Title: Forecasting realized volatility: A hybrid model integrating BiLSTM with HAR-type models
Authors: Yi Luo - Lancaster University (United Kingdom) [presenting]
Marwan Izzeldin - Lancaster University (United Kingdom)
Mike Tsionas - Lancaster University (United Kingdom)
Abstract: In the last few decades, artificial neural networks have been extensively used as a forecasting method for financial time series in both academia and industry. The major advantage of this approach is the possibility to approximate any linear and nonlinear behaviors without knowing the structure of the data generating process. This makes it suitable for forecasting time series which exhibit long-memory and nonlinear dependencies, like conditional volatility. However, much literature has found that the quality of features fed into the neural networks is crucial to the success of the performance of such models. A hybrid methodology is proposed that combines both heterogeneous autoregressive (HAR)-type models and deep feedforward neural network (DFN) model as well as bidirectional long short-term memory (BiLSTM) model in predicting realized volatility. The results show that BiLSTM-based hybrid model outperforms all other models in out-of-sample forecasting. Additionally, the performance of both DFN-based and BiLSTM-based hybrid model beat their single model counterparts, indicating HAR-type components can be served as effective features in DFN and BiLSTM structure.