Decomposing, Learning, and Predicting Realized Volatilities: A Comparison Analysis From the Global Stock Markets

Accurate volatility prediction is essential for guiding investor decision‐making, assessing financial stability, and managing risk. The efficacy of volatility prediction is an important issue that hinges on selecting relevant factors and applying robust analytical tools. Therefore, we propose a novel decomposition‐learning method that integrates deep‐learning techniques for volatility prediction. Specifically, the study employs convolutional neural networks (CNN) and long short‐term memory (LSTM) networks to capture the nonlinear features in time series data. To enhance the model's predictive capabilities, we introduce singular spectrum analysis (SSA) and develop a feature contribution evaluation algorithm to identify and filter out the factors exerting the greatest influence. Building on this foundation, we construct the SSA‐CNN‐LSTM model that supports dual volatility prediction and evaluates each feature's contribution. We design and implement the framework and algorithms for this new approach, applying it to volatility prediction for major global stock indices. The results show that: (1) the trend, cycle, and perturbation components extracted from realized volatility outperform external factors in prediction; and (2) eliminating the features with the lowest contribution significantly enhances the model's predictive performance, thus providing financial markets with a more accurate volatility prediction tool.