and 56.0% (GLM). And the
RF algorithm is still better when considering the
i = 10
case. The percentage of bullish periods
is 57.0%, while the RF obtains an average accuracy ratio of 80.8% and a standard deviation of
0.005. Despite GBM, DL and GLM get accuracy ratios above the 57.0% threshold, the successful
ratio of RF is by large much better than the ones obtained by its competitors. The low values
reported for the standard deviation of the accuracy ratios demonstrate the consistency of the
algorithms in the computation of the market trend.
ML alg. GBM RF DL GLM % I
t ,i
= 1
i Scenario A
1 0.533 (0.020) 0.532 (0.020) 0.533 (0.020) 0.533 (0.021) 0.523
5 0.520 (0.018) 0.520 (0.018) 0.519 (0.018) 0.521 (0.018) 0.551
10 0.523 (0.018) 0.522 (0.019) 0.523 (0.020) 0.522 (0.020) 0.570
i Scenario B
1 0.529 (0.022) 0.529 (0.023) 0.530 (0.022) 0.529 (0.023) 0.523
5 0.522 (0.020) 0.522 (0.020) 0.522 (0.019) 0.522 (0.020) 0.551
10 0.523 (0.022) 0.522 (0.022) 0.523 (0.023) 0.524 (0.022) 0.570
i Scenario C
1 0.523 (0.000) 0.523 (0.000) 0.523 (0.000) 0.523 (0.000) 0.523
5 0.550 (0.000) 0.550 (0.000) 0.550 (0.000) 0.550 (0.000) 0.551
10 0.577 (0.006) 0.582 (0.009) 0.574 (0.005) 0.574 (0.005) 0.570
i Scenario D
1 0.523 (0.004) 0.523 (0.004) 0.523 (0.004) 0.523 (0.004) 0.523
5 0.575 (0.005) 0.724 (0.008) 0.562 (0.004) 0.560 (0.004) 0.551
10 0.606 (0.007) 0.808 (0.005) 0.582 (0.006) 0.577 (0.004) 0.570
Table 4. Mean accuracy values obtained by the machine learning methods
Note: Values represent the mean accuracy obtained by each machine learning method in the dierent scenarios.
Values in parenthesis are the standard deviation of the accuracy values. The experiments were carried out for 3
dierent lagged closing prices in the trend indicator (
i ∈ {1, 5, 10 }
). Last column represents the percentage of bullish
trend indicator.
Roberto Cervelló-Royo and Francisco Guijarro 46
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
4 Conclusions
Machine learning techniques have become of special interest in financial operations. In this
vein, new technologies oer new business opportunities for investors in the financial markets.
This paper deals with the application of machine learning techniques in a well-known
technological index like the NASDAQ 100. Thus, our main goal is the forecasting of market trends
by comparing the performance of well-established machine learning algorithms. Technical
analysis assumes that stock markets tend to repeat themselves, thus its main objective is
to accurately identify patterns through dierent technical indicators. Our experiments are
focused to assess the prediction ability of these indicators through the use of machine learning
algorithms. The mathematical formulas used to calculate technical indicators include past
price and volume, which summarizes the past behaviour of the market. Literature suggests that
this can help to predict the future trend. Our results provide empirical evidence that technical
indicators have poor predictive power for the short-term –the following trading day trend–.
However, results significantly improve when the time horizon is expanded to forecast the 10
trading day trend. Random forest results in the best performance algorithm with an average
accuracy ratio of 80.8%.
References
Abbas, S. A. (2018). Entrepreneurship and information technology businesses in eco-
nomic crisis. Entrepreneurship and Sustainability Issues, 5(3), 682–692. doi:
https://doi.org/10.9770/jesi.2018.5.3(20)
Arévalo, R., García, J., Guijarro, F., y Peris, A. (2017). A dynamic trading rule based on filtered flag
pattern recognition for stock market price forecasting. Expert Systems with Applications,
81, 177–192. doi: https://doi.org/10.1016/j.eswa.2017.03.028
Ballings, M., den Poel, D. V., Hespeels, N., y Gryp, R. (2015). Evaluating multiple classifiers for
stock price direction prediction. Expert Systems with Applications, 42(20), 7046–7056. doi:
https://doi.org/10.1016/ j.eswa.2015.05.013
Bekiros, S. D. (2010). Fuzzy adaptive decision-making for boundedly rational traders in spec-
ulative stock markets. European Journal of Operational Research, 202(1), 285–293. doi:
https://doi.org/10.1016/ j.ejor.2009.04.015
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.
