Worcester Polytechnic Institute

Digital WPI
Major Qualifying Projects (All Years) Major Qualifying Projects

April 2015

Trading in the Financial Market Using Data Mining

Tyler James Stone
Worcester Polytechnic Institute

Follow this and additional works at: https://digitalcommons.wpi.edu/mqp-all

Repository Citation
Stone, T. J. (2015). Trading in the Financial Market Using Data Mining. Retrieved from https://digitalcommons.wpi.edu/mqp-all/3368

This Unrestricted is brought to you for free and open access by the Major Qualifying Projects at Digital WPI. It has been accepted for inclusion in
Major Qualifying Projects (All Years) by an authorized administrator of Digital WPI. For more information, please contact digitalwpi@wpi.edu.
 Trading in the Financial Market
Using Data Mining
A Major Qualifying Project Submitted to the faculty of

Worcester Polytechnic Institute

In partial fulfillment of the requirements for the

Degree in Bachelor of Science in

Computer Science

by Tyler Stone

and

the Degree in Bachelor of Science in

Electrical and Computer Engineering

by Ryan McKenna

4/30/2015

Project Advisors:

Professor Carolina Ruiz Professor Hossein Hakim Professor Michael Radzicki

Computer Science Electrical and Computer Social Science and Policy Studies
Engineering

This report represents work of WPI undergraduate students submitted to the faculty as evidence of a
degree requirement. WPI routinely publishes these reports on its web site without editorial or peer
review. For more information about the projects program at WPI, see
http://www.wpi.edu/Academics/Projects.
Abstract
This project developed a replicable process to associate stocks into clusters based on time series data,
and selects an appropriate automated trading strategy for each cluster for use in trading. This process
included an exploration of data pre-processing methods, selection of a clustering algorithm suited to
this application, identification of an optimal investment strategy for each cluster, and the application of
strategies on the algorithmically generated portfolio. Efficacy was determined through empirical
comparison of gains seen in each test, with the goal of beating the market, or generating percentage
greater than the change observed in the S&P 500. This process will serve as a basis for future research
and development in the field of applied data mining within the financial domain.

i
Executive Summary

Background
The proliferation of personal computing in the financial realm has provided investors with
unprecedented access to real-time data, allowing for decisions to be made faster and on a broader
scope than ever before. To make use of this data, automated trading strategies, which are coded criteria
that can signal the purchase or sale of assets to be made when fulfilled, are set up by the user to
monitor and trade target assets. The process of determining which assets an investor may focus on is
time consuming and non-trivial, relying on extensive domain knowledge. It may be possible, however,
to streamline this selection through the application of data mining.

Goals and Objectives

This project aims to develop a process to associate behaviorally similar stocks into portfolios through
clustering techniques, and trade the clustered portfolios to beat the performance of the market in terms
of percentage gains of the S&P 500 in congruent time frames. Although stock data is analyzed herein,
the process used for creation of algorithmically generated portfolios can be used to increase the ease
and scope with which an investor can participate in any asset class of the financial market for which time
series pricing data is available. In order to accomplish this goal, the following objectives are defined:

 Devise an approach to produce effective clustering of assets.

 Provide a replicable process to implement in a live market
 Evaluate profitability, benchmark against market conditions
 Make recommendations for further process improvement

Methods
A library of Java, Matlab, Tradestation Easy Language, and Python scripts, to be deployed as described in
Appendix C, was created to acquire data, preprocess and cluster this data into portfolios, and apply
automated trading strategies to said portfolios. Acquisition of data was performed with a Java
application that queries the Yahoo! Finance API, creating a CSV file of daily closing prices as a time series
for all stocks whose average yearly trade volume is greater than 1 million. Matlab is then used to
preprocess each 260-day time series through normalization by Z-score. Matlab then generates clusters
from this preprocessed data using the K-Medoids algorithm, outputting a TXT file of cluster centers and
all stocks contained in the cluster, as sorted by ascending distance from center.

Five automated trading strategies, selected to perform well in different market conditions, are
then applied to the cluster centers in Tradestation. A Python script is used to determine the optimal
strategy for each center through a weighted averaging of performance metrics output by Tradestation.
A portfolio is then created for each cluster, containing the center and the 20 stocks that are closest to
the center. Tradestation’s Portfolio Maestro is then used to apply the optimal strategy selected earlier,

ii
to all stocks in the corresponding portfolio. These portfolios were then traded on historical market data
in two tests (one over 120 contiguous trading days with portfolios re-clustered at different intervals, and
another on five disjoint clustered portfolios spread over five years) to expose the automated trading
systems to varied market conditions.

Conclusion
For clustering performed on 260 days of preprocessed closing prices for all stocks, the optimal number
of clusters, K, was found to be 10 in most cases. Thus, all clustering was performed for 10 clusters, to
enable repeatability and direct comparison of the tests herein. Upon testing, a re-clustering period of 60
days was found to be most viable, performing profitably on average across all market conditions
experienced and consistently beating the performance of the S&P 500 in all tests. This suggests that
stocks that behave similarly for 260 days will continue to behave within the margins of profitability for
the automated trading systems used for 60 days following the end date of the 260-day clustering
window, and that a portfolio curated as such should be re-clustered every 60 days to remain profitable.
Further research to verify the conclusion of 60-day viability should be performed, as the cumbersome
nature of the manual data entry necessary to perform these tests prevented the generation of enough
samples to determine the statistical significance of the results. Additionally, refinement of the
automated trading strategies used, which were kept simple to limit the variables at play in the tests
performed, can greatly augment the profitability of this system for practical implementation.

iii
Acknowledgements
Our team would like to thank our advisors for their support and contributions to our project:

 Professor Carolina Ruiz of the Computer Science Department for sharing her data mining
expertise
 Professor Hossein Hakim of the Electrical and Computer Engineering Department for sharing his
experience in the financial domain and automated trading systems
 Professor Michael Radzicki of the Social Science and Policy Studies Department for sharing his
experience in the financial domain and automated trading systems

iv
Table of Contents
Abstract .......................................................................................................................................................... i
Executive Summary....................................................................................................................................... ii
Background ............................................................................................................................................... ii
Goals and Objectives................................................................................................................................. ii
Methods .................................................................................................................................................... ii
Conclusion ................................................................................................................................................ iii
Acknowledgements...................................................................................................................................... iv
Table of Contents .......................................................................................................................................... v
List of Figures .............................................................................................................................................. vii
List of Tables .............................................................................................................................................. viii
1 Introduction .......................................................................................................................................... 1
2 Background ........................................................................................................................................... 2
2.1 Introduction to the Financial Domain and Portfolio Management .............................................. 2
2.1.1 Stock Market Basics .............................................................................................................. 2
2.1.2 Stock Portfolios and Trading ................................................................................................. 2
2.1.3 Beating the Market ............................................................................................................... 2
2.2 Decision Making in Trading ........................................................................................................... 2
2.2.1 Automated Trading Systems ................................................................................................. 3
2.3 Fundamentals of Cluster Analysis ................................................................................................. 4
3 Methodology......................................................................................................................................... 5
3.1 Goal ............................................................................................................................................... 5
3.2 Objectives...................................................................................................................................... 5
3.3 Overview ....................................................................................................................................... 5
3.4 Explanation of Data ....................................................................................................................... 8
3.4.1 Types of Data Acquired ......................................................................................................... 8
3.4.2 Filtering ................................................................................................................................. 8
3.5 Data Pre-Processing ...................................................................................................................... 8
3.5.1 De-trending ........................................................................................................................... 8
3.5.2 Normalize by Z-Score ............................................................................................................ 9
3.5.3 Scaling ................................................................................................................................... 9

