Dimensionality Reduction for Time Series Decoding and Forecasting Problems

Abstract

The paper is devoted to the problem of decoding multiscaled time series and forecasting. The goal is to recover the dependence between an input signal and a target response. The proposed method allows receiving the predicted values not for the next timestamp but for the whole range of values in forecast horizon. The prediction is a multidimensional target vector instead of one timestamp point. We consider the linear model of partial least squares (PLS). The method finds the matrix of a joint description for the design matrix and the outcome matrix. The obtained latent space of the joint descriptions is low-dimensional. This leads to a simple, stable predictive model. We conducted computational experiments on the real dataset of energy consumption and electrocorticograms signals (ECoG). The experiments show significant reduction of the original spaces dimensionality, while the models achieve good prediction quality.