See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/357288587

A stock trading algorithm based on trend forecasting and time series

classification

Conference Paper in Anais do Encontro Nacional de Inteligência Artificial e Computacional (ENIAC) · November 2021
DOI: 10.5753/eniac.2021.18272

CITATIONS READS

0 54

2 authors, including:

João Roberto Bertini Jr

University of Campinas
49 PUBLICATIONS 418 CITATIONS

SEE PROFILE

All content following this page was uploaded by João Roberto Bertini Jr on 04 January 2024.

The user has requested enhancement of the downloaded file.

A stock trading algorithm based on trend forecasting and time
series classification
Matheus Rosisca Padovani 1 , João Roberto Bertini Junior 1
1
Faculdade de Tecnologia – Universidade Estadual de Campinas
Limeira – SP – Brasil
m222360@dac.unicamp.br, bertini@ft.unicamp.br

Abstract. Algorithm trading relies on the automatic identification of buying and

selling points of a given asset to maximize profit. In this paper, we propose the
Trend Classification Trading Algorithm (TCTA) which is based on time series
classification and trend forecasting to perform trade. TCTA first employs the
K-means to cluster 5-days closing price segments and label them according to
its trend. A deep learning classification model is then trained with these label
sequences to estimate the next trend. Trading points are given by the alternation
on trend estimates. Results considering 20 shares from Ibovespa show TCTA
present higher profit than buy-and-hold and trading schemes based on Moving
Average Converge Divergence (MACD) or Bollinger bands.

1. Introduction
Stock trading is the process of negotiating shares to make a profit in the short term. A
trader can make a profit by operating with long positions and short positions. In the
former, the trader buys assets expecting their value will increase to sell. In the latter, the
trader sells an asset without owning it and purchases them as the price drops. Trades can
be done within a day or in a period of a few days or weeks, namely day trade and swing
trade, respectively [Brogaard et al. 2015]. A trading strategy supports trader decisions by
signaling entries, exits, and size positioning. Such strategies rely on financial indicators
and forecasting models to estimate market movements [Gandhmal and Kumar 2019].
Forecasting models can be used to estimate both price and trend. Trend forecast-
ing aims to estimate the price direction within a given period. It is an alternative to stock
price prediction stated as a classification task. Such models can be built with a time series
of historical data. Time series are categorized as stationary or non-stationary regarding
their underlying distribution [Box et al. 2016]. If the time series has a constant mean and
standard deviation, it is characterized as a stationary time series. In this case, statisti-
cal models like the auto-regressive (AR), the moving average (MA), or the mixed auto-
regressive moving average (ARMA) can be successfully applied to perform forecasting.
Non-stationary time series, on the other hand, do not present constancy to data over time.
In such cases, transformations can be applied to match the stationary conditions, thus the
transformed time series can be modeled with Auto-regressive integrated moving average
(ARIMA). Still, auto-regressive models implicitly assume that the future will resemble
the past, and may not be the best model for non-stationary data [Rachev et al. 2007].
Given the restriction of statistical models non-stationary time series, machine
learning algorithms, and most recently, deep learning methods, have been proposed to
address forecasts [Fawaz et al. 2018]. A branch of these algorithms transforms the task
of trend forecasting on a time series classification. The classifying technique chosen de-
pends on the type of problem [Bagnall et al. 2016].
Time series Classification has been a challenging problem of data mining, one of
the challenges is deal with noise on series, which can lead to not consistent results. Even
noise removing using filters does not resolve, instead introduces a problem of lag in the
data [Yang and Wu 2005].
In this paper, we propose a Trend Classification Trading Algorithm (TCTA) that
employs machine learning forecasting methods to signal entrance and exit points. Specif-
ically, we use K-means to group 5-days closing prices into three clusters, which we inter-
pret as up, down, and constant trends, respectively. Then, two neural networks are trained
with sequences of those cluster labels, to predict the next. Trade signals are then emitted
when the time series trend label alternate.
The remainder of the paper is organized as follows. Section 2 describes in a brief
some related works. Section 3 details the proposed methodology employed in the ex-
periments and presents the proposed trading algorithm. Section 4 presents the conducted
experiments and discusses the obtained results. At last, Section 5 concludes the paper the
suggests some future works.