v
 3.5.4 Choosing the Optimal Pre-Processing Technique ................................................................. 9
3.6 Clustering .................................................................................................................................... 10
3.6.1 K-Medoids ........................................................................................................................... 11
3.6.2 Determining the Number of Clusters .................................................................................. 11
3.6.3 Presentation of clustering results ....................................................................................... 12
3.7 Trading Systems .......................................................................................................................... 12
3.8 Testing ......................................................................................................................................... 13
3.8.1 Determining the Optimal Strategy ...................................................................................... 14
3.8.2 Constructing the Portfolio................................................................................................... 14
3.8.3 Tests .................................................................................................................................... 15
4 Results and Analysis ............................................................................................................................ 16
4.1 Backtesting Results ..................................................................................................................... 16
4.2 Comparison to S&P 500 .............................................................................................................. 19
5 Conclusions and Future Work ............................................................................................................. 20
5.1 Need for Improved Data Input Method ...................................................................................... 20
5.2 Improvement of Trading Strategies ............................................................................................ 20
6 Bibliography ........................................................................................................................................ 22
Appendix A: Single Point Data .................................................................................................................... 24
Appendix B: Platform Information .............................................................................................................. 25
Appendix C: Detailed Process Breakdown .................................................................................................. 26
Appendix D: Tested Portfolios 2008-2012 .................................................................................................. 30
Appendix E: Trading Strategies ................................................................................................................... 40

vi
List of Figures
Figure 1 - Stocks clustered over 260 days ..................................................................................................... 4
Figure 2 - Flowchart of Clustering Process .................................................................................................... 7
Figure 3 - Description of K-Medoids ........................................................................................................... 11
Figure 4 - Sum of squared error per Tested Number of Clusters ............................................................... 12
Figure 5 - Automated strategies and respective optimal behaviors (Graphs from Google) ...................... 13
Figure 6 - Visualization of 5, 10, 20, 40 60 and 120 day testing Timeframes ............................................. 15
Figure 7 - Net Profit Over Time per Test ..................................................................................................... 17
Figure 8 - S&P 500 Index Chart from 4/4/2008 to 11/2/2012 .................................................................... 17
Figure 9 - Optimal Strategies per Year ........................................................................................................ 18
Figure 10 - S&P and Clustered Portfolio gains for 10/14/2013 to 3/28/2014 ............................................ 19
Figure 11 - Cumulative S&P 500 and 60-day portfolio gains over 5 years.................................................. 19

vii
List of Tables
Table 1 - Correlation & Sum of squared error of the Clustered Data ........................................................... 9
Table 2 – Example of selected Strategy Performance Attributes and their Weights ................................. 14
Table 3 – 20, 40, 60, and 120-day interval tests for 2013 financial year starting from a $100,000
investment .................................................................................................................................................. 16
Table 4 - 5 year Average Revenue per Trading Interval.............................................................................. 17

viii
1 Introduction
The financial market, once a vision of instant riches and Wall Street, has become easier to access than
ever due to the proliferation of personal computers and the Internet. With online brokering applications
such as Tradestation, investors can trade from the comfort of home. Along with this newfound
availability of data, automated trading systems have become popular as a means of interpreting and
making trading decisions upon market data. A trader can develop a hard-coded trading system that uses
the rules and criteria that they had previously followed manually to automatically buy and sell stocks.

While it may seem ideal to automate a computer to make money for an individual, there are
many caveats that complicate the process of implementation. For instance, automated trading
strategies only perform well under certain market conditions as defined by the user, however the
market conditions that the strategy is trading upon can change frequently. An automated trading system
alone is not intelligent enough to identify those conditions unless it is programmed to do so. An optimal
trading strategy is difficult to quantify, due to the varied conditions under which different strategies
perform best, and the differences in what individual investors may seek to reap from the market. The
selection of an automated trading strategy to use must be unique to a target stock, or a group of stocks
that exhibit the same behavior over time.

This project developed a process by which stocks are grouped into behaviorally similar clusters
based on their time series, and traded by automated trading strategies identified to be optimal for each
cluster’s behavior. This process encompasses extensive research and experimentation, leading to a final
process using a Java application for data acquisition, Matlab scripts for data pre-processing and
clustering, Tradestation and Python scripts for optimal strategy identification, and Tradestation Portfolio
Maestro for backtesting of the portfolios, assessing how the clusters would have traded in historical
market conditions. This process will act as a building block for future research and development in
applied data mining for the financial domain.

1
2 Background

2.1 Introduction to the Financial Domain and Portfolio Management

2.1.1 Stock Market Basics

Capital, or the financial reserves used to purchase assets, is the driving force of investment for any
entity, from corporations to an individual. Corporations may seek capital to fund infrastructure
improvements, new hires, or further investment in other companies. Corporations that opt to be
publicly traded have the power to generate capital through the sale of shares in the company to
investors in the general public in the form of stocks (Investopedia, LLC). Each stock represents a small
sliver of the corporation’s equity, represented as a percentage of the company. Thus, when the net
worth of a company increases, the value of each share increases accordingly, or decreases if the
company loses value over time. Shareholders may opt to buy shares in a corporation that shows
promise in increasing in value, or to invest in a company that they may be otherwise invested in the
future of. In the case of the individual investor, personal capital may be used to purchase shares in
hopes of generating financial gains, thus augmenting the individual’s equity.

2.1.2 Stock Portfolios and Trading

The individual assets held by an entity at any given time compose a portfolio. These assets flow in or out
of a given portfolio through the purchase, holding, or sale of specific assets, through the transaction of
trades in the financial marketplace. Stock assets held in a portfolio can be in the form of a long position,
representing that the investor has purchased shares of a corporation in the hopes that they will increase
in value to be then be sold at a profit, or a short position, wherein the investor sells shares borrowed
from a broker at current value in hopes that the asset will decrease in value, at which point the investor
would buy an equivalent number of shares to cover their debts to the broker at the new, lower price
(Milton).

2.1.3 Beating the Market

A well curated portfolio should minimally “beat the market,” or be more profitable on average than the
overall market in comparable time periods. The S&P 500 index is a time series value, reflective of the
500 most valuable stocks in the US, which is commonly accepted as an indicator of market health,
making it a good benchmark against which to benchmark the performance of a portfolio. Thus, to beat
the market with a portfolio, the curated collection of assets must show gains larger than those of the
S&P 500 for the same time period.

2.2 Decision Making in Trading

Due to the vast number of qualitative variables at play, market behavior is non-causal, in that although
correlations can be found between different assets and indicators, the Gambler’s Fallacy often tricks
investors into believing these correlations represent causation. In other words, the individual is led to

2
believe that events in the current time frame will influence the outcome of random events to come
(Phung).

To navigate the complex financial realm, and guide investment decisions, investors rely on many
established theorems and rules to make sense of the available data. These rules are often published by
established financial institutions, or by successful traders who turn to education and selling knowledge
as an additional revenue stream. Generally speaking, these rules and methods are composed of criteria
that a stock must meet to be deemed favorable for purchase under different investing schemes.
Furthermore, every time a trade is made on a stock, a tick is generated, driving the price up or down by
some amount. With roughly 6000 stocks to consider, and approximately 700 billion shares traded per
year in the NYSE alone, the full scope of the available data is far more than an individual can monitor
(U.S. Census Bureau, 2012). Thus, investors generally limit their scope accordingly, utilizing filters to
manage the data, such as looking at weekly values rather than daily values, or participating in a
particular sector.

2.2.1 Automated Trading Systems

Automated trading systems, in which strategies are implemented via coded rules that signal the system
to buy or sell a given number of shares if a set of conditions is met, offer increased speed and discipline
of decision making. Thus, more trades can be made across multiple assets simultaneously than with
manual trading, with minimal slippage, defined as “the difference between the expected price of a
trade, and the price that the trade executes at” (Investopedia, LLC).

Though a manual trader may use rules to guide oneself in trading, the mind is susceptible to the
Gambler’s Fallacy or other psychological pitfalls, whereas the deterministic nature of computing allows
for coding a set of rules, which will be followed regardless of emotional influence. For example, if a
particular position is held, but the value of the asset has declined a significant amount, a manual trader
may cave to emotional pressure and withdraw from the position to prevent further losses, closing the
trade at a loss. However, the drawdown noticed might be within the realm of statistical possibility,
wherein the asset could rebound in value. If the strategy had been well calculated and hard coded into
a program, the system would have held onto the position, allowing the value to rebound, making the
trade ultimately profitable.

