Realized Volatility Prediction

Abstract

Objective: The purpose of this paper is to use machine learning and time series models in the context of high frequency trading to forecast stock prices. Method: We analyzed time series models such as ARMA and GARCH, gradient boosting tree model – which is a deep learning model – and machine learning models FFNN and GBM to compare each models’ benefits and drawbacks. To determine the accuracy of each models’ stock forecasting, we calculated the root mean square percentage error (RMSPE). The RMSPE reveals the magnitude of error in relation to the actual values; a lower value is wanted. Results & Conclusion: After evaluating the models against real market data, we found that the machine learning models outperformed the time series models. Machine learning models FFNN, and GBM have an RMSPE of roughly 0.20 and 0.21, respectively, while time series models Garch 1 and 2 had a RMSPE of 0.32 and 0.37, respectively. Therefore, feed-forward neural network and GBM forecast stock prices more accurately than LSTM and time series models. Unsupervised algorithms improve prediction accuracy.