High Frequency Trading System Design

High Frequency Trading System Design and Process Management
By
Xiangguang Xiao
Bachelor of Engineering, Tianjin University (1996)
Master of Science in Computer Science, Boston University (2006)
Submitted to the System Design and Management Program in Partial Fulfillment of the
Requirements for the Degree of
Master of Science in Engineering and Management
At the
Massachusetts Institute of Technology
May 2009
0 2009 Massachusetts Institute of Technology
All rights reserved
ARCHIVES
MASSACHUSETTS INSTITUtTE
OF TECHNOLOGY
SEP 2 3 2009
LIBRARIES
Signature of Author
Xiangguang Xiao
System Design and Management Program
May 2009
Certified by
Prdfssor Roy E. Welsch
Thesis Supervisor
Professor of Management Science, Statistics, and Engineering Systems
Sloan School of Management and Engring Systems DiviiAp
Certified by
Director
System Design and Management Program
High Frequency Trading System Design and Process Management
By
Xiangguang Xiao
Submitted to the System Design and Management Program
on May 22, 2009 in Partial Fulfillment of the Requirements for the Degree of Master of Science
in Engineering and Management
ABSTRACT
Trading firms nowadays are highly reliant on data mining, computer modeling and software
development. Financial analysts perform many similar tasks to those in software and
manufacturing industries. However, the finance industry has not yet fully adopted high-standard
systems engineering frameworks and process management approaches that have been successful
in the software and manufacturing industries. Many of the traditional methodologies for product
design, quality control, systematic innovation, and continuous improvement found in engineering
disciplines can be applied to the finance field.
This thesis shows how the knowledge acquired from engineering disciplines can improve the
design and processes management of high frequency trading systems. High frequency trading
systems are computation-based. These systems are automatic or semi-automatic software
systems that are inherently complex and require a high degree of design precision. The design of
a high frequency trading system links multiple fields, including quantitative finance, system
design and software engineering. In the finance industry, where mathematical theories and
trading models are relatively well researched, the ability to implement these designs in real
trading practices is one of the key elements of an investment firm's competitiveness. The
capability of converting investment ideas into high performance trading systems effectively and
efficiently can give an investment firm a huge competitive advantage.
This thesis provides a detailed study composed of high frequency trading system design, system
modeling and principles, and processes management for system development. Particular
emphasis is given to backtesting and optimization, which are considered the most important parts
in building a trading system. This research builds system engineering models that guide the
development process. It also uses experimental trading systems to verify and validate principles
addressed in this thesis. Finally, this thesis concludes that systems engineering principles and
frameworks can be the key to success for implementing high frequency trading or quantitative
investment systems.
Thesis Supervisor: Professor Roy E. Welsch
Title: Professor of Management Science, Statistics, and Engineering Systems, Sloan School of
Management and Engineering Systems Division, Massachusetts Institute of Technology
Acknowledgments
I would like to thank IBM for giving me the opportunity to study at MIT. I owe a debt of
gratitude to Masha Maule and Steffi Diamond at IBM for their support and confidence in me.
I especially would like to thank Professor Roy E. Welsch for providing guidance and knowledge.
I feel grateful and fortunate to receive his advice and to listen to his wisdom.
The System Design and Management (SDM) faculty and staff have been very helpful throughout
the program. I must acknowledge Pat Hale and Bill Foley for being there for me as always. My
classmates and friends at MIT have made my experience at MIT an enjoyable, memorable and
rewarding one. I learned a great deal from both professors and my fellow classmates.
Finally, I would like to thank my fiancee, Yu Wei for her love, support and advice. Without that
support my academic work and thesis would not have been possible.
Table of Contents
High Frequency Trading System Design and Process Management......................................................1
Acknowledgm ents ....................................................................................................................................... 5
Table of Contents ........................................................................................................................................ 7
Table of Figures...........................................................................................................................................9
1. Introduction......................................................................................................................................11
1.1. W hat is a trading system ?............................................................................................... 12
1.2. W hy are system s im portant?....................................................................................... 14
1.3. H igh frequency financial data ...................................................................................... 15
1.4. Contributions of this thesis......................................................................................... 16
1.4.1. Trading system design fram ew ork and processes................................................... 16
1.4.2. Backtesting and optim ization strategies.............................................................. 16
2. Literature reviews ............................................................................................................................ 17
2.1. D iscretionary versus nondiscretionary system s ......................................................... 17
2.2. W hat is a good trading system ................................................................................... 18
2.3. Sim ple system versus com plex system .......................................................................... 19
2.4. Data sources and data characteristics .......................................................................... 20
3. System design and developm ent process................................................................................... 21
3.1. Design with Model-Based Systems Engineering (MBSE) ........................................ 22
3.1.1. System Com ponents................................................................................................ 24
3.1.2. U sers ....................................................................................................................... 25
3.1.3. Live trading Process............................................................................................. 27
3.1.1. System Developm ent Process overview ................................................................. 28
3.1.2. Research phase: generate ideas............................................................................ 29
3.1.2.1. Research m ethods ........................................................................................ 30
3.1.2.2. Design docum ent ........................................................................................ 31
3.1.2.3. Case studies ................................................................................................. 32
3.1.3. Prototype phase.................................................................................................... 37
3.1.3.1. Gather data..................................................................................................... 39
3.1.3.2. Data type....................................................................................................... 40
3.1.3.3. Data cleaning process ................................................................................... 42
3.1.3.4. Datapartition ............................................................................................... 43
3.1.3.5. Coding in m odeling softw are ....................................................................... 44
3.1.3.6. Prototype quality assurance .......................................................................... 45
3.1.3.7. Backtest and Optim ization.......................................................................... 46
3.1.4. Production........................................................................................................... 46
3.1.4.1. Production program m ing............................................................................ 47
3.1.4.2. Developm ent team and tool............................................................................. 49
3.1.4.3. Production quality assurance ....................................................................... 49
4. System backtest .......................................................................................................... ............ 50
4.1. Testing process...............................................................................................................51
4.1.1. Decide value range................................................................................................ 51
4.1.2. Decide value distribution...................................................................................... 52
4.1.3. Rank param eters.................................................................................................. 53
4.1.4. Decide trading cost ............................................................................................. 53
4.2. In-sam ple and out-of-sam ple test ............................................................................... 54
4.3. Avoid overfitting ......................................................................................................... 55
4.3.1. Save out-of-sample/test data ................................... 55
4 .. . pif r n t m h o i n s . ....................................................................................... 5
4.3.2. Different tim e horizons........................................................................................ 55
4.3.3. Differenttim e fram es........................................................................................... 56
4.3.4. Different Instrum ents .......................................................................................... 56
4.3.5. Different asset classes................................................ ........................................ 56
4.3.6. Student T-test ..................................................................................... ....... 56
4.4. Trading statistical data.......................................................................................... 59
5. Optimization ........... ................................................................................................... 59
5. 1. M ethods of optim ization. ................................................................................................ 60
5.1.1. W hole s aopoei i ai ..................................................................................... ... 60
5.1.2. Out-of-Sam ple analysis......................................................................................... 61
5.1.3. Step forward analysis........................................................................................ 61
6. Systems m easurem ents ...... .............................................................................................. 62
6.1. Robustness of system ...................................................................................... ........ 62
6.2. Profit ne s of s e ................................................................................................... 62
6.2.1. Absolute m etrics . .................................................................................. .... 63
6.2.2. Trade statistics .................................................................................................... 64
6.2.3. Benchm arking with index................................................................................... 66
6.2.4. Correlations with index........................................................................................ 66
6.3. Investm ent capacity.................................................................................................... 67
7. Conclusion.........................................................................................................................................68
Appendix I: Trading System Developm ent Process........................................................................... 70
Appendix II: Trading System Modeling with Object-Process Methodology...................................71
Bibliography..............................................................................................................................................78
8
Table of Figures
Figure 1: Six Rings of a Trading System................................................................................... 13
Figure 2: Trading System Components ..................................................................................... 24
Figure 3: Trading system user cases .......................................................................................... 26
Figure 4: Live trading internal process diagram....................................................................... 27
Figure 5: R esearch phase ........................................................................................................... 31
Figure 6: Short term trend......................................................................................................... 35
Figure 7: Prototype phase ........................................................................................................ 38
Figure 8: Most influential market moving indicators ................................................................ 41
Figure 9: Times of important economic releases ..................................................................... 41
Figure 10: Production Phase ...................................................................................................... 47
Figure 11: Three trading types.................................................................................................. 50
Figure 12: Test time (minutes) for USDCHF on short term trend trading ............................... 52
Figure 13: Different time frames for USDJPY on short term trend trading .............................. 57
Figure 14: Returns of different currency pairs.......................................................................... 58
Figure 15: Returns of different currency pairs after eliminating hard-coded values................. 59
Figure 16: Sample distribution of trading order types .............................................................. 65
Figure 17: Portfolio performance comparing to stock indexes.................................................. 66
Figure 18: Correlations with stock indexes .............................................................................. 67
Figure 19: Market depth for EURUSD...................................................................................... 68
Figure 20: Trading system development framework and process (SysML).............................. 70
Figure 21: Trading system development framework and process (OPM)................................ 71
Figure 22: Trading system development framework and process - Unfolded (OPM)............. 72
Figure 23: Trading diagram - zoomed-in (OPM) .................................................................... 73
Figure 24: Trading strategies developing diagram - Zoomed-in (OPM) ................................. 74
Figure 25: Trading application - Unfolded (OPM).................................................................. 74
Figure 26: Data cleaning - Zoomed-in (OPM).......................................................................... 75
Figure 27: Trading strategies/models developing - Zoomed-in (OPM).................................... 76
Figure 28: Backtesting process - Zoomed-in (OPM)................................................................ 77
Figure 29: Production/Deploy process - Zoomed-in (OPM) .................................................... 77
1. Introduction
The financial market is getting more competitive every day. In the trading sector, a competitive
edge is razor thin. To achieve even a 50 basis point excess return is often extremely difficult.
Trading firms today depend on data mining, computer modeling and software development.
Quantitative analysts in the financial industry conduct many day-to-day tasks similar to those in
the software and manufacturing industries. Although the term "financial engineering" is well
known, there still exists a large gap between financial and engineering disciplines. Quantitative
finance has not yet fully adopted the high quality engineering framework and processes that have
succeeded in the software and manufacturing industries.
Financial firms that do not value and implement software engineering processes, such as version
controls or quality assurance processes, often prove to be ineffective and inefficient in the
production and management of their code bases. In academics, quantitative finance classes prefer
to teach students mathematical theories and financial modeling. Also, in engineering school,
system architecture and product development classes focus on physical products such as a robot
design or a new mobile application. This thesis demonstrates that many ideas of product design,
quality control, systematic innovation and continuous improvement processes used in systems
engineering disciplines can be applied to improving the business of trading firms.
This thesis shows how the processes employed by engineering disciplines can improve the
design of high frequency trading. It aims to integrate methods and concepts from various fields,
including quantitative finance, system science, software engineering and data mining. Using
cases of high frequency trading system design and foreign exchange markets, the findings in this
paper can also be applied to non-high frequency trading system design and other financial market
segments.
This paper demonstrates how the guiding system design principles and process management,
especially backtesting processes, are crucial elements in the development of a successful
financial trading system.
1.1. What is a trading system?
Quoting from Wikipedia, "A system is a set of interacting or interdependent entities, real or
abstract, forming an integrated whole." The concept of an "integrated whole" can also be
expressed in terms of a system embodying a set of relationships that are differentiated from
relationships of the set to other elements, and from relationships between an element of the set
and elements that are not part of the relational regime. Sharing the same common characteristics,
most systems:
" Are abstractions of reality
* Have a structure that is defined by its parts and their composition
* Have behavior, which involves inputs, processing and outputs of material, information or
energy
* Have parts with functional as well as structural relationships between each other
* May also refer to a set of rules that govern behavior or structure
A financial trading system is generally considered to be a set of trading rules that control when
and at which price you open and close the trade. "A system is simply a plan or set of rules of
when to buy and sell securities". (Charles D. Kirkpatrick, Julier R. Dahlquist. Technical
Analysis) An example of a rule would be "buy when a moving average crosses above another".
