基于Granger因果关系的多变量时间序列预测模型

MULTIVARIATE TIME SERIES PREDICTION MODEL BASED ON GRANGER CAUSALITY

时间序列数据包含内在的时序结构,而传统的针对多变量时间序列的预测方法没有考虑变量序列的历史观察值的影响。为此,提出一种基于Granger因果关系挖掘的多变量时间序列预测模型。通过选择有效的因变量并加入其滞后观测期来提高支持向量回归对目标序列的预测,同时也提供了较好的因果解释性。理论推导和实验结果表明,该方法不仅能获得比传统方法更精确的预测效果,而且减少了参与运算的变量时间序列。

Time series data contains inherent temporal ordering. Traditional prediction methods for multivariate time series do not consider the influence of their historical observations. In this paper, we propose a multivariate time series prediction model which is based on Granger causality mining. By selecting effective cause variables and adding their lagged observations, the model improves the prediction of target time series when using support vector-regression. Meanwhile it provides a good interpretable causal-effect relationship as well. Theoretical inference and experimental results demonstrate that our method not only achieves better prediction effects than the traditional methods, but also reduces the number of variable time series involved in the computation.