Echo State Networks for predicting financial time series

Abstract

A general problem occurring when training the recurrent neural networks (RNN) is that the solution space is extensive and the chance of choosing a local minimum instead of a global minimum is high. This is due to the fact that the weights among the neurons are variable. ESN networks are solving this issue by training the weights of the connections among the reservoir and the neurons on the output layer. The reservoirs containing fewer neurons are generalizing better with new data that the reservoirs with high number of neurons, indicating the fact that, as for other FNN networks, the overspecialization phenomenon may occur

Keywords

• recurrent neural networks
• echo state networks
• liquid state machine
• time series prediction