UCLA

UCLA Electronic Theses and Dissertations

Title
Stock Trend Prediction: Based on Machine Learning Methods

Permalink
https://escholarship.org/uc/item/0cp1x8th

Author
Song, Yuan

Publication Date
2018

Peer reviewed|Thesis/dissertation

eScholarship.org Powered by the California Digital Library

University of California
 UNIVERSITY OF CALIFORNIA
Los Angeles

Stock Trend Prediction:

Based on Machine Learning Methods

A thesis submitted in partial satisfaction

of the requirements for the degree
Master of Applied Statistics

by

Yuan Song

2018
c Copyright by
Yuan Song
2018
 ABSTRACT OF THE THESIS

Stock Trend Prediction:

Based on Machine Learning Methods

by

Yuan Song
Master of Applied Statistics
University of California, Los Angeles, 2018
Professor Yingnian Wu, Chair

Nowadays, people show more and more enthusiasm for applying machine learning methods
to finance domain. Many scholars and investors are trying to discover the mystery behind
the stock market by applying deep learning. This thesis compares four machine learning
methods: long short-term memory (LSTM), gated recurrent units (GRU), support vector
machine (SVM), and eXtreme gradient boosting (XGBoost) to test which one performs the
best in predicting the stock trend. I chose stock price indicators from 20 well-known public
companies and calculated their related technical indicators as inputs, which are Relative
Strength Index, the Average Directional Movement Index, and the Parabolic Stop and Re-
verse. Experimental results show that recurrent neural network outperforms in time-series
related prediction. Especially for gated recurrent units, its accuracy rate is around 5% higher
than support vector machine and eXtreme gradient boosting.

ii
The thesis of Yuan Song is approved.

Nicolas Christou

Hongquan Xu

Yingnian Wu, Committee Chair

University of California, Los Angeles

2018

iii
 TABLE OF CONTENTS

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1 Previous Stock Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.2 Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.3 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Financial Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.1 Stock Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.2 Stock Trend Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4 Math Theory and Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

4.1 Recurrent Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

4.1.1 Long Short-Term Memory . . . . . . . . . . . . . . . . . . . . . . . . 9

4.1.2 Gated Recurrent Unit . . . . . . . . . . . . . . . . . . . . . . . . . . 11

4.2 Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4.3 eXtreme Gradient Boosting . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

5 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

5.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

5.2 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

5.3.1 RNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

5.3.2 SVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5.3.3 XGBoost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

iv
6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

7 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

v
 LIST OF FIGURES

4.1 Memory Cell of LSTM [23] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

4.2 Memory Cell of GRU [23] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4.3 Illustration of the Maximum Margin for SVM . . . . . . . . . . . . . . . . . . . 13

5.1 Candlestick Chart of Apple Inc. Stock . . . . . . . . . . . . . . . . . . . . . . . 17

5.2 Illustration of the Data Pre-processing before Feeding into the Models . . . . . . 21

5.3 Grid Search for Window Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5.4 Window Sliding Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5.5 RNN Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

6.1 LSTM: Classification Error of Training and Testing Performance Over Number
of Epochs. Case Study of Apple Inc. . . . . . . . . . . . . . . . . . . . . . . . . 27

6.2 LSTM: Predicted and Real Percentage Change of Adjusted Close Price on Train-
ing Set. Case Study of Apple Inc. . . . . . . . . . . . . . . . . . . . . . . . . . . 28

6.3 LSTM: Predicted and Real Classification. Case Study of Apple Inc. . . . . . . . 29

vi
 LIST OF TABLES

5.1 Description and Implication of the Input Variables . . . . . . . . . . . . . . . . 19

5.2 Description and Implication of the Input Variables (cont.) . . . . . . . . . . . . 20

6.1 Accuracy Rate of Four Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

vii
 CHAPTER 1

Introduction

The stock market is one of the most significant approaches for public companies to raise
funds, and it is also the main component of the capital market. The stock market has been
known as 'the barometer of the economy', which means its fluctuation is closely related to
the prosperity or recession of the economic market.

Stock as for the public companies is the way of financing. Through issuing shares and
bonds in the stock market, public companies can raise idle funds to solve the problem of
temporary capital scarcity. For private investors, the stock is one of the easily accessible
financial products. Through holding suitable stocks and doing appropriate portfolio man-
agement, personal assets will be appreciated to some degree. And also, stock trading is
the main business for securities companies and investment banks. With the development
of computer science technology, more and more companies are applying the technology of
big data and Artificial Intelligence to financial investment area. For example, according to
the news report, in 2000, Goldman Sachs used to have more than 600 traders of U.S. cash
equities trading desk in New York; while in 2017, there were only two human equity traders
left, with the computer doing the rest of work 1 .

Stock trend and price prediction is the act of making judgment on the future value of
a company stock, which has long existed since the establishment of stock market. Efficient
prediction of future stock price can yield considerable profits. However, the stock market
is a↵ected by many macro and micro factors, such as interrelated political and economic
indicators, supply-demand relationship of stock, operation conditions of the company and
so on.
1
http://www.newsweek.com/2017/03/10/how-artificial-intelligence-transform-wall-street-560637.html

1
 In 1965, Fama [1] raised the Efficient Market Hypothesis (EMH). It posits that the
stock market is an efficient market, which the price of stocks is the reflection of all publicly
known or newly revealed information and rational expectations. This theory established
solid foundation and basis for follow-up stock prediction study. Based on the size of market
information and the reaction speed of market on information, Fama categorized the efficient
market hypothesis into three levels, which are weak form, semi-strong form, and strong
form. The U.S. market is defined as a semi-strong form market that the market is efficient
and stocks can adjust instantly to absorb new information. Burton Malkie [2], the leading
proponent of EMH, stated that the stock prices therefore are unpredictable and random.