Automated systems are not without pitfalls however, in that the learning curve of
implementation may be very steep for individuals unfamiliar with coding, and the conditions under
which the algorithms must function is constantly changing with market conditions. Though many tools
and pre-coded strategies are available through online brokerage platforms to any individual looking to
trade the stock market, the finer points of implementation can be difficult to master, with vast amounts
of observation and research necessary to determine when to enter a market and refine a strategy
(Wright, 1998). Furthermore, the argument can be made that a well-curated portfolio represents an
Ergodic system, wherein the system “forgets” previous states due to ever-changing statistical conditions
driving asset values (Veysov, 2012).

3
2.3 Fundamentals of Cluster Analysis
Automated financial systems provide unprecedented access to data streams and statistical calculation,
yet tools must be employed to make sense of this data. One such tool is that of cluster analysis, in
which data is grouped in such a way that items within a cluster are similar to one another, but different
than those contained in other clusters. In the context of time series data, clustering separates data
based on behavioral patterns in the stock. The behavioral similarities found by time series clustering can
be visually identified, as shown in Figure 1, which displays a time series clustering of stocks, which can
be seen to follow the same upwards and downwards trends over the 260 day period.

Figure 1 - Stocks clustered over 260 days

The K-means clustering algorithm, of which a variation called K-Medoids was employed herein
(as described in Section 3.6.1), is well suited for clustering of time series. Both methods are heuristic, in
that their solutions may not be optimal, as demonstrated by the random initial selection of centroids in
the algorithm. A hill-climbing method is then employed to iterate toward a more suitable centroid, in
which each new cluster is more closely associated by inter-cluster distance (or further associated from
other clusters in the case of K-Medoids) than in the previous iteration (Keogh & Lin, 2005).

4
3 Methodology

3.1 Goal
This project aims to develop a process to associate behaviorally similar stocks into portfolios through
clustering techniques, and trade the clustered portfolios to beat the performance of the market in terms
of percentage gains of the S&P 500 in congruent time frames. Although stock data is analyzed herein,
the process used for creation of algorithmically generated portfolios can be used to increase the ease
and scope with which an investor can participate in any asset class of the financial market for which time
series pricing data is available.

3.2 Objectives
 Devise an approach to produce effective clustering of assets
 Provide a replicable process to implement in a live market
 Evaluate profitability, benchmark against market conditions
 Make recommendations for further process improvement

3.3 Overview
Figure 2 provides a flow chart of the process used to acquire, pre-process, and cluster asset data for this
project. The description below details this same process, as well as the subsequent trading and
benchmarking methods.

Data Sample Acquisition:

1. Retrieve data from Yahoo! API (Java application)

2. Filter data
3. Select window to cluster upon (Matlab)

Clustering:

4. Pre-process data using the Z-Normalize method (Matlab)

5. Determine optimal number of clusters by elbow method (Matlab)
6. Perform clustering, store cluster results and cluster prototypes (Matlab)

Trading:

7. Test upward trending strategy on prototype and export Strategy Performance report to Excel for
each cluster (Tradestation)
8. Repeat step 7 for downward trending, upward and downward trending, directionless, and
volatility strategies

5
 9. Determine the best performing strategy using a weighted average of strategy performance
attributes (Python script)
10. For each cluster, create a portfolio of the 20 stocks with the shortest Euclidean distance to the
prototype, including the prototype (Portfolio Maestro)

 Test 1: Repeat steps 4-9 six times at 20 day intervals.

 Perform step 10 to generate 4 portfolios covering 120 days: one clustered six times and traded
at 20 day intervals, one clustered three times and traded at 40 day intervals, one clustered twice
and traded at 60 day intervals, and one clustered only once and traded for the full 120 days.

 Test 2: Trade using the best performing strategy for each of the portfolio of stocks for 5, 10, 20,
40, 60, and 120 days after the clustered duration (Portfolio Maestro)
 Repeat steps 4-10 and for 4 more instances, using disjoint date ranges for clustering, and repeat
test 2.

Benchmarking:

11. Compare percentage gains of each test against S&P 500 index (SPY) percentage gains for
identical time frames.

6
Figure 2 - Flowchart of Clustering Process

7
3.4 Explanation of Data

3.4.1 Types of Data Acquired

Both single point data and time series data were acquired through a Java script that queries the Yahoo!
Finance API. Single point data consisted of fundamental data about a company that is calculated over a
longer period of time than one day, such as quarterly or yearly. For a complete list of single point data
collected, please see Appendix A. Time series data consisted of the daily closing price of a stock.

The stocks for which data was collected were selected from a vast pool of over 6000 total stocks
traded in the NASDAQ, NYSE, and AMEX exchanges. These were limited by means of filtering by single
point data and the availability of 2016 days of closing price, as described below in Section 3.4.2.
However, due to lack of available single point data relevant to historical dates (such as those used for
backtesting), this project was unable to include the acquired single point data in the clustering scheme.
Ultimately, the data used for clustering was an 880 by 2016 matrix, representing 8 years of daily closing
prices for 880 stocks. 260-day subsets were input to the clustering algorithm (described in Section 3.6).
Given a data acquisition date of March 20, 2015, this assumes availability of price data from March 19,
2007 to the date of collection. A complete list of unfiltered and filtered data is available in the github
repository listed in Appendix C.

3.4.2 Filtering
This project made use of current single point data for filtering, to eliminate stocks that demonstrate
little movement in price. Low volume stocks, with volume being the number of total trades of said
stocks, were removed from the dataset by filtering out any stocks with an average yearly volume less
than 1,000,000. In order to provide enough data for backtesting, stocks in the dataset were filtered to
have at least 2160 days of data. This data represents roughly 10 trading years and allowed for testing on
a variety of market conditions. The dataset, after filtering, contained the 2160-day time series for of 880
stocks.

3.5 Data Pre-Processing

In order for the stocks to be compared and clustered accurately, methods for data pre-processing had to
be applied to the time series data using Matlab to fit all values to an equivalent range. However,
choosing the pre-processing method for this project was not arbitrary. The following methods were
tested to determine the best pre-processing technique for the time series data.

3.5.1 De-trending
The process of de-trending the time series data consists of taking the average closing price over a
chosen clustering period and subtracting the average from each daily value, removing any upwards or
downwards trends (The MathWorks, Inc).

8
3.5.2 Normalize by Z-Score
Normalizing by z-score is the process in which the average closing price for a given stock over the
clustering period is subtracted from each daily value of a given stock and divided by the standard
deviation of the closing price for each daily value of a given stock, following the equation below
(Wikimedia Foundation, Inc, 2015).

3.5.3 Scaling
The time series data could be scaled from 0 to 1 by finding the minimum and maximum of each closing
price of a given stock. The minimum closing price of the stock is subtracted from each value in the
dataset and then divided by the result of subtracting the minimum from the maximum, following the
equation below (Wikimedia Foundation, Inc, 2015).

3.5.4 Choosing the Optimal Pre-Processing Technique

The above pre-processing techniques had to be tested in order to determine the best pre-processing
technique or group of techniques to use on the dataset. The techniques were tested using a dataset
consisting of 1300 days of 880 stocks. The examined techniques and groups of techniques can be found
in Error! Reference source not found. below.