2. Related works
This section presents some recent reviews and related works to the concepts used
in the proposed TCTA algorithm. Regarding input data, TCTA expects the clos-
ing price processed by the Rate of Change (ROC) indicator. Stock closing price
and volume are the most frequent entry for forecasting algorithms, although un-
structured data such as news and texts from social media has also been employed
[Bustos and Pomares-Quimbaya 2020].
TCTA uses the K-means clustering algorithm to help identify and label the trends
as upward, downward, and constant, before train a forecasting model. Clustering algo-
rithms have been used in market analysis. Most frequently, fuzzy clustering approaches
are used together with neural networks as forecasting method [Rajab and Sharma 2019].
Authors in [Shao et al. 2017] have used K-means to cluster the stock price sequences.
Then a Long short-term memory (LSTM) neural network was trained for each obtained
cluster. Their model first predicts the next day’s closing price, then a refined prediction
is made by the LSTM corresponding to the shortest distance clustering from the input
sequence. The proposed model showed higher prediction accuracy when compared to the
traditional feed-forward neural network and single LSTM neural network prediction. A
two-stage fusion model based on clustering has been proposed in [Xu et al. 2020]. The
authors used K-means to cluster several technical indicators, then ensemble learning has
also been applied to improve the prediction accuracy. Their experiment with daily stock
data of Bombay Stock Exchange, CNX Nifty, and S&P 500 stock indices showed the pro-
posed model obtains a better balance between accuracy and interpretability than the other
four prediction methods.
Once the price sequences have been grouped and labeled, we have a time series
of trend labels. Thus, trend forecasting is addressed as a time series classification task.
Fawaz et al. [Fawaz et al. 2018] and Bagnall et al. [Bagnall et al. 2016] present recent
surveys on time series classification. The former focuses on deep learning methods for
time series classification while the latter presents an extensive comparison with time series
classification algorithms proposed up to five years before 2017.
The proposed algorithm uses deep learning methods to learn the label sequence
patterns and estimate the next trend. Deep learning is a subset of neural networks
that uses multiple layers and provides a high-level abstraction for data modeling. In
[Ozbayoglu et al. 2020], the authors present a review of deep learning models in finan-
cial applications, such as algorithm trading, risk assessment, portfolio management, and
related tasks. From their study, they concluded that most of the trading algorithms concen-
trate on stock price prediction and that LSTM is the most preferred deep learning model.
Thakkar and Chaudhari [Thakkar and Chaudhari 2021] presented a review of deep learn-
ing models for the stock price and trend prediction from 2017 to 2020. They provide
an experimental evaluation with nine deep learning-based models. According to their
comparisons, the deep Q-network model obtained the best result in five days of trend
forecasting.

3. Methodology
This section presents the methods employed to design the TCTA algorithm. Section 3.1
presents how the stock prices are pre-processed and transformed in a sequence of trend
labels. Then, Section 3.2 presents the methods used to learn from the trend label time
series to predict the next trend. Section 3.3, then details the proposed TCTA algorithm to
perform trade through trend classification.

3.1. Data pre-processing

We collected the closing prices of each stock by a determinate period and made some
transformations. We separated closing prices using a sliding window method dividing
them into five days of lag variables. These parts have a four-day overlap since the window
moves just one day to the future per iteration. Each row of the data set ends up with the
date of the first-day value, and five days of values.
To limit the variation and to help to elucidate the trends, a transformation was
made on data replacing closing price values by the Rate of Change (ROC). ROC is a mo-
mentum indicator that compares current and past prices and shows the percentage change
between them. If this value is positive, then the value has increased by that percentage
and, if it is negative, the value had decreased. ROC is calculated by comparing the inter-
val between the prices. For example, ROC 200, compares today’s price with the price of
200 days ago [Achelis 2001]. In this work, we use ROC values instead of closing prices
to better capture changes in prices. The base of comparison is always the first day of the
series, so if we use a five sized interval, we will have the base, noted as ROC 0 that always
will be zero so we took it off, followed by ROC 1, ROC 2, ROC 3 and ROC 4, as shown
in Table 1.