While the Efficient Market Hypothesis is highly accepted among the financial academia,
the real-world market experience gives di↵erent opinions and challenges the hypothesis of
unpredictability. Nowadays, analysts are trying to unveil the mysteries of the stock market,
and some large companies devote a tremendous amount of money on hiring experts to build
statistical models for prediction. There are many reasons to explain this phenomenon. First,
with the rise of big data technology and cloud computing, people can handle and analyze
mass stock exchange datasets with great ease. In many fund companies and securities invest-
ment companies, quantitative analysis has moved from manual work to artificial intelligence.
Second, deep learning achieved great progress on natural language processing. Thus, the
ability of computer to collect and analyze public information of stock market is improved
greatly. Third, the more machine learning techniques the computer absorbs, the 'smarter'
it will be. Artificial intelligence could be the new solution to tackle the stock prediction
problem down. In May 2017, the latest version of Google AlphaGo2 beat the world's best
human Go player, Jie Ke, through the perfect deep learning method. We can speculate that
computer also has the power of analyzing and predicting stock market over human securities
analysts one day.

The important position of stock in the economic market makes it meaningful to predict
its trend and future price. In the following chapters, under the assumption that stock

2
https://deepmind.com/research/alphago/alphago-china/

2
prices can be predicted, I will develop how di↵erent machine learning methods will perform
on stock trend prediction. The rest of this paper is organized into 7 chapters. Chapter
2 presents previous academic work on forecasting the stock market and its trend. This
chapter mainly focuses on works by using traditional statistical methods, neural networks,
and classification. Chapter 3 gives the brief definition of the U.S. stock market. Chapter 4
presents the mathematical concepts, which are at the core of the recurrent neural network,
support vector machine, and eXtreme gradient boosting. Chapter 5 is the introduction of
the dataset, dataset pre-processing and methodology for models. Chapter 6 summarizes the
observed results and Chapter 7 makes the conclusion and discusses future work.

3
 CHAPTER 2

Literature Review

2.1 Previous Stock Prediction

Financial time series forecasting, such as stock market prediction, is usually considered a
challenging task due to its volatile and chaotic characteristics [3]. Precisely predicting stock
trend and price movement is still an open question nowadays. During the past decades, var-
ious methodologies have been raised to perfect the prediction. In 1952, American economist
Markowitz [4], in his paper Portfolio Selection, first raised the concept of portfolio theory.
In this theory, he introduced two quantitative investment indexes, the mean and variance of
portfolio asset. Markowitz formalized investors preferences mathematically, used formulas
to explain the investment diversification, and discussed portfolio selection and management
systematically. This achievement represents the beginning of using mathematical ways to
understand and master the trend of economic market. Besides predicting the overall trend
of stock market, some scholars also dig into the price fluctuation of each stock by statistical
methods. For example, Fung et al. [5] used hypothesis testing method together with piece-
wise segmentation algorithm to predict the trends on the time series. The rise or drop trend
is categorized by the slope of each piecewise segment and the coefficient of determination.

2.2 Neural Network

Given the sophistication of financial time series data, introducing deep learning into
financial market prediction becomes such a popular topic [6]. In technological methods area,
one of the most promising technique is Artificial Neural Networks (ANN). ANN is the non-

4
linear computational models, which is capable of doing prediction based on market data,
without previous learning of the relationships between input and output variables [7]. For
stock prediction, the most popular type of ANN is using the feed-forward neural network and
back-propagation for weight training. Another form of ANN suitable for stock prediction is
the recurrent neural network (RNN).

Many scholars have devoted using the neural network to predict the financial market.
Wilson and Sharda [8] compared the predictive capability of neural networks and classical
multivariate discriminant analysis for firm bankruptcy. The result turned out that neural
networks performed significantly better than the other models at predicting bankruptcy.
Hsieh et al. [9] applied an integrated model of wavelet transforms and recurrent neural
network to forecast stock markets based on four stock indexes: the Dow Jones Industrial
Average, FTSE 100 Index, Nikkei 225 Index, and Taiwan Stock Exchange Capitalization
Weighted Stock Index. They proved that the model works great on four indexes and could
be used in real life to earn profits. Tanigawa and Kamijo [10] used recurrent neural networks
to recognize the stock pattern and 15 out of 16 of the recognition experiments were accurate.

Unlike conventional recurrent neural network, long short-term memory (LSTM) [11] and
Gated Recurrent Unit (GRU) [12] are more suitable to learn from past experience to clas-
sify, process, and predict time series with arbitrary size of time steps. These two models
are designed specifically to deal with the problem of vanishing gradients. Akita et al. [13]
confirmed that LSTM is able to capture the time series influence on stock price than other
models. Persio and Honchar [14] used LSTM model on stock price prediction and concluded
that with the algorithm, the reliability of prediction has been significantly increased. Simi-
larly, Chen et al. [15] conducted stock market prediction using GRU, which also achieved a
good prediction performance with small error.

Thus, in this experiment, I will use LSTM and GRU to predict the stock trends.

