Predictive Root Based Bootstrap Prediction Intervals in Neural Network Models for Time Series Forecasting

Time series (TS) modelling is an important area in the domain of statistics, as it enables us to comprehend the dynamics underlying a particular phenomenon. In the spectrum of non-linear TS data analysis, neural network (NN) models are one of the dominant methods due to their several advantages over statistical methods. However, NN models are unable to provide prediction intervals (PIs) which is an important part of forecasting to capture uncertainties. The predictive root concept earlier used by researchers for both linear and non-linear autoregression models has been extended to ANN models for constructing PIs. Two bootstrap approaches (with and without rescaling) for constructing PIs in ANN models for non-linear TS have been proposed. The performances of the proposed methods have also been evaluated by comparing them with the existing methods using both simulated and real datasets. The proposed methods can be considered as a viable alternative for computing PIs in TS.