Candel, A., Parmar, V., LeDell, E., y Arora, A. (2016). Deep learning with H2O. H2O. AI Inc.
Cervelló-Royo, R., Guijarro, F., y Michniuk, K. (2015). Stock market trading rule
based on pattern recognition and technical analysis: Forecasting the DJIA index
with intraday data. Expert Systems with Applications, 42(14), 5963–5975. doi:
https://doi.org/10.1016/ j.eswa.2015.03.017
Chen, T.-l., y Chen, F.-y. (2016). An intelligent pattern recognition model for supporting
investment decisions in stock market. Information Sciences, 346-347, 261–274. doi:
https://doi.org/10.1016/ j.ins.2016.01.079
Efendi, R., Arbaiy, N., y Deris, M. M. (2018). A new procedure in stock market forecasting based
on fuzzy random auto-regression time series model. Information Sciences, 441, 113–132.
doi: https://doi.org/10.1016/j.ins.2018.02.016
Friedman, J., Hastie, T., y Tibshirani, R. (2000). Additive logistic regression: a statistical view of
boosting (With discussion and a rejoinder by the authors). The Annals of Statistics, 28(2),
337–407. doi: https://doi.org/10.1214/aos/1016218223
García, F., González-Bueno, J., Guijarro, F., y Oliver, J. (2020). Forecasting the environmental,
Roberto Cervelló-Royo and Francisco Guijarro 47
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
social, and governance rating of firms by using corporate financial performance variables:
A rough set approach. Sustainability, 12(8), 3324. doi: https://doi.org/10.3390/su12083324
García, F., Guijarro, F., Oliver, J., y Tamoši
¯
unien
˙
e, R. (2018). Hybrid fuzzy neuralnetwork to predict
price direction in the German DAX-30 index. Technological and Economic Development of
Economy, 24(6), 2161–2178. doi: https://doi.org/10.3846/tede.2018.6394
Göçken, M., Özçalıcı, M., Boru, A., y Dosdoğru, A. T. (2016). Integrating metaheuristics and Artifi-
cial Neural Networks for improved stock price prediction. Expert Systems with Applications,
44, 320–331. doi: https://doi.org/10.1016/j.eswa.2015.09.029
Hu, H., Tang, L., Zhang, S., y Wang, H. (2018). Predicting the direction of stock markets using
optimized neural networks with Google Trends. Neurocomputing, 285, 188–195. doi:
https://doi.org/10.1016/ j.neucom.2018.01.038
Huang, C., Yang, D., y Chuang, Y. (2008). Application of wrapper approach and composite
classifier to the stock trend prediction. Expert Systems with Applications, 34(4), 2870–
2878. doi: https://doi.org/10.1016/j.eswa.2007.05.035
Kara, Y., Boyacioglu, M. A., y Ömer Kaan Baykan. (2011). Predicting direction of stock price index
movement using artificial neural networks and support vector machines: The sample
of the Istanbul Stock Exchange. Expert Systems with Applications, 38(5), 5311–5319. doi:
https://doi.org/10.1016/ j.eswa.2010.10.027
Kim, K.-j. (2003). Financial time series forecasting using support vector machines. Neurocom-
puting, 55(1-2), 307–319. doi: https://doi.org/10.1016/s0925-2312(03)00372-2
Kim, K.-j., y Han, I. (2000). Genetic algorithms approach to feature discretization in artificial
neural networks for the prediction of stock price index. Expert Systems with Applications,
19(2), 125–132. doi: https://doi.org/10.1016/s0957-4174(00)00027-0
Kirkpatrick II, C. D., y Dahlquist, J. A. (2010). Technical analysis: the complete resource for
financial market technicians. FT press.