Table 1. Example of ROC transformation of price values

PRICES VAL1 VAL2 VAL3 VAL4 VAL5 Transform ROC ROC1 ROC2 ROC3 ROC4
02/01/2019 87.90 88.00 88.50 88.48 88.83 → 02/01/2019 0.11 0.68 0.66 1.06
03/01/2019 88.00 88.50 88.48 88.83 90.32 → 03/01/2019 0.57 0.55 0.94 2.64
04/01/2019 88.50 88.48 88.83 90.32 90.42 → 04/01/2019 -0.02 0.37 2.06 2.17
 At the end of this process, we have a data set where each row represents one
day and the percentage change from 4 subsequent days. At this point, the data is 5-
dimensional and represents the percentage variation of 5 days period of one stock. To
classify future values, the data must be divided into classes. Thus, the next step is label-
ing each data instance with the trend it represents, i.e., upward (U), downward (D) and,
constant (C). To automate labeling and make it an unbiased process we employed the
K-means algorithm and have abstracted the group is found to the corresponding class.
The batches of percentage change were fed into K-means and the labels were
created. The values necessary to be classified as upward and downward vary along with
all stocks, depending on their history.
The K-means clustering algorithm is unsupervised, meaning that it does not re-
quire data with labels. Its objective is grouping elements that are like each other in K
clusters, the K value is predefined. At each iteration, a given distance metric between
an element and the centroid of each cluster is measured. Each element is assigned to
cluster with the nearest centroid. At the end of each iteration, all clusters’ centroids are
recalculated as the mean position between all elements of the cluster.
Table 2 shows the inputs as percentage changes or ROC and the output of K-means
as the classes. The visualization of these classes is depicted in Fig. 1 and 2. Values were
taken from a run of the ’BOVA11’, an Exchange-traded fund (ETF) that follows the index
Ibovespa.

Table 2. Table Roc into Classes

ROC 1 ROC 2 ROC 3 ROC 4 OUTPUT
0.24 2.63 4.17 4.06 Upward
-1.95 0.79 0.13 0.04 Constant
-0.57 -0.09 -1.88 -3.17 Downward

Figure 1. Closing price chart classified into Upward (U), Downward (D), or Con-
stant (C) trends
 Figure 2. Examples of the three classes found in the ROC transformed price data

In Figure 1 we can see the division with the prices that each day represents. The
figure shows closing price data from 2020-09-01 to 2020-11-24 of ’BOVA11’ share. Axis
x shows the classes that each day represents, Upward (U), Downward (D), or Constant
(C), Axis y shows closing price at day. There is an example of each class in the colored
boxes. The first day on the red box is a D, meaning that from this day to the next 4 days
the class is Downward according to K-means. The same is applied to the other boxes.
Figure 2 plots the same values presented in Table 2 and Fig. 1 with the same starting
point.

3.2. Trending forecasting methods

As described in the previous section, we transform the stock closing price data into a se-
quence of trending labels, denoted as U (upward), D (downward), and C (constant). Now,
given a sequence of five such labels, we want to estimate the next. i.e. C, C, C, U, U → U .
For doing so, we considered two deep learning models, the Deep Neural Network model
(DNN) and Long Short-Term Memory model (LSTM), as described next.
Before input the label sequence to the deep learning models, however, the cate-
gorical variables need to be transformed into one-hot variables. That transformation is
done to remove the possibility of a model creating ordinal relations between categorical
data. Our model has categorical values representing classes of each batch of days, so we
transform all classes using one-hot encoding. Creating a data set capable of being fed into
the neural networks.

3.2.1. Deep Neural Network

Deep Neural Networks (DNN) are artificial neural networks composed of several layers.
Among the many DNN possible architectures, Feed Forward DNN are, in essence, multi-
layer Perceptrons with a larger number of layers [Aggarwal 2018].
The DNN architecture used in this paper was created with the try and error
method. The idea was to create larger layers at the beginning and shrink them at the
end, we prioritize layers with multiples of the input. The network topology consists of
seven layers. The first is an input layer with 15 neurons, which receives a sequence of five
trend labels (either up, down, or constant) transformed into one-hot encoding. The first
layer is followed by two dense layers of 300 neurons, two dense layers of 150 neurons,
two dense layers of 50 neurons, a layer of 15 neurons, and the final layer with three neu-
rons. All neurons of the dense layers have Relu as an activating function, and the output
layer has a softmax activation. The weights were initialized by the glorot-normal method,
Adam optimizer was used with a learning rate of 0.1 and the categorical cross-entropy
was set as the loss function. We used a method of training that is called early stopping, it
stops the train after 7 generations without improvement on the loss function. We made a
train/test division of 0.85/0.15, and trained with mini-batches of size 30.

