Stock Market Prediction using Ensemble Learning

by

Rakin Bin Rabbani

16101213
B. M. Fahad-ul-Amin
16101251
Sumaiya Tanjil Khan
16101212
Farjiya Benta Mahbub
16301033

A thesis submitted to the Department of Computer Science and Engineering

in partial fulfillment of the requirements for the degree of
B.Sc. in Computer Science and Engineering.

Department of Computer Science and Engineering

Brac University
April 2020

c 2020. Brac University
All rights reserved.
Declaration
It is hereby declared that

1. The thesis submitted is our own original work while completing degree at Brac
University.

2. The thesis does not contain material previously published or written by a

third party, except where this is appropriately cited through full and accurate
referencing.

3. The thesis does not contain material which has been accepted, or submitted,
for any other degree or diploma at a university or other institution.

4. We have acknowledged all main sources of help.

Student’s Full Name & Signature:

B. M. Fahad-ul-Amin Rakin Bin Rabbani

16101251 16101213

Sumaiya Tanjil Khan Farjiya Benta Mahbub

16101212 16301033

i
Approval
The thesis titled “Stock Market Prediction using Ensemble Learning” submitted by
1. Rakin Bin Rabbani (16101213)
2. B. M. Fahad-ul-Amin (16101251)
3. Sumaiya Tanjil Khan (16101212)
4. Farjiya Benta Mahbub (16301033)
Of Spring, 2020 has been accepted as satisfactory in partial fulfillment of the re-
quirement for the degree of B.Sc. in Computer Science on April 7, 2020.

Examining Committee:

Supervisor:
(Member)

Mahbub Majumdar, PhD

Senior Professor and Chairperson(CSE), Interim Dean(School of Sciences)
Department of Computer Science and Engineering
BRAC University

Program Coordinator:
(Member)

Md. Golam Rabiul Alam

Associate Professor
Department of Computer Science and Engineering
Brac University

Head of Department:
(Chair)

Mahbub Majumdar, PhD

Senior Professor and Chairperson(CSE), Interim Dean(School of Sciences)
Department of Computer Science and Engineering
Brac University

ii
Abstract
An unpredictable sector of finance market which involves three major roles: in-
vestors, buyers and sellers is called a stock market. Also stock prices may not only
change the future economy of a country but also have direct effects on the current
economic activities of the country. Forecasting stock market is acquiring more at-
tention due to its expected high profit. But the prediction part of stock markets is
considered as quite a challenging task. Though there are various techniques avail-
able for forecasting stock price but the number of methods for forecasting the stock
market accurately is less than usual. Another way of determining future value as
in rise or fall of the future stock price is known as data analysis. The purpose of
this paper is to discuss how accurately the price of stocks in the US stock market
can be predicted, by generating the best possible factors for particular stocks in the
US stock market using machine learning algorithms. After conducting our prelimi-
nary research and then some, we found that it is quite difficult to predict the price
fluctuations of stocks as the market is highly volatile. Furthermore, the number of
uncertain variables in the equation which makes it hard to isolate any one or few
factors that can be used to accurately predict price fluctuations. Therefore, for our
trial runs, we tried to isolate the best factors that can be used to predict the prices of
stocks with sufficient accuracy. For our approach, we implemented Gradient Boost-
ing, Random Forest, Naive Bayes, AdaBoost, Logistic Regression and SVM to run
on our dataset. Based on the outcome of these algorithms we will take the decision
whether to go long or short for a particular stock.

Keywords: Stock Prediction, Machine Learning, Time Series, US Stock market,

stocks, Gradient Boosting, Random Forest, Naive Bayes, AdaBoost, Logistic Regres-
sion, SVM, Feature reduction

iii
Dedication
We dedicate this project to Allah Almighty our creator, our strong pillar, our source
of inspiration, wisdom, knowledge and understanding. Allah has been the source of
our strength throughout this program. We also dedicate this work to our respected
teachers and our parents who has encouraged us all the way and made sure that we
give it all it takes to finish that which we have started.

iv
Acknowledgement
At first, all praise to the Great Allah for whom our thesis have been completed
without any major interrupt.

We wish to express our sincere thanks to our supervisior Professor Mahbub Majum-
dar of the Computer Science and Engineering department at BRAC University for
providing us with all the necessary facilities for the research.He always encouraged
us to think innovatively and bring new ideas.He has been a constant source of en-
couragement and enthusiasm throughout the time of our thesis.

We are also hugely indebted to Golam Rabiul Alam, Associate Professor of BRAC
University, for sharing his ideas and experiences which helped us to make right
choices in the implementation phase.

We take this opportunity to express gratitude to all of the Department faculty mem-
bers for their help and support. We also thank our parents for the encouragement,
support and attention.Also the cooperation and understanding of our group mate
has helped us to finish our thesis without any hassel. We also like thank all of our
friends who always inspired and motivated us.

v
Table of Contents

Declaration i

Approval ii

Abstract iii

Dedication iv

Acknowledgment v

Table of Contents vi

List of Figures viii

List of Tables ix

Nomenclature x

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Classification of stock . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Stock Price Setting Process . . . . . . . . . . . . . . . . . . . 2
1.1.3 Effect of stock price . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.4 Strategies of Investment . . . . . . . . . . . . . . . . . . . . . 3
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Paper orientation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Literature Review 5
2.1 Previous Researches . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Our improvements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3 Machine Learning Algorithm 9

3.1 Support vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Recurrent neural network . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Naive Bayes Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4 Adaboost Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.5 Gradient Boosting Algorithm . . . . . . . . . . . . . . . . . . . . . . 13
3.6 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.7 Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

vi
4 Methodology 16

5 Data Processing and Implementations 18

5.1 Primary and Secondary Factors . . . . . . . . . . . . . . . . . . . . . 18
5.2 Application of Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 20
5.3 Factor Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

6 Results 28
6.1 Returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
6.2 Comparison On performance . . . . . . . . . . . . . . . . . . . . . . . 32

7 Conclusion 33
7.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
7.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

Bibliography 37

vii
List of Figures

3.1 Possible Hyperplane . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.2 Adaboost Algorithm Working Process . . . . . . . . . . . . . . . . . . 12
3.3 Sigmoid Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4.1 Workflow of Project . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

5.1 System structure of the Model . . . . . . . . . . . . . . . . . . . . . . 21

5.2 Mean period was Return by factor Quantile . . . . . . . . . . . . . . . 22
5.3 Factor Weighted Long Short Portfolio Cumulative Return 1D . . . . . 23
5.4 Factor Weighted Long Short Portfolio Cumulative Return 5D . . . . . 23
5.5 Factor Weighted Long Short Portfolio Cumulative Return 22D . . . . 23
5.6 Period Wise Return by Factor Quantile . . . . . . . . . . . . . . . . . 24
5.7 Cumulative Returns by Quantile 1D . . . . . . . . . . . . . . . . . . . 25
5.8 Cumulative Returns by Quantile 5D . . . . . . . . . . . . . . . . . . . 25
5.9 Cumulative Returns by Quantile 22D . . . . . . . . . . . . . . . . . . 25
5.10 Top Minus Bottom Quantile Mean for 1D . . . . . . . . . . . . . . . 26
5.11 Top Minus Bottom Quantile Mean for 5D . . . . . . . . . . . . . . . 26
5.12 Top Minus Bottom Quantile Mean for 22D . . . . . . . . . . . . . . . 26
5.13 Normal distribution Quantile for 1D . . . . . . . . . . . . . . . . . . 27
5.14 Normal distribution Quantile for 5D . . . . . . . . . . . . . . . . . . 27
5.15 Normal distribution Quantile for 22D . . . . . . . . . . . . . . . . . . 27

