Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com

ScienceDirect
Procedia Computer Science 00 (2019) 000–000
Procedia
Procedia Computer
Computer Science
Science 17500 (2019)
(2020) 000–000
603–608 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

International Workshop on Artificial Intelligence & Internet of Things (A2IoT)

International Workshop on Artificial
August Intelligence
9-12, 2020, Leuven, & Internet of Things (A2IoT)
Belgium
August 9-12, 2020, Leuven, Belgium
Stock
Stock Market
Market prediction
prediction on
on High
High frequency
frequency data
data using
using Long-Short
Long-Short
Term Memory
Term Memory
Zineb Lanbouri*∗∗ , Said Achchab
Zineb Lanbouri* , Said Achchab
National School for Computer Science and Systems analysis, Mohamed V University, Rabat, Morocco
National School for Computer Science and Systems analysis, Mohamed V University, Rabat, Morocco

Abstract
Abstract
High Frequency Trading (HFT) is part of algorithmic trading, and one of the biggest changes that happened in the last 15 years.
High
HFT or Frequency Trading
nanotrading (HFT)the
represents is part of algorithmic
ability, for a trader, trading, and one
to take orders of thevery
within biggest
shortchanges that happened
delays. This in thealast
paper presents 15 years.
model based
HFT
on technical indicators with Long Short Term Memory in order to forecast the price of a stock one-minute, five-minutes andbased
or nanotrading represents the ability, for a trader, to take orders within very short delays. This paper presents a model ten-
on technical
minutes indicators
ahead. First, wewith
getLong Short Term
the S&P500 Memory
intraday in order
trading data to forecast
from the then
Kaggle, pricewe
of acalculate
stock one-minute, five-minutes
technical indicators andand ten-
finally,
minutes
we train ahead. First, we
the regression get the S&P500
Long-Short intraday model.
Term Memory tradingBased
data from Kaggle,
on the then wealongside
price history, calculatetechnical
technicalanalysis
indicators and finally,
indicators and
we train the
strategies, regression
this model is Long-Short Term
executed, and the Memory
results aremodel. Based
analyzed on the
based price history,metrics
on performance alongside
andtechnical analysis
profitability. indicators
Experiment and
results
strategies, this model is executed, and the results are analyzed based on performance metrics
show that the proposed method is effective as well as suitable for prediction a few minutes before. and profitability. Experiment results
show that the proposed method is effective as well as suitable for prediction a few minutes before.
© 2020 The
© 2020 The Authors.
Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V.
© 2020
This The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license
is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open access article under
Peer-review under responsibility
responsibility of the
ofthe CC BY-NC-ND
theConference license
ConferenceProgram
Program (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Chairs.
Chair.
Peer-review under responsibility of the Conference Program Chairs.
Keywords: High Frequency trading; Long-Short Term Memory; Forecasting; Technical Analysis
Keywords: High Frequency trading; Long-Short Term Memory; Forecasting; Technical Analysis

1. Introduction
1. Introduction
Thanks to the rapid development of computing, training data can now be obtained in a very regular way. Traders
dealThanks to the rapid development
with minute-based, or sometimesofeven
computing, training data
nanosecond-based can
data. now be obtained
Therefore, in a very
it is particularly regular way.
important Traders
to determine
deal with minute-based, or sometimes even nanosecond-based data. Therefore, it is particularly important
how to analyze useful information and decide if a technical approach can be relevant. Predicting and Forecasting the to determine
how to analyze
closing useful
price a few information
minutes ahead isand decidethe
exactly if problem
a technical
thatapproach
the LSTM canmodel
be relevant. Predicting
is trying to solve.and
By Forecasting the
definition, high
closing price a few minutes ahead is exactly the problem that the LSTM model is trying to solve. By definition,
Frequency Trading, nanotrading or intraday trading can be defined as a complex type of algorithmic trading that relies high
Frequency Trading,
on low latencies. nanotrading
According or intraday
to the Securitytrading can be defined
and Exchange Commissionas a complex
[1][2], type
HFTofis algorithmic trading
characterized by a that
veryrelies
high
on low latencies. According to the Security and Exchange Commission [1][2], HFT is characterized
number of orders, proprietary trading and a very short holding period. Therefore, nanotrading presents the following by a very high
number of orders, proprietary trading and a very short holding period. Therefore, nanotrading presents the following
challenges:
challenges:

