Microsoft Word – Project_Writeup.doc
1
Machine Learning Techniques for Stock Prediction
Vatsal H. Shah
2
1. Introduction
1.1 An informal Introduction to Stock Market Prediction
Recently, a lot of interesting work has been done in the area of applying Machine
Learning Algorithms for analyzing price patterns and predicting stock prices and index
changes. Most stock traders nowadays depend on Intelligent Trading Systems which help
them in predicting prices based on various situations and conditions, thereby helping
them in making instantaneous investment decisions.
Stock Prices are considered to be very dynamic and susceptible to quick changes because
of the underlying nature of the financial domain and in part because of the mix of known
parameters (Previous Days Closing Price, P/E Ratio etc.) and unknown factors (like
Election Results, Rumors etc.)
An intelligent trader would predict the stock price and buy a stock before the price rises,
or sell it before its value declines. Though it is very hard to replace the expertise that an
experienced trader has gained, an accurate prediction algorithm can directly result into
high profits for investment firms, indicating a direct relationship between the accuracy of
the prediction algorithm and the profit made from using the algorithm.
1.2 Motivation behind the Project
In this paper, we discuss the Machine Learning techniques which have been applied for
stock trading to predict the rise and fall of stock prices before the actual event of an
increase or decrease in the stock price occurs. In particular the paper discusses the
application of Support Vector Machines, Linear Regression, Prediction using Decision
Stumps, Expert Weighting and Online Learning in detail along with the benefits and
pitfalls of each method. The paper introduces the parameters and variables that can be
used in order to recognize the patterns in stock prices which can be helpful in the future
prediction of stocks and how Boosting can be combined with other learning algorithms to
improve the accuracy of such prediction systems.
Note: The main goal of the project was to study and apply as many Machine Learning
Algorithms as possible on a dataset involving a particular domain, namely the Stock
Market, as opposed to coming up with a newer (and/or better) algorithm that is more
efficient in predicting the price of a stock.
3
1.3 Overview of the Document
In Section 2 we try to briefly cover the background which is essential to the study of the
domain of financial prediction systems. In Section 3 we discuss the results obtained from
the application of the algorithms described in Section 1.2. Section 4 presents the
concluding remarks for the experiment. Section 5 and 6 cover the Software Tools used
and a list of research papers that were referenced (and that might serve as further reading
material for those who are interested in this topic and want to explore it further.)
4
2. Background
2.1 Stock Prediction in Detail
In practice, there are 2 Stock Prediction Methodologies:
Fundamental Analysis: Performed by the Fundamental Analysts, this method is
concerned more with the company rather than the actual stock. The analysts make their
decisions based on the past performance of the company, the earnings forecast etc.
Technical Analysis: Performed by the Technical Analysts, this method deals with the
determination of the stock price based on the past patterns of the stock (using time-series
analysis.)
When applying Machine Learning to Stock Data, we are more interested in doing a
Technical Analysis to see if our algorithm can accurately learn the underlying patterns in
the stock time series. This said, Machine Learning can also play a major role in
evaluating and forecasting the performance of the company and other similar parameters
helpful in Fundamental Analysis. In fact, the most successful automated stock prediction
and recommendation systems use some sort of a hybrid analysis model involving both
Fundamental and Technical Analysis.
The Efficient Market Hypothesis (EMH)
The EMH hypothesizes that the future stock price is completely unpredictable given the
past trading history of the stock. There are 3 types of EMH’s: strong, semi-strong, and
weak form. In the weak EMH, any information acquired from examining the stock’s
history is immediately reflected in the price of the stock.
The Random Walk Hypothesis
The Random Walk Hypothesis claims that stock prices do not depend on past stock
prices, so patterns cannot be exploited since trends to not exist.
With the advent of more powerful computing infrastructure (hardware and software)
trading companies now build very efficient algorithmic trading systems that can exploit
the underlying pricing patterns when a huge amount of data-points are made available to
them. Clearly with huge datasets available on hand, Machine Learning Techniques can
seriously challenge the EMH.
5
Indicator Functions
We now take a brief look at the attributes and indicators that are normally used in the
technical analysis of stock prices:
Indicators can be any of the following:
Moving Average (MA) : The average of the past n values till today.
Exponential Moving Average (EMA) : Gives more weightage to the most recent values
while not discarding the older observation entirely.
Rate of Change (ROC) : The ratio of the current price to the price n quotes earlier.
n is generally 5 to 10 days.
Relative Strength Index (RSI): Measures the relative size of recent upward trends against
the size of downward trends within the specified time interval (usually 9 – 14 days).
