An Adaptive Time-Variable Weight Combination Forecasting Method for Time Series

A Bayesian combined forecasting model through using the method of Bayesian time-variable weight is established for forecasting time series in this paper, where the weights are computed using an adaptive update of the Bayesian posterior probability of each predictor, based on their past predictive performance. This method can avoid the singular forecasting model's performance change over different time intervals, respectively, which track the prediction performance of the combined models and adjusts their credit values (weights) according to a proposed set of assignment algorithm so as to select the models with higher accuracy for combination. We use this probabilistically motivated predictor for time series forecasting with the combination of autoregressive integrated moving average (ARIMA), gray model (GM(1,1)) theory and radial basis function (RBF) neural network. Compared for transportation freights volume of China, according to error statistics evaluation index system, the performance of time-varying weights combination forecast model outperforms the linear and nonlinear models used individual or combining, and therefore, this approach offering a vigorous support decision to transportation enterprises and management organization