Time series granulation-based multivariate modelling and prediction

The typical characteristics of time series data exhibit a large data size, high dimensionality, and high correlation. To better extract high-level representative information for time series, this study proposes a novel granular vector autoregressive (GVAR) model, which incorporates granular computing with vector autoregressive (VAR) models to predict the main varying ranges of the multivariate time series. The proposed model first utilises the principle of justifiable granularity to construct information granules, which capture the cardinal information hidden in the time series. Then, the granular VAR model is built based on the upper and lower bounds of the constructed information granules simultaneously. Here, the interval least squares (ILS) algorithm is employed to estimate the model's coefficients, and the regressive order is determined by the Bayesian information criterion (BIC). Finally, experimental studies are conducted to illustrate the effectiveness and practicality of the proposed prediction model.