3.2.2. Long Short-Term Memory

Long Short-Term Memory (LSTM) is a Recurrent Neural Network (RNN) that introduces
a special neuron called long short-term memory. A recurrent neural network has connec-
tions between hidden units associated with a delay, which allows to retain information
of past and discover correlations of distant data. The information remains in the internal
memory of the neuron because the output is transmitted to the next layer and the neuron
itself. This model of network (RNN) is hard to train because of the vanishing gradi-
ent problem when the gradient used to update weights become so small that changes in
weights don’t reflect on the next layers of the network. This problem occurs often with
temporarily distant events and may stop a network further training [Pascanu et al. 2013].
The LSTM was proposed to address the vanishing gradient problem. The LSTM
units are linear and contain a set of cells inside it, whose activation is controlled by three
gates; 1) input gate, whose role is deciding which data of short term memory should be
activated, 2) forget gate, employed to decide which parts of the cell state remains relevant,
and 3) output gate, which tells what part of the cell state is relevant for the current data,
producing an output [Graves et al. 2009].
For creating the LSTM architecture we used the method try and error, the logic
was the same as described in the DNN section 3.2.1. The network has seven layers; the
input layer has 15 neurons, which receives a sequence of 5 trend labels transformed into
one-hot encoding. The hidden layers are composed of two LSTM layers of 60 neurons,
two dense layers of 30 neurons, and a layer of 15 neurons. All neurons of the dense layers
have Relu as an activating function. The output layer has three neurons with softmax acti-
vation. For training the LSTM, we employed Adam optimizer with a learning rate of 0.08,
the categorical cross-entropy as loss function, and glorot-normal as weight initialization
method. We used a method of training that is called early stopping, it stops the train after
7 generations without improvement on the loss function, we made a train/test division of
0.85/0.15 and trained with mini-batches of size 30.
3.3. The Trend Classification Trading Algorithm
This section presents a proposed algorithm named Trend Classification Trading Algo-
rithm (TCTA). After creating both networks and train with the pre-processed data of 3.1
transformed into one hot encoding classes, as explained in Section 3.2. We use each net-
work to generate a list of predictions for all days of the testing data set. We join that
predictions list with a prices list and feed into TCTA.
 TCTA works running through all days analyzing predictions and values, using
them to trade. The output of TCTA is the final Balance after all trades, and a data set
containing all trades made. The current Status of TCTA is important to the run because
it tells if the model is bought, sold, or out of the trades. All trades occur with the entire
balance, so the model can buy just once until it sells. The logic used is described in the
Algorithm 1.

Algorithm 1 Trend Classification Trading algorithm

Require: P rice {Price}; P rediction {Prediction of the Neural Network}; Status{Status indicating if
the model has a short position, a long position or no position} {Status 0 indicates no position, status 1
indicates long position and -1 indicates short position}
if P redicion = U ptrend and Status = 0 or Status = −1 then
Buy(P rice){Buy at current Price}
Status ← Status + 1{update status indicating the buy}
4: else if P redicion = Downtrend and Status = 0 or Status = 1 then
Sell(P rice){Sell at current Price}
Status ← Status − 1{update status indicating the sell}
end if

The strategy used by the neural consists of buying when the prediction is an up-
ward trend and selling when the prevision is a downward trend, if it is a stable prevision,
keeps its position like before. This algorithm repeats every day until the end of the train-
ing data. On the last day, the model makes a last transaction leaving the trades. The final
balance is saved in a data set and compared to others afterward. The TCTA strategy works
with both DNN and LSTM predictions, one at a time. We train and test the networks with
TCTA 3 times for each stock and selected the best run to use as a result.

4. Experiments
The stocks used in the trading experiment were selected from Ibovespa, the main index
tracking stocks performance in Brazil’s stock exchange (B3). We collected closing values
of 19 stocks and one ETF (Exchange-traded fund) ’BOVA11’, that mirrors Ibovespa index
performance. All companies were chosen based on their relevance in the index at the start
and end dates of the time series, for diversification proposes there are no more than two
stocks per Market sector. We collected 10 years of daily closing values from begin of
2010 to the end of 2020.