6.1 Returns for 1D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

6.2 Returns for 5D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6.3 Returns for 22D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6.4 Total Performance for 1D . . . . . . . . . . . . . . . . . . . . . . . . 31
6.5 Total Performance for 5D . . . . . . . . . . . . . . . . . . . . . . . . 31
6.6 Total Performance for 22D . . . . . . . . . . . . . . . . . . . . . . . 31

viii
List of Tables

6.1 Outcomes of The Ensemble Model . . . . . . . . . . . . . . . . . . . . 32

ix
Nomenclature

The next list describes several symbols & abbreviation that will be later used within
the body of the document

 Epsilon

Θ Theta

AN N Artificial Neural Network

ARIM A Autoregressive Integrated Moving Average

D Day

GA Genetic Algorithm

GB Gradient Boosting

GM M Generalized Methods of Moments

HM M Hidden Markov model

LDA Linear Discriminant Algorithm

LS − SV M Least square Support vector Machine

LST M Long-short term Memory

M Month

ML Machine Learning

OOB Out Of Bag

P N N probabilistic Neural Network

P SO Particle Swarm Optimization

SV M Support Vector Machine

SV R Support vector Regression

x
Chapter 1

Introduction

1.1 Background
Public markets which are for issuing, buying and selling stocks that can be traded
on a stock exchange or over-the-counter are referred as the stock market [15]. Stocks
refer to equities which represent fractional ownership in a company and the stock
market is a platform where ownership of investable assets can be bought or sold by
the investors. An efficient stock market can create impact on economic development
since the companies can influence the public and make capital from them.

1.1.1 Classification of stock

There are different kinds of parameters depending on which stocks can be classified.
Some of these parameters are -size of the company, industry, risk, dividend payment,
volatility as well as fundamentals.

One of the most common parameter is Stocks which are based on the ownership
rules. Here, the issuing company takes decision whether it will issue preferred, com-
mon or hybrid stocks. Though preferred and common stocks are quite similar yet
the difference between these two is that investors will get certain amount of divi-
dends per year from the preferred stock whereas no such amount is given is common
stock. As a result, preferred stock is more volatile in nature than the common stock
in terms of price. Moreover, Hybrid stock is the combination of both preferred stock
and common stock which means preferred shares that can be converted into a fixed
number of common stocks at a particular time[19].

Dividend payments are another parameter for classifying stock market. Dividends
are percentage of income that is generated when shares are sold at a profit. First
of all, Income stocks are the stocks that provide a higher dividend in relation to
their share price. So, a higher dividend means larger income. Stable companies
that distribute consistent dividend are referred as income stocks. Since these com-
panies are not that stable so the risk associated with the stock price is less. Risky
stocks are more potent in making greater profits. On the contrary, Low-risk stocks
generate lower returns. The second type of stock under this category is Blue-chip
stocks which are considered to be safe since renowned companies hold these type of
stocks. As the companies does not have much liabilities which help them to have

1
regular dividends. Furthermore, Beta stocks are a measure to indicate the risk or
volatility of stock in terms of beta. A positive value of beta refers it is in sync with
the market and a negative beta value implies that it is against the market. Though
the absolute value of beta is more important. Higher the beta, greater the volatility
and thus more the risk. A beta value over 1 means the stock is more volatile than
the market [22].

1.1.2 Stock Price Setting Process

The worth of a company which is also known as market capitalization is determined
by the product of the company’s stock price and the number of shares outstanding.
[28] Thus setting the price for stocks is very important. There are many factors by
which share price can be determined. Among these factors the most used factors
are herd instinct and market news.

• Herd Instinct: The price of stock can also be influenced by what is known as
herd instinct, a tendency for individuals to mimic the behavior of a larger group.
For instance, as more and more people purchase a stock, driving the price higher
and higher, other people will jump on board, thinking all the other investors have
to be right (or know something not everyone else knows).

• Market News: Other factors may have impact on prices, and can result in sudden or
temporary price changes. Types include quarterly releases, political activities, busi-
ness material activities and economic news. All sorts of news or economic events do
not influence all securities as in sometimes it is affected by weekly news but not by
monthly status [26].

These are some of the ways to set the stock price. Apart from that Supply and
Demand is also another factor to set stock price.

1.1.3 Effect of stock price

Economics demand and supply law affects stock prices. The demand and supply
law says that if there is high demand, the price of that material will increase and
the other way around for high supply. Likewise, if there is a lot of asking price for
a stock, the price of that stock increases and if a stock has more than usual offering
price that stock price will decrease. The price of a company’s stock is more affected
by the law of supply and demand rather than the company’s success. Whether a
company’s stock is desirable or not depends on a variety of reasons such as it can
be strength of an industry sector or the popularity of a brand.

• Cyclical stocks: In a disrupted economy some companies’ growth is moderat-

ing or fastens in a booming economy thus some companies are affected by economic
patterns. As a result, prices of these stocks continue to fluctuate more with changing
economic conditions. Companies are rising during economic booms and dropping as
the economy slows. automobile companies are the best example of cyclical stocks.

2
• Defensive stocks: Unlike cyclical stocks, defensive stocks are not affected by eco-
nomic conditions issue. Best examples are stocks of food, beverage, pharmacy and
insurance firms. These stocks are usually preferred in weak economic conditions,
while cyclical stocks are preferred in booming economies.

1.1.4 Strategies of Investment

As the nature of the stock market is dynamic which makes the share market very
vulnerable. So for any investor it is very important to make the right investment at
the right time. The convention is to sell a share when the price is high and buy it
when the price is low. There are five strategies which investors follow while investing
in any share. Those are :-

• Buying Cheap and Selling Dear: Coming into the market when prices and senti-
ment are depressed and selling out when both are exalted.

• Long-Pull Selection: Picking out companies which will prosper over the years
far more than the average enterprise also known as ”growth stocks”.

• Bargain Purchases: Selecting shares which are selling considerably below their
true value, as measured by reasonably dependable techniques.

• General Trading: Anticipating or participating in the moves of the market as

a whole, as reflected in the familiar ”averages”.

• Selective Trading: Picking out stocks which over a period of a year or less will do
better than the market.

• Short Selling: Short-selling also known as shorting is when an investor borrows

shares from a shareholder and sells them immediately, hoping that he or she will be
able to buy them again at a cheaper price later on, return them to the lender and
pocket the profit [17].

Each approach requires a rational, disciplined, consistent, systematic application.

In a nut shell the whole idea of stock market is to make profit. Thus investors
are so much focused so that they can anticipate the stock market and make their
Investment accordingly.

1.2 Motivation
To make any business decision investors are required to analyze their stock market.
But doing that all by their self is quite challenging and sometimes are not capable
to do that. Alternatively, they have to focus on observing others investors stocks
too. Again if the investors want to do it on their own, they face the difficulties
of collecting the data, analytical tools as it is difficult to predict the share market.
But in today’s world using machine learning techniques and algorithms used in
artificial intelligence it is possible to forecast the stock market quite well and the
accuracy of this prediction largely depends on the available amount of data as well

3
as how well the model has been trained. For this purpose, we are using quantopian
platform which provides huge amount of financial data. To predict the dynamical
stock market quantopian is used where we can provide our algorithms and test the
accuracy. So the main goal of our thesis is to help investors to predict the stock
market correctly.