For this Project, the EMA was considered as the primary indicator because of its ability
to handle an almost infinite amount of past data, a trait that is very valuable in time series
prediction (It is worth noting that the application of other indicators might result in better
prediction accuracies for the stocks under consideration).
EMA (t) = EMA (t-1) + alpha * (Price (t) – EMA (t-1))
Where, alpha = 2/ (N+1), Thus, for N=9, alpha = 0.20
In theory, the Stock Prediction Problem can be considered as evaluating a function F at
time T based on the previous values of F at times t-1,t-2,t-n while assigning
corresponding weight function w at each point to F.
F (t) = w1*F (t-1) + w2*F (t-2) + … + w*F (t-n)
6
The technical analysis charts below show how the EMA models the actual Stock Price.
7
2.2 The Learning Environment
The Weka and YALE Data Mining Environments were used for carrying out the
experiments. The general setup used is as follows:
Since the attribute space we are operating on consists of a very limited number of
attributes (< 10) the Attribute Selection step can be skipped for some of the Machine
Learning methods.
Historical Stock
Data
Data Preprocessing
(Cross Validation)
Attribute Selection
Learning
Algorithm
(Learn Rules)
Learning
Algorithm
(Make Predictions)
Evaluate Results
Another
Learning
Algorithm
Training
Data
Test Data
8
2.3 Preprocessing the Historical Stock Data
For this experiment, the historical data was downloaded from the yahoo finance section.
In particular, the stock prices of two companies were studied, namely Google Inc.
(GOOG) and Yahoo Inc. (YHOO)
The dataset available has the following attributes:
Date Open High Low Close Volume Adj. Close
Intuitively, based on the EMH, the price of the stock yesterday is going to have the most
impact on the price of the stock today. Thus as we go along the time-line, data-points
which are nearer to today’s price point are going to have a greater impact on today’s
price. For a time-series analysis we can take the Date as the X-Axis with integer values
attached to each date, such that the most recent Date Tag in the dataset gets the highest
value and the oldest Date Tag gets the lowest value.
We add one more attribute to the above attributes, this attribute will act as our label for
predicting the movements of the stock price. This attribute will be called “Indicator” and
will be dependent on the other available attributes. For our experiments we use the EMA
(Exponential Moving Average) as the indicator function.
9
3. The Machine Learning Techniques
In this section we evaluate the results generated on applying different learning
algorithms.
3.1 Decision Stump
On applying a simple Decision Stump to predicting the EMA, we found the following
results:
Correlation coefficient 0.8597
Mean absolute error 46.665
Root mean squared error 57.8192
Relative absolute error 46.8704 %
Root relative squared error 50.9763 %
Total Number of Instances 681
3.2 Linear Regression
On applying Simple Linear Regression (with only numeric attributes taken under
consideration) the following results were obtained while predicting the EMA
Correlation coefficient 0.9591
Mean absolute error 12.9115
Root mean squared error 32.0499
Relative absolute error 12.9684 %
Root relative squared error 28.2568 %
Total Number of Instances 681
3.3 Support Vector Machines
Using C-Class Support Vector Machines which use RBF Kernels with the Cost Parameter
C ranging from 512 to 65536, the accuracy in predicting the Stock Movement was as
follows:
Root mean square error: 0.485 +/- 0.012
Accuracy: 60.20 +/- 0.49%
10
3.4 Boosting
The AdaBoostM1 Algorithm was applied to the DataSet after applying the C-SVC
Algorithm. The results show a significant boost with respect to the Accuracy.
Root mean squared error: 0.467 +/- 0.008
Acuracy: 64.32% +/- 3.99%
The following confusion matrix was extracted from the output of the YALE Program
(after applying a combination of C-SVC and AdaBoostM1)
True: 1 -1
1: 37 9
-1: 234 401
False Positive: 23.400 +/- 2.417
True: 1 -1
1: 37 9
-1: 234 401
False Negative: 0.900 +/- 1.221
True: 1 -1
1: 37 9
-1: 234 401
True Positive: 40.100 +/- 1.513
True: 1 -1
1: 37 9
-1: 234 401
True Negative: 3.700 +/- 2.410
11
3.5 Stock Prediction based on Textual Analysis of Financial News Articles
Nowadays, a huge amount of valuable information related to the financial market is
available on the web. A majority of this information comes from Financial News
Articles, Company Reports and Expert Recommendations (Blogs from valid sources can
also act as a source of information) Most of this data is in a textual format as opposed to a
numerical format which makes it hard to use. Thus the problem domain can now be
viewed as one that involves Mining of Text Documents and Time Series Analysis
concurrently.