5
2.3 Classification

Support Vector Machine (SVM) has been widely used to predict the direction of finan-
cial time series and gained high accuracy. Prasaddas and Padhy [16] showed that when
predict futures prices in Indian stock market, the performances of support vector machine
are better than the results of back-propagation technique. Tay and Cao [17] compared the
support vector machine and multi-layer back propagation neural network for financial time
series forecasting. The experiment proved that SVMs showed superior performance over
back propagation in forecasting. Jain et al. [18] researched on the performances of the
Extreme Gradient Boosting (XGBoost) on sales forecasting, and found that XGBoost algo-
rithm outperforms traditional regression algorithms with higher accuracy. Sawon and Hosen
[19] performed a research of prediction on large scale of data to compare the performance of
XGBoost and other traditional models like linear regression and random forest regression.
The result confirmed that XGBoost outperforms the traditional ones with higher accuracy.

So I will also conduct support vector machine and extreme gradient boosting in the
experiment to compare with the performance of recurrent neural network.

6
 CHAPTER 3

Financial Definition

3.1 Stock Market

Stock prediction is using historical price, related market information and so on to forecast
exact price or price trend of the stock in the near future. According to the time granularity
of price information, stock trading can be divided into low-latency trading based on daily
basis and high-frequency trading, which market exchanges in a matter of hours, minutes,
even seconds. High-frequency trading analysis is more common in hedge funds, investment
banks, and large institutional investors. It masters the trading signals before prices' ups and
downs through analyzing great amount of trading data [20]. In this thesis, only low-latency
trading is taking into consideration, which is more common in academia. Its core concept is
to increase the accuracy of stock prediction based on the related market information.

The U.S. stock market mainly studies stocks traded at National Association of Securities
Dealers Automated Quotations (NASDAQ) and New York Stock Exchange (NYSE). Stocks
in the U.S. market are issued by world well-known companies, such as Apple Inc., Google,
Microsoft, etc.

The trading rule of NASDAQ and NYSE is T+0, which is also called same-day affirma-
tion. It means that completing the entire trading is processed on the same day. Investors
can buy or sell stocks at any time on trading days. Also, U.S. stock market does not set
limit up and limit down, rather it adopts circuit breakers/ trading curb to reduce market
volatility and prevent stock prices from increasing or decreasing suddenly.

7
3.2 Stock Trend Definition

In this thesis, I will mainly focus on predicting ups and downs of stock. Stocks leave
some important trading data after each trading day, such as open price, close price, adjusted
close price, highest price, lowest price, volume, etc. Among all, adjusted close price usually
represents the stock price of that trading day. In the trading period, there will be a series
of adjusted close prices. Let us denote it as:

p1 , p2 , p3 ,..., pT .

Here, pt is the close price on t trading day, T is total trading days in this period.

Stock price of a certain trading day will rise or drop comparing to previous trading day,
thus, here I used the change of closing price of two consecutive trading days as the judgment.
Let us denote trading situation as:
8
>
<1 if pt > pt 1 ,
yt = (3.1)
>
:0 if p  p .
t t 1

If it is 1, it means the price goes up on the second trading day. Otherwise, if it is 0, it

means the price goes down or remains the same.

8
 CHAPTER 4

Math Theory and Definition

4.1 Recurrent Neural Network

Recurrent Neural Network is a category of neural networks, which is similar to the concept
of feedback loop. After the data have been processed by the hidden layers, the original data
will be put back to the hidden layers together with the new input. RNNs are very powerful
and suitable for time serial prediction. They have the following two properties: (1) the
distributed hidden states make them to improve the efficiency of storing past information;
(2) the non-linear dynamics can help to update the hidden state in a more sophisticated
way [21]. There are several types of recurrent neural networks with di↵erent structures. In
the following, I will briefly introduce two similar models, long short-term memory and gated
recurrent units.

4.1.1 Long Short-Term Memory

long short-term memory is one of the variations of recurrent neural networks. Though
RNNs do great on time series data, it is hard to avoid the problem of long-term dependencies
due to the vanishing gradient [22]. Then LSTM [11] was introduced as an e↵ective way to
solve vanishing gradient by having memory cells to retain the time related information. A
memory cell includes in the input gate it , output gate ot , and forget gate ft . The gates are
used to control the interaction between itself and neighbor memory cells.

At time t, the input is xt , output is ht , and historic information that memory cell
memorizes is Ct . Input gate controls the portion of information at time t; forget gate
chooses to remember or forget historic information at time t; and output gate decides how
9
 Figure 4.1: Memory Cell of LSTM [23]

much information will be delivered to the next layer. The structure of memory cell block is
shown in Figure 4.1. Since the memory cell has the sequential feature, Ct contains all input
information from time 0 to time t, which represents the status at time t.

The forget gate ft is determined by input xt and output ft 1 of last time, and then
wrapped in the sigmoid function. See (4.1) for the formula. The value of forget gate bounces
between 0 and 1. When ft is 0, that means the previous value is forgotten in the calculation.
Otherwise, if it is 1, the forget gate keeps previous information.

With input xt and previous output ht 1 , through transformation of tanh function, we

can get Ct where its value is between -1 to 1; see (4.2). This means how much information
will be added to Ct after introducing input xt .

Similar to the forget gate, the input gate it regulates the information of input xt and
previous output ht 1 ; see (4.3). If it is 0, then Ct can be neglected; otherwise, if it is 1,
then Ct will be counted into Ct .

