Abstract—This work uses deep learning methods for
intraday directional movements prediction of Standard &
Poor’s 500 index using financial news titles and a set of technical
indicators as input. Deep learning methods can detect and
analyze complex patterns and interactions in the data
automatically allowing speed up the trading process. This paper
focus on architectures such as Convolutional Neural Networks
(CNN) and Recurrent Neural Networks (RNN), which have had
good results in traditional NLP tasks. Results has shown that
CNN can be better than RNN on catching semantic from texts
and RNN is better on catching the context information and
modeling complex temporal characteristics for stock market
forecasting. The proposed method shows some improvement
when compared with similar previous studies.
Keywords: Deep learning; Recurrent neural network;
Convolutional neural network; intraday stock forecasting
I. INTRODUCTION
The aspiration of any investor is to forecast the market
behavior with the aim of making the best decision when he
comes to buying or selling shares of stocks seeking to
maximize his profits. This is a difficult task because market
behavior is volatile and influenced by many factors such as
global economy, politics, investor expectation and others.
The random walk theory [1] introduces a hypothesis that
stocks prices are defined randomly and for these reasons they
are impossible to forecast. However, advances in artificial
intelligence and the growth of available data have made
possible to forecast the stock price behavior with a better
performance than a random process [2]-[8].
There are three approaches related to the information
required to make a prediction. The first approach, technical
analysis, is based on the premise that the future behavior of a
financial time series is conditioned to its own past. The second
approach, fundamental analysis, is based on external
information as political and economic factors. This
information is taken from unstructured data as news articles,
financial reports or even publishing in microblogs by analysts.
Nofsinger [9] shows that in some cases, investors tend to buy
after positive news resulting in a stress of buying and higher
stocks prices; and after negative news, they sell, resulting in a
decrease of prices. Finally the third approach considers as
relevant all information coming from both, financial time
series and textual data.
Prior works in this area focus on technical analysis. These
works use different statistical techniques and artificial
intelligence models to make a prediction based only in
technical information [10][11]. This approach has a limitation
since the market reacts to external information that is not
contained in the historical data used to extract the technical
information.
Inspired by fundamental analysis, many authors propose
the use of text mining techniques and machine learning
techniques to analyze textual data and take out information
that can be relevant to the forecast process [12]-[15]. The
most relevant works in the area are reviewed in [16][17].
Other works [3][18] use hybrid models by combining text
mining techniques with the technical information. This
approach outperforms other baseline strategies.
Recently, with more computational capabilities and the
availability to handle massive databases, it is possible to use
more complex machine learning models, such as deep
learning models, which presents a superior performance in
traditional Natural Language Processing (NLP) tasks. The
outstanding deep learning models are: Convolutional Neural
Network (CNN) [19]-[22], Recurrent Neural Network
[23][24], specifically the Long Short Term Memory
architecture (LSTM) [25][26], and Recurrent Convolutional
Neural Network (RCNN) [27][28].
Some examples of deep learning models for financial time
series forecasting are shown in [29][30]. Those authors apply
a deep neural network model that use as input events taken
from financial news articles to forecast the direction of prices
of a set of stocks and the S&P 500 index. The main
characteristics in the work described in [30] is the event
representation method and the convolutional neural network
which models the influence of these events on stock prices
behavior in short-term, middle-term and long-term.
From the works cited above it is possible to identify three
key points for the construction of deep learning models. The
first one is the definition of the prediction horizon, the second
one is the temporal effect of a news document and the third
one is the representation type of the information. Regarding
the first point, daily prediction (intra-day) is the most used.
Deep learning for stock market prediction from
financial news articles
Manuel R. Vargas, Beatriz S. L. P. de Lima and Alexandre G. Evsukoff
COPPE/Universidade Federal do Rio de Janeiro,
P.O. Box 68506, 21941-972, Rio de Janeiro, RJ, Brazil
978-1-5090-4253-1/17/$31.00 ©2017 IEEE