One method which has been used involves defining the news impact on a particular
stock: Positive, Negative, and Neutral.
A news is considered to have a positive impact (or negative impact) if the stock price
rises (or drops) significantly for a period, after the news story has been broadcasted. If the
stock price does not change dramatically after the news is released, then the news story is
regarded as neutral.
Another method which we study in this paper relates to detecting and determining
patterns in the news articles which correspond directly to a rise or fall in the stock price.
The general architecture is as follows:
A crawler continuously crawls news articles and indexes them for a particular stock
portfolio. The learning environment requests the news since the last T minutes from the
indexer. The learning environment consists of several base learners which look for
specific information in the text document (i.e. patterns like “profits rise” inside a just
released news article, or “share prices will go down” on the blog of a veteran Wall Street
Trader/Speculator etc.). A Bag-Of-Words consisting of Positive Prediction Terms and
Negative Prediction Terms and Phrases is used by the learning environment. Each time a
word/phrase from the Positive Prediction Term occurs in a particular news article, a
PostiveVote is assigned to the article.
12
The Diagram below shows the general architecture of such a system:
As can be seen, this method is very crude in making an accurate prediction.
To enhance the predictions, more weightage can be assigned to articles which come from
credible sources. Also, more weightage can be assigned to a news Headline which
contains a Positive Prediction Term or a Negative Prediction Term.
Boosting can then be applied to the base learners to see if the accuracy can be boosted
further.
Financial News
Crawler
News
Dataset
Preprocessor
Text
Analyzer
Learning
Algorithm
Evaluation
Stocks to
be
surveyed
Bag of
Words
(20 min)
Time
Range
Articles
Indexer
13
3.6 A brief discussion on Applying Expert Weighting to Stock Prediction
Star Analysts
Get Star Analysts for:
GO
This is a list of top research analysts based on the accuracy of earnings estimates on
GOOG, according to StarMine. Analysts that appear here are limited to those covering
GOOG for a significant period of time. Learn More.
Total Ranked Analysts: 31
EPS ACCURACY FOR GOOG - Trailing Two Fiscal Years and Four Quarters
Top-Ranked Analysts GOOG Overall Research Reports
Westerfield, Leland
BMO Capital Markets
Wolk, Marianne
Susquehanna Financial Group
Rohan, Jordan
RBC Capital Markets
Jain, Pratik
First Global Stockbroking Ltd.
Squali, Youssef
Jefferies & Co.
Garcia, Denise
A. G. Edwards & Sons, Inc.
Quarles, Christa
Thomas Weisel Partners
Brown, Derek
Cantor Fitzgerald
As shown in the table above, we are given the opinions of stock market experts (someone
with an opinion, not necessarily the right one!) as input. At each round, we make our
14
prediction based on the predictions of the experts. For the subsequent rounds, we increase
the weights of those experts who predicted the stock correctly and decrease the weights
of those experts who were not correct in their predictions. (Another, often used variation
to this method of expert weighting is to completely disregard those experts who were
incorrect in their previous round prediction, this might yield in lower efficiencies (since
even an expert is bound to make mistakes))
Thus the Expert Weighting Algorithm can be described as follows:
Given: A vector E = { e1, e2,….eN} of Stock Market Experts and their predictions.
Assign W(e(i)) = 1 For Each Expert e(i).
For Round t in 1…T
Make a Prediction based on the Weighted Majority Algorithm.
For experts who made a correct prediction W(e(i))(t) = 2*W(e(i))(t-1)
For experts who made an incorrect prediction W(e(i))(t) = ½ * W(e(i))(t-1)
Store the Expert Ratings for future weight assignments.
The topic of expert weighting based on expert opinions can be considered as a hybrid
technique with influences from both the Fundamental Analysis and Technical Analysis
domains since the experts make their opinions based on the principles of Fundamental
Analysis and out expert weighting algorithm uses that data to do a technical analysis.
It should be noted that the methods described in 3.5 and 3.6 can be considered as a hybrid
combination of Online Learning and Weighted Majority Algorithms because of the
inherent characteristics like
1) Fetching Information One Piece at a time for a given time period.
2) Decisions are based on the past performance without knowing the future.
3) Adapt and learn as we go further.
15
4. Conclusion
Of all the Algorithms we applied, we saw that only Support Vector Machine combined
with Boosting gave us satisfactory results. Linear Regression gave lower mean squared
errors while predicting the EMA pattern.