Table 3. List of shares and label distribution

Stock Upward Constant Downward Stock Upward Constant Downward
BOVA11 703 1397 620 ITUB4 570 1377 773
ABEV3 537 1573 610 JBSS3 602 1402 716
BBDC4 505 1525 690 LAME4 706 1398 616
BRFS3 507 1630 583 LREN3 600 1415 705
CCRO3 400 1572 748 KLBN4 506 1754 460
CIEL3 581 1625 514 MRVE3 571 1431 718
CMIG4 580 1677 463 PETR4 498 1678 544
CYRE3 571 1542 607 TIMS3 358 1582 780
ELET6 371 1510 839 VIVT3 668 1430 622
GOLL4 194 1566 960 VALE3 464 1559 697
 Table 3 shows all stocks used for testing and distribution of classes in each training
data set. The label Constant is by far the most common.
Tests of trade started on 02/01/2020, the first market day of that year. The Moving
Average Convergence Divergence (MACD) Model needed 10 days of market to build the
exponential mean, so all the trades started on day Jan 16, 2020, and go on until June 21,
2021. In total, they traded for 350 market days.
All trading algorithms follow the same pattern, starting with R$10, 000 for trade,
on the same dates. They have functions to buy or sell a stock and each Model has its
logic. At each transaction of a model, the balance, quantity of shares owned, number of
total trades, and actual status of transactions are updated, but this part was taken off the
example’s code to make evident just the logic. All trades have a fee of 0.03%, this value
is the fee, for all transactions made by investors in general at B3 (http://www.b3.com.br).
All models trade with the entire amount of money at each transaction. All project code
was made in Python, neural networks were made using Keras API, share values were
collected using Yahoo Finance API, only the daily closing value was used.

4.1. Comparative Algorithms

We choose three well-stated trading methods to compare the proposed algorithm. A brief
explanation of them and their trading rule is described ahead. They are Buy and Hold
method, MACD trading pattern, and Bollinger Bands trading pattern. Both trading pat-
terns used 5 days moving average, so all models have access to the same data.

4.1.1. Buy and Hold

Buy and hold is a common name for a passive investment strategy, it implies that in the
long run that stock will raise its prices and you should hold it and wait for the gains. This
strategy reduces costs of fees and taxes, which are paid when trading and take a small
portion of the investments. For testing this strategy, the model buys on the first day of
tests and sells the stock on the last day, making only these two transactions.

4.1.2. Moving Average Convergence Divergence

Moving Average Convergence Divergence (MACD) is a trend indicator that uses the rela-
tion between two exponential moving averages (EM A) of different sizes. An exponential
moving average differs from the simple moving average because the exponential gives
more relevance to ending days, unlike the simple that gives all days the same weight. The
values chosen for EM A are typically two values indicating the same strategy, like two
short-term indicators or two long-term indicators.
When the model shows an average divergence also known as a gap between two
EM A values, it is an entry point for trading, because as time passes the means should
converge again. The trading strategy consists in buying when the nearest moving average
(EM A5) is lower than the farthest one (EM A10) expecting the rise in prices returning
to mean. When the nearest average is bigger than the farthest is time to sell, expecting
falling prices returning to mean.
 You can use MACD in multiple ways, long-only, short-only, or long and short
strategies. In this project, we decided to use long and short. The EMA values chosen
were EM A5 and EM A10. In this strategy either you have a long or short position all the
time [Huang and Kim 2006].
• Buy Rule
– If EMA5< EMA10
• Sell Rule
– If EMA5>EMA10

4.1.3. Bollinger Bands

Bollinger Bands are a type of price envelope that uses moving average(M A) and standard
deviation(std), this is the band limit formula (M A + −n ∗ std). It plots the middle line
being the simple moving average, an upper limit n deviations up, creating an upper band,
and a lower band with n deviations down. It’s a parameter to known when the market is
devalued or overvalued, using standard deviation in the formula makes volatility part of
the equation, so if the market is too volatile, the bands will be widen indicating a new
range of prices for the market.
Trading using Bollinger Bands consists in selling when the values are high, above
the upper limit, and buying when the prices are low, below the lower limit. Since a value
tends to return to the mean, typically in some time the values converge to the mean and
the strategy made money with this volatility [Williams 2006].
In this project we are using 5 days as Moving Average and 1.5 standard deviations,
representing a short-term strategy. We trade using long and short to use Bollinger’s full
potential.
• Buy Rule
– If price < Bollinger Lower Band
• Sell Rule
– If price > Bollinger Upper Band
4.2. Results and Discussion
The results of trading models are shown as Percentage over initial capital on Figure 3 and
Table 4.

Figure 3. Boxplot distribution of trading results (percentage)

 Figure 3 shows the distribution of results, each ball represents a stock return and
each row represents a model. TCTA models were more consistent than others and had
more positive values overall.

