Enhancing Volatility Prediction: Comparison Study Between Persistent and Anti-persistent Financial Series.

Abstract

Predicting financial volatility is crucial for managing risks and making investment decisions. This research introduces a novel method for creating a prediction model that effectively handles the intricate dynamics of financial time series data. By utilizing the advantages of both time series models and recurrent neural networks, we present two hybrid models: Vanilla-RGARCH and LSTM-RGARCH. These models are designed to overcome the shortcomings of Realized GARCH (RGARCH) and HAR models in representing various stylized facts of financial data. While RGARCH models are proficient in capturing asymmetry, they fail to address long-term memory. Conversely, HAR models are adept at capturing long-term memory. The innovative model combines forecasted values from the RGARCH model with components from the HAR model, including daily, weekly, and monthly realized volatility, within a neural network framework. This combination helps to bypass the complexities involved in directly merging the HAR model with RGARCH. Through this method, our hybrid models provide a thorough depiction of the characteristics of financial data. The proposed approach is evaluated on two distinct types of financial series; persistent and anti-persistent, to demonstrate its robustness and capacity to generalize in different contexts. The performance of hybrid models is compared to that of conventional RGARCH and HAR models, demonstrating their superiority in precise prediction of financial volatility and their ability to capture complex trends observed in real data. In addition, a principal component analysis (PCA) is used to visualize the results and facilitate their interpretation.