1.3 Objective
The process of making decision based on the previous experience and available in-
formation is known as prediction. And because of the dynamic nature of the stock
market predicting stock market accurately is a very difficult task to do as we dis-
cussed before. The easier way to decide for the future on an option totally depends
on the accuracy of the prediction. Moreover, stock market has a positive relation
with the country’s finance cause if stock markets rise, the financial growth of a
country would be high and vice-versa. As stock market has great impact on the
economy many has tried to predict its nature previously. Also there is many avail-
able machine learning methods such as Neural Network, Evolutionary Algorithms,
SVM, Neuro-Fuzzy, HMM, ARIMA, Decision Tree, Random Forest, Adaboost, GB,
Linear regression, Logistic Regression, Naive Bayes etc. In this paper we have fo-
cused to create a system which can provide more accurate result and able to predict
share market with less amount of hassle. To achieve our goal, we intend to use data
which cost nothing as we are using quantopian platform.

1.4 Paper orientation

This section mainly focusses on the brief description of our entire paper and the steps
which we have taken to implement machine learning algorithm and related works
to predict the stock market. In the next chapter Literature review is described
and the information from the past papers have helped us to acquire vast knowledge
on this topic. Then the following section is about the machine learning algorithm
. Chapter four is about the data processing and inplementation of the required
algorithm. The result and the conclusion of our thesis is discussed in chapter six
and seven accordingly.

4
Chapter 2

Literature Review

In this chapter we will discuss about the study we did on previous papers of stock
market prediction and how our work is improved considering the limitations from
previous papers.

2.1 Previous Researches

Stock market prediction has been the ultimate objective for many investors. The
changes in the companies are dependent on how the market reacts [20]. There were
quite a few different implementations of machine learning algorithms for the pur-
poses of making stock market price predictions. Different papers experimented with
different machine learning algorithms that they implemented in order to figure out
which models produced the best results [29].

In [32] Tianxin Dai et al. experimented with two machine learning algorithms SVM
and Logistic regression. They trained the model iteratively to make the prediction
as accurate as possible. After running the baseline model of Logistic regression
algorithm they got 0.0123 profit, 55.07% precision 30.05% recall 38.39% accuracy.
Similarly, λ -HL of Logistic regression algorithm gave them 0.0186 profit, 59.39%
precision 27.43% recall 41.58% accuracy. After training the model for one hour by
the help of SVM, they discovered 0.0926 profit, 56.45% precision, 38.10% recall and
43.92% accuracy. The use of machine learning algorithm for the forecasting stock
market is quite challenging since various factors and market trends must to be taken
into account.

In another paper [21] Alice Zheng and Jack Jin came up with the idea that traders
can use multiple analysis techniques such as fundamental analysis, technical analysis
and quantitative analysis to future stock price of their company for the purpose of
training. To predict stock market they choose to work with four machine learning
algorithms which are SVM, Bayesian Network, Simple Neural Network and Logistic
Regression. In SVM they were forced to work with smaller dataset since Gaussian
Kernel maps infinite dimensional features for their predictors which lead to time
complexity of O(n2 samples ). They found out in their output, the rate of error were
40% to 50% which implies that their choice of predictors were not appropriate.

5
Guanting Chen et al. [31] mentions that Deep Learning has been used widely and
demonstrated to be an amazing machine learning tool now a days. Among the deep
learning algorithms they selected LSTM algorithm to forecast the future returns of
the stock price. They used TensorFlow to experiment on their dataset. They face
some constraints as they could not articulate with different neurons and layers in
LSTM.

SVM is used to reduce the input dimension for load forecasting [8]. On the other
hand Least Square Support Vector Machine is a very famous SVM algorithm for
predicting the stock market. In [14] paper, we found that they used LS-SVM along
with PSO in some stock sectors of various companies to predict the daily stock
price. They concluded by stating that the error rate was less because of using both
algorithms together rather than using only LS-SVM.

In [16] they build a predictive model by combining four machine learning algorithms
which are Artificial Neural Network (ANN), Support Vector Machine (SVM), ran-
dom forest and Naive Bayes in the Indian stock market. Firstly, to input data they
considered the computation of ten technical parameters. Secondly they tried to rep-
resent these technical parameters as trend deterministic data. After the evaluation
of that predictive model it has been seen that random forest algorithm dominated
on the overall performance.

To determine how a financial market behaves, a hybrid model was proposed in [9]
MR Hasan’s paper by accommodating HMM, ANN and GA which helps to analyze
the stock market in depth and optimize the required parameters.

To make stock market prediction and trend analysis automated Abraham, Nath
and Mahanti established a hybrid computing method. They mainly focused on pre-
dicting one day ahead by using Nasdaq-100 index of the Nasdaq stock market along
with neural networks. An encouraging result was found when they used a neuro-
fuzzy system to analyze the anticipated values [5].

Chen introduced investment strategies based on a feed forward neural network,

PNN. Historical data has been used to train this model to generate better results
in comparison to the one which were estimated after using the random walk and
parametric GMM models [7].

Besides using machine learning algorithms, data mining techniques had also been
used for the prediction purpose of the dynamic stock market. Ou and Wang tried to
predict the Hong Kong stock market by the help of ten data mining techniques in
total. Some of those are LDA, K-nearest neighbor classification, kernel estimation
based Naive Bayes, SVM, LS-SVM etc [11].

Another Prediction System was introduced for Taiwan Stock market which is ba-
sically a recurrent neural network trained by extracting features through ARIMA
[3]. Autocorrelation and partial correlation function plots are examined just after
differencing the raw data of the TSEWSI series. Second difference data are used to
train neural networks for better prediction. The result showed that network which

6
were trained using four-year weekly data can forecast accurately for six weeks.

In another paper [10] Sheng-Hsun Hsu et al. talked about the dynamic relationship
between stock price series and their predictors. To identify this problem with the
help of only one artificial technique is quite challenging. So, a two-stage architecture
came into help. The first stage is SOM which is used to break the input dataset
into clusters as in data points are grouped together on the basis of similar statisti-
cal distributions. Afterwards SVR is applied on these clusters to foresee financial
indices. The two-stage architecture proved to provide more enhanced outcomes.

Jyotirmoy [18] in his paper proposed sentiment analysis and supervised machine
learning to predict the security movements in stock market and the profit from it.
Sentiment analysis is run on data collected from sources such that newspapers, ar-
ticles and blogs to interpret the sentiment from the data. The limitation they face
is to collect minute data because it is expensive and not available in public domain.
This helps to find out the link between the market sentiment and market movement.

Marshall [25] observed best performance using a Support Vector Machine (SVM)
with a radial basis kernel, when compared with Logistic Regression, Bayesian Net-
work, and a Simple Neural Network. Due to their limited processing power, they
were only able to use a subset of their data for training their model, and recom-
mended that a more powerful processor be used to achieve better results.

On [27] it was discussed that when performing stock price prediction, [30] it came
out to be that ANN the algorithm that was once popular for prediction suffers from
overfitting due to large numbers of parameters that it needs to fix. This is where
support vector machine (SVM) came into play and it was suggested that this method
could be used as an alternative to avoid such limitations, where according the VC
theory [13] SVM computes globally obtained solution unlike those that are obtained
through ANN which almost in all the cases tends to fall in the local minima. It was
seen that using an SVM model the accuracy of the predicted output came out to be
around 57% [4].

Since a random forest ensembles multiple decision trees. In [12] it has proposed
ensemble techniques in order to lessen the over-fitting problem of training data.
This technique is mainly used for the analysis purpose of Stock returns.

Moreover, K. Sharif and M. Abu-Ghazaleh [33] came up with an advancements

in the machine learning field which lead to ensemble methods used for forecasting.
Quantopian has its simulator which allows ensemble methods for prediction and
trading on a daily basis. The output from each simulation is compared against an
appropriate standard and performance evaluation is done accordingly.

7
2.2 Our improvements
We have observed various algorithms which were implemented to predict stock mar-
ket but all these algorithm had some shortcomings. So, we worked on it and tried to
improvise it by using a different approach which will be efficient for the prediction.
Since we chose ensemble learning so for that we selected five machine learning algo-
rithms. These are - Support Vector Machine(SVM), Random Forest, Naive Bayes,
Logistic Regression, Adaboost. All these are highly used to financial time series.

At first we chose SVM but later we did not work with it further because SVM
takes a longer time and required smaller dataset is used to give the result. Then we
moved to Random Forest and this gave us a more accurate result as it deals with
bootstrapping and each time it works on different set of data which increases the
accuracy overall. After that we tried other two algorithms as in Logistic Regression
and Adaboost. But Adaboost had some limitations in providing the result up to
the mark. So, for that we worked with an advanced version of Adaboost which is
Gradient Boosting as it is much more flexible in terms of giving the result. Assump-
tion of independent predictors is the main constraint of Naive Bayes. It assumes
implicitly that all the attributes have mutual independence which is not possible in
real life scenario. Considering all these we decided to mix and match the algorithms
and then calculate the average of the results given by each of the algorithms.

8
Chapter 3

Machine Learning Algorithm

In this chapter side the machine learning algorithm we have chosen to use will
be discussed. As there are various machine learning algorithms are available but
we decide to work with Logistic Regression, Random Forest, Adaboost, Gradient
Boosting, Stochastic Gradient Descent and Naive Bayes Algorithm.

3.1 Support vector Machine

Support vector machine in short SVM is a classification and regression based al-
gorithm. It is used to maximize predictive accuracy whilst avoiding the overfitting
of data. It is used for applications such as handwriting, face, text and hypertext
classification, Bioinformatics etc [5]. SVM is used to achieve maximum separation
between data points. Hyperplane is a part of SVM that maximize the separation of
data points by increasing the line width with increments. It starts by drawing a line
and two equidistant parallel lines. Next the algorithm picks a stopping point so that
the algorithm does not run into an infinite loop and also picks an expanding factor
close to 1 example 0.99. The main purpose of SVM is to determine a hyperplane in
an N-dimensional space that distinctly classifies the data points, here n is the num-
ber of chosen features. Although there is various possible hyperplane in support

Figure 3.1: Possible Hyperplane

9
vector algorithm but the best one is that hyperplane which has maximum margin.
In other word, the best hyperplane is the one which provides maximum distance
between data points for both classes. As these will result in some reinforcement so
that future data points can be classified with more confidence. The loss function
that helps maximize the margin is hinge loss. To maximize the margin between
hyperplane and data points the following equation is used:
c(x, y, f (x)) = 1 − y ∗ f (x) (3.1)
The hinge loss function is equal to 0 when y*f(x) becomes greater or equal 1. All
though the support vector algorithm is very useful but it took lots of time to run in
quantopian. So in spite of many advantages we chose not to use it in our ensemble
learning.

3.2 Recurrent neural network

Recurrent Neural Network (RNN) is a form of deep supervised learning. RNN uses
the previous memorized memory makes sense on the current data as in the output of
the previous step is fed as the input for the current step. The same process continues
and results in a loop. The network that is created by the help of the loop is called
the recurrent network. There are three types of RNN:
1. One to Many : For one input there are multiple outputs.
2. Many to One : For many input there only one output.
3. Many to Many : For many input there are multiple outputs.
All inputs and outputs are independent of each other in conventional neural net-
works, but in cases such as when it is important to predict the next word of a
sentence, the preceding words are needed to be remembered. Thus RNN came into
being with the aid of a Hidden Layer which solved this problem. Hidden layer is
the key and most significant aspect of RNN which retains certain details about a
sequence. RNN is best for time series data as it can work on both stationary and
non-stationary data. Traditional neural networks do not have persistence whereas
RNN address this issue since they are networks with loops that allow information
to persist. RNN occurs with recursive formula as it works in loop.
s(t) = fw (st−1 , xt ) (3.2)
here, st =state at time step t
xt =input at time step at t
fw =recursive function
st = tanh(ws st−1 + wx xt ) (3.3)
here, tanh is the activation function
wx and ws = weight But there are some shortcomings in RNN as in when the time
gap grows the density of information increases so it becomes impossible to learn
to collect these information altogether. Another limitation is vanishing gradient
where each layer’s weight is multiplied but after certain time the updated weight
will become close to zero as in it will get vanished so the whole training network
will suffer for this.

10
3.3 Naive Bayes Algorithm
Naive Bayes, a classification technique based on Bayes’ Theorem. This classifications
assumes independent predictors. In simple terms, a Naive Bayes classifier assumes
that the presence of a particular feature in a class is unrelated to the presence of
any other feature. Naive Bayes classifiers are formed by a group of classification
algorithms based on Bayes’ Theorem. It is not represented as a single algorithm
rather a common principle is shared by a family of algorithm, i.e. every pair of
features being classified is independent of each other.

One example will help to understand this clearly: a flower may be considered to
be a rose if it is red, have round petals, and have good fragrance. Even if these
features are dependent on each other as in on the existence of the other features,
all of these features contribute independently to the probability that this flower is
a rose and that is why it is known as ‘Naive’.

The construction of Naive Bayes model is easy and it can be used for very large
data sets. Along with simplicity, Naive Bayes performs really well in comparison
to highly sophisticated classification methods. Bayes rule is used in Naive Bayes
classification.

P (X|CK )
P (Ck |X) = P (Ck ) ∗ (3.4)
P (X)
The above equation is defined as the probability of a particular event Ck which is
occurring for feature vector x. To estimate the value for P(Ck |x) we first need to
compute P(x| Ck ). To find this we need the following formula:
d
Y
P (X|ck ) = P (Xj |cK ) (3.5)
j=1

We assume that a particular value of xj is independent of the occurrence of all other

xj . Then the final formula which forms the x feature vector for a particular event
Ck is given below: Qd
j=1 P (Xj |CK )
P (Ck |X) = P (Ck ) ∗ (3.6)
P (X)
This formula is mainly used for the classification and to minimize the percentage
error we should keep the value of P(Ck |x) higher.