Variables are the quantities used in the rules - two moving averages, and the parameters are the
actual values used in the variables - 5 minutes and 15 minutes. "A system lets us determine a
priori how it will react to particular market situations." In an ever-challenging financial market, a
trading system is more complex than just a set of trading rules. The system must also include the
means of controlling the risks of losing capital, mathematical models, historic data, and
sometimes specialized software. A system is an "integrated whole" that consists of key elements,
relationship between internal elements and relationships with external elements. A U.S. military
strategist, John A. Warden III, introduced the Five Ring System model in The Air
Campaign contending that any complex system could be broken down into five concentric rings.
Each ring - Leadership, Processes, Infrastructure, Population and Action Units - could be used to
isolate key elements of any system that needed change. The model was used effectively by Air
Force planners in the First Gulf War. Using Warden's book as a model, I will break down a
trading system into the following levels:
. Investment philosophy
. Trading strategies and mathematical models
. Risk management and money Management
. Disciplined processes
* Backtest and Optimization
* Financial data and trading software
Consider each level of system or "ring" as one of the trading system's centers of gravity. The
principal idea behind the six rings is to develop a trading system that earns a desired profit in a
targeted time frame. To accomplish this goal, we need to address each of the rings individually
and collectively to build a robust and profitable trading system. The system developer needs to
engage as many rings as possible with more emphasis on the inner rings.
Figure 1: Six Rings of a Trading System
1.2. Why are systems important?
Among the various reasons that traders employ trading systems, the ability to backtest for
hypothetical performance using historical data is a key reason. Quantitative funds particularly
like the fact that they can quickly validate their new trading ideas. Although there can be a
significant difference between hypothetical and actual results, the backtest result can give the
financial engineers some idea how their trading systems would work in reality. If the trading
system is properly built, financial engineers can leverage the optimization and feedback process
to refine their investment ideas, and eventually improve the trading systems.
Another important reason that traders utilize trading systems is that it helps to reduce the
emotions involved in making difficult trading decisions. The human emotions of greed and fear
sometimes prevent human beings from making timely decisions in the financial market. Trading
systems provide a disciplined way to overcome mistakes that can be caused by greed and fear.
With a well-designed trading system, you can expect consistency in trading execution and
repeatable trading results.
Arguably, a trading system is more important than profit models. Many traders believe that
simple and well known trading models can work with successful implementations - good risk
control and emotion control. Conceivably, a trading system that implements a naive model could
make profits if it includes a solid risk management strategy and reacts to the market in a timely
manner.
For high frequency trading firms that depend on the "instant" analysis of data ticks, the use of a
trading system is the only way to implement successful investment strategies. Using a trading
system fully engages the computing power of IT systems, saving time and reducing human
mistakes. Many high frequency trading firms hold the philosophy of "do not predict the market,
react to the market". Realizing this philosophy requires a mechanical trading system that
captures short time market movements and makes corresponding decisions rapidly.
1.3. High frequency financial data
The original form of market data is tick data which shows every price change and volume.
Finance markets generate enormous tick data every trading day. "The number of observations in
one single day of a liquid market is equivalent to the number of daily data within 30 years."
(Michel M. Dacorogna, Ramazan Gengay, Ulrich MUller, Richard B. Olsen, Olivier V. Pictet.
2001) With advances in information technology, large amounts of high frequency financial data
including intraday tick data have become more available. We can now analyze financial time
series and markets at almost any desired frequency. Many traders make their trading decisions
based upon high-frequency data such as one-minute bar' data or even tick-by-tick data. "High
frequency data can provide a desirable experimental bench for practitioners to understand market
microstructure and for analyzing financial markets." (Dacorogna et al., 2001)
Processing large amounts of data offers both opportunities and challenges for a trading system
designer. Statistically, larger amounts of available data result in higher degrees of freedom and
bring a new level of significance.2 From the backtest perspective, more sample trades are
generated in the backtest using shorter time scales with the same models. However, challenges
also come with the large data size. High frequency data is noisy which can cause problems for
modeling as the model may reflect the noise as opposed to the underlying market dynamics.
Dealing with large amounts of high frequency data requires the right mathematical tools, models
and techniques, which the system designer has to consider when developing their high frequency
trading system and backtesting strategies.
The foreign exchange (FX) market is the largest and most liquid financial market in the world. It
is attractive for high frequency trading because of its high liquidity, small spreads and low
I A bar is a graphical representation of a financial instrument's movement that usually contains the open, high, low and closing
prices for a set period of time.
2 A thorough discussion of high-frequency data can be found in Michel M. Dacorogna, Ramazan Gengay, Ulrich Miller, Richard
B. Olsen, Olivier V. Pictet. An introduction to High-Frequency Finance. San Diego: Academic Pres. 2001
transaction cost. Since 2008, I have worked with a few other students on building high frequency
trading systems on the foreign exchange. This thesis will use the FX market in most studies to
leverage the research we have done and the lessons we have learned.
1.4. Contributions of this thesis
1.4.1. Trading system design framework and processes
In the financial market, extensive research is conducted to develop sophisticated mathematical
trading models. Equally important, if not more, is the framework of implementation - the design
of the production system surrounding the existing financial research/models. The design and
processes is the linkage of best practices in engineering disciplines and quantitative finance. In
today's trading world, "the key determinant of sustainable competitive advantage is the ability to
continually discover, build, and operate better trading systems." (Andrew Kumiega and
Benjamin Van Vliet, 2008) Converting an investment idea into mathematical models and then a
working trading system in a fast and high quality manner is critical for one to compete in the
market place. This thesis identifies the guiding principles and processes for a high frequency
trading system design with emphasis on the quality of design and the robustness of the system.
1.4.2. Backtesting and optimization strategies
Among all the components, backtesting and optimization without overfitting are the most
important parts in building a trading system. Often misunderstood and misused, backtest and
optimization with right strategies to avoid overfitting in the processes are the keys to building a
successful trading system. Many trading systems backed by great investment ideas are
abandoned during the backtesting phase as they were not tested properly. Other strategies/models
with good backtesting performance failed miserably in live trading and caused a huge capital loss
because they were improperly optimized. Many financial engineers argue that backtesting and
optimization are more important than the mathematical and financial theories behind the trading
strategies when building high frequency trading systems. You can build a good trading system
with a very simple and straightforward model if backtested well and optimized to control risks
properly. With inappropriate testing and optimization, you can convert even a good trading
model into a capital disaster. This thesis discusses effective testing and optimization methods to
avoid over-fitting with backtesting and optimization.
2. Literature reviews
2.1. Discretionary versus nondiscretionary systems
Trading systems can be discretionary and non-discretionary (mechanical), sometimes both.
Discretionary trading systems require traders to determine entries and exits, by intuition, or by
their own judgment of the importance of technical or fundamental signals that they receive. The
number of signals is potentially unlimited. So in a discretionary trading system, traders exercise
some discretion in making trades. Nondiscretionary trading systems operate mechanically.
Entries and exists are determined mechanically by the system as opposed to traders. The trading
decisions are based upon a fixed number of pre-defined technical or fundamental signals without
the participation of the trader. Discretionary trading systems require the knowledge and
experience from individual traders who must constantly apply their creativity under changing
market conditions. A non-discretionary trading system, however, requires creativity from the
trader and system designer only in the system design and development phase.
Discretionary trading systems are best used by highly experienced traders with an abundance of
practical market knowledge that lets them determine the validity and true meaning of market
signals. These traders typically have intuition and have internalized a large number of different
historical patterns that they can compare to the current market conditions. They generally have to
maintain their mental power and self-discipline to perform consistently under different situations.
In essence, discretionary traders develop their brains as a natural trading system.
Nondiscretionary traders, however, often take highly systematic approaches. They build the
seasoned market insights and knowledge of traders into a working strategy in a software program
format. A fully nondiscretionary trading system is one that runs on its own. This type of trading
system reads the live market data which is continually fed in by brokers and makes trading
decisions automatically based on the pre-determined rules. Trading system designers tend to
have an engineering background - good at modeling and programming and familiar with
statistics and systems. These designers study the market, read research papers and build models
and test their trading ideas using statistical methods. In many firms, financial engineers work
with skillful programmers on system design and development and they leverage each other's
strengths. The developed trading system usually has been fully backtested on both in-sample and
out-of-sample data before it is put into live trading. Starting a mechanical trading system in
production without full backtesting is considered to be highly risky. Having a mechanical trading
system in place that adheres to a well built plan can prevent haphazard emotional trading. A
nondiscretionary trading system enforces discipline. Without strict adherence to a proven
nondiscretionary trading system, human emotions can enter into the trading decisions and cause
unquantifiable errors. The comparison between discretionary and nondiscretionary systems was
performed by Kirkpatrick and Dahlquist in their book Technical Analysis (FT press, 2008) and
summarized in this section. Authors in this book fully discussed the benefits of nondiscretionary
systems. "Researchers have showed that the majority of successful traders and investors use
nondiscretionary systems" (Kirkpatrick and Dahlquist, 2008). "A mechanical approach to the
markets can be successful and this is backed up by the fact that approximately 80% of the $30
billion in the managed futures industry is traded by exact systematic methods" (John R. Hill,
George Pruitt, and Lundy Hill. The Ultimate Trading Guide, 2000).
2.2. What is a good trading system
There are common characteristics of good trading systems. In Beyond Technical Analysis (Wiley
2001), Tushar Chande, the author, discusses "Six Cardinal" Rules in trading system design.
Kirkpatrick and Dahlquist, in their book Technical Analysis (FT press, 2008), consider these
rules as the characteristics of a good trading system:
" Positive expectation
* Small number of trading rules - ten rules or less
* Robust parameter values, usable over many different time periods and markets. Being
able to trade on different markets is a great indication of trading system robustness.
* Able to trade multiple contracts
* Use risk control, money management, and portfolio design
* Fully mechanical.
These characteristics are regarded as common criteria of a good trading system. However, most
trading systems are not evaluated individually. The later sections of this thesis consider trading
systems from other perspectives such as their portfolio diversification potential, the liquidity they
require and the investment capacity they can provide.
2.3. Simple system versus complex system
Many traders believe that simple trading systems with fewer parameters give comparably good
performance and are more robust than the complex trading systems that have many parameters.
In The Alchemy of Trading Systems (BookSurge Publishing, 2005), Matthew Hanson, the author,
explains why simple trading systems work better. With little data, one tends to favor the simpler
trading models because they contain fewer parameters and because tests like the likelihood ratio
test would strongly penalize the increase of parameters.
In An introduction to High-Frequency Finance (Dacorogna et al., 2001), authors had discussion
on simplicity versus complexity. "If simplicity is a desirable feature of theoretical models, one
should not necessarily seek simplicity at the cost of missing important features of the data-
generating process. Sometimes, it is useful to explore more complicated (nonlinear) models that
contain more parameters" (Dacorogna et al., 2001). According to the authors, increasing
complexity is strongly penalized when explored with low-frequency data because of the loss of
degrees of freedom. With high-frequency data, however, the penalty is relatively small because
of the abundance of samples even in a limited sampling period thus high-frequency studies can
be done for short periods. The results are less affected by structural change in the overall
economy than low-frequency studies with samples of many years. This gives us a better
opportunity to explore complex trading models. "High-frequency data paves the way for
studying financial markets at very different time scales, from minutes to years, and represents an
aggregation factor of four to five orders of magnitude." (Dacorogna et al., 2001). Market
behaviors observed on low-frequency data can be better validated by the fact that the same
behaviors are also observed with high significance on intraday data. This enables the
development of more complex trading models, such as those that simultaneously characterize
market patterns on different time scales.
Although it is feasible to develop more complex trading systems on high frequency data, it also
introduces the risk of overfitting, which increases when factoring in more rules, variables and
different timeframes.
2.4. Data sources and data characteristics
High frequency trading is equivalent to high frequency analysis. High frequency trading is more
about the analysis of high frequency data than it is about trading frequently.
In An introduction to High-Frequency Finance (Dacorogna et al., 2001), the authors provided a
thorough discussion of high frequency financial data. A few key points of high frequency data
presented by the authors are summarized in the discussion of previous sections (1.3 and 2.3).