The state vector Ct is consisted of two parts. The first part is the state vector of previous
time node Ct 1 , which is controlled by forget gate. And the second part is Ct , which is

10
determined by how much input gate takes; see (4.4).

The output gate ot is similar to input gate and forget gate; see (4.5). The size of output
ot decides how much likely the state vector Ct will be outputted and received by other layers.

The output vector ht is the combination of the output gate ot and the cell state Ct ; see
(4.6). Ct needs to be transformed by tanh function, and then let output gate ot decide the
portion of output.

ft = (Wf ⇤ [ht 1 , xt ] + bf ) (4.1)

Ct = tanh(WC · [ht 1 , xt ] + bC ) (4.2)

it = (Wi · [ht 1 , xt ] + bi ) (4.3)

Ct = ft ⇤ Ct 1 + it ⇤ Ct (4.4)

ot = ⌃(Wo · [ht 1 , xt ] + bo ) (4.5)

ht = ot ⇤ tanh(Ct ) (4.6)

4.1.2 Gated Recurrent Unit

Gated Recurrent Unit is very similar to long short-term memory. The main di↵erence is
how they apply gates. For GRU, there are only two gates instead of three, which are the reset
gate and update gate . Figure 4.2 illustrated the procedure of GRU memory cell. The reset
gate decides how to combine the new input and previous memory, and controls the extent of
neglecting previous memory. The smaller value of reset gate is, the more previous memory
will be neglected. The update gate determines how much of previous memory will be kept.
The larger value of update gate is, the more previous information will be brought. When we
set all reset gates into 1 and update gates into 0, then we back to the plain recurrent neural
network model.

11
 Figure 4.2: Memory Cell of GRU [23]

zt = (Wz · [ht 1 , xt ]) (4.7)

rt = (Wr · [ht 1 , xt ]) (4.8)

het = tanh(W · [rt ⇤ ht 1 , xt ]) (4.9)

ht = (1 zt ) ⇤ ht 1 + zt ⇤ het (4.10)

There are few more di↵erences between LSTM and GRU. First, LSTM can control on
the exposure of memory, while GRU exposes the memory, which is not controllable. Then,
GRU does not have the output gate like in LSTM. Also, the input gate and forget gate in
LSTM are substituted by the update gate, and thus the reset gate is directly applied to
the previous hidden state. Since GRU has fewer parameters than LSTM, we could make a
preliminary guess that GRU will be faster than LSTM and need less data. However, if the
data scale is large, LSTM may lead to a better result.

12
4.2 Support Vector Machine

Support Vector Machine algorithm is a supervised learning method proposed by Vapnik

[24]. It is based on the statistical learning theory for both classification and regression. In
the binary classification case, the goal is to find an optimal separating hyperplane, which is
a line here. It separates the data into two categories and each class lays on either side of
the line. The SVM is also called the maximal margin classifier, where margin is the distance
between the nearest training spots of each category [25].

Figure 4.3: Illustration of the Maximum Margin for SVM

Let {x1 , x2 , x3 ,..., xn } be the dataset and let yi 2 {0,1} be the class label of xi . A
weight vector w = {w1 , w2 , w3 ,..., wn } determines the contribution of each input vector to
the prediction.

13
 The binary linear classifier separates all points correctly by using

yi (wi ⇤ xi + b) 1, 8i. (4.11)

Since we want the decision boundary to be as far away from the data points (support
2
vectors) of both categories as possible, we need to maximize the margin, m = kwk
. In other
words, the larger the margin is, the less generalization error of the classifier will be. So, we
need to
1
minimize kwk2 ,
2 (4.12)
subject to yi (wi ⇤ xi + b) 1, 8i.

The solution to it can be solved by the Lagrungian form. The process will not be shown
in this paper.

4.3 eXtreme Gradient Boosting

eXtreme Gradient Boosting (XGBoost) is developed by Tianqi Chen [26], which uses the
gradient descent algorithm to ensemble a set of classification and regression trees (CART).
Comparing to only one tree in the CART mode, the tree ensemble model shows stronger
predictive ability. These trees will vote for the most promising result. In other words, the
gradient descent could minimize the loss by adding new models and correct the errors of
existing models [27].

A decision tree is the model that begins with a single non-leaf node that branches into
di↵erent outcomes. Then the outcomes lead to more additional nodes. Each non-leaf node
represents the test on one particular feature, each branch represents the outcome of this
feature, and each leaf node stores a classification. Once the split for each feature is made,
the one with the minimum loss is viewed as the best split criteria and set it as the rule for
that node. The splitting process keeps going until the termination condition is met.

Boosting technique holds the principle that a set of weak classifiers can create a single
strong classifier. Weak classifiers are the classifiers only perform slightly better than random
classifiers, and strong classifiers are the arbitrarily well-corrected with high accuracy [28].
14
For the boosting methods, additive training method is applied in each step, which a week
classifier is added to the model. In XGBoost, the weak classifier is the new decision tree.
Equations below show this hallmark.
F0 = 0 (4.13)

Ft (x) = Ft 1 (x) + h(x) (4.14)

Here, h(x) is the new decision tree after Ft 1 (x) and Ft (x) is the full model after t 1
steps. The objective of XGBoost model is to find the tree Ft (x) that minimizes the equation
(4.15) at the tth step.

Obj(Ft ) = L(Ft 1 + Ft ) + ⌦(Ft ) (4.15)

Here, L is the loss function that decides the predictive power, and ⌦ is the regularization
function controlling overfitting.

