Technical Rules Optimization Using Intelligent Hybrid Systems

The optimization approaches based on evolutionary computations and artificial neural networks have conquered wide popularity in problems, whose solution space is so complex and large that it is impossible to use traditional methods of optimization. Each approach individually possesses the advantages and deficiencies. For example, Giles and Lawrence (2001) showed that ANN could be successfully used in short-term series prediction. The application of ANN is especially effective, when the links between influencing and dependent variables are nonlinear and badly noised, what is typical for a stock market. However the main drawback in the ANN application is that they do not give us understanding of the underlying process and obtaining of specific collection of rules, therefore decision-making completely relied on ANN functioning is not desirable. In turn, the application of genetic programming (GP) eliminates these problems and allows creating more complex rules from simple ones, but it requires great computational capabilities. The use of hybrid systems could help to avoid the weaknesses inherent in each methodology, whilst capitalizing on their individual strength. Stock market is an example of complex system where the process of decision-making is concerned with great difficulties, since numerous factors and high level of noisiness influence future changes of stock indexes. Technical analysis (TA) is one of the most popular methods, which is widely used by stock traders (Dempster and Jones, 2001). It proposes a set of numerous investment rules to investor. Use of individual trading rules is not profitable from investor’s point of view (Dempster and Jones, 2000) while numerous rules application is complicated by separate rules optimization and their effective combination problem. In this work the authors propose their own optimization approach of investment rules on a stock market combining GP, ANN and TA. The developed system was tested on real stock market indexes for 20 years period in economically realistic mode. The results obtained were checked with regard to the statistical significance and compared with the results of the systems solely based on each of the above methods.