Yan and Zivot, in Analysis of High-Frequency Financial Data with S-Plus (2003), also discussed
high frequency data characteristics "These high-frequency financial data sets have been widely
used to study various market microstructure related issues, including price discovery,
competition among related markets, strategic behavior of market participants, and modeling of
real time market dynamics. Moreover, high-frequency data are also useful for studying the
statistical properties, volatility in particular, of asset returns at lower frequencies."
According to Yan and Zivot (2003), high-frequency financial data possess unique features absent
in data measured at lower frequencies:
e The number of observations in high-frequency data sets can be enormous. The average
daily number of quotes in the EURUSD spot market could easily exceed 20,000. You
must handle and analyze a vast amount of data for any high frequency trading system.
* High frequency data are very noisy. High-frequency data is noisy and the data needs to
be cleaned skillfully before it can be used for modeling.
" Tick data by nature is irregularly spaced and with random daily numbers of observations.
This tremendously increases the complexity of data cleaning and system development.
The above characteristics of high frequency financial data substantially complicate the process of
statistical analysis and trading systems development. This may require specialized statistics and
econometrics software packages, more powerful development platforms and rigorous
implementation processes.
3. System design and development process
Trading firms can face many obstacles when managing the development and operation of trading
systems and the challenges are not unique to financial systems.
3
A trading system in essence is a
complex software system. Software engineering is a discipline that leverages systematic and
quantifiable approaches to the design, development, operation and reengineering of software
systems. Traders and financial analysts can learn to be engineers, although trading models are
even less tangible and constructs and stresses are measured in different ways.
3 Andrew Kumiega, Benjamin Van Vliet, Quality Money Management, San Diego: Academic Press. 2008. See quality principles
and best practices of managing the development and operation of investment systems in this book.
3.1. Design with Model-Based Systems Engineering (MBSE)
Systems engineering is an interdisciplinary field of engineering that focuses on how complex
engineering projects should be designed and managed. A high frequency financial trading system
is complex yet needs speed, reliability and robustness. It inherently requires high design
precision and design integration. Systems engineering deals with work-processes and tools to
handle this kind of project and it overlaps with both technical and human-centered disciplines,
such as control engineering and management processes.
Other engineering disciplines are transitioning from a document-based approach to a model-base
approach. Friedenthal, Moore and Steiner (A Practical Guide to SysML, 2008) asserts that
significant benefits of MBSE are to "enhance communications, specification and design
precision, design integration, and reuse that can improve design quality, productivity and reduce
the development risk" and MBSE places emphasis on "producing and controlling a coherent
system model, and using this model to specify and design the system."
Many MBSE system modeling languages are available. I have used SysML
4
and Object-Process
Methodology
5
during my research for high frequency trading system architecture. Both are
capable of modeling high frequency trading systems and have different logical processes but
serve the same goal: calibrate, bound, and validate your designs.
SysML is a general purpose graphical modeling language that supports the analysis,
specification, design, verification and validation of complex systems. SysML provides a
means to capture the system modeling information as part of an MBSE approach without
imposing a specific method on how this is performed. The selected method determines
which activities are performed, their ordering, and which modeling artifacts are created to
represent the system.
4 See more details of SysML in Sanford Friedenthal, Alan Moore, Rick Steiner. A Practical Guide to SysML: The Systems
Modeling Language. Morgan Kaufmann . 2008
5 See more details of OPM in Dov Dori, Object-Process Methodology. Springer. 2002
* Object-Process Methodology, or OPM, is a new approach to modeling complex systems
that consist of humans, physical objects and information. It is a formal paradigm for
system development, lifecycle support, and evolution. OPM combines formal yet simple
graphics with natural language sentences to express the function, structure, and behavior
of systems in an integrated, single model. The name Object-Process methodology comes
from its two major building blocks: objects and processes. A third OPM entity is state,
which is a situation at which an object can be. Object, processes and states are the only
bricks involved in building systems.
SysML is useful for people with a software design background. I found it is easy to describe
system-wide processes and function loops in the system design. It is easier to use than OPM
when explaining high level processes to financial engineers, traders and other stakeholders who
do not necessarily have the knowledge of system architecture. OPM is a great tool that identifies
all building blocks (objects) of the system and the internal relationship (process) among these
objects. It drills down, zooms in and out so that we can capture the dynamics between different
components and sub-systems. OPM may scare you a little with the appearance of complexity, but
it is very easy to learn and understand. OPM is especially good for building conceptual models.
Both modeling languages are capable of modeling the system architecture and processes of a
trading system design. Which system modeling tool is better for you? It depends. You may
prefer one to the other, depending on different backgrounds, the preference of thinking processes
and the nature of your project. The most important thing is to adopt a system modeling
methodology in your trading system and process management. A high quality coherent system
model and design framework helps maintain consistency and achieve repeatable success.
Working with Robert Scanlon, a student of system design and management program'09 at MIT, I
modeled a high frequency trading system using Object-Process Methodology (OPM). The model
was split into several separate diagrams, each representing a process in different levels of the
OPM. A high level view of the separate diagrams shows the relationships that exist between
them. Each successive diagram is a zoom-in of a process represented in a higher level diagram.
All diagrams are in Appendix II for reference. This system model is an experimental model for
research purposes and not applicable to building a proper trading system. This thesis uses SysML
modeling as the tool to explain system components, processes and design framework. Appendix
I has the complete diagram of a trading system design framework and its processes. The purpose
of this model is strictly for research and is not applicable to any real world trading systems.
3.1.1. System Components
bW POoc*] Tradng System ISystem Componerts )
Figure 2: Trading System Components
In Chapter 2.1, a trading system is decomposed into the following six levels:
. Investment Philosophy
. Trading Strategies/Mathematical models
. Risk Management/Money Management
. Disciplined processes
. Backtest and Optimization
. Financial data and trading software
The six levels of a trading system can be absorbed into the system design - risk management,
trading strategies, financial data and trading software are components/object of a system and the
investment philosophy, backtesting and optimization are processes in the design. Figure 2,
following the MBSE approach, displays a first level decomposition:
* Trading strategies/model (the implementation of the investment ideas)
" Data (used to develop trading strategies and backtest)
" Risk Control model
" Application platform (either proprietary software or commercial products)
3.1.2. Users
There are many users directly or indirectly involved in every trading system. The four direct
users who are involved into trading system design process are:
Traders, who generally execute trades and may contribute trading ideas and have special domain
knowledge that is built into the trading system. Traders are generally the end users of a live
trading system. Automatic mechanical trading systems require very minimal human intervention
and traders will only monitor the systems and trade by phone as a back-up plan under certain
circumstances. Traders may work with programmers to generate a regular basis trade report that
contains trading statistics. Later chapters discuss how to measure trading system performance by
analyzing trade statistics.
Financial Analysts/Engineers spend most of their time on research and developing strategies,
backtesting and optimization. In some firms they are responsible for cleaning data but with help
from other teams such as IT support. They primarily focus on algorithm development and
prototyping in modeling languages such as SAS, MATLAB, and S-PLUS. They may leave pure
technology issues such as error handling and database design for programmers as they usually do
not program for production.
Programmers perform production programming. They are skillful with programming in C++, C#
or Java. They convert proofed prototypes into real time trading systems. They do not need to
understand financial markets, only to create fast, high quality production code. Programmers in
have a relatively flexible positions in the trading firm. They may help traders to generate trade
reports, or help financial engineers with data preparation.
IT support provides support functions, such as installation and configuration of the trading
system and ensuring network stability and security.
ut P~ocki Tr.a*g System DOmebhu( MEJob Roim I I
smkTOut
Tft-e
Figure 3: Trading system user cases
Figure 3 describes the major user cases for building and running a trading system in an
investment firm. A real firm does not necessarily categorize the team the same way as we do in
Figure 3, and job tasks can be overlapped. For instance, financial engineers may code for
production in a hedge fund startup, while an established trading firm may even add a separate
quality assurance team to test production programs.
..... .... ................................................ ..... ... . ........
3.1.3. Live trading Process
ibd [Elock] Tradng System Domain[ IAppcaion Layout ]
(4bocka
Aplcto dm
<4ck: c~ck ccblcb
:DaCleaning
listoric Da Proider API - iri Da Providers
:d DknkaA
<1cbkx'- Broker API
-k I ~ o h - - - - - - - - -- - a er
Traing Models
Trader
Figure 4: Live trading internal process diagram
The diagram in figure 4 displays a high level overview of the live trading process: the historical
data provided by data providers (it can be the firm's proprietary data or the data from live
traders) goes through a data cleaning process, which can be used to calculate realized volatility
and other market behaviors. The mathematical models of calculation are built into trading
models and well tested. Trading models must read the live trading data from the brokers. Usually
the live trading data goes through the exact same data cleaning process as for historical data. The
live data carries the current market information and the trading model reacts to that information
by factoring the quantified historical market behaviors. All trading decisions are evaluated by the
risk control model that is also based on historical and live market data. Ultimately, all orders are
sent to the Brokers API and executed.
In general, the high-frequency trading system is "a fully-automated process of electronically
receiving data, processing that data through decision logic, generating orders, communicating
those orders electronically and finally, receiving confirmation of transactions."' The trader adds
no value in terms of normal operation of the world of systematic trading and investment and
6 Thomas Neal Falkenberry, High Frequency Data Filtering, Tickdata Inc. 2002, [Online]. Available:
http://www.tickdata.com/pdf/TickDataFilteringWhitePaper.pdf
automated executions. The trader's job is to monitor and oversee the trading/investment systems.
When a machine breaks down and the algorithm stops working, the trader must steps in and
decide either to stop trading or process orders manually using a back-up system or by phone. The
trader role is critical to the firm.
In addition, Figure 4 shows the components of the application platform, which can be
decomposed into its own objects, including the API Module, Data Cleaning Module in live trade,
Visualization Module, and Execution Engine. The API Module exists simply to connect the Data
Providers to the Trading System. The Data Providers object can be broken into two other objects,
Historic Data Providers and Live Data Providers, which each are used as instruments to
differentiate sub-processes of Trading.
3.1.1. System Development Process overview
Trading system development can be divided into three major phases: research, prototype and
production. Each phase has its own internal processes and sub-systems respectively (See
Appendix I):
" Research phase - Every trading system starts with its own investment philosophy and
trading ideas that are generated by research. There are a variety of method one could use to
perform this research. Typically, one generates trading ideas based on market study,
literature review, reviewing existing models, or reverse engineering. The outcome of the
research phase is a design document of your trading system.
* Prototype phase - Prototyping is the phase to verify and validate your investment ideas. In
this stage you prepare and partition the historical high frequency data. You also build your
models in modeling software, such as MATLAB or S-PLUS. Finally, you backtest your
models and optimize them on different data partitions. By analyzing the backtesting
performance, you decide to either accept or reject an investment idea. If a trading model is
accepted, a detail design specification must be created so that the production team can code
precisely.
* Production phase - Production is the product realization process of the trading system. It
usually follows a software development and quality process and the implementation is based
on the detail design specification from the product prototyping. In addition to checking the
software quality attributes, paper trading and experimental trading will be performed during
the production phase and before launching the live trading service.
Trading system development does not necessarily follow a strict waterfall process. There can be
iterations or loops over each design phase as problems arise or new discoveries are made. While
Six Sigma or Agile Development may be useful, there is no standard or best approach to
implement the trading systems. As in every design project, the approach used and the tools
chosen is dependent on the problem, skill of the engineers, development time available, and
budget. However, the chosen design approach should provide structure and discipline while
being compatible with the abilities of financial engineers and programmers. Failure is most likely
to result without such design discipline.
3.1.2. Research phase: generate ideas
The investment philosophy and investment ideas generated by sound research are the foundation
for building any investment/trading system. There are various ways to perform this research.
Before discussing research approaches, let's look into more fundamental ideas that are the
principles and mindset of researching and designing a high frequency trading system:
e The investment philosophy and investment idea is the foundation for any trading system. If
there is a flaw in our investment philosophy or any logic error with our trading ideas, you are
risking capital.
e Understand the difference between discretionary and nondiscretionary systems. The high
frequency trading system design leans toward a nondiscretionary, automatic system that can
be quantified precisely with trading rules and parameters that are explicit and constant.
