Evaluation of established and new deep learning models for time series equity securities forecasting

Sammanfattning: In this bachelor thesis we investigate the importance of feature selection when making predictions on time series data. We compare how well different deep neural network models perform within equity securities time series prediction, namely the models RNN (Recurrent Neural Network), LSTM (Long Short Term Memory), LSTM with a peephole connection and last but not least GRU (Gated Recurrent Unit). We also briefly look at a simpler prediction model, a regression application of the Support Vector Machine, SVR. We compare the different techniques in terms of prediction accuracy with two commonly used error metrics, the mean absolute error (MSE) and the root mean square error (RMSE). The GRU cell had the highest accuracy out of the models on all of the used stocks, not the LSTM. This could be due to the dataset size being to small for the LSTM implementations to perform at their best. There were no significant improvements in prediction accuracy for feature selection over multiple stocks and time periods. The initial plan for this research was to investigate times series prediction accuracy of Neural ODEs. Extensive research wad made on the topic of Neural ODEs, but we failed to implement it properly on time series data as a result of knowledge-gaps and time constraints. Hence the study was re-framed, but it is still largely connected to Neural ODEs as we have conducted a brief literature study on one of the most trending statistical learning subjects of recent years.