Forecasting Stock Returns Using Artificial Intelligence

Neural networks have been increasingly utilized for different tasks in the financial sector. The main goal of this master thesis is to compare neural networks’ short-term forecasting ability against a conventional time series forecasting model. A portion of the work addresses the theoretical foundations of neural networks and the associated machine learning concepts. In the empirical part, we applied two types of LSTM networks and a GARCH model on five daily stock returns data sets and compared their out-of-sample prediction accuracy using RMSE and Diebold-Mariano test. Additionally, we re-stated the task into a classification problem by transforming the models’ forecasts into binary trading signals and applied them in a backtest, where the performance of each strategy was evaluated by its profitability.