MULTILAYER PERCEPTRONS AS APPROXIMATIONS TO PROBABILITY DENSITY FUNCTIONS IN TIME SERIES FORECASTING
The paper presents the method of utilisation of multilayer perceptron neural networks to probability densiity function approximation in the problem of time series forecasting. The theoretical background has been given and the specification of neural prediction model, which generates the probability distribution of the forecasted variable in the issue of financial time series predicition, has been described. Next, the research concerning the performance of such model designed for the forecasting of the Polish stock index WIG has been discussed. Two versions of the model have been applied: first - comprised of 12 perceptron networks with single output each, second - based on one network with 12 outputs. Three test cases (for subsequent stock exchange sessions ) have been analysed. Obtained probability distributions are somewhat similar to empirical distribution (achieved for model development data), but they clearly indicate predicted tendency of index change and show specific uncertainty of the forecast.
CEJSH db identifier