A simple probabilistic model which is based on the Bayes rule with a strong con-
ditional independence assumption is known as a Naive Bayes classifier. The Naive
Bayes model involves a simplifying conditional independence assumption. That is,
given a class (positive or negative), the words are conditionally independent of each
other. Due to this simplifying assumption the model is termed as “naive”. It makes
fast classification algorithm applicable for the problem without being affected by the
assumption in the accuracy of text classification. A simple Naive Bayes classifier can
be enhanced to match the classification accuracy of these more complicated models
for sentiment analysis.

11
The advantages of using Naive Bayes as our classifier are:

• Naive Bayes classifiers due to their conditional independence assumptions are

extremely fast to train and can scale over large datasets.

• They are robust to noise and less prone to over-fitting.

• Ease of implementation is also a major advantage of Naive Bayes.

3.4 Adaboost Algorithm

AdaBoost also known as “Adaptive Boosting”, is the first practical boosting algo-
rithm which were proposed by Freund and Schapire in 1996. It focuses on classifi-
cation problems and the main objective is to transform a set of weak classifiers into
a strong one.

Figure 3.2: Adaboost Algorithm Working Process

The final equation for classification can be represented as

m
X
f (x) = sign( Θm fm (x)) (3.7)
m=1

where fm stands for the mth weak classifier and θm is the corresponding weight. It
is exactly the weighted combination of M weak classifiers.

The main mechanism of adaboost is quite simple which can be represented for a
given data set containing n point as such- xi Rd ; yi {−1, 1}
Here -1 denotes the negative class while 1 represents the positive one. Initialize the
weight for each data point as:
1
w(xi , yi ) = (3.8)
n

12
Here, i=1,2,3.....n. Boosting algorithm increases the accuracy of weak learners [1].
A weak learner uses a simple assumption to bring out a hypothesis that comes from
an easily learnable hypothesis class and performs at least slightly better than a ran-
dom guess. If each weak learner is properly implemented, then Boosting aggregates
the weak hypothesis to provide a better predictor which will perform well on hard
to learn problems.

The AdaBoost algorithm outputs a “strong” function that is a weighted sum of

weak classifiers. The algorithm follows an iterative process where in each iteration
the algorithm focuses on the samples where the previous hypothesis gave incorrect
answer. The weak learner is returns a weak function whose error is t such that-
m
X
t = LD t (ht ) = Dit [hi (xi 6= yi )] (3.9)
i=1

Then a specific classifier is assigned a weight for ht as follows: So, the weight given
to that weak classifier is inversely proportional to the error of that weak classifier.

3.5 Gradient Boosting Algorithm

Gradient Boosting is an advanced version of AdaBoost which trains many models
in a gradual, additive and sequential manner to find out the limitations of weak
learners. They find the limitations by taking loss function into account. The loss
function is calculated by using gradients. The equation for loss function is:

y = ax + b + e (3.10)

here, e needs a special mention as it is the error term. The loss function is a way of
measuring indicating how effective model’s coefficients are at fitting the underlying
data. For example, in order to predict the sales prices by using a regression, then
the loss function can be based on the error between true and predicted house prices.
The main advantage of using gradient boosting is it allows to optimize user specified
cost function. It mainly converts weak learners into strong learners.

3.6 Random Forest

Random forest is a decision tree based non-linear technique [1]. Here the word ran-
dom means we need to select the data randomly this method is called bootstrapping.
Usually decision tree has a problem of over-fitting so to overcome this problem ran-
dom forest came into help. An advantage of using Random Forest is that it can be
used for both classification and regression [6]. For our purposes, we applied Random
Forest for classification. The algorithm works as follows:
I. At first subset from the entire data-set is selected as in bootstrapping is done.
This subset is used to make a decision tree. In this way multiple decision trees are
built from multiple subsets.
II. Next, by constructing different subsets we get different decision trees and each

13
tree gives a result and the final classification is done by a method called bagging.
In order to classify, bagging choose that class which gets majority votes according
to the classification given by each of the decision trees.

We can then easily estimate the error in our results in the following manner:
I. The data sample which were left out from the original dataset during bootstrap-
ping are called ”Out of bag” (OOB). Then this sample is being predicted by using
the decision trees previously constructed out of bootstrapped sample.
II. We then sum the out of bag sample predictions and determine the error rate. It
is called the OOB estimation of error rate. There is also no need for cross-validation
or separate test set to get an accurate estimate of the test set error in random forests.

Given enough trees have been grown, the OOB estimate of error rate is quite accu-
rate.

3.7 Logistic Regression

Logistic regression is a supervised classification algorithm. For a classification prob-
lem, the output (target variable) or the class y, can take only discrete values for a
given set of inputs(features), X. Logistic regression can be represented as a sigmoid
function which is basically an S-shaped curve that can take any real-valued number
and map it into a range between 0 and 1.

Figure 3.3: Sigmoid Function

Input values (x) are combined linearly using weights or coefficient values represented
as Beta) to predict an output value (y). The logistic regression equation is:

y = eb0+b1∗x /1 + eb0+b1∗x (3.11)

14
Where y is the predicted output, b0 is the bias or intercept term and b1 is the
coefficient for the single input value (x). Each column in your input data has an as-
sociated b coefficient (a constant real value) that must be learned from your training
data.
The actual representation of the model that you would store in memory or in a file
are the coefficients in the equation (the beta value or b’s).

15
Chapter 4

Methodology

In this chapter, we have discussed about how we used quantopian platform to make
our Custom-Factor and using the best possible machine learning algorithms we tried
to predict the stock price.

Figure 4.1: Workflow of Project

16
One of the main advantage of quantopian is that it provides US stock market data
from 2002 to till date. Total 39 factors are provided by quantopian. Though all of
the factors cannot be incorporated in our algorithms because some of the factors
can give positive result and negative result. So basically quantopian helps us to
select the best factors among the 39 factors which get best fit in our algorithms
and as a result helps us to make our customized factor. As some of the factor
gives positive outcomes and some gives negative in our algorithm. Moreover, it pro-
vides us the option to implement our own algorithm in the live trading environment.

In the quantopian, platform we implemented our own strategy incorporating several

machine learning algorithms into it. We began by collecting the data of 1500 of
the top stocks in the market using the QTradableStocksUS() function provided by
Quantopian. Next, we needed to import the factors that would go into our algo-
rithm. We also had to make our own custom factor which will be described in later
part. Some of the factors are Asset Growth 3M, Asset to Equity Ratio, Mean Rever-
sion 1M, EBIT to Assets, EBITDA Yield, Moneyflow Volume 5D, Price Momentum
3M, Return on Invest Capital and more. These are generally considered among the
best features to use of our algorithms, but we narrowed down our list through trial
and error.

We defined our universe to be the US stock market, and performed a preliminary

preprocessing of data to remove any NaN values in our datasets. From the dataset
provided we chose only the top 70% of stocks which had the highest increase, and the
lowest 30% of stocks which dropped the most to run our algorithms on. The upper
70 percent were all labelled as ‘1’ and the lower 30% were labelled as ‘-1’ and the rest
were labelled as’0’. We will mainly focus on the ’1’ and’-1’ part which will eventually
tell us if we should go long or short. Next we trained our algorithm based on this
criteria and selection. We were trying to predict the outcome of n+5 days where
n is considered as the current day, given today’s opening price and our list of factors.

