Evolutionary Computation in Financial Engineering: A Roadmap of GAs and GP

In this and previous issues, R. Baker and S. Smith has introduced the techniques of evolutionary computation and their application to nancial engineering. In this article, we would like to rst summarize a few major insights revealed in those two articles and then provide a roadmap for readers to initialize their own trails. First of all, why evolutionary computation? Where is the motivation? In engineering or applied mathematics, evolutionary computation has been extensively applied to problems whose solution space is irregular, i.e., too large and highly complex, so that it is di cult to employ conventional optimization procedures to search for the global optimum. For those who are interested in nancial econometrics, Dorsey and Mayer (1995) is a good reference to help one have a taste of this kind of solution space. In this study, Dorsey and Mayer listed several irregular features of the likelihood function (the search space) frequently encountered in econometrics. They selected 11 test problems, and showed the advantages of genetic algorithms over many general-purpose direct-search algorithms such as simplex, adaptive random search, and simulated annealing. Now, are solution spaces for most nancial optimization problems irregular? My answer is yes, and I am not alone to entertain such a view. General acceptance of this property has in fact fostered the growth of the eld nancial engineering. Take nancial forecasting for example. There is a tendency to jettison the early well-accepted random walk hypothesis or the unpredictability of nancial time series. In the literature, nonlinear models such as arti cial neural networks have proved to be capable of beating random walks not only statistically signi cantly but also economically signi cantly (Rauscher, 1997). Su ce it to say, many nonlinear models have the potential to beat the random walk but the search space of nonlinear models is doubtlessly large and complex. Similarly, in technical analysis, search over the space of trading rules is obviously a daunting task. But, is evolutionary computation the only choice for us to cope with the irregular search space? Of course not. There are many equally promising tools available in the machine learning literature. However, evolutionary computation should not be considered as a kind of optimization technique to compete with other alternative techniques but an optimization principle to be incorporated with existing techniques. It is more a complement to than a substitute for many existing techniques. For example, the application of evolutionary computation to arti cial neural networks has generated one of most promising research areas known as evolutionary articial neural networks, (EANNs) (Yao, 1993), to which we shall come back later. For now let me introduce evolutionary computation as an optimization principle.