Recurrent networks appears to performs better than more commonly used tapped delay line feeforward, network in prediction f finanical time series, as stock price variations. We present an incrementally built network that improves predictions significatively as several recursive hidden layers are added to a conventional multilayer with tapped delay line iputs. Other improvements, produced by the use of adequaate preprocessing of data, are sown as well. We discuss the results in terms of the inflence upon the prediction of additional variables, like general indexes. Also we show the generalisation capabilities of the neural network, in particular the ability to make predictions on different stock than the one used for training.
展开▼
