Využití strojového učení při predikci volatility

Abstract

This thesis focuses on modeling the realized volatility of the S&P 500 index by employing both traditional statistical models (EWMA, GARCH, HAR) and selected machine learning methods (XGBoost, CatBoost). The primary objective is to compare the accuracy of volatility forecasts using historical minute data and to evaluate whether advanced machine learning algorithms can outperform classic approaches. The theoretical part defines key concepts such as volatility and realized volatility and describes the main models used in financial markets. It also introduces the concept of machine learning, with an emphasis on supervised learning methods. The practical part includes data analysis, the calculation of daily, weekly, and monthly realized volatility, and the training of individual models. Their predictive capabilities are assessed using selected statistical metrics. The results indicate that machine learning models can, in some cases, offer higher accuracy compared to traditional models, thus serving as an effective tool for portfolio management and risk control.