Shared state space model for background information extraction and time series prediction

Abstract

Time series prediction is important for financial analysis, climate forecasting, and so on. Existing works mainly focus on the prediction of target time series based on the prior of the sequence itself, but ignore the background information behind the sequences. Such ignored information is usually essential to build a robust prediction model in complex real-world applications. However, how to extract the shared background information behind multiple sequences and how to incorporate the extracted information in the prediction model are two main challenges. To address the above two challenges, we propose a shared state space model (SSSM) by introducing a shared background information component into the state space model. In SSSM, we consider all sequences as a whole and model each target series by utilizing a state space model with shared same parameters and background information. First, we employ two recurrent neural networks to extract the temporal characteristic of the target sequence as well as the background information. Second, the above extracted information is integrated into a state space model in the form of a linear Gaussian component, whose inference procedure is accomplished by Kalman Filter. Finally, the model is optimized following a log-likelihood of the model with the above two components. Experiments on real-world applications show that our model can extract the share information behind the data and outperforms the state-of-the-art methods.