Table 1 - Correlation & Sum of squared error of the Clustered Data

Correlation Sum of squared error

Processing Technique Random Actual Random Actual
No Pre-Processing -0.1320 -0.0230 1004700000 158990000
Detrend Only -0.1341 -0.0195 974030000 511200
Z-Normalize Only -0.1557 -0.4523 257050 81562
Scale (0 to 1) Only -0.1345 -0.4186 23238 7067
Detrend & Z-Normalize -0.1561 -0.2919 252370 63993
Detrend & Scale (0 to 1) -0.1372 -0.3120 22502 4152
Z-Normalize & Scale -0.1338 -0.4255 23258 7002
Scale & Z-Norm & Detrend -0.1356 -0.3088 22470 4156

First, the optimal number of clusters for the dataset was determined using the sum of squared
error elbow method described in Section 3.6.2. Then, the selected pre-processing techniques were
applied to the dataset, and the dataset was clustered using the k-means clustering algorithm. The
correlation between the proximity matrix and the incidence matrix of the clustering, and the of sum of
squared errors of intra-cluster distances were then collected for each test and compared. A proximity
matrix is an n x n (with n being the number of stocks in the dataset to be clustered) representation of
the calculated Euclidean distance, described in Section 3.6.1, from every stock in the dataset. For
example, the value in row i, column j of the matrix contains the Euclidean distance between Stock i and

9
Stock j. An incidence matrix is an n x n matrix of binary values, indicating whether each corresponding
pair of stocks belongs to the same cluster. For example, if Stock i and Stock j belong to the same cluster,
the value within the matrix at row i, column j would contain ‘1’. If Stock i and Stock j did not belong to
the same cluster, the value within the incidence matrix at row i, column j would contain ‘0’. Both
correlation and sum of squared error are commonly-used metrics in cluster performance evaluation,
and were therefore chosen to evaluate the performance of pre-processing methods. A correlation value
of -1.0 is optimal, as it describes clusters that are entirely similar within themselves, and different from
other clusters (Tan, Steinbach, & Kumar, 2005). It is important to note that the sum-squared error
depends on the input data, as the low sum squared of distances of all scaled trials in Table 1 portrays.

The equation for sum of squared error (SSE) is above, in which there are K clusters and xi
represents each individual instance within a cluster. Ck represents the k-th cluster. The variable k
represents the mean vector of each cluster, represented by the equation below.

In the equation above, Nk represents the number of instances within cluster k and each is
represented as a vector (Rokach & Maimon, 2010).

The Normalize by Z-score pre-processing method performed best out of the examined pre-
processing techniques with the greatest negative correlation, and was thus selected as the sole means
of pre-processing our time series data.

3.6 Clustering
After pre-processing, stocks were clustered using the K-Medoids algorithm again using Matlab. Before
clustering however, the optimal number of clusters had to be determined, as the number of clusters, K,
must be specified for the K-Medoids algorithm. For all clustering, 260-day time series (representing
roughly one trading year) were used as the input to represent each stock.

10
3.6.1 K-Medoids
K-Medoids clustering was used to associate stocks based on time series behavior via the Partitioning
Around Medoids Algorithm, as described in Figure 3 (Theodoridis & Koutroumbas, 2006).

Figure 3 - Description of K-Medoids

To determine the clusters, Euclidean distance was used as the distance metric between time
series data for a pair of stocks. Euclidean distance is a metric to find the distance between two points. It
is simple to apply to multi-dimensional data, making it a common and useful metric. An example of
Euclidean distance to find the distance between the time series of stocks A and B with 260 days of data
is shown in the equation below (Wikimedia Foundation, Inc, 2015), where Ai and Bi are closing prices on
each day for stocks A and B, respectively.

Although testing with single point data was omitted for this project, a distance metric was
implemented to combine the Euclidean distance of time series data and that of the single point data. To
use both single point data and time series data in a distance, a weighted average would be applied to
the Euclidean distance of time series data (t) and the Euclidean distance of the single point data (s),
following the equation below. The sum of weights (w1 + w2) would total 1.0.

K-Medoids provides an advantage over other clustering algorithms for this project in that the
defined centroid of the cluster is an existing data point rather than an averaged point in the feature
space, as with k-means. The defined centroid, known as a Prototype, provides a stock, characteristic of
the cluster’s behavior, upon which to test for an optimal strategy, which can then be applied to the
other stocks in the cluster. Setting the seed of the random number generator in Matlab to a fixed value
yields consistent cluster centroid selection for comparison, but means that centroids will be found on
local optimums rather than global optimum (Keogh & Lin, 2005).

3.6.2 Determining the Number of Clusters

The sum of squared error elbow method was used to determine the optimal number of clusters
applicable to the data. The elbow method is a method that iterates from 2 to n clusters, finding the
average sum of squared error between all the clusters for each iteration. The sum of squared error will

11
quickly decrease, but as the number of clusters increases the sum of squared error will change in smaller
decrements. The “elbow” of this curve is chosen as the optimal number of clusters (Tan, Steinbach, &
Kumar, 2005). Figure 4 below illustrates a graph of the sum of squared error and the selected optimal
number of clusters, in which there is optimal similarity within clusters, and dissimilarity between
clusters.

A script was written to find the optimal number of clusters by from this graph. A line is created,
starting at the first point on the graph and connecting to the last point on the graph. Then, the
perpendicular distance between this line and each point on the graph is calculated. The point with the
largest distance from the line is determined to be the optimal number of clusters.

Figure 4 - Sum of squared error per Tested Number of Clusters

The optimal number of clusters for the time series data in the tested time intervals was most
commonly 10 clusters. This number was chosen as the default number of clusters for all future work on
the dataset.

3.6.3 Presentation of clustering results

To assist in decision-making and the manual entry of clustering results into the trading platform, the
clustering results are output to a .csv file, each column representing a cluster. The first element in each
column is the prototype of the given cluster, and the successive elements moving down are sorted by
distance to the centroid of the cluster. Thus, the closest n stocks can be selected for use, omitting
outliers that display a high degree of deviance from the prototype, relative to other stocks in the given
cluster.

3.7 Trading Systems

Once clusters are established, trading strategies can be applied to cluster prototypes using Tradestation
to determine which trading strategies might perform well on the given cluster. A cluster prototype is the
centroid of all stocks in a cluster, meaning that its behavior over time is indicative of average behavior
witnessed in other stocks of the same cluster. If a prototype stock performs well using a certain strategy,

12
it is hypothesized that other stocks behaving similar to that prototype will perform just as well with that
strategy.

Five simple trading strategies were used. Additional information about these trading strategies,
including the Easy Language code for implementation, can be found in Appendix E. The purpose of
simple strategies was to prove that the clusters performed well as a portfolio regardless of strategy
complexity. The chosen strategies were diverse, therefore performing better or worse depending on the
market conditions that a stock is experiencing (Wright, 1998). By applying each of these trading systems
to each cluster prototype and comparing the results, the system that performs best can be determined,
thus characterizing the cluster.

Figure 5 - Automated strategies and respective optimal behaviors (Graphs from Google)

Error! Reference source not found. displays the five trading strategies chosen for application to
the clusters, and examples of the respective market conditions that they perform well under. Trend
following strategies perform well in market conditions that consistently move upward or downward,
showing little deviation from the average rate of change, well suited for trading over long intervals. The
up trend and down trend systems developed for this project are very similar, but the downward
trending system signals sell short and buy to cover orders instead of buy long and sell orders, as the up
trend system signals. In addition, both the upward and downward trend following systems can be
applied simultaneously for long term, bi-directional movement. Directionless strategies perform well in
market conditions that trade sideways, showing little gain or loss on average across a time period.
Similarly, the volatility strategy is profitable in markets that are classified as noisy, capitalizing on the
chaotic movements in stock price (Wright, 1998).

3.8 Testing
For the tests performed, backtesting was used in place of real-time testing. In finance, the term
backtesting refers to testing a trading strategy over a period of time in history in order to determine
whether or not the strategy might perform well in the future. While it would take a long time to test a
strategy in real-time, years of past data can be used to benchmark strategy performance and make

13
informed choices about what strategies may perform best in subsequent days. Tradestation and
Tradestation Portfolio Maestro were used for data acquisition and backtesting on market data utilizing
daily bars, which each represent the open and close price of a stock over for a given day.

3.8.1 Determining the Optimal Strategy

Using the same time frame upon which each cluster was generated, the optimal strategy for each cluster
prototype was tested using a custom Python script. Strategy performance was assessed by selecting
eight attributes from the Tradestation Strategy Performance Report that investors commonly use as
measures of how well a strategy performs. A weighted average was then applied to the attributes to
determine the final “strategy score,” with weights set arbitrarily, again according to how an investor
may judge efficacy. The attributes and weights can be re-configured to reflect the investor’s trading
style. An example of this selection process, including the selected attributes and their weights, as well as
the weighted average scores for each strategy can be found in Table 2 below.

Table 2 – Example of selected Strategy Performance Attributes and their Weights

The bottom row of Table 2 indicates the “strategy score” for each strategy, as defined in the
equation below, which was used to select the optimal trading strategy for the cluster. In this case, the
Volatility strategy was selected as it had the most favorable results, shown by the Strategy score being
the highest.