Table 4. Percentage over initial capital for trading models under comparison
Stock BuyHold TCTA-DNN TCTA-LSTM MACD Bollinger
BOVA11 9.6 107.87 83.2 55.97 58.66
ABEV3 -1.06 6.53 -34.56 2.04 -44.62
BBDC4 -8.64 2.22 8.4 21.9 49.64
BRFS3 -19.42 17.93 17.93 26.05 219.77
CCRO3 -26.87 21.94 29.08 -62.24 -36.81
CIEL3 -44.53 -32.24 -30.05 -49.28 -4.92
CMIG4 6.39 -9.69 -6.52 64.98 16.76
CYRE3 -21.43 21.27 2.38 14.03 -46.86
ELET6 37.07 7.93 40.35 11.75 -21.61
GOLL4 -31.94 -0.78 26.99 -19.81 30.41
ITUB4 -1.15 63.19 3.16 -2.58 -5.02
JBSS3 0.24 -36.37 -58.51 -37.92 -39.99
LAME4 -25.58 -25.58 -25.58 -38.3 -24.61
LREN3 -22.13 -27.23 19.97 -33.8 -25.82
KLBN4 21.92 -23.97 15.94 -18.61 -15.21
MRVE3 -19.07 2.26 12.74 -8.84 -50.17
PETR4 2.19 302.87 280.79 18.73 2.8
TIMS3 -25.3 -4.12 -5.68 -46.3 -57.41
VIVT3 -10.57 12.53 9.88 -64.94 -42.46
VALE3 119.57 9.27 9.27 57.83 17.14
Mean -3.03 20.79 19.95 -5.46 -1.01
Std 34.79 74.25 68.3 39.39 61.75
Results removing outliers (two best and two worst runs of each model)
Mean -4.88 4.60 8.01 -6.56 -11.94
Std 17.41 21.43 18.04 29.29 28.11

Table 4 shows the best result of each share in bold, we can see the victories lead
by, the TCTA-DNN with 7 wins, Bollinger Bands with 5, TCTA-LSTM with 4, Buy and
Hold Strategy with 3, and MACD with just one.
Talking more deeply about the TCTA, they had the biggest deviation of results.
Otherwise, they had more constant profits. Both TCTA-DNN and TCTA-LSTM had near
results, they diverge in few cases, surprisingly their mean was almost the same, overall
TCTA-DNN, showed more constant wins and had a higher deviation. Even if we remove
the two biggest results of both, it is still better than the other models, even though the
difference would be smaller.
The TCTA works predicting the next day trend and trading, if it succeeds in buying
a stock before a raise or selling before a drop, it increases profits over time. However, this
guess is not always right and besides that, sudden variations in prices caused by changes
in a company or the economy cannot be predicted, these variations lead to movements out
of trends provoking losses for TCTA. MACD and Bollinger Bands react to sudden price
changes in less time, a cross in the means of MACD, and prices hitting bands in Bollinger
bands are faster than a change on the 5-day trend, used by TCTA, sometimes leading to
minor losses to them, yet these short-term fast trades can become noise on the long run.
The common statistical trading methods in general create a somewhat bigger re-
turn than Buy and Hold, but trading fees decrease their return over time, besides that they
come with much higher risk and volatility.
MACD was the least volatile model, and had the least wins, this method can start
a short position when prices increasing suddenly, expecting a return to the means and if
prices stay a long time far from the mean, it loses money.
Bollinger Bands method showed variated results, it was slightly better than Buy
and Hold in the mean, but worse in volatility. The logic of the method can lead to early
trade exits not benefiting from long-term appreciations.
If we remove both the two best and two worst stocks of each model we will end
up with just the two TCTA algorithms with positive mean results, TCTA-LSTM being the
best and Bollinger Bands being the worst. The standard deviation of TCTAs would be
near the deviation of buy and hold but with a mean 8% bigger for TCTA-DNN and 12%
bigger for TCTA-LSTM. The difference between LSTM and DNN shows that a specific
neural network can surpass a general network with proper data and training. Removing
these outliers brings the models into an average field, that may be expected for other
stocks or periods.