Krauss, C., Do, X. A., y Huck, N. (2017). Deep neural networks, gradient-boosted trees, random
forests: Statistical arbitrage on the S&P 500. European Journal of Operational Research,
259(2), 689–702. doi: https://doi.org/10.1016/j.ejor.2016.10.031
Lee, M.-C. (2009). Using support vector machine with a hybrid feature selection method to
the stock trend prediction. Expert Systems with Applications, 36(8), 10896–10904. doi:
https://doi.org/10.1016/ j.eswa.2009.02.038
Leung, M. T., Daouk, H., y Chen, A.-S. (2000). Forecasting stock indices: a comparison of
classification and level estimation models. International Journal of Forecasting, 16(2),
173–190. doi: https://doi.org/10.1016/s0169-2070(99)00048-5
Li, D., Li, Z., y Li, R. (2017). Automate the identification of technical patterns: a k-
nearest-neighbour model approach. Applied Economics, 50(17), 1978–1991. doi:
https://doi.org/10.1080/00036846.2017.1383596
Liaw, A., y Wiener, M. (2002). Classification and regression by randomforest. R News, 2(3), 18–22.
Nguyen, T. H., Shirai, K., y Velcin, J. (2015). Sentiment analysis on social media for stock
movement prediction. Expert Systems with Applications, 42(24), 9603–9611. doi:
https://doi.org/10.1016/ j.eswa.2015.07.052
Ni, L.-P., Ni, Z.-W., y Gao, Y.-Z. (2011). Stock trend prediction based on fractal feature selection
and support vector machine. Expert Systems with Applications, 38(5), 5569–5576. doi:
https://doi.org/10.1016/ j.eswa.2010.10.079
Oppel, S., Meirinho, A., Ramírez, I., Gardner, B., O’Connell, A. F., Miller, P. I., y Louzao,
M. (2012). Comparison of five modelling techniques to predict the spatial dis-
Roberto Cervelló-Royo and Francisco Guijarro 48
Finance, Markets and Valuation Vol. 6, Num. 1 (January-June 2020), 37–49
tribution and abundance of seabirds. Biological Conservation, 156, 94–104. doi:
https://doi.org/10.1016/ j.biocon.2011.11.013
Patel, J., Shah, S., Thakkar, P., y Kotecha, K. (2015a). Predicting stock and stock price index move-
ment using trend deterministic data preparation and machine learning techniques. Expert
Systems with Applications, 42(1), 259–268. doi: https://doi.org/10.1016/j.eswa.2014.07.040
Patel, J., Shah, S., Thakkar, P., y Kotecha, K. (2015b). Predicting stock market index using fusion
of machine learning techniques. Expert Systems with Applications, 42(4), 2162–2172. doi:
https://doi.org/10.1016/ j.eswa.2014.10.031
Sayegh, A., Tate, J. E., y Ropkins, K. (2016). Understanding how roadside concentrations of
NOx are influenced by the background levels, traic density, and meteorological con-
ditions using boosted regression trees. Atmospheric Environment, 127, 163–175. doi:
https://doi.org/10.1016/ j.atmosenv.2015.12.024
Touzani, S., Granderson, J., y Fernandes, S. (2018). Gradient boosting machine for modeling
the energy consumption of commercial buildings. Energy and Buildings, 158, 1533–1543.
doi: https://doi.org/10.1016/j.enbuild.2017.11.039
Tsinaslanidis, P. E. (2018). Subsequence dynamic time warping for charting: Bullish and bearish
class predictions for NYSE stocks. Expert Systems with Applications, 94, 193–204. doi:
https://doi.org/10.1016/ j.eswa.2017.10.055
Weng, B., Lu, L., Wang, X., Megahed, F. M., y Martinez, W. (2018). Predicting short-term stock
prices using ensemble methods and online data sources. Expert Systems with Applications,
112, 258–273. doi: https://doi.org/10.1016/j.eswa.2018.06.016
Zhang, J., Cui, S., Xu, Y., Li, Q., y Li, T. (2018). A novel data-driven stock price
trend prediction system. Expert Systems with Applications, 97, 60–69. doi:
https://doi.org/10.1016/ j.eswa.2017.12.026
Roberto Cervelló-Royo and Francisco Guijarro 49