∗ Corresponding author. Tel.: +212-668-78-38-60

∗ Corresponding
E-mail address:author. Tel.: +212-668-78-38-60
[email protected]
E-mail address: [email protected]
1877-0509 © 2020 The Authors. Published by Elsevier B.V.
1877-0509
This © 2020
is an open Thearticle
access Authors. Published
under by Elsevier B.V.
the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
1877-0509 © 2020
This is an open Thearticle
access Authors. Published
under by Elsevier B.V.
the Conference
CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Program Chairs.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Conference Program Chairs.
Peer-review under responsibility of the Conference Program Chairs.
10.1016/j.procs.2020.07.087
604 Zineb Lanbouri et al. / Procedia Computer Science 175 (2020) 603–608
2 Lanbouri et al. / Procedia Computer Science 00 (2019) 000–000

• Speed while taking and cancelling orders;

• Difficulty to predict through the noise of volatile markets; ;
• Profit generation from small accumulated margins.

In our ongoing work, we are considering Deep Learning to tackle these challenges and pro-
pose a model with the objective of Forecasting the Closing price one to ten-minutes ahead.
Forecasting and Times Series predictions have been the object of theoretical and empirical studies for many decades.
The Stock Exchange, in particular, tends to be directly influenced by any improvement in that domain. Therefore, the
number of variables and the amount of information are very significant in a way that increases profit and decision
time as well. Such criteria are the topic of this paper. Forecasting is the prediction of some future event by analyzing
the historical data [3] while Time Series is a chronological sequence of observations for given variables X over time
n
(t)
x [4](see equation 1). Many researchers have exploited the area of Stock market prediction using Deep Learning
n
in order to improve forecasting and generate profits for the investors. Deep Learning is a part of Machine Learning
characterized by its non-linearity and representation learning approach while exploring high levels of abstraction
that make algorithms close to human brain. These algorithms achieve positive results in domains such as image
recognition, natural language processing, self-driving cars, and a number of other areas. .
Deep Learning for stock market prediction: [5] opts for an autoencoder composed of stacked RBMs to extract
features.The study is conducted using daily returns and monthly returns in order to enhance momentum which consists
in stocks with high past returns that perform months later. M et al. [6]’s objective is to predict the stock price for two
companies in the IT sector and another one in the pharmaceutical sector. The Convolutional Neural Network has
showed the best results in comparison with the Recurrent Neural Network and Long-Short Term Memory since it
doesn’t rely on past data. Fischer and Krauss [7] uses Long Short Term Memory algorithm on S&P 500 stock prices
and obtain 0.46% as a result. The training data consists of 250 days prior to the prediction, the objectif is to predict
one-day ahead with keras. [8] adopts LSTM in comparison with other Machine Learning algorithms using Open,
High, Low, Close, Volume and 175 TA features. This paper shows how LSTM performs in terms of accuracy, average
and return per operation. And in this study, the selection of LSTM is justified by its ability to handle sequences whilst
distinguishing between recent and early examples [9].
Technical indicators: Most of the automated trading systems adopt Technical Analysis [10]. Unlike Fundamental
Analysis that is based on external factors like macroeconomy and financial analysis, TA is attributed to historic data
and relies on assumptions, for example, prices are defined by the supply-demand relation and changes in supply and
demand cause tendencies to reverse, and can be identified in[11]. [12]’s work includes a combination of momentum,
volume, volatility and cycle based indicators with a total of 200 features, then the f-test is performed for feature
selection. The study showed that the resulting accuracy was highly dependent on the parameters. [13]uses technical
indicators for variance, range and bar information to capture market shocks. The author designs an ensemble model
ARIMA-GARH-NN to discover hidden intraday patterns. The algorithm of choice here is Long-Short Term Memory
which is a special kind of RNN, introduced by Hochreiter and Schmidhuber in 1997. The main contributions of this
work are the following: (1)a forecasting model for high frequency trading using LSTM technique; (2)evaluation of
the model by comparing Technical Indicators’ impact on Forecasting values. The remainder of this paper is organised
as follows: Section II presents the methodology and experiment that we are adopting for this project, High Frequency
Trading with LSTM, section III shows the results and discussion, and finally section IV concludes the article.

