Modelling Subjective Time Series Prediction
van Heusden · International Journal of Modelling and Simulation · 1982
Human supervision of slowly responding dynamic systems includes prediction of the output variables. In experiments, subjects were asked to make predictions of graphically displayed time series. The purpose was to find the influence of the statistical properties of the processes, used to generate the time series.For this reason, the “output” of the subjects is modelled with respect to the input. The following models were investigated:- discrete transfer funtion model of MA- (Moving Average) and ARMA-type (mixed Auto-Regressive Moving Average);- extrapolation by means of splines; two versions were tested;- optimal prediction model; in which the order and parameters were updated;- limited memory model, a variant of the former model;- fading memory model, a model elaborated by Rouse;- pattern recognition models.Here, the selection of the most suitable model depends on the objectives of the application: the ARMA-model leads to the best description of the actual output (minimal sum of squared residuals). However, if the question is: how do people ,predict in such situations, in other words is the validity of the model good for new situations, then other models become more important: the results show that the residuals of the other models do not differ very much; the last th~ee models were better to some extent. Furthermore, the subjective prediction errors were analysed and compared with the properties of the time series.