3.8.2 Constructing the Portfolio

Once the optimal strategy was selected for each cluster, portfolios were constructed using Portfolio
Maestro from the 20 stocks with the shortest Euclidean distance from the prototype, including the
prototype. Preliminary testing indicated that portfolios containing the entire cluster contents did not
generate profitable results, due to the relative dissimilarity of the behavior of stocks furthest from the
prototype in the cluster. This additional step of filtering was performed because K-Medoids clustering
classifies all input data into one of the K clusters that it generates, leading to the presence of relative
outliers within the clusters. Therefore, the number of stocks used in a trading portfolio had to be limited
to buffer against the presence of outliers. The number of stocks to include from each cluster into each

14
portfolio was selected to be 20, as an arbitrary value large enough to show convergent behavior
between a group of stocks, but small enough to not capture outliers from clusters that may contain as
few as 40 stocks.

3.8.3 Tests
Tests were then carried out on the clustered portfolios using Portfolio Maestro, the results of which can
be seen in section 4.1. The optimal strategy for each portfolio was applied on various durations starting
at the end of the cluster period. Test 1, of 6x20, 3x40, 2x60, and 1x120 clustering intervals, was
performed within a single 120-day period to asses the portfolios’ (re-clustered at the described
intervals) performance under equivalent market conditions. Test 2 traded 5 disjoint sets of clustered
portfolios for 5, 10, 20, 40, 60, and 120 days each, with 5 trading periods for each clustered portfolio, to
subject the system to the varied market conditions seen from 2007 to 2012, and assess at which interval
the system is most profitable on average. One example of cluster-training and test durations is
visualized in Figure 6 below.

Both tests are then benchmarked against the S&P 500 to gauge if the market was successfully
beaten in the respective time frame, as seen in section 4.2.

Training Test

Figure 6 - Visualization of 5, 10, 20, 40 60 and 120 day testing Timeframes

15
4 Results and Analysis
For reproducibility, all platform information is included in Appendix B, with the exact step-by-step
explanation available in Appendix C.

4.1 Backtesting Results

Results of test 1 are shown in Table 3, representing portfolio testing on prototypes and the nearest 20
stocks in clusters generated from the 2012 financial year, and traded during the 2013 financial year with
a $100,000 initial investment. The state of the market throughout 2012-2014 has been what is
described as a “Bull market”, wherein the market is trending upwards consistently. This presents
favorable conditions for trading, hence the profitable results across all tested time intervals. The 60-day
interval is the most profitable in this test, however, showing 31% gains after trading for 120 days.
Table 3 – 20, 40, 60, and 120-day interval tests for 2013 financial year starting from a $100,000 investment

The notion that a 60-day interval for re-clustering is most effective is further supported by the
results seen in Table 4, which displays the results of test 2. Portfolio information is available in Appendix
D. This range of testing dates, selected to include the recession of 2008-2009, presents more diverse
market conditions those from which the results in Table 3 were derived. As expected, all tested time
intervals, when averaged across each year with an initial investment of $100,000, were far less
profitable than results from 2013-2014. However the 60-day test stood out as the only interval that was
profitable on average.

16
 Table 4 - 5 year Average Revenue per Trading Interval

The results above indicate that the portfolios are susceptible to negative market conditions,
resulting in a major loss of profit during the recession of 2008. The downturn and reversal of value can
be seen in Figure 7, where each clustered portfolio’s value is plotted linearly from 2008 to 2012. For
comparison, the S&P 500 index’s performance for the same time period is included in Figure 8 to show
the correlation to the health of the market. As the market rebounded, the net worth of the clustered
portfolios increased as well.

Figure 7 - Net Profit Over Time per Test

Figure 8 - S&P 500 Index Chart from 4/4/2008 to 11/2/2012

17
 Market conditions also played a major role in the optimal strategy chosen for each cluster,
which was expected due to the diversity of conditions favored by each strategy. Figure 9, showing the
distribution of strategy selection for each test clustered yearly from 2007 to 2011, illustrates the
reduction of the diversity of strategies chosen surrounding the recession. In particular, tests carried out
in the years 2008 and 2009 indicated that seven of the ten clusters performed best when using the
Trend Down strategy, which would falter as the market subsequently rebounded. More detailed
information regarding specific clusters and their associated strategies can be found in Appendix D.

Optimal Strategies per Year

8

7
# Clusters Using Strategy

5 Tend Upwards
4 Trend Downwards

3 Trend Both
Directionless
2
Volatility
1

0
2007 2008 2009 2010 2011
Test Year

Figure 9 - Optimal Strategies per Year

18
4.2 Comparison to S&P 500
Test 1, whose results are described in Table 3, represents four portfolios with different re-clustering
periods traded on a contiguous 120-day range. Figure 10 portrays the gains of the S&P 500 SPY index
and the four clustered portfolios over the 120-day range of October 14, 2013 to March 28, 2014. All
clustered portfolios performed above the market, with the 60-day portfolio showing 21.68% gains over
the 9.08% change in the SPY index. For this “Bull market” period, all clustered time frames beat the
market, with the 60-day portfolio performing best.

Figure 10 - S&P and Clustered Portfolio gains for 10/14/2013 to 3/28/2014

Informed by the promising initial results of the 60-day portfolio, additional analysis was
performed on the data from Table 4 (test 2) to compare against S&P 500 performance during the same
intervals that the clustered portfolio was traded upon. Figure 11 displays the outcome of this analysis,
with the 60-day portfolio showing 15% total gains when the SPY index dropped 1.1% over equivalent
date ranges, lending more support to the efficacy of the 60-day trading interval.

Figure 11 - Cumulative S&P 500 and 60-day portfolio gains over 5 years

19
5 Conclusions and Future Work
Having developed a library of Java, Matlab, Tradestation Easy Language, and Python scripts over the
course of this project, recommendations for effective implementation and future academic research
must be defined. The above results indicate that although promising in some conditions, room exists for
improvement with regards to time series clustering of stock value data with the K-Medoids algorithm.
Although statistical significance of both tests conducted cannot be established due to the limited
number of tests, these preliminary results indicate that a 60-day trading period may be the most
consistently profitable trading window for clusters generated on 260 days of data, having effectively
out-performed the S&P 500 in both tests. Further testing of these results is recommended.
Furthermore, the use of more robust trading strategies with the inclusion of stop losses and exit
strategies can minimize loss seen in periods of downturn in the market, making the proposed system
more practical for implementation.

5.1 Need for Improved Data Input Method

Due to the time consuming nature of manual data entry to input clustering results from Matlab into
TradeStation, the volume of tests possible to complete in the time available for this project was limited.
Testing to find the optimal strategy in TradeStation, and portfolio backtesting in Portfolio Maestro
require the manual creation of clusters and strategy application due to a lack of automated interface for
rendering the .csv or .txt files of the clusters found by Matlab into portfolios that can be acted upon in
these proprietary programs. Thus, testing for the optimal strategy of each of the five options mentioned
in Section 3.7 on ten cluster prototypes requires fifty separate strategy performance reports to be
generated and then analyzed by a Python script to score the strategies according to the weightings seen
in Section 3.8.1. Although initially, the intention was to perform the aforementioned steps over thirty
different clustering results, adjacent in time, the process proved too time consuming, thus resorting to
an abbreviated test over five different re-clustering periods. The results were not tested for statistical
significance due to an insufficient number of samples, and can not be used to make anything more than
a cursory deduction and support the need for future testing. Possible future work on the topic should
include work on automation of this process to expand the application abilities of Matlab, and improve
the ease of performing advanced calculation in TradeStation.

