In todays, worldwide stock markets have experienced dramatic volatility in their returns. Traditionally, two main approaches- time series analysis and fundamental analysis - exist in predicting stock price. But the results of statistical analysis are not quite satisfactory to meet our expectation. Besides, they have some limitation of applications according to the data characteristics and also require comparatively strict assumptions on the distribution. As a result, artificial intelligence (AI) technologies are introduced in this area.
Over the past four decades, the field of AI has made a great progress toward computerizing human reasoning. Especially, neural network (NN) have newly received special interests. Because of its non-linear learning and smooth interpolation capabilites, NN are supplementing or taking the place of statistical and conventional expert system (ES) approaches in many financial decision making. Stock market prediction is one of the such problems. So, we choose NN as a means to test whether it could produce a successful model in which their generalization capabilities could be used for stock market prediction.
Five types (NN1, NN2, NN3, NN4, NN5) of independent modular networks are developed to predict KOSPI (Korea Composite Price Index)'s up/down direction after four weeks. These type of network only differ in learning period. 156 number of networks have learned and predicted.
NN5 - arithmetic average derived by varying period - showed higher accuracy (79.45%) than MLRs. In buying and selling simulation, buying-and-selling using NN5 and MLR2 produced higher return than buy-and-hold strategy, particularly, NN5 can attain 2.4 times higher return than buy-and-hold's one.
Today, Korean investor institutions are bustling to face with open market environment and the departure of stock price index futures on KOSPI 200. The development of trading system which supplement the limitation of this study can lighten their endeavor.