GMDH-based financial forecasting on a hypercube parallel computer

摘要

In this paper we describe the Group Method of Data Handing (GMDH) as a data mining and forecasting technique for financial markets. Stock prices and international indices may depend on many other stock prices and indices; the question is: which other stocks and indices, taken from a large set of possible candidates, and what are approximately the relationships (preferably in mathematical terms). On the basis of these deduced relationships, predictions could be made with respect to future behavior of stocks and indices. GMDH is a neural-network like approach, which is subject to intensive training, testing and trimming, based on a set of training and test examples. This is a very time-consuming process, which needs parallel and/or distributed computing in order to reduce training time. In this paper, an object-oriented implementation on a hypercube parallel computer (64-vertices nCUBE2) is presented. The object-oriented set-up and fast response through parallel computing together guarantee the necessary experimentation flexibility in order to find the appropriate GMDH-configuration dependent on the problem at hand. Some illustrative prediction results and parallel performance results are presented.