Forecasting stock volatility with neural network on time varying transition probability

Abstract

Forecasting volatilities of financial security returns are important for many financial applications e.g., portfolio investment construction, risk management and trading strategy. The GARCH model has long been refined to capture the true dynamic of volatility on a security return. By applying the Markov switching to the GARCH model, the source of the temporary high volatility and high persistence of a shock to the volatility can be captured. In this study, we refine the Markov switching GARCH model further by applying the notion of the neural network to approximate the time varying transition probabilities. We aim to achieve a model that provides better return volatility prediction. The realized volatility of the SET index is investigated while many financial-macro data are used as input factors for the neural network from January 2012 to December 2022. From the empirical results, the models of one hidden layer with 3 to 6 nodes are suggested since a minor improvement can be observed. However, the significant improvement is still unclear and the models are exposed to the overfitting problem.