Evolutionary optimization of trading strategies
- Publication Type:
- Issue Date:
NO FULL TEXT AVAILABLE. This thesis contains 3rd party copyright material. ----- A good many technical trading strategies have been proposed in financial literatures and trading houses to support trading investment decisions. In the stock market, a good trading strategy may make great profits while a bad one could result in disastrous losses. Therefore, it is critical for stock traders to find or tune trading strategies to maximize the profit and/or to minimize the risk. Unfortunately, it is very challenging for financial professionals to work out trading strategies dependable in the real world, not to mention non-professional investors. This is due to many reasons, for instance, the dynamic market environment, comprehensive constraints, huge quantities of data, sensitive and inconstant strategy performance, and complicated computational settings and development. Consequently, it is not practical for traders to work out an 'optimal' trading strategies in a constrained market by themselves. Therefore, the backtesting and optimization of trading strategies are very significant for stock market participants. To this end, data mining and related approaches have been demonstrated to play prominent roles in optimizing trading strategies. The existing data mining related research on the optimization of technical trading strategies mainly concentrates on extracting interesting trading patterns of statistical significance, demonstrating and pushing the use of specific data mining algorithms. As a result, many interesting trading strategies are found, while few of them are dependable in the market. The gap between the academic findings and business expectations comes from a few reasons, such as the over-simplification of optimization environment and evaluation fitness. In particular, no work has been on providing a comprehensive analysis and processing of trading strategy optimization, which consist of problem definition, data. storage and pre-processing, parameter tuning, computational efficiency, and system implementation, etc. This thesis presents a systematic view of the above issues and corresponding solutions in trading strategy optimization through developing new and varying storage and evolutionary computing and system implementation techniques. One of key characters of this work is its practicability of the developed techniques and resulting findings, which have been trained and tested in terms of huge amounts of market data, various organizational factors, and realistic evaluation metrics. In particular, the following problems have been addressed in this thesis, which more or less have not been analyzed or tackled very well in the existing research and development. First, the problem definition of trading strategy optimization is addressed in terms of considering not only attributes enclosed in the optimization algorithms and trading strategies, but also constraints in the target market where the strategy to be identified and used. This presents a multi-attribute and multi-objective constrained environment for trading strategy optimization. Second, very detailed and enormous stock transaction data challenges efficient data storage, access and preparation for trading strategy optimization. This problem is answered by a new dynamic storage method which provides a flexible mechanism to balance between storage space and access efficiency. Third, genetic algorithm (GA) based evolutionary computing is used to deal with the high dimension reduction of multi-attribute and multi-objective trading strategy optimization. However, the parameter variations and discrete fitness surfaces of GA-based optimization make it very time and computing consuming. To this end, several techniques have been developed to fine tune the GA-based optimization algorithms towards a more efficient and effective way. Fourth, the back testing and optimization of time consuming trading strategies is very challenging. To address this challenge, parallel GA algorithms have been developed, which can efficiently extract target trading strategies in changing market conditions. Fifth, the results given by GA are crisp and subject to the disturbance of market dynamics and user preferences. This makes the findings unde-pendable in the real markets. In order to obtain more stable solutions, an enhanced CA algorithm with more robust deployment capability is proposed. Finally, it is very difficult for a normal trader to build a system to deal with the above challenges and support her/his optimization jobs. A multi-agent system is proposed to address this issue. It integrates all the abovementioned technologies, and can help users easily specify and execute optimization jobs to their advantages. In summary, this thesis explores in detail how evolutionary computing can be effectively and efficiently applied to the optimization of technical trading strategies. It provides not only a systematic processing of this optimization problem, but also innovative and practical solutions to various impediments that arise out of the optimization process.
Please use this identifier to cite or link to this item: