Volatility forecasting with Hybrid‐long short‐term memory models: Evidence from the COVID‐19 period
研究将传统时间序列模型与LSTM模型结合，构建混合LSTM模型预测美国三大股指的日内已实现波动率，发现其在COVID-19期间显著提升预测性能，并分析了构建方法对效果的影响。
Abstract Volatility forecasting, a central issue in financial risk modelling and management, has attracted increasing attention after several major financial market crises. In this article, we draw upon the literature on volatility forecasting and hybrid models to construct the Hybrid‐ long short‐term memory ( LSTM) models to forecast the intraday realized volatility in three major US stock indexes. We construct the hybrid models by combining one or multiple traditional time series models with the LSTM model, and incorporating either the estimated parameters, or the predicted volatility, or both from the statistical models as additional input values into the LSTM model. We perform the out‐of‐sample test of our Hybrid‐LSTM models in volatility forecasting during the coronavirus disease 2019 (COVID‐19) period. Empirical results show that the Hybrid‐LSTM models can still significantly improve the volatility forecasting performance of the LSTM model during the COVID‐19 period. By analysing how the construction methods may influence the forecasting performance of the Hybrid‐LSTM models, we provide some suggestions on their design. Finally, we identify the optimal Hybrid‐LSTM model for each stock index and compare its performance with the LSTM model on each day during our sample period. We find that the Hybrid‐LSTM models' great capability of capturing market dynamics explains their good performance in forecasting.