" Do not have an opinion of the market. For a lot of high frequency mechanical trading
systems, profits are made from reacting to markets quickly, not by predicting the future of
the market.
e Understand the downside of your investment idea and consider risk management in the
research phase. Start building your risk control model when you begin generating investment
ideas.
" Discipline is a key. An automatic trading system helps you stay disciplined and away from
greed and fear.
" Develop and customize your own processes. No existing process fits your research, design
and development precisely. Use frameworks and general guidance but customize them and
develop your own processes.
" Test often and without overfitting. Any research results must be backtested and proved. Test
often and avoid overfitting.
3.1.2.1. Research methods
The first decision you make for trading system design is that of trading philosophy and premises
built from research. There are various ways to do your investment research. Figure 5 shows that
research may include the literature review of academic papers, new perspectives on existing
models, market study, and even reverse engineering. "Reverse engineering initiates the redesign
process, wherein a product is observed, tracked and analyzed and tested in terms of its
performance, trade statistics and other characteristics." (Kumeiga and Van Vliet, 2008) The new
design could be a replica of the original or a new adaptation of its underlying trading system(s).
Backtesting and optimization will never discover new trading methods. Simply trying different
combinations of parameters and trade rules in backtesting typically results in overfitting your
strategy to historical data and are destined to fail in live trading.
Figure 5: Research phase
3.1.2.2. Design document
The outcome of the research phase is the design document that describes all relevant aspects of
our trading ideas. The design document will be used as the blueprint for prototyping in the next
phase. The design document can include:
" Description of your investment philosophy and investment ideas
" The intended market
* Specific instruments you will trade on
" How much volatility and liquidity is required
* Trade selection algorithms for both position opening and closing
" Trade execution algorithms
e Data requirements
" Optimization routines
" Time horizon for the system
* Risk management logic
* Performance metrics
ucsly Research[ Resech )
4sAN*MIN
............ .........
" Design alternatives
e Design pitfalls
" Future enhancements
When the design document is created, a team meeting can be held to go through the details.
Financial engineers may present their multiple designs in the team meeting and team members
can help each other validate the trading philosophy and premises and verify the design. The team
meeting acts as a gate meeting - control the quality of the design and decide whether it is good
enough to go to the next phase.
3.1.2.3. Case studies
As discussed in 3.1.2.1, there are several approaches to researching and generating
trading/investment ideas. For intraday trading strategies, the primary market study approach for
is to examine the intra-day patterns of the exchange rate behavior, using the "firm" bid-ask
quotes and transactions of currency pairs recorded in the electronic broking system of the spot
foreign exchange markets. Once trading ideas emerge, you can document them and generate
design documents for prototyping. The following sections discuss two case studies for trading
strategies formation.
3.1.2.3.1. Trading the non-farm payroll report
This trading idea originates from the article of "Trading the Non-Farm Payroll Report" by Cory
Mitchell (http://www.investopedia.com/articles/forex/09/non-farm-payroll-report.asp).
The non-
farm payroll (NFP) report is a key economic indicator for the United States. It represents the
total number of paid workers in the U.S. exclusive of farm employees, government employees,
private household employees and employees of nonprofit organizations.
Of all news announcements, the NFP report consistently causes one of the largest rate
movements in the forex market. Upon release, the market is very volatile. Traders usually wait
for the wild rate swings to subside, and then attempt to capitalize on the real market move after
the speculators have been wiped out or have taken profits or losses. Traders attempt to capture
rational movement after the announcement, instead of the irrational volatility that pervades the
first few minutes after a news release. The release of the NFP generally occurs on the first Friday
of every month at 8:30am EST. This news release creates a favorable environment for active
traders in that it provides a near guarantee of a tradable move following the announcement. As
with all aspects of trading, whether you make money on it is not assured. Approaching the trade
from a logical standpoint based on how the market is reacting can provide you with more
consistent results than simply anticipating the market movement.
This section expands the outline of the design document as follows:
" Description of your investment philosophy and investment ideas:
o Of all news announcements, the NFP report consistently causes one of the largest rate
movements in the forex market. Trading the NFP report is based on waiting for a
small consolidation- the inside bar. After the initial volatility of the report has
subsided, the market chooses which direction it will go. By controlling risk with a
moderate stop, you are poised to make a potentially large profit from a huge move
that almost always occurs each time the NFP is released.
o Foreign Exchange
e Specific Instruments you will trade on
o Three currency pairs: USDJPY, USDEUR and GBPUSD
o Realized volatility < 10 basis points. The 10 basis points is the initial default value
based on market observation.
o Accumulated tradable size > 1OM US$
" Trade execution algorithms
o Nothing is done during the first bar after the NFP report (8:30-8:45am in the case of
the 15-minute chart).
o The bar created at 8:30-8:45 will be wide ranging; wait for an inside bar to occur after
this initial bar. Wait for the most recent bar's range to be completely inside the
previous bar's range.
o This inside bar's high and low rate set potential trade triggers. When a subsequent bar
closes above or below the inside bar, make a trade in the direction of the
breakout. You can also enter a trade as soon as the bar moves past the high or low
without waiting for the bar to close.
o Place a 30 pips' stop loss order on the trade you entered.
* Data requirements
o 15-minutes data from January 2, 2000 to April 30, 2009
o Economic data
o Calendar data
" Optimization routines
o Quarterly
o Optimize on G7 currency pairs
o 15 minutes
" Risk management logic
o Stop loss order: Place a 30- pips stop loss order on the trade you entered.
o Time based hard exit: Most of the move occurs within four hours. Exit four hours
after your entry time.
o Maximum trades: Make up to a maximum of two trades. If both get stopped out, do
not re-enter. The inside bar's high and low are used again for a second trade, if
needed.
o Leverage: No leverage
* Performance metrics
o Compounded Total Return
o Average net return per trade
o Maximum Intraday Drawdown
o Sharpe Ratio
o Sortino Ratio
o Use 5 minutes bar instead
o A trailing stop is an alternative to exit
*J. A pip is the last decimal point of a quotation. In the Forex market prices are quoted to the fourth decimal point. If the
EURUSD moves from 1.4319 to 1.4320, that is one pip.
3.1.2.3.2. Trading on short term momentum/trend
Trend analysis is the basis for many successful trading programs. Whether these programs are
trend following, taking positions in the direction of the trend, or mean reverting, taking positions
contrary to price moment, identification of the trend is an essential component. Trend system
works because "Prices are not normally distributed but have a fat tail. The fat tail means that
there are an unusually large number of directional price moves that are longer than would be
expected if prices were randomly distributed." (Perry J. Kaufman. New Trading Systems and
Methods). The fat tail is a statistical phenomenon caused by a combination of market
fundamentals and human behaviors. The net effect is that prices persist in movement in one
direction much longer than can be explained by random distribution. Based upon this theory, a
"naYve" high frequency strategy can be created - capture the short term momentum/trend in 5
minutes bar. See Figure 6 below. The logic behind this strategy is that whenever the price moves
very fast in a short time, there may be a positive feedback of the price action to push the price
further.
Figure 6: Short term trend
This section is my outline of the design document based upon current thinking:
* Description of your investment philosophy and investment ideas:
...........
o Whenever the price moves very fast in a short time, there may be a positive feedback
of the price action to push the price further. Capture the short term price trend.
o Foreign Exchange
" Specific Instruments you will trade on
o USDCHF
o Realized volatility < 12 basis points. The 12 basis points is the initial default value
based on market observation.
o Accumulated tradable size > 15M US$
e Trade execution algorithms
o Estimate the historic volatility of the past 2 hours by using simple moving average
o If the historic volatility < 12 basis points, trade is allowed
o Enter long position when price > 2 hour's high
o Enter short position when price < 2 hour's low
o Place a 15-pips stop loss order on the trade you entered
e Data requirements
o 5-minutes price data from January 2, 2000 to April 30, 2009
o Economic data
o Calendar data
e Optimization routines
o Monthly
o Optimize on G7 currency pairs
o 5 minutes
" Risk management logic
o Stop loss order: Place a 15-pips stop loss order on the trade you entered.
o Time based hard exit: Exit 2 hours after the entry time.
o Maximum trades: One trade at most for each direction at any given time
e Performance metrics
o Compounded Total Return
o Average net return per trade
o Maximum Intraday Drawdown
o Sharpe Ratio
o Sortino Ratio
o Use 1-minute bar instead
o A trailing stop is an alternative to exit
" Implementation Shortfall
o Stop orders may have big slippage
o The spread and slippage during the news may be huge, so avoid trade around the
economic news release
" Future enhancements
o Implement an accelerator indicator in order to reduce return volatility - focus not only
on the momentum (trend in the underlying), but also on the trend in the momentum
signal itself (acceleration).
o Perform trend survival analysis and statistical calculations to determine if there is a
way to find an optimal point (duration or return) to get out of the current position and
wait for the next signal.
3.1.3. Prototype phase
Before the prototype, the idea in its abstract format may contain logic errors and flaws in the
math. Prototyping proves investment ideas and mathematics before moving on. Prototyping
allows you to validate investment philosophy, detect logic errors and evaluate alternative
methods and help others understand the math behind investment/trading ideas. The primary
reason for prototyping is to resolve uncertainties early in the process. With prototyping, you
know whether the mathematical models are feasible and if the system is robust enough to profit
under different market conditions. And you can also get a good estimation about the workload of
the implementation for live trading applications. As Figure 7 shows, prototyping starts with data
preparation and partition, goes through the trading model coding process and finally the
backtesting and optimization.
"Prototyping is an exploratory and interactive process." (Kumeiga and Van Vliet, 2008) It
follows the requirements of a design document that consists of the investment philosophy,
mathematics, data requirements, and so on. The prototyping process forces the team to clearly
define the mathematical models and addresses all related issues, such as data issues in the early
stage of the process. The result of the prototyping will be a clear definitive requirements
specification on which the development of production software will be based. (Kumeiga and Van
Vliet, 2008) The document of detail design specification is created in this phase.
Prototype Phase
Clean Historical Data Data Parition Coding Backtest and Optimize
Production
Begins
<ddatastore>J
Raw Data
Clean Data <Adatastore> Data Paritioning Coding based Trading Modellstrategy
Cleaned Data
on Design Doc
result <<structuredo
datastore~ n-Sample Test
Train Data
<edatastore>. Optimization
Validation Data
dat-castore>> Out-of-Sample
Test Dat Test Yes
Preformance
-- - - - - - Review
Acceptdal?
1'es
Create Detal
Specification
jYes
Managemert
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - Review -
Accepted?
Figure 7: Prototype phase
3.1.3.1. Gather data
The cleaned historical data is the foundation of successful backtesting and optimization. Backtest
may itself create problems due to known and unknown errors in the data. Gathering high quality
historical data is not just a case of buying and integrating them into the system. For a trading
firm, once they decide which asset class they are going to invest or trade, the important first step
is to investigate the availability of data needed. They need to decide what kind of data they need,
where to get and how to clean the data so that the data fits the needs of their trading systems.
There are three sources for historical data:
Purchasing data from historic data providers, sometimes the historical data can be
provided by your brokers if you meet certain criteria. When purchasing data, you need to
identify the need for data (tick based, 1-minute based, etc.) and data availability. "You
also need to consider the experience of vendors, their references from their past
customers, their financial viability and compliance" (Kumeiga and Van Vliet, 2008)
Lastly, you must judge data quality independent of the claims of data providers. If you
have existing trading systems with some known high quality data, run a backtest on both
your data and the data from the vendors and do a comparison.
e There are free data to download from internet. Some companies or organizations offer
free data downloads. This approach is not recommended since you do not have control of
the data quality. Unreliable data can lead to the wrong microstructure calculation
resulting in bad trades in which the money lost would be more than the price of the better
quality data. Using high quality, more expensive data can be cost efficient. However,
even high quality data may have problems so checking data quality and the cleaning
process cannot be skipped.
* You can also start collecting data by recording the market price using programs and
cleaning and converting the data into the formats you really need. Brokers often provide
an application programming interface (API) that can feed in live trading data. Some
commercial trading platform provide a function called record market data and replay so
you can 'save' today's data and test your trading system in the future using today's data.
