Memetic Input Variable Selection in Neuro-Genetic Prediction System

摘要

This paper describes a hybrid neuro-genetic system applied to short-term prediction of a stock index. Prediction is made by an ensemble of three simple neural networks, each of which processes data selected by the evolutionary algorithm. The input data comes from the three stock markets and additionally includes two exchange rates. Besides the review of results, which have been published in detail elsewhere, some aspects of the system which appeared to be vital for accomplishing high efficiency axe thoroughly examined and discussed. These include: autonomous mechanism for extraction of technical analysis patterns, memetic-type local improvement phase in the evolutionary process, and incorporation of domain knowledge into memetic-evolutionary variable pre-selection.