5.2 Improvement of Trading Strategies

The five automated trading strategies (upward trending, downward trending, upward & downward
trending, directionless, and volatility) applied are extremely basic, meaning that very few conditions
need to be met in order to make a trade. In particular, the strategies are defined in such a way that they
are always looking to make trades, employing a single exit condition per strategy. These are criteria that,
when met, signal the automated system to terminate any positions held. More robust (and potentially
cautious) trading systems would employ multiple exits per strategy, to detect a variety of unfavorable
market conditions. The most basic of exits that could be employed to greatly benefit this system’s
behavior is that of a stop loss, which terminates the position held on a stock when a drawdown of

20
equity, defined by a dollar amount or percentage, is detected. For example, if 100 stocks are purchased
at a total of $1000 with a 10% stop loss, the stocks would be sold if the value of the 100 stocks drops
below $900. Such an improvement would improve the resilience of the trading system in conditions
such as those seen in the recession of 2008.

21
6 Bibliography
Investopedia, LLC. (n.d.). Backtesting. Retrieved from Investopedia.com:
http://www.investopedia.com/terms/b/backtesting.asp

Investopedia, LLC. (n.d.). Equity Curve. Retrieved April 3, 2015, from Investopedia.com:
http://www.investopedia.com/terms/e/equity-curve.asp

Investopedia, LLC. (n.d.). Law of Supply and Demand. Retrieved April 1, 2015, from Investopedia.com:
http://www.investopedia.com/terms/l/law-of-supply-demand.asp

Investopedia, LLC. (n.d.). Public Company. Retrieved April 1, 2015, from Investopedia.com:
http://www.investopedia.com/terms/p/publiccompany.asp

Investopedia, LLC. (n.d.). Slippage. Retrieved April 1, 2015, from Investopedia.com:

http://www.investopedia.com/terms/s/slippage.asp

Keogh, E., & Lin, J. (2005). Clustering of Time-Series Subsequences is Meaningless: Implications for
Previous and Future Research. Knowledge and Information Systems, 154-177.

Milton, A. (n.d.). Long and Short. Retrieved April 3, 2015, from About.com:
http://daytrading.about.com/od/daytradingglossary/g/FuturesLongShor.htm

Phung, A. (n.d.). Behavioral Finance: Key Concepts - Gambler's Fallacy. Retrieved April 1, 2015, from
Investopedia.com:
http://www.investopedia.com/university/behavioral_finance/behavioral7.asp

Rokach, L., & Maimon, O. (2010). Clustering Methods. In Data Mining and Knowledge Discovery
Handbook (p. 326). New York City: Springer.

Tan, P.-N., Steinbach, M., & Kumar, V. (2005). Introduction to Data Mining. New York City: Pearson.

The MathWorks, Inc. (n.d.). Detrending Data. Retrieved from Mathworks.com:

http://www.mathworks.com/help/matlab/data_analysis/detrending-data.html

The MathWorks, Inc. (n.d.). kmedoids. Retrieved from Mathworks.com:

http://www.mathworks.com/help/stats/kmedoids.html

Theodoridis, S., & Koutroumbas, K. (2006). Pattern Recognition 3rd ed. Waltham: Academic Press.

U.S. Census Bureau. (2012). Statistical Abstract of the United States. 750.

Veysov, A. (2012). Financial System Classification: From Conventional Dichotomy to a More Modern
View. Munich Personal RePEc Archive.

22
Wikimedia Foundation, Inc. (2015, March 27). Euclidean Distance. Retrieved from Wikipedia.com:
http://en.wikipedia.org/wiki/Euclidean_distance

Wikimedia Foundation, Inc. (2015, February 17). Feature Scaling. Retrieved from Wikipedia.com:
http://en.wikipedia.org/wiki/Feature_scaling

Wikimedia Foundation, Inc. (2015, February 24). Standard Score. Retrieved from Wikipedia.com:
http://en.wikipedia.org/wiki/Standard_score

Wright, C. F. (1998). Trading as a Business. Charlie F. Wright.

23
Appendix A: Single Point Data
The single point variables available for acquisition from the Yahoo! API are listed below. The only
variable used in this project was Average Volume, used during the filtering stage of the process.

 Sector (nominal)
 Market Capitalization (numerical)
 Average Volume (numerical)
 200 Day Moving Average (numerical)
 50 Day Moving Average (numerical)
 Change from 200 Day Moving Average (numerical)
 Change from 50 Day Moving Average (numerical)
 Earnings Per Share (numerical)
 Estimate EPS Next Quarter (numerical)
 Estimate EPS Next Year (numerical)
 Next Year EPS Price Estimate (numerical)
 Current Year EPS Price Estimate (numerical)
 Year High (numerical)
 Year Low (numerical)
 PEG Ratio (numerical)
 PE Ratio (numerical)
 Short Ratio (numerical)

24
Appendix B: Platform Information
Matlab Version 2014b
Tradestation Version 9.1
Portfolio Maestro Version 3.0

25
Appendix C: Detailed Process Breakdown
This appendix references files in the zip file located on the github repository here:

https://github.com/tyler-stone/financial-data-mining

The entire process described in the project is as follows:

Data Acquisition

1. Run the TWNN.jar file under the data_acquisition folder and type ‘1’ and press enter. Then,
import a CSV file with the list of stocks of interest. The default list used for this project is titled
stock_list.csv within the same folder.
2. After, re-run TWNN.jar. This time, choose to retrieve either real-time or historical data. This
project only used historical data, but real-time data (or single-point data, as it is referenced
within this project) is available to retrieve as well. Choose ‘4’ and press enter, and the
application will connect to the Yahoo! API and retrieve all historical data for each stock imported
in the database.
3. Re-run TWNN.jar once more, but this time to export either real-time or historical data. Choosing
‘5,’ a CSV file will appear in C:\ with historical data for each stock up to 2016 days. Note: this
number is equivalent to approximately 5 years of trading data. The data is a constant that can
be modified in the source code of the Java project.

Data Pre-Processing and Clustering

4. To load the CSV files generated above into Matlab, include the single point and time series CSV
files in the Matlab directory, and call the loadStocks() function, with the following parameters:

Inputs:

timeSeriesFilename – [str] name of CSV of raw time series price data for N stocks

singlePointFilename – [str] name of CSV file of single point data for N stocks

Duration – [int] number of days (up to 2160) to include in ‘ts’ file output

Outputs:

ts – array of raw time series data for N stocks over duration

sp – array of single point data for N stocks

tickers – array of N stock tickers (symbols, eg: AAPL) corresponding to time series in ts

26
 Example:

>> [ts,sp,tickers] =
loadStocks(‘timeSeriesFilename',’singlePointFilename’, Duration);

5. To pre-process time series data, use the prepWindow() function with the following parameters:

Inputs:

Series – [int array] array of N time series to be preprocessed ( ts from above)

Duration – [int] number of days to preprocess over (260 for all tests)

Offset – [int] number of days between start date of ts and the 260 day period to be
preprocessed

Outputs:

tsPrep – array of 260 days of normalized time series values for N stocks

Example:

>> [tsPrep] = prepWindow(ts, 260, offset);

6. To evaluate the number of clusters to use (if 10 is no longer valid), use the findOptimalK()
function:

Inputs: tsPrep – N x 260 pre-processed time series

Outputs: K – Optimal number of clusters found at elbow of SSE plot

>> [K] = findOptimalK(tsPrep)

7. To create the weighted distance matrix for use in clustering, use the weighted_multi_distance()
function with the following parameters:

Inputs:

sp – [array] single point data (from step 4, above)

tsPrep – [array] pre-processed 260-day time series for all stocks (from step 5)

w1 – [float] weight applied to sp in calculating distMat (Equal to 0 in all tests)

w2 – [float] weight applied to tsPrep in calculating distMat (Equal to 1 in all tests)

Outputs:

distMat - N x N distance matrix

27
 Example:

>> [distMat] = weighted_multi_distance(sp, 0, tsPrep, 1);

8. To generate clusters based on the distance matrix, call the weightedClustPlot() function with the
following parameters:

Inputs:

distMat – [array] N x N distance matrix (from step 6, above)

tsPrep - [array] pre-processed 260-day time series for all stocks (from step 5)

numClust – [int] number of clusters, K, to generate (Equal to 10 for all tests)

tickers – [cell array] list of tickers (from step 4)

Outputs:

cVect – N x 1 array indicating what cluster each ticker is associated with

midx – 10 x 1 array of indices of cluster centroids

- A file nearest.txt is generated in the directory, listing the stock prototypes for each cluster
across the top row, with each column representing a cluster, as sorted low-high by distance
to the prototype. This file is what is used as a reference for subsequent tests in Tradestation
and Portfolio Maestro.