This approach requires the financial engineers to have sophisticated programming and
database administration skill.
Olsen & Associates has a high frequency currency database. NYSE TAQ provides Equity
historical high frequency data, and Berkeley Options database provides Options high frequency
data. There are also commercial redistributors such as (wrds.wharton.upenn.edu) and QAI Fast-
Tick (www.qaisoftware.com)
3.1.3.2. Data type
"Sophisticated trading/investment systems may potentially incorporate several different types of
data." (Kumeiga and Van Vliet, 2008) Data types include:
" Price data
7
usually consists of the read bid/ask price, trade, and volume data. For high
frequency trading system, the tick data might be flat ticket which means multiple trades
going off at the same price. Some financial derivatives don't have historical trade price
data, instead, they use valuation data
8
.
* Economic data tends to be one of the most important factors for short-term market
movements and this is particularly true in the currency market. High frequency trading
systems are sensitive to short term price movements so economic data is important to the
systems. For currency trading, the most important economic releases are: Non-Farm
Payrolls, Interest rates, retail sales, inflation (CPI), trade balance and manufacturing PMI.
Some economic data is extremely important to short time price movement. The list in
Figure 8 ranks the most market-moving data for the U.S. dollar in 2007, 2006 and 2004,
on a 20-minute basis.
7 Definition of Price data is based on Quality Money Management (Kumeiga and Van Vliet, 2008). See more details
of different data types from this book.
8 Definition of Valuation data (next page) is based on Quality Money Management (Kumeiga and Van Vliet, 2008)
Ti isiipiors -Ca S -0us T _p kdci Ci1 ANrs (Zi uisTF-T pI 0Miior-s Mnl64 b0 s)
Non-FaIn Payrok 69pips 1 Non-Fm Paynds - 1 FaP s
T- kiOnestRas (Fi) 57 Pips 2 kinirestais(F0MC) 2 WiieRsFond
3 in(CPOi 39 pips 3 Trade Balanw 3 Foreign Pudiases US Treas (11C)
4 Reba Sales 35pips 4 innafn(CP) 4 TradeBalace
5 PrdurPdidR 35 pips 5 Reil Sales 5 CureAOunt
New HOe Sales 34pips 6 FoigPurdiaseiU li Duale Goods
Exg Home Sales 34Pps 7 ISMMamfadiing 7 Rel Sales
8 Durabl Goods Ors 33 Psi 8 PrduePin IdeX 8 hdlan(CPI)
9 Non-F Pards 32 pips 9 Nonfain Papnls 9 NonFan Payrs
Figure 8: Most influential market moving indicators
Source: John Kicklighter, Top 5 Most Market Moving Indicators for the US Dollar, DailyFx.com, 2008
Calendar data consists of the release date/time of the most important economic data. Some
high frequency trading systems are purely built based on price data. Even these systems need
to be aware of economic calendars by either avoid trading on that period, or take advantage
of it such as "trading on new" strategies. Figure 9 lists the approximate times (EST) at which
the most important economic releases for most traded currencies. These are also the times at
which you and your trading models should be paying extra attention to the markets.
U.S. USD 8:30 - 10:00
Japan JPY 18:50 - 23:30
Canada CAD 7:00 - 8:30
U.K. GBP 2:00 - 4:30
Italy EUR 3:45 - 5:00
Germany EUR 2:00 - 6:00
France EUR 2:45 - 4:00
Switzerland CHF 1:45 - 5:30
New Zealand NZD 16:45 - 21:00
Australia AUD 17:30 - 19:30
Figure 9: Times of important economic releases
Source: Kathy Lien, Trading On News Releases, Investopedia.com, 2008
e Valuation data is different from price data. For over-the-counter (OTC) derivatives, no
historical trade price data exists. The price of these assets exists only in theory, so it is a
valuation price.
* Fundamental data consists of key business items, such as earnings, sales, inventories, and
rents. It can be calculated data based on fundamental facts such as P/E Ratio, ROE, and
Free Cash Flow etc.
3.1.3.3. Data cleaning process
Algorithm creation, backtesting and risk management requires good and clean data. High
frequency tick raw data is noisy, and the data cleaning and filtering process is partially important.
The importance of data cleaning is underestimated quite often by financial engineers who might
spend days on backtesting before realizing the data is not good enough. Bad data can lead to bad
models and losing investment capital. High-frequency trading is data sensitive so spending time
on cleaning is vital.
Price data is most often used in high frequency data. A high frequency currency trading system
relies on four major sub-processes to clean the price data.
" Check timestamps. Some timestamps are not in the correct order and require sorting. .
* Eliminate outliners. To reduce the effect of data errors and outliners, search and
replace/delete bad ticks in the original time series.
" Filter out non-trading days, including weekends, holidays and special events. For foreign
exchanges, filter weekend data (5pm Friday - 5pm Sunday).
* Converting into different timeframes such as 1 minute, 5 minutes, 15 minutes, and 30
minutes so traders can use this data based on the requirements of their trading strategies.
According Kirkpatrick and Dahlquist (2008), the historical data vendor should always be
consistent with the data vendor of live trading in practice because different vendors may have
different data feeding; The authors also pointed out that "dirty" data can happen on a live feed
thus data cleaning is also necessary in live trading. So, in practice, the live trading system also
has an internal function for data cleaning, as Figure 4 shows.
3.1.3.4. Data partition
When using the entire data set for developing the model, you may introduce overfitting. Given a
set of random data, you may be able to create a model that generates good performance. Using
the whole data set to build out the model, you may be able to generate a superior model, but in
most case you are creating overfitting. To address this problem, use data partitioning - divide
cleaned historical data into three partitions that include training data, validation data and test
data.
" Training data (sometimes called the In-Sample data) typically is the largest partition and
contains the data used to build models. The amount of data required depends on the
periods of the system. A general rule is that the data set can "generate at least 30 to 50
trades and cover where the market traveled up, down and sideways" (Kirkpatrick and
Dahlquist, 2008) After confirming the parameters and trading rules with training data, run
optimization on it. Alternatively, re-divide data by following the optimization methods
described in Chapter 6.
" Validation data is used to assess the performance of each model so you can compare
different models. You can also compare the performance on validation data to the one
generated based upon training data. If a model with good performance on training data
immediately fails with validation data, you know there is overfitting. Use the validation
data for fine-tuning to improve the model.
e Test data (sometimes called the Out-of-Sample data) is used to re-verify the models and
ensure everything is working properly, with no adjustment anymore.
For high frequency systems, the in-sample test is on training data. Optimization is done on both
training and validation data.
3.1.3.5. Coding in modeling software
For practitioners, MATLAB, SAS, S-PLUS are great tools for prototype programming. Many
financial engineers use Excel, but it may not be powerful enough to handle vast amount of high
frequency data. If your models require high computation power, you may have to either upgrade
your hardware or have your in-house developed system support the modeling. Usually in-house
systems leverage the power of C++ or C# to develop specific software tools to support certain
file handling or computation in a much faster manner. However, C++ and C# require higher
programming skills compared to MATLAB and other typical modeling languages. There are
always decisions and tradeoffs to make when considering different aspects of your project.
In general, prototyping coding is totally separated from production programs. People who
perform this task can be in totally different groups. The first group of people can be traders,
financial analysts and financial engineers. The prototyping will be based on widely used
financial modeling tools, such as MATLAB and SAS. The second group can be software
engineers, who do not have any opinions on the models. These people implement the production
code in C++ and C# based on the definitive requirements specification generated from the
prototyping. The argument is that prototyping focuses on speed and financial engineers need to
explore individual requirements, design and implementation options as quickly as possible. Fast
exploration is often at the expense of reliability, program performance, robustness and program
maintainability, but the whole prototyping process involves many rapid iterations. These
sacrifices (the lack of reliability, robustness, etc.) actually enable developers to explore design
alternatives and detect design flaws in a cheap way. Another argument to support this approach
is that the financial engineers' strength is in the quantitative side and may not have the
programming skill to generate high quality production code with the robustness and reliability
required in live trading. Conversely, highly experienced programmers may not have sufficient
investment knowledge to generate good investment ideas. Some trading firms prefer to have
good "pure" programmers who do not have much financial knowledge so "they are able to resist
the temptation to change your idea and add some extra 'features' in production". This seems
more like an organization management issue to me.
Another approach is called "Evolutionary Prototypes" (Andrew Kumiega, Benjamin Van Vliet.
Quality Money Management) which means the product team prototypes selected parts of a
system first, and then evolves the rest of the system from those parts. This approach does not
discard the prototyping code. Rather, it evolves the code into the production application that will
be used in live trading. That is, prototyping evolves into the production software. The working
system is the final product of a series of evolutionary prototypes. In this approach, prototyping
must be built on high quality code from the very beginning. Evolutionary prototyping requires
strong management and it works well for a highly experienced and well organized team. It may
slow down the process if the team lacks necessary skills. According to Quality Money
Management, the evolutionary prototyping should start with riskiest areas. Prototyping the
riskiest part first helps you identify the biggest obstacle so you can estimate the efforts and
feasibility of the project. This approach may work well with a small agile team with solid skills
in both finance and software engineering. Or the project is a cumulative development project.
The team starts with a simple working model and gradually builds more comprehensive models
on the top the working model in the previous iteration by either modifying/adding new features
or integrating other systems.
Modeling language such as MATLAB provide features that convert prototyping codes into C++
or C# format. This kind of feature enables you to reuse the prototyping code in production and
facilitates evolutionary prototypes.
3.1.3.6. Prototype quality assurance
During the prototyping phase, the most important aspect of programming quality assurance is to
be sure that all calculations are correct. This sounds simple, but it is surprisingly overlooked.
Before backtesting and optimization, always ensure the correctness of modeling code.
Backtest and Optimization
According to Kumiega and Van Vliet (Quality Money Management, 2008), a backtest is a test of
the ability of trading/investment strategies to meet the requirements of the design document and
"simulation and statistical analysis of a trading/investment strategies' input and output based on
cleaned historical data by factoring in transaction costs and slippage in execution." A trading
strategy will be accepted and sent to production if it generates acceptable performance, otherwise
it will be rejected. For the 'rejected' trading models, either it will be re-investigated at the
research phase, or it is discarded for good.
Optimizing is a process to achieve the best result (may not necessarily be the best return) by
changing the parameters of a trading strategy/model. Designing experiments is one of the most
important benefits of optimization. The designer might find parameters and trading rules that do
not work under any circumstance and can be eliminated.
Backtest and optimization can be performed at two stages - the prototyping phase and the
production phase. Backtesting and optimization at the prototyping phase is "essentially the make
or break stage of the trading/system design and development project." (Kumiega and Van Vliet
2008) The project either gets killed or moves forward to the prototyping stage. The backtesting
and optimization at the production phase is a verification and validation process for the quality of
the production - the implementation of prototype. The right methodologies and processes of
backtesting and optimization are vital and are a focus point of this thesis. Chapters 4 and Chapter
5 will discuss backtesting and optimization respectively.
3.1.4. Production
Once the prototypes of trading and risk models are accepted, production is the last procedure to
convert a trading/investment ideas into live trading strategies/models. The detailed design
specification from the prototyping phase and the accepted prototypes will be the material the
implementation is based on.
3.1.3.7.
a-ce y Producton[ =Production 1)
Figure 10: Production Phase
3.1.4.1. Production programming
Production programming should follow software engineering processes and the best practices of
the software industry. Some companies may depend on highly skilled software engineers for the
implementation/coding without a systematic design approach. The assumption is that good
engineers can make the design work without any process. Extraordinary people, who are highly
experienced and motivated, can get the job done, but dream teams are rare and companies risk
burning out their best employees. Heroics in software development are an indication of process
ILsuesTrn TcaT
PIser Treenr&T
b Od
failure that can lead to dysfunctional behavior in both organizations and individuals (Yourdon,
Edward. Death March: The Complete Software Developer's Guide to Surviving "Mission
Impossible" Projects, Prentice Hall, Englewood Cliffs, NJ, 1997.) Software engineering
practices are designed to make the development of software less chaotic and reliably repeatable.
