Robust Deep Time Series Model Selection for Effective Decision Support

Time series data consists of sampled data points taken over time. Time series modeling and prediction are one of the most fundamental problems and have been subject to intensive researches. The reliable statistical analysis of economic   time series data that feed into accurate GDP estimation is extremely important in key decision making across businesses and government. The objectives of the proposed research programmes are as follows:  to develop a fundamental understanding of nonlinear time series deep learning models, and  model structure optimisation algorithms and investigate probabilistic interpretation of deep learning  time series models;  to investigate and design statistical measures that are based on deep learning models for the detection of  anomaly in time series,  and make sense of the detected anomalies (classification); to  propose deep learning models based data imputation (data cleaning)  and analyse the statistical properties of the proposed approaches;   to investigate algorithms for traditional time series problems such as multiple step predictions, nonstationary tracking, etc,  but based on deep model;  and extensively perform data experiments for the proposed approaches against baseline algorithms, e.g. linear regression models  and parametric models in statistics. This work is jointly supported by EPSRC and Office for National Statistics.