Echo State Networks and Existing Paradigms for Stock Market Prediction

Stock market is essentially a chaotic and a highly unstable time-series. People have used approaches such as clipping gradient i.e., stopping the gradient from getting too small or too high, via gated recurrent units and LSTM units, to predict the stock prices. However, information is still lost and hence is not an ideal approach. This paper aims to compare the efficiency of standard approaches of stock prediction namely, regression, recurrent neural networks using long short-term memory units, with the echo state network (ESN) which uses reservoir computing to model chaotic non-linear systems such as daily closing stock prices in a stock market. My experiments on the NY Stocks dataset demonstrates as to how the novel echo state network outperforms the conventional regression and neural network approach.