2. Methodology and experiment

We want to predict the Closing price using High Frequency data by com-
bining classical financial models and Long-Short Term Memory algorithms.
(t) (t) (t) (t)
X = {X , X , .., X }wherex = { x , x , .., x } The overall structure of our data-driven model includes three major com-
1 2 n 1 2 n
(1) ponents (see figure 1): Data collection, Data preprocessing and Prediction
model.
 ZinebetLanbouri
Lanbouri et al.Computer
al. / Procedia / ProcediaScience
Computer Science000–000
00 (2019) 175 (2020) 603–608 605
3

Fig. 1: Workflow

2.1. Data preprocessing

First of all, we collect S&P500 intraday trading data from Kaggle (www.kaggle.com/nickdl/snp-
500-intraday-data). The original data files comprise 484 observations. For each observation, every
time stamp, Open, High, Low, close price and volume are available from 11/09/2017 to 16/02/2018
with a total of 43148 sequence data. The average spacing between values is 1 minute. Fig-
ure 2 plots Amazon stock behaviour, Open, high, Low, Close prices and Volume over one day.

Our preprocessing approach follows several steps (see figure 3):

step1 for Data cleaning by detecting and removing major errors,
step2 for feature creation, consists in calculating some TA met-
rics, then step3 for feature scaling and normalization.
Data cleaning or cleansing [14] is about detecting invalid and
missing values, then handling them. We are considering data that
are Missing Completely At Random (MCAR) [15]. The dataset
contains missing data that can be split into two categories: val-
ues recorded as missing for a a small period and missing values
for a long period. Then, we applied two techniques: 1. Value im-
putation: missing values for a small period (between 1 and 5 in-
Fig. 2: Amazon Stock Behaviour stances) are replaced with previous values. 2. Instances discard:
missing values for a long period are discarded.
Feature creation Traders are usually confronted with different strategies. We can then distinguish between technical
analysis and fundamental analyis. Technical analysis relies only on the stock market, taking into consideration history
of the price and some indicators like Moving average and MACD among others. When we take a closer look into
works related to stock market forecasting, there are some studies like [16] that believe that ”momentum indicators
capture the duration or the turning point of a trend, the MA is often combined with a momentum like MACD”. This
work relies, also, on calculated technical indicators EMA (Exponetial Moving Average) - MACD (Moving Average
Convergence Divergence) - Bollinger bands and resulting decisions. Exponential Moving Average (EMA) is an indi-
cator of trend which, unlike Simple Moving Average, gives more importance to recent prices.
Moving Average Convergence Divergence (MACD) is a trend-following momentum indicator and when MACD
crosses its signal, it can function as a buy and sell signal.
Bollinger bands [17] is a common heuristic used by traders in order to identify turning points in stock prices. Correc-
tion of overbought/oversold security signals reflects a change in directional movement.
Feature scaling: The train and test data are standardized where x̃ corresponds to x after standardization and Min is the
606 Zineb Lanbouri et al. / Procedia Computer Science 175 (2020) 603–608
4 Lanbouri et al. / Procedia Computer Science 00 (2019) 000–000

minimal value of x over time and over the observations, Max is the maximal value of x. After obtaining the predicted
output, destandardization is applied in order to get the real profitability value (see equation 2).
x−Min
x̃ = Max−Min (2)

2.2. LSTM Prediction model

The objective of the research is to evaluate the performance

of the LSTM network on High Frequency Trading. In case of
memory and sequential information, the Recurrent Neural Net-
work is a widely used modeling approach by [18]. A recurrent
neural network (RNN) can be considered as multiple copies of
the same NN, each copy sends the information to its successor
Fig. 3: Preprocessing steps using BackPropagation Through Time. RNN is simple but it has
a drawback which is known by Curse of dimensionality where
gradient can be either vanishing or exploding. LSTM is a solution to that problem, it is suitable because of its ability
to have memory and to distingush between recent and older data using gates, see figure 4 below adapted from [19].