One of the most important characteristics of a trading system is it must have high quality. No
functional bug is allowed. A small error can cause a large capital loss or simply a disaster. Some
coding strategies, such as 'code, release and fix later', used by some internet startups will not
work with trading firms. You need adopt a suitable programming process and apply the best
practices of software engineering to form your own (most comfortable) approach to achieve
superior quality of coding. The methodology and practices that were used to code experimental
systems of this research are Test-driven development (TDD) and Pair trading.
" TDD is a software development technique that uses short development iterations based on
pre-written test cases that define desired improvements or new functions. Each iteration
produces the code necessary to pass that iteration's tests. Finally, programmers re-factor the
code to accommodate changes. A key TDD concept is that preparing tests before coding
facilitates rapid feedback changes (Quoted from Wikipedia).
" Pair programming is a software development technique in which two programmers work
together at one work station. One types in code while the other reviews each line of code as it
is typed in. The person typing is called the driver. The person reviewing the code is called the
navigator. The two programmers switch roles frequently (possibly every 30 minutes). While
reviewing, the observer also considers the strategic direction of the work, coming up with
ideas for improvements and likely future problems to address. This frees the driver to focus
all of his or her attention on the "tactical" aspects of completing the current task, using the
observer as a safety net and guide (Quoted from Wikipedia).
At first glance, TDD and Pair programming appears to take longer and cost more. Research,
however, has shown the opposite result - these processes and techniques increase discipline,
provide better time management, and the coding quality is much improved. Again, there is no
standard best approach for production programming. You need to create your own development
process that best fits your team.
As Figure 10 shows, besides implementing software quality processes, financial engineers run
backtest for the production version of trading system to ensure it achieves the expected
performance on historical data. For some cases, financial engineers run optimization against the
production code to see if there is any difference between the prototypes and the production.
3.1.4.2. Development team and tool
The competitive advantage of real time high frequency trading system is speed. For some high
frequency trading systems, fast computation and execution are critical. Unlike the programming
in prototypes, the programming during the production phase shifts from fast iterations that
validate mathematical models, to robustness, reliability and fast execution in live trading.
Programming tools used will be C++, C# or others that generate fast-execution at runtime.
Prototyping usually is done by financial engineers while production is performed by 'pure'
programmers with the right skill set.
3.1.4.3. Production quality assurance
For production code, there are three phases of quality assurance. First, the software coding
quality has to be assured. Software quality assurance methodologies apply here and can include
test-driven development, scenario based testing and test automation. As is true for development
process, you can adopt the right approach for software quality assurance, considering the
extremely high quality requirement for trading systems. Second, the testing will include the
backtesting and sometimes optimization. The major purpose of the backtesting at this stage is not
to prove the validity of investment ideas and the math as you do in prototyping. Instead,
backtesting compares the performance of the production code to one of the prototypes so you
know if the production code was implemented successfully or not. With the same historical data,
one would expect similar performance. During certain circumstances, the testing result can be
different between prototyping and production. However, the gap should be in an acceptable
range and the reason for the gap should be fully investigated. Finally, there is a live trading test
before you start the full trading. This test includes:
" Paper trading test. This is simulated trading where a trading system generates real-time
signals based on the live data, but only trades with virtual money and executes dummy orders.
Paper backtesting is helpful to uncover algorithm failures under extreme market conditions.
" Experimental trading test. This is live trading with small amounts of money. Experimental
trading helps you detect algorithm flaws in the live market and helps identify operational
issues.
Figure 11: Three trading types
After evaluating the performance generated by paper trading and experimental trading
(performance measurement is discussed in Chapter 6) and deciding to go or kill, the trading
system will be either sent back for rework, or be launched for a full live trading. For high
frequency trading strategies, it is best to start paper trading first, and then small money
experimental trading before final launch of full live trading (see figure 11). For low-frequency
data, it may not be feasible to go through all three types of trading because it may wait too long
to accumulate enough paper trades and experimental trades that represent a good sample.
4. System backtest
A backtest is a simulation of a trading/investment strategy on historical data. It is a validation
and verification process for model assumptions and parameters. It is essentially a quality
assurance test against the detailed design document generated in the prototyping phase. After
financial engineers determine what data to use, it is then necessary to partition data into test data,
validation data, and test data, run an in-sample test, optimize and finally run an out-of-sample
test. Financial engineers will then analyze the outputs of tests in order to validate against the
design document for the strategy developing. The outcome of tests will help you decide whether
to accept the trading strategies or reject them. 9
4.1. Testing process
The testing process must be carefully defined before testing can begin. The process is more
important than the actual testing. Correctly defined, the final system will have realistic goals and
predictive qualities. Incorrectly done, it will look successful but fail in live trading. Although this
chapter discusses backtest, we start with emphasizing the importance of investment philosophy -
everything starts with ideas. Backtesting should be the process of validating your investment
ideas. If the test results confirm your ideas and investment philosophy then you can have
confidence in the trading strategy. If you indiscriminately test all indicator combinations, you
will have no idea whether you are finding a good trading strategy or simply overfitting the data.
Re-emphasizing the importance of investment philosophy is to underline the importance of
backtesting to validate mathematical models and the investment ideas behind the math. Numbers
are like people; torture them enough and they will tell you anything (Anonymous - quoted by
John Ehlers in Rocket Science for Traders). So before fully kicking-off the testing, you should
have the design document in place and fully understand the investment ideas and expectations.
The following ideas of deciding value range, value distribution and ranking parameters were
adopted from Perry J. Kaufman's book, New Trading Systems and Methods (Wiley, 2005).
4.1.1. Decide value range
There are a large number of tests with high frequency data if you have multiple parameters. Due
to the vast amount of data, each sample run on minutes-based high frequency data may take
much longer than on daily data. If the trading strategy is based on the tick data, the time each run
takes is significantly larger than the 1-minute data. See Figure 12, the estimated test time for
USDCHF short term momentum trading system.
9 Definitions of backtest are based on Quality Money Management (Kumiega and Van Vliet, 2008)
40000
35000
30000
25000
20000
15000
10000
5000
0
Daily Hourly 30 min15 min 5 min Ticket
Figure 12: Test time (minutes) for USDCHF on short term trend trading
You must select fewer values to test for each parameter so choose them carefully and wisely
based on the statistics from market studies. Tests with many parameters on high frequency data
can be extremely time-consuming. In New Trading Systems and Methods (Wiley, 2005),
Kaufman pointed out that limiting the test is actually an advantage because the value selection
process forces you to find the most reasonable value ranges for the trading strategy. "Without
these restrictions, the test process gets closer to indiscriminant exploration, and less of a
validation process."
4.1.2. Decide value distribution
After deciding the value range, you should also examine the values in the range and decide if you
really need them all. For most cases, it is not necessary or practical to test all the values for a
valid parameter, considering you have large amounts of intraday data. The values do not need to
be spaced evenly apart. 10 For instance: 1, 2, 3, 5, 10, 15, 30, 60 would be a better choice of
values than using all numbers between 1 and 60, because, in terms of percentage of change, the
10 See more details of identifying parameters including deciding value range, value distribution and ranking parameters (next
page) from New Trading Systems and Methods (Perry J. Kaufman, 2005)
difference between 1 and 2 will be significant, but the difference between 59 and 60 minutes is
negligible. Using equal increments will weight this set of tests heavily towards the long end;
therefore, the comparison is not very reasonable to some extent. In practice, you can use methods
like exponential smoothing to reduce the number of tests. You will achieve significant benefits
by limiting the values used in the test.
4.1.3. Rank parameters
Testing one parameter at a time can give you a better understanding of the dynamics of a system.
It is hard to understand the dynamics between different parameters when testing a few
parameters simultaneously. The most important parameter should be tested first, and a parameter
ranking system is needed. This ranking should be based on the trading model logic and the
correlations with historical performance. For instance, risk control using the stop-loss, or time-
based exit should improve the final model at the later stage of trading model research. However,
it is not likely that simply cutting losses will generate net profits, so it does not make sense to
rank the stop-loss point as the first parameter tested. Ranking parameters also gives you a good
chance to rethink and validate all parameters used by the trading algorithm. When ranking
parameters in practice, you can close out some parameters if they turn out to be insignificant
statistically. You should capture relevant parameters that impact the trading algorithm but be
aware that the more parameters you have, the greater the risk of overfitting.
4.1.4. Decide trading cost
High frequency trading has the advantages of smaller actual loses and more trades. A high
frequency trading system that performs well with more trades will give more confidence of
future performance. On the negative side, high frequency trading has to fight against slippage
and overall trading cost. The trading cost has a significant impact on high frequency trading
models and can change the appearance of the multiple-test performance pattern. The trading cost
usually includes the spread, slippage and brokers' commission. Before you run backtesting, you
should have a reasonable estimation for the trading costs you may encounter during live trading.
Unrealistic estimation of trading costs can lead to the failure of live trading.
4.2. In-sample and out-of-sample test
"Performing proper in-sample test and out-of-sample tests is perhaps the most critical step in the
development process of a trading system."(Kumiega and Van Vliet, 2008) Once you decide the
test size, the parameters and value range, you can start performing the in-sample and out-of-
sample tests. In-sample will be run on both train and validation data. Out-of-sample will be
allowed on out-of-sample/test data only. "Financial engineers are aware of the extent to which
in-sample results may differ from out-of-sample results." (Kumiega and Van Vliet, 2008) When
this happens, the trading algorithm must be examined and its parameters and trading rules need
to be re-verified. A robust system should be profitable on both in-sample and out-of sample.
Always save the test partition and never tune your system based on test data. A backtest result
will yield one of three possible outcomes.'"
* Profitable in both in-sample and out-of-sample. In this case, trading models will be
accepted and the production will start as soon as possible.
* Profitable only for in-sample but not out-of-sample. A three-step approach needs to be
taken. First, check the prototype code and make sure there are no programming errors.
Next, tune the parameters and re-test the models to see if you can achieve profitability in
both in-sample and out-of-sample. After a few rounds of testing and if the model is still
profitable in one test and not the other, you reject the model and move to the third step
which sends the 'rejected' model back to the strategy developing phase and continues the
research.
e Unprofitable in both in-sample and out-of-sample. The model will be rejected and for
most cases it will be discarded for good, unless the investment/trading idea is very solid
and supported by premier research. In this case, you may send the models to the strategy
developing phase for continuing research.
" See the discussion of three possible outcomes of a backtest in Quality Money Management (Kumiega and Van Vliet, 2008)
4.3. Avoid overfitting
4.3.1. Save out-of-sample/test data
As discussed previously, you need to partition the data into separate pieces. You can then build a
model using one dataset ("training dataset" or "In-Sample") and evaluate its fit by applying the
model to another set ("test dataset" or "Out-of-Sample"). Invariably the out-of-sample
performance will be considerably less than the performance generated in the optimization. A
tendency for system designers when they see unsatisfactory out-of-sample results is to repeat the
optimization process until the out-of-sample results 'become' better. Eventually the out-of-
sample data becomes the same as in-sample data and the process gets closer to overfitting. Thus
you might alternatively want to use three sets: training dataset (to build the initial model),
validation dataset (to guide the progression of the model), and a test dataset to gauge the
accuracy of the final model. A method to check whether a good performance is truly generated
from the model, or simply just because of overfitting, is that you can test the same model with
different currency pairs, or even different asset classes. If the same parameter set works in
different markets then you have a reliable and robust system. This seems counter-intuitive
because you would expect that each currency pair or each market is different and requires
different parameters and values. However, if a high frequency trading system needs different
parameters for different currency pairs, it might be an indication of overfitting.
4.3.2. Different time horizons
During a backtest, returns over different periods should be consistent. For instance, a high
frequency model with parameters optimized over a 3 year holding period should have a similar
return over 6 months, 1 year, 2 year and 5 years. Other measurements such as sharpe ratio,
standard deviation and daily maximum drawdown should be consistent in different holding
periods. If a high frequency trading model does not work or is not consistent in different time
horizons, it is a sign of over-fitting and you should reject the model. Many high frequency
models are market neutral and are able to deliver consistent returns during different time
horizons.
