Stock Selection and Prediction using Deep Learning
Manasi Naik
Student of MTech Data Science (NMIMS university)
Mumbai, India
manasinaik.nmims@gmail.com
Nilambar Mishra
Team Manager (L & T Financial Services)
Mumbai, India
nmmishra@gmail.com
Abstract— This paper offers systematic analysis on
selection process of stocks and prediction of same using Deep
Learning. The ability to learn huge raw dataset and extract
important data from it, without any prior knowledge makes it
attractive. The performance of Deep Learning Model is highly
dependent on optimizer, loss function, network structure,
activation function, and other parameters. This paper is an
attempt to study the performance of Autoencoder Neural
Network in stock selection process, and future prediction of
these stocks using Long Short-Term Memory (LSTM) model.
The study involves static 5 years daily data from 2013 to 2018
of each company in S & P 500. Assumption for this study is
that user is willing to invest a fix amount in stocks and owns a
stock already. Autoencoder has been applied on stocks which
are primarily filtered according to the amount user wants to
invest and the last day closing price of all the stocks to consider
affordability. RMSE score, between original data and
recreated data, will be used to select stocks. Considering
behaviour of user is neutral, top 50 stocks with least RMSE
score and bottom 50 stocks with most RMSE score have been
selected. Latent Features of these selected stocks will be
examined. Among these 100 stocks, 20 stocks from different
sectors will be selected by checking the correlation of them
with the stock user already owns. LSTM model will predict the
next day future price of these 20 stocks with high accuracy
using adam optimiser.
Empirical results suggest that top 50 stocks with least
RMSE score will give low return with low risk. This implies
that top 50 stocks from the result of autoencoder gives blue
chip stocks. Similarly bottom 50 stocks with high RMSE score
will give high return with high risk. This implies that top 50
stocks from the result of autoencoder gives small cap stocks.
Correlation estimation is noticeably improved after
Autoencoder Neural Network. Future one day stock prediction
using LSTM gives high accuracy with Root Mean Square
Error 2.22. Our study offers handful insights and useful
roadmap for further investigation of stock market analysis and
prediction using Deep Learning Networks.
Keywords— Autoencoder Neural Network, Correlation
Matrix, Future Prediction, LSTM, RMSE, Stocks Selection,
Stocks Prediction.
I. INTRODUCTION
Autoencoder Neural Network is a data compression
model. Autoencoder is a tricky three layered neural network,
known as auto-association before, constructs the "building
block" of deep learning, which has been demonstrated to
achieve good performance in various domains [1]. It has
been used for image reconstruction and face recognition [2].
In modern finance the major challenge we face is
selection of stocks for investment, since there are many
stocks in the market. The compression model of
Autoencoder will help to compress huge data of companies
and recreate it. The Root Mean Square Error of original and
recreated data will give the stocks that explains the market.
Latent Features will have the most important data. It will
help us to understand better correlation between stocks. This
will help to select appropriate stocks from the market. As
correlation between stock and return will provide important
strategic information. It will help us in dynamic allocation
[3].
Risk Return analysis is an important factor in financial
studies, as it plays a key role in most individual decision
making process [4]. The risk return analysis will show the
complexity involved in stocks. Beta is important in CAPM
model. It measures the volatility of the expected return of the
asset in comparison to the market. In this study we will
check the top 50 companies having least RMSE score after
processing data through Autoencoder. After plotting risk-
return plot of top 50 companies, it will show that these
companies have low risk low return. This indicates these
companies might be large cap companies. Similarly risk
return plot of bottom 50 companies with higher RMSE score,
will tend to have high risk and high return. It indicates these
companies might be small cap companies. we have used
CAPM methodology to calculate risk return array.
Prediction of the stocks is another major problem faced
by investors. It has been observed that there are regular and
repetitive changes in stock market in some particular time
interval [5]. This seasonal change in the market is due to
production and sale of seasonal utilities. For example
woollen cloth sale will increase in winter [6]. To overcome
this, we require model which will handle long term
dependency while predicting price in stock market. Long
Short-Term Memory networks (LSTM) are special type of
Recurrent Neural Network (RNN). LSTM can learn long
term dependency. Due to this it learns long term seasonal
effect and gives high accuracy in prediction of next day
price. LSTM model used in this paper is giving around
0.0031 loss while predicting next day price.
II. LITERATURE REVIEW
The research paper has used CORN-K algorithm,
RACORN-K algorithm for portfolio selection. In this paper
DJIA, MSCI, SP500(N), HSI datasets have been used. It
shows promising performance on several stock markets. But
the risk is largely ignored when optimizing portfolios, which
may lead to unreliable profits, particularly in volatile
markets. [7].
The research paper has used Naive Bayes Algorithm.
Author has automated stock purchase model algorithm. It is
useful to create online web portal for trading.[8]
The research paper has used Linear Regression and
Neural Network. Authors have proved that ANN techniques
give more accurate result than Linear Regression. The
drawback is, it is applicable to only single company [9].
The research paper has used GARCH (1,1), EGARCH
(1,1), TGARCH (1,1) algorithms to predict the future 19
Electronic copy available at: https://ssrn.com/abstract=3753033
评论0