Another technique which looks promising but which we did not cover the evaluation of
was Expert Weighting. More recently, the linguistic analysis of Financial News Results
to predict stocks has been a topic of extensive study.
The choice of the indicator function can dramatically improve/reduce the accuracy of the
prediction system. Also a particular Machine Learning Algorithm might be better suited
to a particular type of stock, say Technology Stocks, whereas the same algorithm might
give lower accuracies while predicting some other types of Stocks, say Energy Stocks.
Moreover, we should also note that while applying the Machine Learning Algorithms for
Technical Analysis, we assumed that the effect of the Unknown Factors (Election
Results, Rumors, Political Effects etc.) was already embedded into the historical stock
pattern. Commercial Trading systems might have a more sophisticated mechanism for
taking the unknowns into account.
While we studied the algorithms discretely, more often than not, a hybrid algorithm is
used for stock prediction. For instance an Algorithmic Trading System might involve
a 3-tier architecture with SVM’s and Boosting at the bottom, an Online Algorithm (For
instance an Expert Weighting scheme that we discussed in section 3.6 as the middle layer
and Textual Analysis of Stock Market News, Financial Reports as the top layer to make
predictions.
As a result of the research conducted for this project, a subset of the algorithms discussed
above have been implemented at the following web address:
http://www.stoocker.com
16
5. Tools//DataSets
Tools
LibSVM
http://www.csie.ntu.edu.tw/~cjlin/libsvm/
YALE (Yet another Learning Environment)
http://rapid-i.com/content/view/26/82/
WEKA
http://www.cs.waikato.ac.nz/ml/weka/
DataSets
Historical Stock Data available from http://www.finance.yahoo.com
Financial Stock Data also available for non-commercial purpose at:
http://www.econ.ubc.ca/whistler/325/dataset.htm
The most commonly available attributes are:
Date Open High Low Close Volume Adj. Close
17
6. References and Further Reading
Forecasting stock market movement direction with support vector machine
http://madis1.iss.ac.cn/madis.files/pub-papers/c&or-hw-hnw-04-1.pdf
On Developing a Financial Prediction System: Pitfalls and Possibilities
http://www.smartquant.com/references/NeuralNetworks/neural20.pdf
Prediction of Stock Market Index Changes
http://citeseer.ist.psu.edu/cache/papers/cs/129/ftp:zSzzSzftp.cs.bilkent.edu.trzSzpubzSzte
ch-reportszSz1992zSzBU-CEIS-9201.pdf/sirin93prediction.pdf
J Moody, M Saffell, Learning to Trade via Direct Reinforcement, IEEE
Transactions on Neural Networks, Vol. 12, No 4, July 2001.
http://www.cs.ucsd.edu/~dboswell/PastWork/Moody01LearningToTradeViaDirectReinfo
rcement.pdf
AUTOMATED TRADING WITH BOOSTING AND EXPERT WEIGHTING
http://www.andromeda.rutgers.edu/~jmbarr/NYComp/CreamerEEA.pdf
Forecasting Stock Prices Using Neural Networks
http://www.andrew.cmu.edu/user/wyliec/project.pdf
Using Neural Networks to Forecast Stock Market Prices
http://people.ok.ubc.ca/rlawrenc/research/Papers/nn.pdf
MACHINE LEARNING IN COMPUTATIONAL FINANCE
http://www.cs.rpi.edu/~magdon/students/boyarshinov_victor/boyarshinov_PhDthesis.pdf
The Predicting Power of Textual Information on Financial Markets
http://www.comp.hkbu.edu.hk/~cib/2005/Jun/iib_vol5no1_article1.pdf
APPLICATION OF MACHINE LEARNING TO SHORT-TERM EQUITY
RETURN PREDICTION
http://publish.uwo.ca/~jnuttall/cooper.pdf
Foreign Exchange Trading using a Learning Classifier System
vagrant
18
http://www.cems.uwe.ac.uk/lcsg/reports/uwelcsg05-007r.pdf
Techniques and Software for Development and Evaluation of Trading Strategies
http://www.cs.umu.se/~thomash/reports/phdthesis.pdf
Data Mining for Prediction Financial Series Case
http://szemke.math.univ.gda.pl/zemke2003PhD.pdf
Reinforcement Learning for Optimized Trade Execution
http://www.cs.ualberta.ca/~sutton/kearnstradeexecution.pdf
Vatsal H. Shah | vatsals@vatsals.com | vhs212@nyu.edu
Foundations of Machine Learning | Spring 2007
Dr. Mehryar Mohri
Courant Institute of Mathematical Science
New York University
19