Training Neural Networks for Financial Forecasting: Backpropagation vs Particle Swarm Optimization

Neural networks (NN) architectures can be effectively used to classify, forecast and recognize quantity of interest in, e.g., computer vision, machine translation, finance, etc. Concerning the financial framework, fore- casting procedures are often used as a part of the decision making process in both trading and portfolio strategy optimization. Unfortunately training a NN is in general a challenging task mainly because of the high number of parameters involved. In particular, a typical NN is based on a large number of layers, each of which may be composed by several neurons , moreover, for every component, normalization as well as training algorithms, have to be performed. One of the most popular method to overcome such difficulties is represented by the so called back propagation algorithm . Other possibilities are represented by genetic algorithms , and, in this family, the swarm particle optimization method seems to be rather promising. In this paper we want to compare canonical back- propagation and the swarm particle optimization algorithm in minimizing the error on surface created by financial time series, particularly concerning the task of forecast up/down movements for the assets we are interested in.