- A figure displaying all stocks on a single plot, and a plot of each cluster is generated for
visual reference.

Example:

>> [cVect,midx] = weightedClustPlot(distMat, tsPrep, 10, tickers);

Optimal Strategy Identification

9. Using Tradestation, set the data begin and end date to the same as the days chosen for
clustering. Because trading days do not correspond with the calendar, finding the date for data
can be difficult. This problem was solved by modifying the Java project (TWNN.jar) source code
and re-exporting the date of data point for each symbol rather than the value.
10. For each cluster prototype, backtest each strategy and save the strategy performance report as
an excel file with the following name structure: x_straty.xlsx, where x stands for the cluster
number (2-n clusters) and y stands for the strategy number (2-n strategies). In order to identify
the optimal strategy, the Python script looks for files that start with the cluster number and then
compares it to all other files with the same cluster number.

28
 11. Repeat steps 9-10 for each clustered sample of data. If there is only one set of data that was
clustered, there is only one sample.
12. Copy the folder containing the files to the location of the Python script
(identify_optimal_strategy) and separate each clustered sample into separate folders (i.e.
sample0, sample1). If there is only one sample, place all files into a folder labeled sample0
within the same directory structure as the Python script
13. Run the Python script using the command ‘python main.py > results.txt’
14. Results.txt should contain the optimal strategy for each cluster.

Trading with Optimal Strategy

15. Using Portfolio Maestro, create a new portfolio and create new strategy groups for each
strategy used.
16. For each cluster, create a custom symbol list by clicking ‘Add Symbol List’ in the strategy group
toolbar, and then clicking ‘Add/Remove Custom Symbol Lists’ in the popup window.
17. Click ‘Create’ and label the list according to the cluster number. Using the results from step 6,
add the first 20 stocks in the column corresponding to the cluster.
18. Repeat step 17 for each cluster.
19. For each strategy group add the symbol lists corresponding to the clusters that performed best
under this strategy, using data from step 14.
20. Add each strategy group to the portfolio.
21. Click ‘Backtest Portfolio’ and choose a begin date that starts on the same day that the clustering
period ended, to the desired end date.
22. Backtest the portfolio and view the results of the backtest.

29
Appendix D: Tested Portfolios 2008-2012
Sample 0
Cluster Begin Cluster End/Test Begin Test Test End Net Profit
5-day 4/4/2008 -$8,162.00
10-day 4/11/2008 $2,858.00
20-day 4/25/2008 -$3,638.00
3/19/2007 3/28/2008
40-day 5/23/2008 $12,801.00
60-day 6/23/2008 $17,847.00
120-day 9/17/2008 -$41,996.00

Cluster # Optimal Strategy Cluster Prototype Portfolio

KLAC XL MTZ
DHT SNDK NAT
ONNN IDTI HSBC
0 Trend Downward LRCX BA URI DOV
ALTR JBL LLTC
RVBD LUV LRCX
SWHC ZQK
GOOGL CY USU
SPWR CME JNS
AAPL NBG AET
1 Directionless GOOGL SCHW VOD MRK
TROW BEAV SAN
MSFT DHR JASO
DECK AON
PNK BYD DRH
SNV SHO S
MSI SLXP AXP
2 Trend Downward PNK HIG CTL DTE
ATML AEG FCS
DISH BEN GPK
MRVL ROVI
FCX BHP BBL
SCCO FLR MT
KBR DRYS ABB
3 Directionless FCX WMB LUK XOM
AME SBS PX
HAL OII CNQ
NOK BRFS
4 Trend Downward SCG SCG XEL WEC

30
 DUK OGE WR
ED TE GXP
CNP AEP CMS
POM SNY AES
SNE GENE ESRX
MDCO UNH
MOS MON BVN
BMRN ANR HES
CNX GG APA
5 Trend Upward MOS RRC SGY KGC
DAR AUXL ATVI
FLS ILMN ABX
OI AEM
CAR COF ETFC
C TRW FDO
HBAN ALU JAH
6 Trend Downward CAR ASNA LB BCS
FITB RAD GCI
TWX JBLU OC
BKD STI
NNN O PLD
WFM DLR NKE
MPEL HCP
7 Trend Downward NNN NYCB ALB
FNFG ANF
XOMA HCP
CERS HCN
HOV KBH PHM
MAT UDR LEN
TOL SPF NRF
8 Trend Downward HOV PPHM SEE RYL
CUBE AIV SNSS
BBBY DHI FL
NKTR SIMG
LNG SPR IR
ARMH ROK STZ
GPOR BWA BHI
9 Trend Downward LNG PWR ITW HLX
LINTA QTM RY
GNTX GE A
GES DE

31
 Sample 1
Cluster Begin Cluster End/Test Begin Test Test End Net Profit
5-day 4/16/2009 -$13,037.00
10-day 4/23/2009 -$12,897.00
3/28/2008 4/8/2009 20-day 5/7/2009 -$30,026.00
40-day 6/5/2009 -$38,061.00
60-day 7/6/2009 -$14,403.00
120-day 9/29/2009 -$39,659.00

Cluster # Optimal Strategy Cluster Prototype Portfolio

ARO KND QCOM
ARMH TJX STJ
SWKS ASNA RVBD
0 Trend Downward ARO SYMC BIG CRIS
PMCS NTRS AUQ
CONN ROST KSS
AMZN DEPO
CCI QQQ AMT
WDR PX BBRY
DISH MRO INFY
1 Trend Downward CCI CRM ADI AMAT
RGP INTC IBM
MXIM IVZ JWN
LLTC AAPL
OCN FDO AMGN
SLXP RCPI GERN
CTRX APOL JBLU
2 Trend Upward OCN COCO NFLX
GMCR ALK
SQNM DLTR
IMGN PBCT
RRD PLD CAT
CBL PGH OMC
SCI ERF MAC
3 Trend Downward RRD URI AIV MFC
LUK MPW LNC
EMN NI ING
AXP WRES
RL AMTD SAP
4 Trend Downward RL LB URBN ECL
SPLS WU SCHW

32
 CVC AVP HAS
FTR GPS JAH
SYY FAST PLCM
LMT VFC
SPN JOY TS
OIS HK BTU
HAL NOV ESV
5 Trend Upward SPN PTEN NBR HLX
BHI ACI KEG
X MT RDC
CAM HP
TC TER CBG
CENX AKS CRZO
GGP JASO CBI
6 Trend Downward TC ANF TEX TXT
RIO VALE PH
FBR SUNE FCX
GNW EBAY
AINV LF COF
UDR WNC FDX
HSBC MAS DUK
7 Trend Downward AINV STM CMCSK BAC
EQR NWL AFL
IP UNM HCN
EL SYK
LEN AEO AXL
RYL MTG ICON
KBH SNDK HOV
8 Trend Downward LEN MAR HDB PENN
DKS AGO CME
CCE CMLS ACAD
DHI KMX
LNG FHN
SPPI WNR
UAL FNF
Trend Upward &
9 LNG MUX IAG
Trend Downward
PPHM MNKD
NKTR
MNST

33
 Sample 2
Cluster Begin Cluster End/Test Begin Test Test End Net Profit
5-day 4/28/2010 -$166.00
10-day 5/5/2010 -$7,825.00
4/8/2009 4/21/2010 20-day 5/19/2010 -$14,520.00
40-day 6/17/2010 -$6,213.00
60-day 7/16/2010 -$11,151.00
120-day 10/11/2010 $13,626.00

Cluster # Optimal Strategy Cluster Prototype Portfolio

DLR BBBY BC
JWN ROK CTRX
ETN ALK FDX
0 Trend Upward DLR BMR SPG CCL
VTR CRM IACI
ALTR PLD RMD
BCE CNI
CY TC BWA
VSH STLD ALKS
FCS CYTR MTG
1 Trend Upward CY EXAS ONNN AMGN
AGCO CAB LVLT
AKS FISV NCR
MDLZ RVBD
HEB FCEL IRM
AGEN FREE FLR
ASTI CTIC BSX
2 Volatility HEB ARNA WFT ATVI
SMFG MTU AEZS
XOMA SUNE NEE
GERN ISIS
BEN SCCO OXY
SGY L CSX
ITW FCX CAT
3 Directionless BEN FLEX XL RY
STZ FLS IP
IR LEG RL
URBN TYC
EXK SDRL HK
4 Trend Upward EXK HL ABEV XCO
SLW HAL IAG