5. Conclusions
In this paper, we proposed the Trend Classification Trading Algorithm which uses time
series classification and trend forecasting to identify entry and exit points. TCTA uses
K-means to group price data and to label the trends. Then we tested two deep learning
neural networks models, DNN and LTSM, to estimate the next trend. We have conducted
trading experiments with 20 shares from Ibovespa comparing the two variants of TCTA,
TCTA-DNN, and TCTA-LTSM, with buy-and-hold and two trading schemes based on the
MACD and Bollinger Bands indicators, respectively. The trading schemes were evaluated
according to the percentage of earns/loses over an initial capital and within the period Jan.
16, 2020, to Jun. 21, 2021. The TCTA variants presented the best results among the trad-
ing strategies. Confirming that neural network models learned a pattern on the prices.
TCTA-DNN presented a slightly better average of 20.8 while TCTA-LTSM of 19.9 per-
cent. However, the TCTA-DNN presented a higher standard deviation than its counterpart,
74.5 and 68.3, respectively. The other trading strategies had negative performance in the
period, being MACD the worst. Removing the outliers would make TCTA-LTSM the best
one. Future works include considering different labeling strategies, more deep learning
models, and comparing TCTA against other deep learning-based trading models.

References
Achelis, S. B. (2001). Technical analysis from A to Z. McGraw Hill, 2nd edition.
Aggarwal, C. C. (2018). Neural Networks and Deep Learning. Springer, 1st edition.
 Bagnall, A., Lines, J., Bostrom, A., Large, J., and Keogh, E. (2016). The great time series
classification bake off: a review and experimental evaluation of recent algorithmic
advances. Data Mining and Knowledge Discovery.
Box, G. E. P., Jenkins, G. M., Reinsel, G. C., and Ljung, G. M. (2016). Time Series
Analysis Forecasting and Control, volume 5. John Wiley and Sons, Inc.
Brogaard, J., Hagströmer, B., and Lars Nordén, R. R. (2015). Trading fast and slow:
Colocation and liquidity. The Review of Financial Studies, 28:3407–3443.
Bustos, O. and Pomares-Quimbaya, A. (2020). Stock market movement forecast: A
systematic review. Expert Systems with Applications, 156:113464.
Fawaz, H. I., Forestier, G., Weber, J., Idoumghar, L., and Muller, P.-A. (2018). Deep
learning for time series classification: a review. Data Mining and Knowledge Discov-
ery, pages ”917–963”.
Gandhmal, D. P. and Kumar, K. (2019). Systematic analysis and review of stock market
prediction techniques. Computer Science Review, 34:100190.
Graves, A., Liwicki, M., Fernández, S., Bertolami, R., Bunke, H., and Schmidhuber, J.
(2009). A novel connectionist system for unconstrained handwriting recognition. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 31(5):855–868.
Huang, B. and Kim, Y. S. (2006). A test of MACD trading strategy. Master’s thesis,
Faculty of Business Administration - Simon Fraser University.
Ozbayoglu, A. M., Gudelek, M. U., and Sezer, O. B. (2020). Deep learning for financial
applications : A survey. Applied Soft Computing, 93:106384.
Pascanu, R., Mikolov, T., and Bengio, Y. (2013). On the difficulty of training recurrent
neural networks. In Proceedings of the 30th International Conference on Machine
Learning, volume 28, pages 1310–1318. PMLR.
Rachev, S. T., Mittinik, S., Fabozzi, F. J., Focardi, S. M., and Jas̆ić, T. (2007). Financial
Econometrics: From Basics to Advanced Modeling Techniques. John Wiley & Sons,
Inc, 1st edition.
Rajab, S. and Sharma, V. (2019). An interpretable neuro-fuzzy approach to stock price
forecasting. Soft Computing, 23:921–936.
Shao, X., Ma, D., Liu, Y., and Yin, Q. (2017). Short-term forecast of stock price of multi-
branch lstm based on k-means. In The 2017 4th International Conference on Systems
and Informatics (ICSAI 2017), pages 1546–1551.
Thakkar, A. and Chaudhari, K. (2021). A comprehensive survey on deep neural networks
for stock market: The need, challenges, and future direction. Expert Systems with
Applications, 177:114800.
Williams, O. (2006). Empirical optimization of bollinger bands for profitability. Master’s
thesis, Faculty of Business Administration - Simon Fraser University.
Xu, Y., Yang, C., Peng, S., and Nojima, Y. (2020). A hybrid two-stage financial stock
forecasting algorithm based on clustering and ensemble learning. Applied Intelligence.
Yang, Q. and Wu, X. (2005). 10 challenging problems in data mining research. Interna-
tional Journal of Information Technology and Decision Making, 5:”597–604”.

View publication stats