We also generated our very own factor that is the custom factor using a combi-
nation of the factors which were chosen through trial and error. With the help
of these customized factor, AdaBoost, Random Forest, Logistic Regression, Gradi-
ent Boosting and Naive Bayes perform the classification. The column we named
‘EnsembleAlgo’, was the average of the five algorithms which was fed into the algo-
rithms when we were trying to make our prediction. With the help of the pipeline
we are merging all the five algorithms to get our final result. We ran two trials using
our model and two sets of training data. One was the time series data of the daily
opening prices of our chosen stocks from ‘2018-01-01’ to ‘2019-03-01’ and the other
was from ‘2015-01-01’ to ‘2016-03-01’.

After performing these, we got to realize that due to the volatility of the mar-
ket just feeding all the factors into the ML algorithm will not give a very consistent
result overall because a given factor can have both a positive or a negative effect on
the prediction itself at different given time period with respect to the market. To
overcome this problem, we had to implement feature reduction dynamically through
trial and error.

17
Chapter 5

Data Processing and

Implementations

In this chapter we discussed about how we processed the data and how the cus-
tomized factors were generated.

5.1 Primary and Secondary Factors

We start with primary data as the daily (1 day) open, close, high price, low price,
and volume of each stock in our tradeable universe. Previously Q500US and Q1500US
were used but nowadays QTradeableStocksUS is in use which is a better tradable
universe, bigger in size and includes both Q500US and Q1500US. Moreover, QTrade-
ableStocksUS provides a better environment for our dataset.

Since we are working with ensemble learning we need to use pipeline to merge all
the algorithm and run it in quantopian by the help of run pipeline. USEquityPric-
ing is our main dataset and we will be working with the latest values from it. By
the help of morningstar we can extract all the financial statement, sheets and also
ratios like balance sheet, cash flow statement, income statement, operation ratios,
earnings report, valuation which acts as the primary factors. We are importing
some built-in factors like simple moving average, AverageDollarVolume, Returns
etc which helps us to customize the factor. Talib library helps us to collect financial
analysis function. Numpy performs different mathematical operations, panda ma-
nipulates the dataset. Alphalen is used to measure the effectiveness of how perfectly
our customized factor works. To manage performance and risk analysis we are using
pyfolio library. Lastly, we are using scipy library to optimize our algorithm to get
an efficient result.

As we have already extracted our primary data with the help of Morningstar so this
primary data will help us to create our own factor. In quantopian though there are
multiple factors available which are stored in all factor dictionary but we are not
using all of them instead of that we are customizing our very own factor by filtering
out from the available factors using the make factor method. This method consists
of the following factors:

18
Asset Growth 3M: It measures the overall growth of a stock in a three-month
period. Here we are considering 22 days as a month.

(Asset value P rior − Asset V alue Current) ∗ 100

AssetGrowth = (5.1)
Asset value P rior
Asset to Equity Ratio: It is a ratio which shows the relationship between the to-
tal assets of the firm to the portion owned by shareholders. A higher ratio indicates
that the firm is in huge debt.

EBIT to Asset: It represents a firm’s profitability before payment of any interest

and taxes. In other word it is also termed as operating earnings and operating profit.

EBIT = N et income + Interest + T axes (5.2)

EBITDA Yield: This value is calculated before firm distribute its earning among
interest, taxes, depreciation and amortization to determine the firms profitability

Return on Invetsed Capital:In short (ROIC),which is used to examine how effi-

ciently a firm use its capital to earn profit.This ratio gives an idea of how effectively
a company is using its money to generate returns.

ROIC = N et operating prof it af ter tax/Invested Capital (5.3)

Mean Reversion 1M: In finance Mean reversion is considered as a theory that

proposes that asset prices and historical returns eventually will return as in revert
to the long-run mean or average level of the entire dataset. Interest rates and the
price-earnings (P/E) ratio of a company are taken into consideration other than
Percentage returns and prices[2].

dP = ηP (M − P )dt + σP dz (5.4)

Here, M is the long-run mean price which the prices tend to revert
and h is the speed of reversion.

Price Momentum 3M: It refers to how the price of a stock changes over a period
of time which helps investors to trade stocks as a stock price can be rising as in
bullish or can be falling as in bearish.

M omentum = Latest priceClosing price ∗ N umber of days (5.5)

Vol 3M: It refers to the daily average of the cumulative trading volume over the
time period of the the last three months.

Moneyflow Volume 5D: This indicator shows the time and price of stock dur-
ing the purchase. It is positive when more security was purchased in uptick time
compared to the downtick time.

19
39 Week Returns: The total money earned or lost during a transaction over a
year or the ratio of profit and investment is known as returns.

(Current V alue − Initial V alue)

RateOf Returns = [ ] ∗ 100 (5.6)
Initial V alue

The shift mask() method is used to process the data as in to divide our data into
upper percentile and lower percentile. The variable upper percentile stores the up-
per 30% of the data as in 70% to 100%. Whereas, the lower percentile stores the
lower 30% of the data as in 0% to 30%. To generate more accurate result, we are
using upper percentile and lower percentile as these data are extreme data which
means it has a lot of fluctuations.

We are considering 3 different time frames which are 1 day, 5 days, 22 days to
predict the future returns. The roll method is used to shift the data according to
the time frame. If the given time frame is 5 then day 6 data is stored at day 1.
Thus, all the data are shifted in this manner. After removing all the NAN values
from both the upper percentile and lower percentile we get two variables upper and
lower. Values those are greater than upper will be assigned to upper mask and values
those are lesser than lower, will be assigned to lower mask. After that in the mask
variable we will combine both upper and lower. Then we created a new column
which is more like a binary array acting as our output column. After running our
very own ensemble learning model considering customized factor the result that is
generated is kept in that binary output column. After the prediction, if the result
is greater than the value ’0.5’ we will consider it as 1 which means it is telling us to
go ’Long’. If the prediction gives us a result less than 0.5 then we will put -1 in the
output binary column which will tell us to go for ’Short’.

5.2 Application of Algorithms

Initially we worked with various algorithms such as SVM, RNN, Naive Bayes, Ran-
dom Forest, Logistic Regression, Adaboost and Gradient Boosting. We were unable
to use some of the algorithm in our classification as those could not provide the ex-
pected results. At first we try to implement SVM algorithm considering the factors
but it gave us Time Limit Exceed (TLE) error. As we already know, the nature of
the stock market is not consistent. SVM algorithm was ineffective for such dynam-
icity in quantopian platform. Again quantopian does not support RNN algorithm.
As a result, we could not use the RNN algorithm as well.

Moreover, Logistic regression requires less time to perform the operation whereas
Adaboost algorithm is more dominant than other algorithms in terms of perfor-
mance. After considering time, performance, platform and factors we chose Logistic
regression, Naive Bayes, Random Forest, Adaboost and Gradient Boosting which

20
will be best for ensemble learning. Finally, the classification is mainly done through
the average of the results provided by each of the five algorithms. So based on this
classification model we do the prediction of the USA stock market.

Figure 5.1: System structure of the Model

21
5.3 Factor Graphs
1. Mean Period Wise Return by Quantile: The positive y-axis of the graph
represents ’long’. Whereas, the negative y-axis defines ’short’. We are considering
five quantiles and each quantile is divided into three parts as in 1D, 5D and 22D.
By observing the figure 5.2 we can say 5D trading gives good returns for both long
and short.22D iS not giving good returns in comparison to 1D and 5D trading.