2.3. LSTM description

hlt−1 hl−1 hlt−1 hl−1

 
Proposed by [21], the LSTM architecture consists of a set of t t

sub-networks, called memory blocks. Each block contains:  

gate  
Input Output
i gate
o
• Memory cell : stores state 
  
hl
• Front door : controls what to learn  Cell
t 
hl−1
 ×
hlt−1  
• Forget door : controls what to forget 
 g ×  ct
 ×
t
• Exit door : controls the amount of content to modify Input


modulation
gate 

The LSTM unit can decide to keep the existing memory via the f Forget gate
doors. This memory therefore allows to: . Forget (clear mem-  
l   l−1
ory) . Input (add to memory) . Output (recover from memory). ht−1 ht
Fig. 4: A graphical representation of LSTM memory cells
The LSTM architecture used in our experiments is given by the used in this paper (inspired from [20]).
l n
equations (3) [22][20]: ct ∈ R is a vector of memory cell, i for
the input gate, f for the forget gate, o for the output gate, h for the hidden state and x for the modulation gate. In these
equations, sigm and tanh are applied element-wise.

3. Results and discussion

We split the dataset into the training set from 11/09/2017 9:30 A.M. to
    17/01/2018 11:50 A.M., and the validation set from 17/01/2018 11:51A.M. to
 i  sigm  l−1  16/02/2018 03:59A.M. Then, experiments were carried out to predict the one-
 f  sigm
  =   T 2n,4n zt minute, five-minutes and ten-minutes ahead price. The key idea is to verify how
 o  sigm
    zlt−1 close each model is to reality in order to know the extent of the risk that the user
x tanh would take while predicting x-minutes ahead. To further improve our analysis, we
clt = f  clt−1 + i  x apply for every observation a model with and without technical indicators (table
1for RMSE values).
zlt = o  tanh(clt )
(3)
 Zineb Lanbouri et al. / Procedia Computer Science 175 (2020) 603–608 607
Lanbouri et al. / Procedia Computer Science 00 (2019) 000–000 5

3.1. Evaluation

Without Technical indicators

The input layer has a dimensionality of 5 features, that consist of Open, High, Low, Close prices and Volume which
is connected to 10 hidden nodes.
With Technical indicators
The number of features is 10: Open, High, Low, Close prices, Volume, EMA12, EMA25, MACD, Bollinger
Up and Bollinger Down. In order to evaluate the network, RMSE is calculated for the standardized test part.

Table 1: RMSE values with and without Technical Indicators

1-minute ahead 5-minutes ahead 10-minutes ahead

3.2. Study Case: Amazon without TI 0.0018 0.0046 0.0046
with TI 0.0721 0.0201 0.0108
We study the effectiveness of our ap-
proach in details in this Amazon study
case with and without Technical Indi-
cators. The forecasting performance is
compared in terms of root mean squared error on the Test part with real values (without standardization) (RMSE,
see fig: 5(b) and fig: 6(b)). As shown in fig: 5(a) and fig: 6(a), we can see that we get a better performance with-
out Technical Indicators specially in Amazon’s case. RMSE is in average 8.24 (see fig 6(a)) while using Technical
Indicators as opposite to 5.74 (see fig 5(a)) without.

(a) Comparison between Forecast and Observed values (b) RMSE for Amazon Test part

Fig. 5: Without Techncal Indicators

4. Conclusion and future work

Traders and stakeholders rely, nowadays, on many complex parameters and financial model calculations in order to
make the right buy-sell-hold decision. The accuracy remains below 50%. That’s why this study is about establishing
the influence of Technical indicators in forecasting. This model is based on the Long-Short Term Memory algorithm
using High Frequency historical data. It confirms that the Closing price can be predicted 10-minutes ahead, 5-minutes
ahead and with a better performance one-minute ahead without the use of Technical Indicators. For future work, we
attempt to focus on Sampling and Back-testing in order to better master this domain. Furthermore, Deep Learning and
LSTM are very promising, we can then expect to improve this model and achieve online prediction for forecasting.
608 Zineb Lanbouri et al. / Procedia Computer Science 175 (2020) 603–608
6 Lanbouri et al. / Procedia Computer Science 00 (2019) 000–000

(a) Comparison between Forecast and Observed values using Technical (b) RMSE for Amazon Test part with Technical Indicators
Indicators

