Forecasting High-Frequency Financial Time Series with Evolutionary Neural Trees: The Case of Heng-Sheng Stock Index

In this paper, the evolutionary neural trees (ENT) are applied to forecasing the high-frequency stock returns of Heng-Sheng stock index on December, 1998. To understand what may con-sistute an eeective implementation, six experiments are conducted. These experiments are diierent in data-preprocessing procedures, sample sizes, search intensity and complexity regularization. Our results shows that ENT can perform more eeciently if we can associate ENT with a linear lter so that it can concentrate on searching in the space of non-linear signals. Also, as well demonstarted in this study, the infrequent bursts (outliers) appearing in the high-frequency data can be very disturbing for the normal operation of ENT. 1 Motivation and Introudc-tion There has been such a rapid increase in the application of artiicial neural networks to-nance that ANNs have now become a standard tool for nance practitioners. However, the credibility of this application crucially depends on the style in which the network is designed. When what concerns us is not just one but many factors which may determine the performance of neural nets, it is very diicult, though not intractable, to design them manually be-cuase of combinatoric complexity. Therefore, it is desirable to know whether an automatic procedure , such as evolutionary computation, can help, and this motivated the many applications of evolutionary artiicial nural nets (EANNs) to nance. However, based on a survey article by 3], who distinguished three kinds of evolution in EANNs, most nancial applications are only concerned with lowest level of evolution, namely, connecting weights. 1], to our best knowledge, is the rst one to explore the potentials of EANNs at a high level of evolution, namely, evolution of archiectures and learning rules. By setting a forecasting competition on forgein exchange rates, they compared the performance of 8 backpropagation feedforward neural nets (BPNNs), 8 EANNs and the random walk model based on ve chosen criteria. It is found that all nerual network models can statistically beat the RW in all criteria at the 1% signiicance level. In addition, among the 16 nerual netowrk models generated in diier-ent designs, the best model is the EANN with the largest search space. In light of the earlier evidence of 1], this paper would like to pursue this line of research, while with a diiernt focus. Instead of compar