A Novel Algorithmic Trading Framework Applying Evolution and Machine Learning for Portfolio Optimization
MetadataShow full item record
The goal of this thesis is to implement an automated trading system able to outperform the benchmark uniform buy-and-hold strategy. Performance is measured in term of multiple risk and return measures. A comprehensive signal processing framework is built to facilitate rapid development and testing of multiple candidate trading systems. A subset of the most promising trading systems, 14 in total, are presented in this thesis.The trading systems are not only designed to function as decision support systems for human trader, but also to function as autonomous traders in their own right. They are built to be applicable in real life, any academic value is a welcomed side effect. Many of the trading systems are heavily influenced by machine learning methods, with a special focus on time series prediction. Support vector machines (SVM), artificial neural networks (ANN), and regression are the primary methods applied for low level machine learning. Evolutionary algorithms are applied as high level meta-parameter tuning. In addition to machine learning methods, classical portfolio optimization methods using modern portfolio theory are applied, some in combination with machine learning methods.The trading systems are tested by means of simulation on data from two stock markets -- the OBX index on Oslo stock exchange and Dow Jones Industrial Average (Dow). Results indicate that the Dow index is highly efficient. No trading system was able to outperform the benchmark when using transaction costs. Results on the OBX index indicate some inefficiencies with potential for exploitation. Many trading systems manage to generate higher return than the benchmark. Some achieve excess return with the same level of risk. The efficient portfolio trading system yields highest return given the same level of risk as the buy-and-hold portfolio. In general results tend to support the efficient market hypothesis. Given the same risk level, experiments from this thesis finds that efficient portfolios can outperform uniform ones in inefficient markets as these are not efficient enough in themselves as in the case of OBX. No strategy was found that can substantially outperform an efficient portfolio.Support Vector Machines are identified to be the best time series predictors among the machine learning techniques used. Evolution is found to be a powerful tool in identifying and training trading systems. Future work will focus on evolution as an integral part of the trading systems.