Time series prediction with signal-to-noise ratio maps and high performance computing

Time series prediction methods applied to chaotic signals affected by noise use a continuous pattern function as the least-squares estimate of an unknown deterministic map. The noise variance around the continuous pattern function is not always constant but may exhibit spatial variability, which directly affects prediction performance. In this paper we propose a novel approach for increasing predictor performance using a multi-resolution signal-to-noise ratio (SNR) map of phase space. We calculate it using a parralel algorithm on a high performance computing cluster and in the final stage of the approach we use a novel feature selection algorithm to build a kernel machine. We show the selected features form sparse kernel machines which outperform existing methods for the prediction of noisy financial data.