15
 CHAPTER 5

Experiment

5.1 Dataset

The historical stock dataset was gathered from the API of Google Finance by applying the
Python library pandas-datareader. I chose 20 well-known public companies' stocks traded
at both NASDAQ and NYSE, ranging from January 1st, 2010 to December 31st, 2017.
There were altogether 1986 trading days. With access to the API, Python can automatically
download the daily time series data. The dataset includes daily open price, daily highest
price, daily lowest price, daily close price, daily adjusted close price, and daily trading
volume.

Figure 5.1 is the candlestick chart with Apple Inc. stock as an example. This figure is
used to give a brief depiction of the gathered data. For this figure, with large scale of data,
it could be hard to recognize the candle shape. However, we can still figure the trend by
noticing that the green means rising and orange means dropping. The red line is the mean
price line. The chunk below draws the trading volume of every trading day.

I used the original raw data to calculate the percentage change of price trend, qi , which
was then used as input. The characteristics of stock price is the most direct and intuitive
way to understand the trend of stock trend. Comparing to the absolute value of stock price,
the percentage change of price trend is more e↵ective in stock prediction. Di↵erent stocks
have di↵erent base prices, which lead to large variation on the absolute value of stock price.
By using the percentage change, prediction will be less a↵ected by the price base.

16
Figure 5.1: Candlestick Chart of Apple Inc. Stock

17
 The percentage change of adjusted price is shown in the following.

pi pi 1
qi = (5.1)
pi 1

After taking the fraction, the adjusted close price sequence was changed to the percentage
change sequence.

q1 , q2 , q3 ,..., qT .

Besides the historical stock trading data, the dataset also included in the technical in-
dicators of stock trading. In this experiment, three indicators that investors care about are
counted in, which are Relative Strength Index (RSI), the Average Directional Movement
Index(ADX), and the Parabolic Stop and Reverse (SAR). These three indicators were devel-
oped by J. Welles Wilder Jr. [29], and have been widely used in technical analysis softwares.
Relative strength index is the momentum oscillator that compares a stock's average recent
gains and losses in the certain time period and uses this fraction to measure the variation
and speed of price movements. Average Directional Movement Index consists of the positive
directional indicator (PDI) and negative directional indicator (NDI). This indicator is used
to quantify the strength of trend regardless of whether rising or dropping. Last, the parabolic
stop and reverse is the indicator for investors not only to decide the place of trailing stop loss
or exit points, but also find the reverse trading position after stopping. Table 5.1 provides
the brief introduction of all variables and related formula.

18
 Table 5.1: Description and Implication of the Input Variables
Name Definition/Implication
Panel A. Daily Trading Data
The first price of a stock traded at the beginning of a
Open Price
specified trading day.
The last price of a stock in the last transaction
Close Price
on a specified trading day.
The close price adjusted based on the reflection
Adjusted Close Price
of dividends and splits.
The highest price when a stock traded on a specified
Highest Price
trading day.
The lowest price when a stock traded on a
Lowest Price
specified trading day.
Percentage Change of The percentage change of adjusted close price on ith
Adjusted Close Price trading day to the previous (i 1)th trading day.
Total amount of shares or contracts of a stock traded
Trading Volume
on a specified trading day.
Panel B. Technical Indicators
100
RSI = 100
1 + RS
n = number of RSI periods
T otalGains
Average Gain =
n
T otalLoss
Relative Strength Index Average Loss =
n
AverageGain
First RS =
AverageLoss
Smoothed RS =
(P reviousAverageGain ⇤ 13 + CurrentGain)/14
(P reviousAverageLoss ⇤ 13 + CurrentLoss)/14

19
 Table 5.2: Description and Implication of the Input Variables (cont.)
Name Definition/Implication
(ADXt 1 ⇤ (n 1)) + DXt
ADXt =
n
where

Average Directional n = smoothing period

100 ⇤ (P DIt M DIt )
Movement Index DXt =
P DIt + M DIt
P DIt = M AX(Ht Ht 1 , 0)

N DIt = M AX(Lt Lt 1 , 0)
Pt = P t 1 + AF ⇤ (EPt 1 Pt 1 )

Pt = current value of the indicator

Parabolic Stop and Reverse Pt 1 = previous value of the indicator

AF = acceleration factor

EPt 1 = Extreme point in the previous period

20
5.2 Data Processing

The figure below briefly illustrates the procedure of data pre-processing.

Figure 5.2: Illustration of the Data Pre-processing before Feeding into the Models

1. Sliding window. Window sliding is the strategy of taking the previous time steps
to forecast the subsequent time step. Then, with this method, I can transform the data to
solve the classification problem. Here I cross-validated the window size, and got the optimal
window size of 5 as shown in Figure 5.3. Thus, in this case, I will use a relative small window
size to let models relearn and adapt the latest market information. The stock information
from the latest week could provide more accurate prediction.
21
 Figure 5.3: Grid Search for Window Size

Figure 5.4 depicts the strategy of sliding window in size of 5. All the inputs with N/A were
omitted.

Figure 5.4: Window Sliding Strategy

22
 2. Normalization.

I used min-max normalization on data, which can scale di↵erent indicators of stocks to
be comparable. It can be realized by using MinMaxScaler(feature range=(-1,1)) in Python
to transform the original data into range of [-1,1]. That means the minimum value of input
is mapped to -1, and the maximum value of input is mapped to 1.

