In this thesis Genetic Programming is used to create trading systems for the EUR/USD foreign exchange market using intraday data. In addition to the exchange rates several moving averages are used as inputs.
The developed evolutionary algorithm extends the framework ECJ. The created trading systems are being evaluated by a fitness function that consists of a trading simulation. Genetic operators have been adapted to support "node weights". By using these on the one hand macromutaion is tried to be reduced on the other hand the interpretability of the created trading systems is tried to be improved.
Results of experiments show that created trading systems are apparently successfull in profitably using informations contained within the exchange rates. Profits of the created trading systems are maximized by using the optimal position size. It is shown that if the minimum investment period is met the achieved results are optimal even when taking into account the used risk adjusted performance figure.
Table of Contents
1 Introduction
1.1 Motivation
1.2 Objective and structure
2 Basic principles and state of the art
2.1 Genetic Programming
2.1.1 Program Structure
2.1.2 Initialization of the GP Population
2.1.3 The Genetic Operators
2.1.4 Fitness Function
2.1.5 Selection
2.1.6 Process of the algorithm
2.1.7 Crossover, building blocks and schemata
2.1.8 Approaches against macromutation
2.1.9 Modularization
2.1.10 Further approaches for improvement
2.2 Artificial Neural Networks
2.2.1 Components of neural networks
2.2.2 Network topologies
2.2.3 Learning methods
2.3 Trading Systems
2.3.1 Tape Reader
2.3.2 Market timing
2.3.3 Position sizing
2.3.4 Comparison of trading systems
2.3.5 Fundamental versus technical analysis
2.3.6 The Currency Market
2.3.7 Approaches for the development of trading systems
3 Draft
3.1 Overview
3.2 Requirements on the software
3.3 Conception of software
3.3.1 The Evolutionary Algorithm
3.3.2 The fitness function
4 Implementation
4.1 Components of the developed software
4.2 Classes of the exchange rate data server
4.3 Classes of the Evolutionary Algorithm
4.4 Overview over the framework ECJ
4.5 Problems during experiments
5 Experiment results
5.1 Results with node weights
5.1.1 Results of the training time period
5.1.2 Results of the validation time period
5.1.3 Results of the test time period
5.1.4 Results as monthly turnovers
5.1.5 Created trading rules
5.2 Results without node weights
5.2.1 Results of the training period
5.2.2 Results of the validation time periods
5.2.3 Results of the test period
5.2.4 Results as monthly returns
5.2.5 Created trading rules
5.3 Identification and application of optimal f
6 Discussion and evaluation
6.1 Outlook
7 Summary
Objective and Research Scope
This thesis aims to apply Genetic Programming (GP) to the automated development of trading systems for the financial market, specifically the currency market (EUR/USD), by analyzing their profitability through historical simulations. The core research question addresses whether GP can generate effective, interpretable trading rules that adapt to changing market conditions, while maintaining a balance between system complexity and practical applicability.
- Application of Genetic Programming for automated trading system generation.
- Use of historical exchange rate data and technical indicators for backtesting.
- Implementation of a modular, software-based framework ("EVAM") for evolutionary development.
- Evaluation of "node weights" to improve interpretability and control macromutation in GP.
- Integration of risk-adjusted performance metrics, including the optimization of "optimal f" for position sizing.
Excerpt from the Book
1.1 Motivation
The natural evolution has turned out to be a most successful mechanism for the engenderment and adaptation of creatures to the environment. Without receiving any particular instructions or even precise objective definitions, it has succeeded in finding sophisticated solutions for problems existing in the real world.
Genetic Programming (GP) is an approach for using the creative power within the natural evolution for the automatic development of computer programs (cf. (Koz92, Chapter 1-6)). It is used to try to simulate mechanisms of the natural evolution in order to generate automatic programs solving a given problem. In a series of applications, GP has been used for solving mathematical problems as well as for solving real-world problems successfully. Among them are counted such problems as symbolic regression, (cf. (Koz92, cf. Chapter 10)), classification (cf. (Koz92, Chapter 17)), the synthesis of artificial neural networks (cf. (Gru94, Chapter 2 following)), pattern recognition ((Tac93, pages 2 to 10)), robot control (cf. (BNO97, pages 2 to 10)) and the generation of images (cf. (GH97, pages 2 to 7)) are counted among these problems.
Automated learning by means of GP can be interpreted as heuristic search algorithm finding out of the set of all possible programs those offering the best solution for the given problem. Dependent on the given problem, the search range is very large and oftentimes neither continuous nor differentiable and thus the search range of all possible programs is ill-fitting for classical search algorithms (cf. (LP02, page 2 following)).
Summary of Chapters
1 Introduction: Discusses the motivation behind using evolutionary algorithms for automated trading and defines the objectives and structure of the thesis.
2 Basic principles and state of the art: Provides a comprehensive overview of Genetic Programming, Artificial Neural Networks, and technical trading systems as the theoretical foundation.
3 Draft: Outlines the conceptual design and software requirements for the evolutionary asset management system (EVAM).
4 Implementation: Details the technical realization of the software components and the framework integration.
5 Experiment results: Presents the findings from simulations with and without node weights, including performance data and created trading rules.
6 Discussion and evaluation: Analyzes the experimental results and provides an outlook on future potential developments.
7 Summary: Concludes the thesis by revisiting the main findings regarding the applicability of GP in currency trading.
Keywords
Genetic Programming, Trading Systems, Currency Market, Artificial Neural Networks, Backtesting, Evolutionary Algorithms, Node Weights, Technical Analysis, Financial Forecasting, Position Sizing, Optimal f, EUR/USD, Automated Trading, Machine Learning, Optimization
Frequently Asked Questions
What is the primary focus of this thesis?
The thesis focuses on the automated development of trading systems for the currency market using Genetic Programming (GP) to generate profitable and interpretable trading rules based on historical exchange rate data.
Which financial market is analyzed in this work?
The study focuses specifically on the EUR/USD currency pair, utilizing high-frequency historical data to develop and test trading strategies.
What is the main objective of the research?
The primary objective is to design a framework that can generate trading systems capable of adapting to changing market conditions, while balancing the trade-off between rule complexity and the user's ability to interpret those rules.
Which machine learning methodology is utilized?
The work employs Genetic Programming (GP) as a heuristic search algorithm to evolve sets of trading rules. Additionally, it integrates "node weights" to control the evolution process and improve rule interpretability.
What is covered in the main body of the work?
The main body covers the theoretical principles of GP and neural networks, the design and software architecture of the "EVAM" framework, the technical implementation details, and an empirical analysis of experiment results comparing different settings.
Which keywords best describe this research?
Key terms include Genetic Programming, Trading Systems, Currency Market, Backtesting, Evolutionary Algorithms, Position Sizing, and Optimal f.
How does the author handle the problem of overfitting in trading strategies?
The author employs a strategy of rolling time periods (training, validation, and test) to ensure that models are evaluated on unseen data, thus avoiding over-optimization and ensuring better generalization.
What are "node weights" in the context of this GP implementation?
Node weights are numerical properties assigned to individual nodes in the rule tree. They are used to influence the likelihood of mutation or crossover occurring at specific points, thereby helping to preserve successful subtrees and enhance interpretability.
Why does the author prefer GP over Artificial Neural Networks for this task?
While both methods perform well, the author argues that the rules generated by GP are more transparent and interpretable than the "black box" nature of neural networks, which improves user confidence in the generated systems.
- Quote paper
- Diplom Informatiker Holger Hartmann (Author), 2007, Development of Trading Systems using Genetic Programming with a Case Study, Munich, GRIN Verlag, https://www.grin.com/document/81369