4.3.3. Different time frames
Like the ideas at different time horizons, a trading system that is built on a 5-minutes bar should
work for on different timeframes such as 1-minute and 15-minutes. Returns over different
timeframes should be consistent. If a high frequency trading model works on 5-minute bar data
but totally failed on 15-minutes bar, a further investigation is needed.
4.3.4. Different Instruments
The consistency of returns on different instruments is a good indication of robustness for a
trading system. If trading system is designed and optimized for USDJPY and the same system
works for EURUSD without changing any parameter, then you know the system is robust. The
characteristic of robustness can help you detect overfitting issues for certain models. If a trading
system works well with USDJPY but fails with EURUSD, it raises a red flag and you might have
to investigate. This is especially true for high frequency trading systems that are price data based
and use technical trading strategies, such as momentum, break-out and resistance/support line.
4.3.5. Different asset classes
Another method of reducing the effect of overfitting is to use more than one market as an out-of-
sample test. It is difficult to have the same parameters set in different markets and at the same
time overfitting. This is counter-intuitive because you would expect that each market has its own
personalities and requires different parameters. However, the high specialization of a trading
system (only works for one asset class) usually indicates that the results are from overfitting, not
live trading performance. A reliable system should work in most markets.
4.3.6. Student T-test
Among tests that help determine whether the results are significant, the student t-test is the most
useful. It helps financial engineers to detect a systematic bias in the data by showing that the
mean of the data is significantly different from zero. This information helps us decide whether
the results of only a few trades represent a good system, or whether a series of losing trades
implies that a system has no value. For a series set of trades produced by backtesting and live
trading, the t-test is:
t= average trade results/1 STD of trade results * (number of trades) A 0.5
The values of t that are needed to be significant can be found in the T-Distribution table. For
instance, if you have 31 trades, the degree of freedom is 30, and the value of T must be greater
than 2.423 to achieve 99% confidence in results.
4.3.7. Case studies
The short term trend trading system is built for USDCHF on a 5-minute bar. To test for
robustness and overfitting, the two following tests were performed:
e Test the trading system on different timeframes. This original trading system is built and
optimized based on a 5-minute bar. Without changing any parameters, the out-of-sample
test was performed for 1-minute and 15-minutes bar data from 2008. See Figure 13. The
monthly returns are relatively consistent among all three time frames. The annualized
return for 1, 5 and 15 minutes are 1.62%, 4.53%, and 4.22% respectively. The trading
system is relatively robust and the chance of over fitting is small.
2.00%
1.50%
1.00%
0.50%
0.00%
-0.50%
-1.00%
-5 minutes
-1 minute
- 15 minute
-1.50%
Figure 13: Different time frames for USDJPY on short term trend trading
.............................................................
* Test on different currency pairs. The trading system was built for USDCHF. Also
performed was an out-of-sample test on the data of 2008 for USDEUR, USDJPY and
USDCAD. See Figure 14 for results. The trading system failed on EURUSD and
USDJPY. The first impression would be that the trading system has an over-fitting issue
because it only 'fits' the currency pair the system was built and optimized upon.
However, taking a closer look at the design document and production code, it seems that
the problems are related to hard-coded absolute values, such as the historical volatility
and stop-loss value.
6,00%
4.00%
2.00%
0.00%
-2.00% IqCF ERSY UDA
-4.00%A
-6.00%
Figure 14: Returns of different currency pairs
When you say "historic volatility < 12 basis points" in the design document, it introduced
the 12 basis point into the system as an absolute and hard-coded value. A better approach
is to use the relative value so the real value can be calculated against each currency in
real time. For instance, instead of saying "historic volatility < 12 basis points", say that
"historical volatility < 0.95 * simple moving average of last 40 bars". Thus the value of
historical volatility is dynamically calculated based on the simple moving average for
each currency pair in run time. The calculations of take-profit and stop-loss values can
use the same approach. See Figure 15 that presents out-of-sample test results after
eliminating a few hard-coded values and re-optimized based on USDCHF. Although the
return of USDCHF is reduced, the returns across different currency pairs are more
consistent. EURUSD and USDCAD have a similar return during the basktest without any
adjustment of parameters. However USDJPY only has a very few trades during 1 year
period in the backtest and this indicates that there is a need for future improvement in the
implementation.
3.50%
3.00%
2.50%
2.00%
1.50%
1.00%-
0.50%-+
0.00%
USDCHF EURUSD USDJPY USDCAD
Figure 15: Returns of different currency pairs after eliminating hard-coded values
Based upon the above two tests, I would argue that the implementation of a trading system
itself is vital to the robustness of a trading system and minimizes the overfitting. So when we
see signs of over fitting, it doesn't necessarily mean the investment/trading ideas are bad. The
failure may arise from the implementation itself.
4.4. Trading statistical data
After backtesting, it is important to analyze the statistical data. Not only will it help to measure
performance, it will also identify the weakest areas, such as strategies taking too much liquidity
and depending on the market conditions. Chapter 6 discusses statistical data and how to measure
systems' performance.
5. Optimization
Optimization is the process of using historical data to test the effects of changes in parameters of
a trading system to produce the maximum value of the test objective. The test objective can be
net profit, sharpe ratio or a robust solution. An important benefit of optimization is that the
designer may find parameters and rules that do not work under any circumstances and can be
eliminated. Another benefit of optimization is to maintain peak performance in a trading system
under the continuously changing market. Although often beneficial, optimization can be one of
the most misused techniques. The principal concern with optimization is the tendency to overfit.
Overfitting occurs when the optimization program finds the absolute best set of parameters to the
training data and the model is twisted to fit the data set being tested. (Kirkpatrick and Dahlquist,
2008) Avoiding overfitting was discussed in Chapter 4. There are many ways to optimize your
trading system. The most basic approach is sequential testing that tests different combinations of
parameters. The complex techniques may include genetic algorithms and Monte Carlo sampling.
For sequential testing, methods including whole sample, out-of-Sample and walk/step forward
optimization are discussed in Technical Analysis (Charles D. Kirkpatrick, Julie R. Dahlquist.
2008). They are simple and effective. I therefore summarized them in the following discussion.
5.1. Methods of optimization
5.1.1. Whole sample analysis
A whole sample analysis is to run optimizations on the entire sample data. Without out-of-
sample data for validation, this approach has a tendency of overfitting. But in some cases whole
sample optimization will be useful:
" Use whole sample on one of the currency pairs to train the model, and then test it against
other currency pairs.
" After determining the optimal parameter sets, divide the optimization period roughly into
tenths to fifteenths and run a test on each period using the optimal parameter set. Then
you can analyze the results from the ten to fifteen different test periods to see if the
system generated consistent results under all conditions. (Kirkpatrick and Dahlquist,
2008) Consistency is important. If the results are not consistent, the system has a major
problem and should be optimized using other approaches. If it still does not work, then
you need to reject this trading system.
5.1.2. Out-of-Sample analysis
Out-of-Sample analysis is a commonly used approach that divides the data into two sections, one
of 70-80% percent of the data to be used to develop the system called in-sample data, and one of
20-30% called the out-of-sample data. This concept of data partitioning is well known. The 20%-
30% out-of-sample data usually are the most recent data. This OOS method optimizes based on
the in-sample data and then tests it on the out-of-sample data. If the result of the out-of-sample
period is similar to the in-sample performance, the system is considered validated.
A key here is to include bull, bear, sideways movements and a good number of price shocks of
various sizes.12 A tendency for the system designer when they see unsatisfactory out-of-sample
results is to repeat the optimization process with different parameter values to make the out-of-
sample results 'become' better. Eventually the out-of-sample data becomes the same as in-
sample data and the process gets closer to overfitting.
For out-of-sample optimization, you can alternatively divide data into three partitions: train data,
validation data and test data. The in-sample test and optimization will never happen on test data
(otherwise, it is not out-of-sample).
5.1.3. Step forward analysis
It is close to OOS but moves the test window (in-sample). The test window moves forward until
the test reaches the most recent data. For each test window, an optimization on in-sample data is
run and an out-of-sample test is performed. The results from all testing windows will be analyzed
for profitability and consistency. The acceptable parameter sets are those that have consistent
performance across different test windows.
12 In New Trading Systems and Methods (Kaufman, 2005) and Technical Analysis (Kirkpatrick and Dahlquist, 2008), authors of
both books pointed out that sample data must cover different market circumstances
6. Systems measurements
When analyzing a trading system, you must look at the system's robustness, the profit, the risk,
and its investment capacity, and its contribution to portfolio performance.
6.1. Robustness of system
Robustness means how strong and healthy your models are - whether it will be able to perform
consistently when the market changes, because the market never exactly repeats the past, which
is already introduced into the system design as a basis. The formal definition for system
robustness from MIT's engineering system division is "Robustness - demonstrated or promised
ability to perform under a variety of circumstances; ability to deliver desired functions in spite of
changes in the environment, uses, or internal variations that are either built-in or emergent" [In
ESD Terms and Definitions, http://esd.mit.edu/WPS/esd-wp-2002-01.pdf]. Particularly for a
high frequency trading system, the system's robustness can be considered as a measurement of
counter-overfitting thus the indicators of overfitting addressed earlier serve well as the indicators
of a trading system's robustness. A quick check of a system's robustness is to determine whether
your high frequency trading system can profit consistently in:
" different time horizons
e different timeframes
" different instruments
e different asset classes/markets
6.2. Profit and risk measures
Measuring and comparing the profitability of various trading systems requires a number of
performance criteria, in order to evaluate any trading strategy during the test phase and to
compare those results to an out-of-sample period or to live trading. The major ways to measure
the profitability and risks fall into four categories described in the following sections.
6.2.1. Absolute metrics
There are a number of metrics of performance and risks, and you need to determine
measurements based on the nature of trading systems and the requirements of investors:
* Compounded Total Return. The total compounded return during the testing period. The
return is compounded in order to compare with another trading system.
* Average Return per Month: It is usually used to analyze the seasonality of return and the
consistency of trading performance. Graphical format is usually used.
* Annualized rate of return. It can be used for relating the results of a system against that of
a market benchmark.
* Maximum Drawdown: The peak-to-trough decline during a specific record period and it
is measured from the time a retrenchment begins to when a new high is reached. It gives
a rough idea of the minimum capital needed to trade in this market. For high frequency
trading systems, you can use intraday maximum drawdown due to the fact that most high
frequency trading positions are closed in a very short time period. A surprisingly small
maximum drawdown is a sign of overfitting or too small a test period. It is safer to take
the average maximum drawdown from a range of tests than the one from the best result.
" Average time to recover: A large drawdown may be inevitable in live trading; a shorter
time to recovery is most desirable.
* Average net return per trade: This is a great indicator about how sensitive the high
frequency system will be to trading cost.
* Sharpe Ratio: It tells us whether a return is due to smart investment decisions or a result
of excess risk. It is a common measure of the return versus risk of a portfolio or system,
but it has several problems when applied to trading systems. First, it penalizes upside
fluctuations as much as downside fluctuations. Second, it does not distinguish between
intermittent losses and consecutive losses.
" Sortino Ratio. It is similar to Sharpe ratio. But it uses downside deviation for the
denominator instead of standard deviation, the use of which doesn't discriminate between
up and down volatility.
" Number of Positive Return Months. During the test period, the number of months with
positive return.
* Number of Negative Return Months. During the test period, the number of months with
negative return.
* Smoothness of returns: Investors always prefer the consistency of return so a smoother
equity curve is always more desirable than one with high volatility.
A general rule with these metrics is that one would expect, in live trading, that the return will be
half of returns generated in backtesting and the risk (maximum drawdown etc) will be two times
the backtesting result.
6.2.2. Trade statistics
Trade statistics are important to understand the reliability and health of the trading systems. They
are also very useful pieces of information that can help us with understanding risk and cash
management.
* Number of trades: This indicates whether your test was statistically reliable. A high
frequency trading system generates a larger number of trades so it has a better chance of
performing up to expectation.
" Percentage winning trades: The more winning trades your trading system has, the better
chance of its being profitable in live trading.
* Average holding period: All else being equal, a trading system that holds positions
shorter than another system is preferable.
" Averages profit each trade: A high frequency trading system is extremely sensitive to
transaction costs and this figure will tell you how sensitive your trading system will be
against the trading cost. If this value is too small, it means that your trading system is
very vulnerable to transaction costs.
