Title: Multiscale risk forecasting: A deep learning based ensemble approach
Authors: Kaijian He - Hunan University of Science and Technology (China) [presenting]
Geoffrey Tso - City University of Hong Kong (Hong Kong)
Yingchao Zou - Beijing University of Science and Technology (China)
Abstract: In the volatile financial markets characterized by complex mixture of different underlying dynamics, a new multiscale ensemble approach has been proposed for forecasting more accurate value at risk measures. It takes advantage of a new multiscale technique called variational mode decomposition as the basis for disensembling the underlying risk factors in the multiscale domain. The individual characteristics of these risk factors are modeled using different econometrics and artificial intelligence models. The forecasts for these risk factors serve as the ensemble members. Based on that, the nonlinear ensemble model is employed to aggregate the ensemble members and produce the optimal forecasts. The time varying dynamic weights for the ensemble members are modeled using the deep learning model, including the widely popular Long Short Term Memory (LSTM) etc. Empirical evaluation of the performance of the proposed model have been conducted using the extensive database, constructed by daily observations in the major financial markets. Experiment results confirm that the proposed forecasting models produce an improved forecasting accuracy for the risk estimates. It is found that different risk factors has different time scale focus. Their influence on the joint risk movement is time varying and dynamic, which linear ensemble models fail to capture.