Reservoir computing vs. neural networks in financial forecasting

Stock market prediction techniques are a major research area, thus, extracting time-dependent patterns for the existing predictive models is of major significance. In this work, we compare forecasting performance of the nonlinear model of recurrent neural networks (RNN) in two implementations, LSTM and CNN-LSTM, to the relatively novel approach of reservoir computing (RC), and in specific, the particular class of the echo state networks (ESN). This comparison focuses on exploiting data latent dynamics, in performing efficient training and high quality predictions of the evolution of real-world financial data. Applying a multivariate scheme to a stock market index without any stationarity techniques, a definite precedence of the ESN-RC over both types of RNN's in computational efficiency as well as prediction quality, emerges. Finally, the implemented approach is friendly to the trader, since specific values of a stock market timeseries provide with a frame allowing for in time forecasting, under real-world circumstances.