3. Split Dataset.

The dataset was split into chronological sets. I took the first 80% of dataset as the
training set, and the rest 20% as the testing set. Afterwards, I reshaped the input to be
three-dimensional, [samples, timesteps, features], to fit the network.

5.3 Method

The purpose of this thesis is to predict whether the stock price in the next few days will
increase or decrease by using the past few days' price indicators and related technical indi-
cators. For this binary classification problem, I implemented long short-term memory, gated
recurrent unit, support vector machine, and eXtreme gradient boosting. In the following
sections, I will introduce how I applied these four models to achieve the purpose.

5.3.1 RNN

I used the Python library Keras to build and train the models. The Keras library can
run on top of TensorFlow and is the simplified interface to TensorFlow. With Keras, the
built-in function LSTM and GRU can be used directly. Besides Keras, Pandas library was
included to read and reframe the data.

Models is the core of Keras neural network. It represents the neural network that we
defined has layers, activation function and so on. The training and testing were also based
on this models. The Sequential model represents that the network is built on a linear stack
of layers, which is the basic structure. Dense, Activation, and Dropout are the core layers
of neural network. Dense is actually the fully-connected layer. Activation is the activation

23
function, which uses ReLu, Softmax, or sigmoid function to amend on the results from last
layer. When there are too many neurons, the e↵ects will be worsening due to overfitting.
Then Dropout will be used to randomly exclude neurons from last layer while training
the network. This is a regularization method that can reduce the problem of overfitting
e↵ectively and improve the performance of model.

Figure 5.5: RNN Architecture

In this experiment, I used the prior 5 values as timesteps to predict the 6th value, and
then compared the predicted and the real 6th value. So, there were altogether 20 units as
inputs. Thus, I defined the RNN with 20 neurons in the input layer, 50 neurons in the first
hidden layer, 100 neurons in the second hidden layer, and 1 neuron in the output layer. Two
20% dropout layers were added after the first and the second layer. I selected Root Mean
Square Error (RMSE) to calculate error, and optimize through adam.

The model fit for 50 training epochs with batch size equaling to 80. At last, I applied the
24
fit function to set the validation data argument in order to record the history of the training
and testing loss. The training and testing loss were plotted and is shown in Figure 6.1.

The next step after defining the model was to forecast the entire test set. According to
whether the predicted qi is larger or smaller than 0, we can turn the problem into classifica-
tion. After that, I calculated the error for the model based on the prediction and real values
in the initial scale.

Above is the methodology for the recurrent neural network. In this experiment, I applied
two special cases of RNN, long short-term memory and gated recurrent units, to the hidden
layers.

5.3.2 SVM

Support vector machine can be achieved by using sklearn in Python. The model is less
complex to achieve comparing to the recurrent neural network. With the built-in function,
I tuned the parameters to perfect the model.

There are three parameters we need to take care of, which are kernel, regularization, and
gamma.

Kernel decides the type of svm. There are di↵erent kernel functions that are available to
choose, including radial basis function kernel (RBF), linear, sigmoid and so on. Depending on
the application and cross-validation method, I could determine which kernel type performs
best. In this experiment, I chose linear kernel, which has lower error rate comparing to other
kernel types. The equation of linear kernel is listed as follows:

K(xi , xj ) = xTi xj . (5.2)

The next parameter needs to be tuned was the regularization, 'C' in sklearn library.
This term determines how much to avoid the error of misclassifying in the training set and
controls the ability of generalization when optimizing. With larger C, the optimization will
make the margin of hyperplane smaller, which probably will lead to lower generalization.
Low generalization means the model only works well on the training set but poorly on new
25
sample. It could be caused by the problem of overfitting. With smaller C, the optimization
will set the margin larger and reduce the risk of overfitting. Here, after grid search on
di↵erent C, I selected 100 for the regularization parameter.

The last parameter is gamma. It determines how far the training sample can influence.
Higher gamma means close that only points near hyperplane will be considered. Lower
gamma means far that far away points will also be counted. However, since I implemented
linear kernel, it is not necessary to tune because the default gamma is 1 for linear.

5.3.3 XGBoost

For XGBoost, there exists python package named xgboost. After splitting training and
testing sets, following parameters need to be modified. For tree booster, there are several
types are available, such as gbtree, gblinear and dart. Since the target problem is the classi-
fication, it would be better to use gbtree. The number of rounds determines the maximum
iteration times. It can be tuned by cross-validation and the result showed 1000 performs the
best. Then for the objective parameter, I chose binary: logistic, which is used for binary
classification. The eta parameter determines the learning rate of grasping the patterns of
data and it ranges from 0.01 to 0.3. Higher eta can make computation quicker. I selected
0.3 in this case. The last parameter is the regularization term, gamma, which controls the
problem of overfitting. The default = 0 means that there is no regularization. The higher
gamma, the higher regularization. I began tuning it with 0 and checked the cross-validation
error rate, then took gamma = 5. Besides the parameters above, all others were set as
default.

26
 CHAPTER 6

Results

The results are divided into two parts. The first part is the recurrent neural network
training results. Since for the 20 stocks, the plots demonstrate similar information, I only
illustrated the LSTM method for Apple Inc. as example. The second part is accuracy rate
comparison among four models.

Figure 6.1 shows the RMSE of training and testing sets over 50 epochs. We can notice
that errors reduce over more epochs. Training set is colored as blue and testing set is colored
as orange. Testing set error is higher than training set error in most cases, which is expected.

