In order to evaluate our approach, the best-generated experts were compared with the passive buy-and-hold strategy. To reduce the time-period influence on the performance of the passive strategy, ten time-series have been randomly selected from the period starting from 2 January 1997 to 8 November 1999. The results demonstrated that genetically evolved experts were in general able to trade much better than the buy-and-hold strategy.
The generated models have simple structures and are composed of relevant easy-tounderstand decision rules. We have tested not only the performance of the models on the learning data, but also their generalization ability on the test data. The results of out-of-sample testing were usually worse than those on the learning phase due to the presence of ‘noisy’ data and incompleteness of information contained in the financial time series. It is also apparent that the generated experts have a finite lifetime of successful trading. The daily expert performance statistics show that the ‘oldest’ experts demonstrated gradually deteriorating performance in subsequent days in comparison with the newly generated ones. On most of the tested time series, the expert lifetime was no longer than two weeks.
A great number of genetically evolved models based on the different financial rules, have been generated. The knowledge base, composed of over 200 decision rules concerning technical and economical situations, was used to generate 400000 experts for each stock. Computer time and cost were not critical issues. The experiments were performed on a PC with K7 AMD processor ð650 MhzÞ with 256 Mbyte memory. Expert generation time ranged from 5 to 7 min:
The evolution-based technology in stock trading is still at an experimental stage. We have shown that genetic algorithms can be a powerful tool for short-term traders. Further research is needed not only to build a solid theoretical foundation in knowledge discovery in financial time series, but also to carry out time-consuming model validation on real-life data.
Our research can be extended in several directions. From the empirical point of view, our approach needs to be tested on other markets; in fact, it is not really clear that market organization is of no significance with regard to the results of technical trading. More precisely, order-driven and price-driven markets may have different characteristics relative to technical trading. However, as long as daily prices are considered, as in our study, this point is of no great importance. In fact, if genetic algorithms are to be applied to intraday data, the bid-ask spread has to be considered; until now, it seemed very difficult to tackle this problem, because of the considerable quantity of data requiring consideration.
Probably, a more promising direction for our research is to consider portfolio management by means of genetic algorithms.[2] In this paper we have developed experts who give advices concerning individual stocks; a natural idea is to use these advices to manage a portfolio in a dynamic way. However, the measurement of performance is more complicated as one has to take into account not only the portfolio risk to evaluate its performance, but also final wealth.
This involves comparing the return to a benchmark portfolio, for example, a stock index (see Reference [23] for a survey on this problem) and searching for optimal solutions in the set of portfolios which exhibit the same risk.
ACKNOWLEDGEMENTS
We thank two anonymous referees for helpful comments.
REFERENCES
1. Fama E. Efficient capital markets: a review of theoretical and empirical work. Journal of Finance 1970; 25(2):383– 417.
Уважаемый посетитель!
Чтобы распечатать файл, скачайте его (в формате Word).
Ссылка на скачивание - внизу страницы.