Figure 5.2: Mean period was Return by factor Quantile

22
2. Factor Weighted Long Short Portfolio Cumulative Return: As per
quantile deceleration figure 5.3 shows different positions of the portfolio during 1D,
5D AND 22D.

Figure 5.3: Factor Weighted Long Short Portfolio Cumulative Return 1D

Figure 5.4: Factor Weighted Long Short Portfolio Cumulative Return 5D

Figure 5.5: Factor Weighted Long Short Portfolio Cumulative Return 22D

23
3. Period Wise Return by Factor Quantile: The violin shaped graphs in
figure 5.6 represents the comparison between the summary statistics of range of
quantiles. It also provides where the returns are concentrated for each time period.

Figure 5.6: Period Wise Return by Factor Quantile

24
4. Cumulative Returns by Quantile: The main purpose of the curves given in
figure 5.7, 5.8 and 5.9 is to observe the separation between the quartiles. The graph
represented that the more distance and the less overlapping there is between the
quartiles the more it performs better. As we proceed with time, the graph showed
us that 5.9 gives better result comparing to 5.7 and 5.8 as there were less overlapping
and more distance between the quartiles.

Figure 5.7: Cumulative Returns by Quantile 1D

Figure 5.8: Cumulative Returns by Quantile 5D

Figure 5.9: Cumulative Returns by Quantile 22D

25
5. Top Minus Bottom Quantile Mean: To smoothen out the result for a partic-
ular trading period the graph in figure 5.10 determines the average of the difference
between top quantile and bottom quantile. The more positive the graph is, the more
return we get during that particular trading period.

Figure 5.10: Top Minus Bottom Quantile Mean for 1D

Figure 5.11: Top Minus Bottom Quantile Mean for 5D

Figure 5.12: Top Minus Bottom Quantile Mean for 22D

26
6. Normal distribution Quantile: Normal distribution is the most commonly
assumed form of distribution in technical stock market analysis and other forms of
statistical analysis. The normal distribution has two parameters: the standard de-
viation and the mean [24]. From the normal distribution graph we can understand
how the data points are distributed and standard deviation indicates the difference
between the observed data point and mean value. Usually the less the difference is
the better the prediction. From the Figure 6.1, 6.2, 6.3 it shows that 1D normal
distribution quantile is providing better outcomes than 5D and 22D.

Figure 5.13: Normal distribution Quantile for 1D

Figure 5.14: Normal distribution Quantile for 5D

Figure 5.15: Normal distribution Quantile for 22D

27
Chapter 6

Results

In this chapter we have discussed the result we achieved through our ensemble
model which we simulated from 05-01-2010 to 31-12-2018 and made comparison
between the result and factor graphs. Also we have divided the result into three
parts: daily(1D), weekly(5D), monthly(22D) and compared among these three to
determine which one provides better returns.

6.1 Returns
Total Return: Total return is the rate of return achieved from an investment for a
particular time period which includes interest, capital gains and dividends[23]. After
running our ensemble model we got variation in total returns for different time-span
(1D, 5D, 22D).
• Total return from 1D trading: 20.90%

• Total return from 5D trading: 40.58%

• Total return from 22D trading: -4.52%

Common Return: Common returns are the total returns which are associated
with the common risk factors modeled by quantopian exposure to market beta,
sectors, momentum, mean reversion, volatility, size, and value. Common returns
should be less then total returns and it signifies the algorithm is running well.
• Common return from 1D trading: 15.06%

• Common return from 5D trading: 24.53%

• Common return from 22D trading: 7.06%

Specific Return: Quantopian defines the returns which are not correlated to any
of the known risk factors like value, size, momentum etc. is known as specific return.

• Specific return from 1D trading: 5.42%

• Specific return from 5D trading: 13.18%

• Specific return from 22D trading: -11.08%

28
Figure 6.1: Returns for 1D

Figure 6.2: Returns for 5D

Figure 6.3: Returns for 22D

29
Sharp Ratio: It is the measurement of performance from an investment after the
adjustment of its risk.

• Sharp Ratio from 1D trading: 0.11%

• Sharp Ratio from 5D trading: 0.54%

• Sharp Ratio from 22D trading: 0.11%

Max Drawdown: The maximum loss that is observed from the observed maximum
point to the observed minimum point of the graph is known as max drawdown.

• Max Drawdown from 1D trading: -16.41%

• Max Drawdown from 5D trading: -12.55%

• Max Drawdown from 22D trading: -25.20%

Volatility: It measures the amount of risk.

• Volatility from 1D trading: 0.08%

• Volatility from 5D trading: 0.07%

• Volatility from 22D trading: 0.07%

30
Figure 6.4: Total Performance for 1D

Figure 6.5: Total Performance for 5D

Figure 6.6: Total Performance for 22D

31
6.2 Comparison On performance
From the below table, we came to a conclusion that the outcome generated during
5D trading is better in comparison with the trading of 1D and 22D time span. But
previously when we generated tear-sheet from the factors we found that 1D gave
better results than 5D and 22D. 5D trading performance deteriorates a little bit
than 1D but better than 22D. The performance of factor’s tear-sheet was based on
the time frame of one year only. When we worked with our ensemble algorithm we
considered the time frame to be ten years. While running our ensemble model for
10 years at first we saw that the daily returns are pretty good for 2 or 3 years but
then the fluctuations started. On the other hand, the weekly returns graph was in
constant increase which performed better than daily return at the end. Though the
daily return was performing well than the weekly return at the beginning but with
time due to the constant increase weekly return outperformed the daily return in
the long run.
The output of this defined that the trading of 5D was better in performance rather
than 1D and 22D.

Returns 1D 5D 22D
Total Returns 20.90 40.58 -4.52
Common Returns 15.06 24.53 7.06
Specific Returns 5.42 13.18 -11.08
Sharp Ratio 0.11 0.54 0.11
Max Drawdown -16.41 -12.55 -25.20
Volatility 0.08 0.07 0.07
Table 6.1: Outcomes of The Ensemble Model

32
Chapter 7

Conclusion

In this chapter, we finish up the paper with a discussion of the outcomes and limi-
tations of results. At long last, we talk about the potential zones in which further
research should be possible.

7.1 Conclusion
Through our attempts we came to the conclusion that the US stock market is much
volatile and there are too many variables to take into account when trying to pre-
dict prices of stocks. This becomes evident as discussed in the previous section, even
with the implementations of all the factors we could not use some of the factors later
on. We concluded with some factors that are effective during a certain period might
be counterproductive during another. Thus, the human factor and the laws of de-
mand and supply that govern the stock market make it too chaotic to predict with
sufficient accuracy.

We faced some limitations such as we had limited computing power available to

us. Due to our limited resources, we were not able to train our models without
exceeding the time limit for some algorithms. We were also unable to implement
any kind of neural networks as quantopian would not allow us to import keras for
tensor flow.

Apart from Gardient Boosting, Logistic regression, Naive Bayes and Random For-
est we also attempted to implement SVM (Support Vector Machine) and RNN as
it is commonly used in making stock market predictions, and gives good results.
We observed that Adaboost outperformed all the other four algorithms. However,
training with only SVM on each dataset takes approximately 18 minutes 33 sec-
onds to train, whereas ensemble algorithm took only 13 minutes 54 seconds to train.
Moreover, RNN was not supported by quantopian platform. Thus we run each of
the five algorithms and average the outputs provided by the algorithms to find out
the final output as in to do the classification. While considering only the factors we
found that 1D trading is giving better results when we were working for only one
year basis but after running the ensemble model for the time span of 10 years we
observed weekly trading gives more accurate returns.