Fig. 6: With Technical Indicators

References

[1] C. Lattemann, P. Loos, J. Gomolka, H.-P. Burghof, A. Breuer, P. Gomber, M. Krogmann, J. Nagel, R. Riess, R. Riordan, et al., High Frequency
Trading, Business & Information Systems Engineering 4 (2012) 93–108.
[2] A. J. Menkveld, High frequency trading and the new market makers, Journal of financial Markets 16 (2013) 712–740.
[3] S. Selvin, R. Vinayakumar, E. Gopalakrishnan, V. K. Menon, K. Soman, Stock price prediction using lstm, rnn and cnn-sliding window model,
in: 2017 International Conference on Advances in Computing, Communications and Informatics (ICACCI), IEEE, 2017, pp. 1643–1647.
[4] P. Malhotra, L. Vig, G. Shroff, P. Agarwal, Long short term memory networks for anomaly detection in time series, in: Proceedings, Presses
universitaires de Louvain, 2015, p. 89.
[5] L. Takeuchi, Y.-Y. A. Lee, Applying deep learning to enhance momentum trading strategies in stocks, in: Technical Report, Stanford University,
2013.
[6] H. M, E. A. Gopalakrishnan, V. Menon, S. Kp, Nse stock market prediction using deep-learning models, Procedia Computer Science 132
(2018) 1351–1362. doi:10.1016/j.procs.2018.05.050.
[7] T. Fischer, C. Krauss, Deep learning with long short-term memory networks for financial market predictions, European Journal of Operational
Research 270 (2018) 654–669.
[8] D. M. Nelson, A. C. Pereira, R. A. De Oliveira, Stock market’s price movement prediction with LSTM neural networks, International Joint
Conference on Neural Networks (IJCNN) (2017) 1419–1426.
[9] I. Sutskever, O. Vinyals, Q. V. Le, Sequence to sequence learning with neural networks, in: Advances in neural information processing systems,
2014, pp. 3104–3112.
[10] B. Huang, Y. Huan, L. D. Xu, L. Zheng, Z. Zou, Automated trading systems statistical and machine learning methods and hardware imple-
mentation: a survey, Enterprise Information Systems 13 (2019) 132–144.
[11] C. D. Kirkpatrick II, J. A. Dahlquist, Technical analysis: the complete resource for financial market technicians, FT press, 2010.
[12] A. Dingli, K. S. Fournier, Financial time series forecasting-a machine learning approach, Machine Learning and Applications: An International
Journal 4 (2017) 3.
[13] J. Sun, K. Xiao, C. Liu, W. Zhou, H. Xiong, Exploiting intra-day patterns for market shock prediction: A machine learning approach, Expert
Systems with Applications 127 (2019) 272–281.
[14] E. Rahm, H. H. Do, Data cleaning: Problems and current approaches, IEEE Data Eng. Bull. 23 (2000) 3–13.
[15] M. Saar-Tsechansky, F. Provost, Handling missing values when applying classification models, Journal of machine learning research 8 (2007)
1623–1657.
[16] K.-Y. Kwon, R. J. Kish, Technical trading strategies and return predictability: Nyse, Applied Financial Economics 12 (2002) 639–653.
[17] M. Butler, D. Kazakov, A learning adaptive bollinger band system, in: 2012 IEEE Conference on Computational Intelligence for Financial
Engineering & Economics (CIFEr), IEEE, 2012, pp. 1–8.
[18] C. L. Dunis, X. Huang, Forecasting and trading currency volatility: An application of recurrent neural regression and model combination,
Journal of Forecasting 21 (2002) 317–354.
[19] M. Dixon, D. Klabjan, J. H. Bang, Classification-based financial markets prediction using deep neural networks, Algorithmic Finance (2017)
1–11.
[20] W. Zaremba, I. Sutskever, O. Vinyals, Recurrent neural network regularization, arXiv preprint arXiv:1409.2329 (2014).
[21] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural computation 9 (1997) 1735–1780.
[22] A. Graves, N. Jaitly, A.-r. Mohamed, Hybrid speech recognition with deep bidirectional lstm, in: 2013 IEEE workshop on automatic speech
recognition and understanding, IEEE, 2013, pp. 273–278.