Figure 6.1: LSTM: Classification Error of Training and Testing Performance Over Number
of Epochs. Case Study of Apple Inc.

27
 Figure 6.2 gives the predicted and real percentage change of adjusted close price on
training set for Apple Inc. It is based on the method of LSTM. The orange one refers to
predicted qi and the blue one refers to the real qi . The x-coordinate represents the days.
0 is the middle line and the wave bounces in between -1 and +1. We can notice that not
all the time the model can predict correctly on the trend. We will see how it performs by
taking a look at the accuracy rate.

Figure 6.2: LSTM: Predicted and Real Percentage Change of Adjusted Close Price on Train-
ing Set. Case Study of Apple Inc.

Figure 6.3 shows the predicted and real classification for Apple Inc. The range is from 0
to 1 where 0 means the trend going down and 1 means the trend going up. The x-coordinate
represents the days. We can see several times the models can predict precisely.

28
 Figure 6.3: LSTM: Predicted and Real Classification. Case Study of Apple Inc.

The following Table 6.1 gives detailed accuracy rate of each stock with di↵erent models.
Accuracy rate is calculated as following formulas.

Error Rate: The error rate is the percentage of the testing sets that are incorrectly
classified.

Error Rate = (Number of Incorrectly Classified Test Sample/Test Sample Size) * 100

Accuracy Rate: The accuracy rate is the percentage of correct classifications.

Accuracy Rate = 1 - Error Rate

From the table, we can notice the accuracy rate is still higher than random guess (50%).
Both LSTM and GRU show higher accuracy rate than other two classification methods,
which encourages us to dig more into applying recurrent neural network to financial times
series problems.

29
 Table 6.1: Accuracy Rate of Four Models
LSTM GRU XGBoost SVM

AAPL 62.93% 63.43% 59.70% 57.58%

AIG 66.97% 67.47% 59.2% 58.79%
AMZN 61.67% 62.17% 60.76% 56.21%
BA 60.66% 60.40% 59.24% 60.15%
COF 63.43% 63.18% 59.24% 57.12%
EBAY 63.69% 63.43% 59.39% 59.70%
FDX 64.19% 64.44% 60.76% 59.55%
GE 68.48% 68.48% 57.88% 61.21%
GOOG 61.41% 61.92% 59.85% 58.33%
HD 65.45% 65.96% 59.24% 59.70%
JNJ 64.19% 64.44% 58.33% 56.52%
JPM 65.96% 66.21% 62.12% 61.67%
KO 62.17% 62.17% 61.67% 59.24%
MSFT 63.69% 64.19% 60.61% 59.09%
NKE 65.71% 65.45% 61.67% 60.61%
ORCL 63.94% 64.44% 58.03% 58.94%
PEP 66.46% 65.45% 61.36% 59.39%
T 67.22% 66.97% 61.82% 58.64%
WMT 64.44% 64.19% 63.94% 60.91%
XOM 62.42% 62.17% 60.15% 60.45%
AVERAGE 64.25% 64.33% 60.25% 59.19%

30
 CHAPTER 7

Conclusion and Future Work

In this thesis, it has shown how four di↵erent machine learning methods: long short-
term memory, gated recurrent units, support vector machine and eXtreme gradient boosting
performed on financial time series trend prediction. In the experiment, 20 U.S. stocks trading
at NASDAQ and NYSE were selected to compare the accuracy rate yielded by various
models. The results turned out that recurrent neural network did the best comparing to
SVM and XGBoost. RNN can have around 5% improvement of accuracy rate. This may
suggest that in the time series related prediction, recurrent neural network outstands among
other methods.

In RNN, gated recurrent unit performed slightly better than long short-term memory.
However, due to limited stock number and relative small data size, we could not arbitrarily
conclude that gated recurrent unit is better than long short-term memory in stock prediction.
It would be better to compare these two models if data scale is large enough.

In the future, more works can be done to improve the performance of stock prediction.

1. Import more Indicators. In this experiment, I only included three indicators raised
by J. Welles Wilder Jr. Actually, there are more indicators to evaluate stocks while trading.
Including more indexes could possibly improve the accuracy for models to analyze.

2. Stack Models. Here, I only compared single models with each other. We could
discover more by stacking models to see if it can improve the prediction ability.

3. Import Textual Information. Some scholars have used sentiment analysis on

Twitter [30] to predict the stock market movement. Besides breaking down social media in-
formation, other qualitative indicators like news, international and domestics policy changes

31
can also be used as inputs to predict price trend.

4. Predict Exact Price Value. John Hull [31] in his book Options, Futures Other
Derivatives introduced mathematical finance solutions, such as Ito's Lemma, Black-Scholes-
Merton model, to study the pricing of financial products. In the future, it could be possible
to combine the mathematical finance ways and machine learning methods to determine and
predict the exact price value.

Although most people hold negative view on using machine learning to predict stock
price as they believe stock market is too volatile to predict, scholars and traders from big
companies are still trying to explore more indicators and models to accurately forecast the
economic market. We can speculate that in the near future, the machine learning methods
will provide us a bright future in learning and mastering the market.