33
7.2 Future work
As we are working with quantopian it has some advantages like we can access it
remotely through internet and also provides US stock market dataset as well as the
opportunity of Live preview of data trading. On the contrary there are some limi-
tations of using quantopian. One of the main constraints was it does not allow to
use RNN which is one of the significant algorithms. But this algorithm has proved
to provide more accurate result to predict stock market previously. So if you could
run this ensemble model in another dynamic platform we could have achieved better
result without maintaining these constraints.

Moreover, if we could incorporate our own resources we could get more efficient
results. Dynamic environment not only can predict the stock market of US but
through this we can predict the stock market of other countries by only processing
the data. That is why we are willing to work with dynamic environment in the
near future. However, in the future if we want least time trading as in hourly and
quarterly that cannot be done in quantopian since the least time trading it supports
is 1D trading. To improve our existing model we can set an upper threshold for
profit and lower threshold for loss. When the predicted price of a particular stock
exceeds the upper threshold the stock will be sold automatically on that predicted
day and when it falls under lower threshold, that particular stock should be auto-
matically sold right at that moment. In a nut shell, even though quantopian has
some limitations but for our proposed model it was able to provide a fair prediction
about the stock market.

34
Bibliography

[1] B. G. Malkiel and E. F. Fama, “Efficient capital markets: A review of theory

and empirical work”, The journal of Finance, vol. 25, no. 2, pp. 383–417, 1970.
[2] J. M. Poterba and L. H. Summers, “Mean reversion in stock prices: Evidence
and implications”, Journal of financial economics, vol. 22, no. 1, pp. 27–59,
1988.
[3] J.-H. Wang and J.-Y. Leu, “Stock market trend prediction using arima-based
neural networks”, in Proceedings of International Conference on Neural Net-
works (ICNN’96), IEEE, vol. 4, 1996, pp. 2160–2165.
[4] V. N. Vapnik, “An overview of statistical learning theory”, IEEE transactions
on neural networks, vol. 10, no. 5, pp. 988–999, 1999.
[5] A. Abraham, B. Nath, and P. K. Mahanti, “Hybrid intelligent systems for
stock market analysis”, in International Conference on Computational Science,
Springer, 2001, pp. 337–345.
[6] A. Liaw, M. Wiener, et al., “Classification and regression by randomforest”,
R news, vol. 2, no. 3, pp. 18–22, 2002.
[7] A.-S. Chen, M. T. Leung, and H. Daouk, “Application of neural networks to an
emerging financial market: Forecasting and trading the taiwan stock index”,
Computers & Operations Research, vol. 30, no. 6, pp. 901–923, 2003.
[8] X. Tao, H. Renmu, W. Peng, and X. Dongjie, “Input dimension reduction for
load forecasting based on support vector machines”, in 2004 IEEE Interna-
tional Conference on Electric Utility Deregulation, Restructuring and Power
Technologies. Proceedings, IEEE, vol. 2, 2004, pp. 510–514.
[9] M. R. Hassan, B. Nath, and M. Kirley, “A fusion model of hmm, ann and ga
for stock market forecasting”, Expert systems with Applications, vol. 33, no. 1,
pp. 171–180, 2007.
[10] S.-H. Hsu, J. P.-A. Hsieh, T.-C. Chih, and K.-C. Hsu, “A two-stage architec-
ture for stock price forecasting by integrating self-organizing map and support
vector regression”, Expert Systems with Applications, vol. 36, no. 4, pp. 7947–
7951, 2009.
[11] P. Ou and H. Wang, “Prediction of stock market index movement by ten data
mining techniques”, Modern Applied Science, vol. 3, no. 12, pp. 28–42, 2009.
[12] C.-F. Tsai, Y.-C. Lin, D. C. Yen, and Y.-M. Chen, “Predicting stock returns
by classifier ensembles”, Applied Soft Computing, vol. 11, no. 2, pp. 2452–2459,
2011.

35
[13] V. N. Vapnik, The nature of statistical learning theory. Springer science &
business media, 2013.
[14] O. Hegazy, O. S. Soliman, and M. A. Salam, “A machine learning model for
stock market prediction”, arXiv preprint arXiv:1402.7351, 2014.
[15] R. Brown, “What is the stock market?”, 2015. [Online]. Available: https://
corporatefinanceinstitute.com/resources/knowledge/trading-investing/stock-
market/.
[16] J. Patel, S. Shah, P. Thakkar, and K. Kotecha, “Predicting stock and stock
price index movement using trend deterministic data preparation and machine
learning techniques”, Expert systems with applications, vol. 42, no. 1, pp. 259–
268, 2015.
[17] D. Philip, “Why you should never short-sell stocks”, 2015. [Online]. Available:
https://scholar.google.com/scholar?hl=en&as sdt=2006&q=Why+you+
should+never+short-sell+stocks&btnG=.
[18] J. Roy and A. A. R. Nayeem, “Using sentiment analysis & machine learning
for security price forecasting”, PhD thesis, BRAC University, 2015.
[19] Sanjay, “Shares stock market”, vol. 4, 2016.
[20] M. Dunne, “Stock market prediction”, Unpublished thesis). University College
Cork. Retrieved December, vol. 26, 2017.
[21] A. Zheng and J. Jin, “Using ai to make predictions on stock market”, Stanford
University, Tech. Rep, Tech. Rep., 2017.
[22] Allen, “Types of stocks – common, preferred, hybrid, etc”, 2018. [Online].
Available: https://cleartax.in/s/stock-types.
[23] G. Kenton, “Total return”, 2018. [Online]. Available: https://www.investopedia.
com/terms/t/totalreturn.asp.
[24] S. Chen, “Normal distribution”, 2019. [Online]. Available: https : / / www .
investopedia.com/terms/n/normaldistribution.asp.
[25] M. hargrave, “Return on assets”, 2019. [Online]. Available: https : / / www .
investopedia.com/terms/r/returnonassets.asp.
[26] F. Jean, “How are share prices set?”, 2019. [Online]. Available: https://www.
investopedia.com/ask/answers/12/how-are-share-prices-set.asp.
[27] W. kenton, “Return on invested capital – roic definition”, 2019. [Online]. Avail-
able: https://www.investopedia.com/terms/r/returnoninvestmentcapital.asp.
[28] K. Leslie, “How is a company’s stock price and market cap determined?”,
2019. [Online]. Available: https://www.investopedia.com/ask/answers/how-
companys-stock-price-and-market-cap-determined.
[29] C. Mitchell, “Percentage price oscillator”, 2019. [Online]. Available: http://
www.investopedia.com/terms/p/ppo.asp.
[30] ——, “Williams %r definition and uses”, 2019. [Online]. Available: https://
www.investopedia.com/terms/w/williamsr.asp.
[31] G. Chen, Y. Chen, and T. Fushimi, “Application of deep learning to algorith-
mic trading”,

36
[32] T. Dai, A. Shah, and H. Zhong, “Cs229 project report automated stock trading
using machine learning algorithms”,
[33] K. Sharif and M. Abu-Ghazaleh, “Investigating algorithmic stock market trad-
ing using ensemble machine learning methods”,

37