Analysis of multivariate time series predictability based on their features

Abstract

In this study we explore the features of time-series that can be used for evaluation of their predictability. We suggest using features based on Kolmogorov-Sinai entropy, correlation dimension and Hurst exponent to test multivariate predictability. Besides we use two new features such as ‘noise measure’ and ‘random walk detection’. Then we experimentally test the accuracy of multivariate time series forecasting models, including vector autoregressive model (VAR), multivariate singular spectrum analysis (MSSA) model, local approximation (LA) model and recurrent neural network model with long short term memory (LSTM) cells. At last we test different causality methods for choosing additional time series as the predictors and claim that the relevance of taking into account additional predictors highly depends on the characteristics of the target time series and can be estimated using the developed method. The results of the work can be used as theoretical and experimental basis for the development of forecasting applications for the short time series using a combination of corporate and open source data as additional data predictors.