MODEL COMPARISON FOR MONTHLY FORECASTS OF THE CAC 40
Michel Crucianu, J. P. Asselin de Beauville, Romuald Boné · 1998
sets of neural networks and we use the mean predictions. Also, we develop models on different training sets and test them on different tests sets before concluding that one class of models is superior to the others. We study the problem of generating monthly forecasts of the CAC 40 financial index. We compare the results obtained by linear models and several feed-forward and recurrent neural networks. We find that the non-linear models significantly outperform the linear ones, as well as a conservative transaction policy, providing thus a basis for profitable monthly transactions. Instead of developing single networks for this task, we develop network ensembles and we use the variance of the individual network outputs to establish confidence limits for the predictions. The availability of ensembles of trained networks also allows us to give an estimate for the confidence limits of each prediction. 2. PRELIMINARY ANALYSES AND EXPERIMENT DESCRIPTION Several time series were available...