32
 REFERENCES

[1] Fama, E. F. The Behavior of Stock-Market Prices. Journal of Business, 1965; 38(1),
34105.
[2] Malkiel, B. G. A Random Walk Down Wall Street. New York: Norton, 1973.
[3] Wang, B., Huang, H., Wang, X. A Novel Text Mining Approach to Financial Time
Series Forecasting. Neurocomputing, 2012; 83(6):13645.
[4] Markowitz, H. Portfolio selection[J]. The Journal of Finance, 1952; 7(1): 77-91.
[5] Fung, G. P. C., Yu, J. X., Lam, W. News Sensitive Stock Trend Prediction[M]. Advances
in knowledge discovery and data mining. Springer Berlin Heidelberg, 2002; 481-493.
[6] Cavalcante, R.C., Brasileiro, R.C., Souza V.L.F., Nobrega, J.P., Oliveira, A.L.I. Com-
putational Intelligence and Financial Markets: A Survey and Future Directions.. Expert
Systems with Applications, 2016; 55:194211.
[7] Maciel, L.S., Ballini, R. Design A Neural Network for Time Series Financial Forecasting:
Accuracy and Robustness Analysis. Instituto de Economia, Universidade Estadual de
Campinas, Sao Paulo-Brasil, 2008.
[8] Wilson R. L., Sharda R. Bankruptcy Prediction Using Neural Networks. Decision sup-
port systems, 1994; 11(5), pp. 545-557.
[9] Hsieh, T., Hsiao, H., Yeh, W. Forecasting Stock Markets Using Wavelet Transforms
and Recurrent Neural Networks: An Integrated System Based on Artificial Bee Colony
Algorithm. Applied Soft Computing, 2011; 11(2), pp.2510-2525.
[10] Kamijo, K., Tanigawa, T. Stock Price Pattern Recognition: A Recurrent Neural Net-
work Approach. In Proceedings of the international joint conference on neural networks
(IJCNN), 1990; (pp. 215221), Washington DC.
[11] Hochreiter. S., Schmidhuber, J. Long short-term memory. Neural Computation, 1997;
9(8):17351780
[12] Cho, K., Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwen, H., Ben-
gio, Y. Learning phrase representations using RNN encoder-decoder for statistical ma-
chine translation. in Conference on Empirical Methods in Natural Language Processing
(EMNLP 2014), 2014; arXiv:1406.1078.
[13] Akita R., Yoshihara A., Matsubara T., Uehara K. Deep learning for stock prediction
using numerical and textual information[C]. Computer and Information Science (ICIS).
IEEE/ACIS 15th International Conference on. IEEE, 2016: 1-6.
[14] Persio, L., Honchar, O. Artificial Neural Networks architectures for stock price predic-
tion: comparisons and applications. International journal of circuits, systems and signal
processing, 2016; 10.
33
[15] Chen, W, Zhang, Y, Yeo, C.K., Lau, C.T., Lee, B.S. Stock market prediction using
neural network through news on online social networks[C]. Smart Cities Conference
(ISC2), 2017 International, 2017: 11-2.

[16] Prasaddas, S., Padhy, S. Support Vector Machines for Prediction of Futures Prices in
Indian Stock Market. International Journal of Computer Applications, 2012; 41(3):226.

[17] Tay, F.E.H., Cao, L. Application of support vector machines in financial time series
forecasting. Omega, 2001; Vol. 29, No. 4, pp.309317.

[18] Jain A., Menon M.N., Chandra S. Sales Forecasting for Retail Chains. India, 2015.

[19] Sawon, M., Hosen, M. Prediction on large scale data using extreme gradient boosting.
[Thesis Report] Thesis Report, BSc (Computer Science and Engineering), BRAC Uni-
versity, Dhaka, Bangladesh, 2016.

[20] Menkveld, A. J. High frequency trading and the new market makers[J]. Journal of Fi-
nancial Markets, 2013; 16(4): 712-740.

[21] Hinton, G.E.: Recurrent neural networks,

https://www.cs.toronto.edu/ hinton/csc2535/notes/lec10new.pdf

[22] Palangi, H., Ward, R., Deng, L. Distributed Compressive Sensing: A Deep Learning
Approach. IEEE Transactions on Signal Processing, 2016; 64(17):450418.

[23] Olah, C.: Understanding LSTM Networks

http://colah.github.io/posts/2015-08-Understanding-LSTMs/

[24] Vapnik, V. N. The Nature of Statistical Learning Theory. Springer-Verlag, 1995.

[25] Boser, B.E., Guyon, I., Vapnik, V.N. A training algorithm for optimal margin classifiers.
In Proceedings of the Fifth Annual Workshop of Computational Learning Theory, 1992;
5, 144152. Pittsburgh, ACM.

[26] Chen, T.Q. Introduction to Boosted Trees. University of Washing Computer Science.
University of Washington, 22 Oct, 2014. Web.

[27] Brownlee, J.: A Gentle Introduction to XGBoost for Applied Machine Learning - Ma-
chine Learning Mastery
https://machinelearningmastery.com/gentle-introduction-xgboost-applied
-machine-learning/

[28] Freund, Y., Schapire, R. A Decision-Theoretic Generalization of On-Line Learning and

an Application to Boosting. Journal of Computer and System Sciences, 1997;55(1),
pp.119-139.

[29] Wilder, J. New concepts in technical trading systems. Greensboro, N.C.: Trend Research,
1978.

34
[30] Yang, S., Mo, S., Liu, A. Twitter financial community sentiment and its predictive
relationship to stock market movement. Quantitative Finance, 2015; 15(10), pp.1637-
1656.

[31] Hull, J. Options, Futures, and Other Derivatives, Global Edition, 2017; Pearson Educa-
tion UK.

35