Comparison of traditional and machine learning methods in modelling financial market assets

This thesis compares methods of traditional statistics and machine learning in predicting logarithmic returns on different types of assets. To better generalize the asset behavior, four ETFs are selected for the paper, namely equity, bond, commodity and real estate. ARIMA and GARCH models are chosen as representatives of traditional methods, while random forests, recurrent neural networks, long short-term memory networks and hybrid neural network model are chosen for machine learning. The theoretical part of the paper introduces the specifics of financial time series, the transformations required to work with traditional statistical models, and then the different modeling methods. In the practical part, the statistical assumptions of the time series are verified and models are developed. Since the ARIMA-GARCH model failed to remove heteroskedasticity from the residuals, only ARIMA models are used for the predictions from the traditional models. Within the machine learning models, grid searches and random searches are defined to handle hyperparameters and appropriate model architectures are selected. The results show that neural network models perform better than ARIMA models, ARIMA and the hybrid model are also the only models that do not outperform constant model which only predicts zero for each period.