A Novel BNN-based Probabilistic Prediction Model for Non-linear and Non-stationary Time Series

Abstract

Improving the prediction accuracy of time series is of great practical significance and theoretical value for people to prevent risks and supervise markets. However, due to the diversity of influencing factors of time series, time series often present nonlinear and non-correlated characteristics. This paper is committed to introducing nonlinear and non-stationary sequence processing methods through deep learning. First, the TCN (Temporal Convolutional Network) model, which is more suitable for sequence prediction, is selected from the basic model architecture. Then, construct BNN (Bayesian Neural Networks) to output the prediction distribution with confidence intervals. We chose two publicly available datasets for our study. Compared with N-BEAT, LSTM, and TCN methods, this paper has achieved higher prediction performance than the original methods. The distribution of confidence in the output enriches the output form and provides data support for subsequent research.