34
 AG CEF AXAS
EGO COST QTM
ZMH ABT KO
GOOGL CVC
OKE XEC MWE
OGE CTSH TE
ENB BTE ROSE
5 Directionless OKE TCK HSP WLT
DTE FOSL CMS
INFY QQQ WLL
ARCC MPW
HOT FOXA AINV
AIV BEAV PH
LB HST GNW
6 Directionless HOT RCL UDR TRP
ALB MAC HON
FOX PCP DIS
EQR SHO
UAL SNDK HD
HAS GGP CLF
PCYC AAL DDD
7 Volatility UAL NYCB CREE SNE
NOC FAST NKTR
LUV ASNA EL
BIG JDSU
PHM SGEN COCO
KBH CVS PBCT
TOL JBLU
8 Directionless PHM RYL NTRS
DHI JEC
WEN MDCO
AON BYD
LNG TCB APP
RCPI SNV SQNM
DNR WNR GENE
9 Trend Downward LNG EAT BAS BK
VLO MBI FNF
SUSQ NBIX NOK
ABC CLDX

35
 Sample 3
Cluster Begin Cluster End/Test Begin Test Test End Net Profit
5-day 5/9/2011 $3,498.00
10-day 5/16/2011 -$3,079.00
4/21/2010 5/2/2011 20-day 5/31/2011 $582.00
40-day 6/28/2011 -$10,656.00
60-day 7/27/2011 $2,074.00
120-day 10/20/2011 -$4,443.00

Cluster # Optimal Strategy Cluster Prototype Portfolio

CAT HAL DD
JOY DE FOSL
CBI ALTR CRZO
0 Trend Upward CAT KBR OIS ACAS
ALXN NVO EMN
PAY BWA CHKP
BTE TRMB
HES QQQ TROW
SLB AA FLR
VSH ERF PAL
1 Trend Upward HES NOV DOW PH
TXN APA PLD
JCI MXIM ADI
HON DAR
NKTR ETR PCYC
TEF MRK BBY
GSS WEC HALO
2 Trend Downward NKTR ADVS AVP MNKD
ABEV CSCO HCBK
LGF LF AGO
CPB RAI
STJ WTW MTZ
PES PDS BAS
CBS SSYS TD
3 Trend Upward STJ MTW SPPI STO
CNC GPOR NBR
SGY CERN ACHN
TRN GLNG
SPLS CME TGT
4 Volatility SPLS MSFT CMA INTC
AMKR ELX CYTR

36
 MGM CSC DRYS
NTRS TASR MTG
ZQK MRVL ISIS
HA PMCS
HOT AMZN KMX
EMR KIM BMRN
HT UPS ARCC
5 Trend Upward HOT VALE NSR EXPD
RCL HST DDR
PCP DRH AIV
CA BEAV
PVA UMPQ ARWR
GNW BIOS CONN
WY APP ARO
6 Trend Downward PVA CLRX VMC FRO
CX BAC FHN
PBCT HPQ VLY
COCO HOV
SCCO WIN BRCM
CTL NTAP FTR
TV CCJ SIMG
7 Volatility SCCO PCAR F BAA
ROVI AMX MTOR
GOOGL GS RHT
LVS FFIV
USG CNQ SUSQ
MAS TOL MWA
KBH PPHM RAD
8 Trend Upward USG FCEL RTN SPF
DHI BBT GPS
ZION ODP RRD
PPC NYMT
VLO OKE STM
DVN TEX DIS
TER GPK TTWO
9 Volatility VLO AON BHI KND
WMB WNR SM
LNG HFC WFM
TSO EXEL

37
 Sample 4
Cluster Begin Cluster End/Test Begin Test Test End Net Profit
5-day 5/18/2012 $9,871.00
10-day 5/25/2012 $5,314.00
5/2/2011 5/11/2012 20-day 6/11/2012 $12,064.00
40-day 7/10/2012 $25,408.00
60-day 8/7/2012 $20,666.00
120-day 11/2/2012 $16,478.00

Cluster # Optimal Strategy Cluster Prototype Portfolio

ALTR JNPR F
RCL ABB AUO
SID VSH STM
Trend Upward &
0 ALTR SEE RVBD MS
Trend Downward
FCS JDSU HES
LUV XRX ING
ONNN GNW
IAG DNDN WTSL
KGC EXC GFI
EA DECK ABX
1 Trend Downward IAG BBG GG PEG
HMY AU COG
END EGO GMCR
DLIA GST
AA ECA PVA
GERN TSL S
TLM ALU FBR
2 Trend Downward AA JASO WLT MT
NAT MTOR CSIQ
FNFG ERIC BHP
AINV CENX
FTI NUAN GOOGL
SPPI NBL AEE
OKE OGE CAM
Trend Upward &
3 FTI GLNG MBI BMRN
Trend Downward
HUM HCP MON
HALO SBGI ACHN
MCD DUK
SPLS
4 Directionless SPLS HLF
DDD

38
 ENB
OII
FNSR
YHOO
BTU FSLR XCO
ACI PPHM UPL
ANR CNX SUNE
5 Trend Downward BTU BHI PTEN CTRP
NFX DHT HAL
VG VALE PGH
SPWR FTR
A SYK OI
TEX UTX DOW
MMM MET LXP
6 Volatility A LH LNC BKD
PCAR BCS THC
MTW HTZ ALB
IDTI CBG
STI IR HIG
CMA PFG TRW
BK JNS PMCS
Trend Upward &
7 STI ZION ZMH NKTR
Trend Downward
BAC SNV HSBC
MWA RF CCK
SCHW HCBK
HON SWK WNC
LHO SHO KEY
DHR IVZ DIS
8 Volatility HON ROK SPR TWX
HST OMC HOT
PAYX PH TROW
FLS BA
HD LF FL
YUM LOW CVS
PIR LEN SPG
9 Trend Upward HD M DLR DHI
URI PAA INTC
USB ABT SBUX
GRMN KATE

39
Appendix E: Trading Strategies
The directionless and volatility strategies used in this project were directly from Trading as a Business,
referenced in the Bibliography. However, the Trend Upward and Trend Downward strategies were
custom strategies developed in EasyLanguage for use in Tradestation and Portfolio Maestro.

Trend Upward
inputs:
Price( Close ),
Run( 10 );

variables:
Rise( 0 ),
AvgWtd ( 0 ),
Slope( 0 ),
result( 0 );

begin
AvgWtd = WAverage( Price, Run );
Rise = AvgWtd - AvgWtd[Run];
Slope = Rise / Run;

Condition1 = Slope > Slope[1] and Slope[1] < Slope[2] and Slope <
0;
Condition2 = Slope < Slope[1] and Slope[1] > Slope[2] and Slope >
0;

if MarketPosition = 0 then begin

If Condition1 then buy at next bar market;
//If Condition3 then sellshort next bar at market;
end;

If MarketPosition = 1 and Condition2 then sell next bar at

market;

end;

Trend Downward
inputs:
Price( Close ),
Run( 10 );

variables:
Rise( 0 ),
AvgWtd ( 0 ),
Slope( 0 ),
result( 0 );

40
begin
AvgWtd = WAverage( Price, Run );
Rise = AvgWtd - AvgWtd[Run];
Slope = Rise / Run;

Condition1 = Slope > Slope[1] and Slope[1] < Slope[2] and Slope < 0;
Condition2 = Slope < Slope[1] and Slope[1] > Slope[2] and Slope > 0;

if MarketPosition = 0 then begin

If Condition2 then sell Short at next bar at market;
//If Condition3 then sellshort next bar at market;
end;

If MarketPosition = -1 and Condition1 then Buy to cover next bar at

market;
end;

41