* The distribution of different types of orders. The shares of the total number of trades for
Limit orders, Stop Orders and Market Orders respectively. In Figure 16, the distributions
of different types of orders are not desirable due to the fact that the majority of orders are
stop orders so the system may have very high slippage when the market become volatile
and liquidity becomes an issue.
0 Limite Orders
0 Stop Orders
a Market Orders
Figure 16: Sample distribution of trading order types
6.2.3. Benchmarking with index
Benchmarking is another way to measure the performance of a trading system or the portfolio.
First, it shows to the investors how much you out-perform or under-perform the stock indexes.
Second, it tells us the correlation between our trading systems and stock indexes. As Figure 17
shows, the performance of a portfolio of FX high frequency trading systems out-performs the
Shanghai Stock Exchange (SSE) and Dow Jones Industrial Average (DJI). The portfolio equity
curve is smooth and it is not impacted by the bullish/bearish stock markets.
1 _$ :P' - 'n'
Figure 17: Portfolio performance comparing to stock indexes
6.2.4. Correlations with index
Another important measurement of a high frequency trading system is its correlation with
traditional assets classes. A lot of investors would consider high frequency trading products as a
diversification of their investment portfolio so it is useful to calculate the return correlation
between a high frequency trading system and major stock markets. As figure 18 shows, a
portfolio of FX high frequency trading systems has very low correlations with major stock
market indexes so it has great diversification potentials for investors.
DJIA -0.04
S&P 500 -0.02
NIKKEI 225 0.04
Hang Seng -0.16
DAX -0.04
FTSE 100 -0.01
Shanghai SE Composite -0.04
Figure 18: Correlations with stock indexes
6.3. Investment capacity
Investors are concerned with the capacity of your trading systems. They want to grasp how much
money you can manage. Calculation of investment capacity is not easy for any asset class. There
are different systematic approaches to calculating the investment capacity for different markets
and different investment products. A simple approach I adopted to calculate investment capacity
for high frequency currency trading systems is based on the liquidity your brokers can provide in
your trading hours and the commission they ask for. There are three factors in the calculations:
" Market Depth in broker's trading book
* Broker's commissions
* Trading hours
For instance, if a new trading system can only accept 2 pips as the maximum trading cost after
the broker's commission and spread, you need to look into your broker's trading book to see how
much liquidity you can get in 2 pips during your normal trading hours. Figure 19 is a screenshot
taken from Interactive Broker's Trader WorkStation (TWS). There is Euro $37,850,000 as the
cumulated tradable size (in Bid) for EURUSD in less than 2 pips. You can average the tradable
size in the normal trading hours of your systems so you can get a good estimation of how much
you can trade/invest for this particular trading system.
Eve Q2ds~s Quigu4 O
a
6 CaoegM Imes n UGmWWe
Figure 19: Market depth for EURUSD
7. Conclusion
The goal of this research was to demonstrate how the application of academic research and the
native processes and well-established, coherent design principles of an IT engineering product
framework could improve the design, development and reliability of high frequency financial
trading systems. By nature, computation-based high frequency trading systems are highly reliant
on software engineering systems. Developing a successful high frequency trading system can
increase the competitive advantage and profitability of investment firms. Being able to convert
academic research results and mathematical equations into a profitable trading system quickly is
critical. From a systems and a business perspective, the ability to implement high frequency
trading algorithms in a real-world trading environment efficiently and effectively is a key
competitive advantage. Investment firms that succeed acquire great leverage in the new world of
global, high-speed trading. A major point of emphasis in my research is to point out and
demonstrate how the application of a coherent system design framework and rigorous
implementation processes, such as backtesting and optimization, is the most important part of
high frequency trading system development.
My research used two approaches (SysML and OPM) of model-based systems engineering
(MBSE) to architect and create development process models for a high frequency trading system.
My thesis shows that well-defined system models and processes are fundamental to ensuring
repeatable success and consistent performance. Because a high frequency trading system is
inherently complex and requires precise calculation and fast execution, MBSE offers significant
potential benefits to facilitate a high quality design, improve productivity and reduce the
development risk. Arguably, without any system modeling, exceptional people may still build a
good trading system. However, making success highly dependent on extraordinary individuals is
risky, unsustainable and does not scale. Although there is no one standard system design
framework and set of processes, as is true for all engineering disciplines, the approach selected
and the tools chosen depend on the nature of the problem, the skill of the engineers, and the
development time and budget. To maximize the chance of success and minimize the risk of
failure, this thesis shows that building a high frequency trading system requires the application of
a proven design framework and process management principles.
High frequency trading systems deal with large amounts of data and are extremely calculation-
intensive. You must have a strategic plan to backtest effectively and efficiently. Without using
appropriate backtest techniques, the differences in system performance can be enormous. My
research also points out that backtesting and optimization is more important than the trading
algorithms used in developing a system. You can build a good trading system with a naive model
if you backtest and optimize it correctly. With the right backtesting approaches and techniques,
you can prove that your models will not work if they are incapable of avoiding the loss of capital
in live trading. Similarly, the misuse of backtesting and optimization can lead to trading failures,
even if the underlying trading algorithms and rules are sound.
Lastly, my research concludes that flexible, continuous system improvement is also fundamental
to success. The design framework, development processes and backtesting techniques employed
in system development need to be constantly reviewed and improved. Rigid design principles
and processes do not optimize chances of success. The inherent nature of high frequency trading
requires systematic innovation, development and continuous improvement to guarantee the
quality of design and performance of processes of high frequency trading systems.
Appendix I: Trading System Development Process
[r r
Figure 20: Trading system development framework and process (SysML)
................ . ........................ . NNOMEMO
Appendix II: Trading System Modeling with Object-Process
Methodology
Trading Strategy Set
FinancialI Engineer Programmer Trader
Trading Application
Strategy Developing and Trading
TaigSrtg
Portfolio
Market Studies & Research Broke Data Vendor IT System
- - - - - - - - - - - - - - - - . . rN Market Value
Figure 21: Trading system development framework and process (OPM)
Figure 21 to Figure 28 represent trading system models generated in Object-Process
Methodology (OPM). This first diagram is the highest level diagram and the following diagrams
are created by zooming in or unfolding on this diagram, each of the following diagrams
represents the sub-systems and their internal process. This highest level diagram represents the
primary process and components (objects) in which it interacts. The system does two things.
First, it creates and modifies Trading Strategies, which is used within the system. Also, it is
possible that the trading strategies/models developed by his or her system could be exported into
another system. The Trading process also alters the Portfolio object, which is a representation of
all the positions held by those using the trading system. The entire purpose of the system is to
alter the Portfolio over time, such that its Market Value (a characteristic exhibited by the
Portfolio) is hopefully being increased over time.
In order to function, Trading also requires IT Systems, Data Providers, and Brokers as
instruments. These are all considered external to the system, since they typically exist and the
people implementing this system do not have control of them. The IT System is managed by the
IT Staff. This relationship is not represented with an explicit process, because it is not of much
importance when describing this system. Financial engineer, programmer and trader also are
involved in the Trading strategies and live trading process.
Figure 22: Trading system development framework and process - Unfolded (OPM)
This diagram displays a first level decomposition of the trading system. The two major functions
of such a trading system are separated. One part focuses on strategies and the other is for live
trading.
Figure 23: Trading diagram - zoomed-in (OPM)
This trading diagram is a zoom-in from figure 23 which represents the live trading process. And
the next figure (Figure 24) is another zoom represents the strategy developing process.
.......... .... ..............
Figure 24: Trading strategies developing diagram - Zoomed-in (OPM)
This diagram is a zoom-in of strategy developing from the upper level diagram. It is composed
of four sub-processes, including Data Cleaning, Strategy Developing, Backtesting, and
Executing. Typically, Executing is constantly acting, while the other three processes are running
in sequence to develop trading strategies.
Trading Application
Date C eaning Module - -
Visualization Module
Backtesting Module
I API Module :
Figure 25: Trading application - Unfolded (OPM)
74
In addition, the Proprietary software is decomposed into its own objects, which include API
Module, Data Cleaning Module, Visualization Module, Execution Engine, and Backtesting
Module. The API Module exists simply to connect the Data Providers to the Trading System,
which we did not believe warranted the introduction of a specific process. Therefore, it is
connected directly to the Data Provider using a structural link. The other four objects each are
directly related to sub-processes of Trading. This is intentional - since we are creating software it
makes the most sense to organize the application around the primary processes of the application
to facilitate development.
The Data Providers object can be specialized into two other objects, Historic Data Providers and
Live Data Providers. Each are used as instruments for different sub-processes of Trading.
Figure 26: Data cleaning - Zoomed-in (OPM)
This is a zoomed-in OPM diagram for data cleaning that contains sub-processes we have
discussed in early chapters based on SysML.
Matket Studie-
Figure 27: Trading strategies/models developing - Zoomed-in (OPM)
This diagram shows the process of how financial engineers generate investment ideas and risk
management models based on research. Then financial engineers create a design document that
will be used by prototyping. This process has been discussed in chapter 3.
The next two diagrams are Backtesting and Production, respectively. The processes themselves
are explained in chapters 3 and 4 and these two diagrams are presentations of the same processes
but in OPM format.
As discussed earlier, both SysML and OPM are powerful and capable. Choosing between them
depends on our background, the nature of the projects and your personal thinking process
preference.
Figure 28: Backtesting process - Zoomed-in (OPM)
Figure 29: Production/Deploy process - Zoomed-in (OPM)
......................
Bibliography
Michel M. Dacorogna, Ramazan Gengay, Ulrich M6ller, Richard B. Olsen, Olivier V. Pictet. An
introduction to High-Frequency Finance. Academic Pres. 2001
Perry J. Kaufman. New Trading Systems and Methods, Wiley. 2005
Andrew Kumiega, Benjamin Van Vliet. Quality Money Management: Process Engineering and
Best Practices for Systematic Trading and Investment. Academic Press. 2008
Charles D. Kirkpatrick, Julie R. Dahlquist. Technical Analysis: The Complete Resource for
Financial Market Technicians. FT Press. 2008.
Lawrence Bernstein, C. M. Yuhas. Trustworthy Systems Through Quantitative Software
Engineering. Wiley-IEEE Computer Society Press. 2005
John R. Wolberg. Expert Trading Systems: Modeling Financial Markets with Kernel Regression.
Wiley. 2000.
Tushar Chande. Beyond Technical Analysis: How to Develop and Implement a Winning Trading
System, 2nd Edition. Wiley. 2001
Bingcheng Yan, Eric Zivot. "Analysis of High-Frequency Financial Data with S-Plus"
Department of Economics, University of Washington. 2003
Galit Shmueli, Nitin R. Patel, Peter C. Bruce. Data Miningfor Business Intelligence: Concepts,
Techniques, and Applications in Microsoft Office Excel with XL Miner. Wiley-Interscience 2006
Sanford Friedenthal, Alan Moore, Rick Steiner. A Practical Guide to SysML: The Systems
Modeling Language. Morgan Kaufmann . 2008
Dov Dori , Object-Process Methodology. Springer. 2002
Edited by Christian Dunis and Bin Zhou. Nonlinear modelling of high frequency financial time
series. Wiley. 1998.
Edited by Pierre Lequeux. Financial markets tick by tick: insights in financial markets
microstructure. Wiley. 1999.
Thomas Stridsman, Trading systems that work: building and evaluating effective trading
systems. McGraw Hill. 2001.
John Kicklighter. "Top 5 Most Market Moving Indicators for the US Dollar". DailyFx.com. 2008
Cory Mitchell. "Trading The Non-Farm Payroll Report". Investopedia.com. 2009
Kathy Lien. "Trading On News Releases ". Investopedia.com. 2008
Free Online Forex Trading Course. FOREXHIT.com
http://www.forexhit.com/learn-forex/free-online-forex-trading-course.html#chart

High Frequency Trading System Design

Uploaded by

Copyright:

Available Formats

High Frequency Trading System Design

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

High Frequency Trading System Design

Uploaded by

Copyright:

Available Formats

High Frequency Trading System Design and Process Management

You might also like