Three model ensemble perspectives on volatility prediction and Model ensemble algorithm based on count weighting

Abstract

Abstract This research is a study of predicting realized volatility of five representative indexes in A-share market. This paper collects 12 classical Volatility Prediction Models and compares their prediction performance systematically. The second step of the research is to collect 9 financial and economic features, which are divided into three feature sets, and make a comparative study on a total of 6 machine learning models of support vector machine (SVM), random forests and four neural networks with different structures. Finally, we propose 3 ensemble perspectives: feature ensemble, structure ensemble, model ensemble. Moreover, in the research of model ensemble, we find the selection of component models in ensemble models is usually performance stably. This can be seen from the counts of component models in ensemble models. As a result, we propose a new algorithm called “Model ensemble algorithm based on count weighting” which performance better than Simple average ensemble model.