USING PERCENTAGE ACCURACY TO MEASURE NEURAL-NETWORK PREDICTIONS IN STOCK-MARKET MOVEMENTS

A speculator on a Stock Market, aside from having money to spare, need s at least one other thing - a means of producing accurate and underst andable predictions ahead of others in the Market, so that a tactical and price advantage can be gained. This work demonstrates that it is p ossible to predict one such Market to a high degree of accuracy. For t he purpose, the Financial Times - Stock Exchange (F.T.-S.E.) 100 Share Index in the UK, known as 'The Footsie', was selected. Neural network predictions were obtained for the daily Market close 5 days ahead, an d 25 days ahead, as measured in mean square error and in root mean squ are error. To measure percentage accuracy, each individual test case p rediction was compared with the actual market outcome, and total perce ntage accuracy for the whole test set was similarly calculated. Compar isons were also drawn with predictions for the same test cases using f our types of Multiple Linear Regression. The neural network results in dicated that predictions based upon the lowest mean square error bear little relationship to the same test cases, when measured in terms of overall percentage accuracy. For the lay person, or a Stock-Market spe culator, it was also shown that predictions can be produced to a high level of accuracy, in a readily understandable format.