Closing comments

Note that this does not necessarily mean that the movement of all stocks, in all industries, can be better predicted through inclusion of social media data. However, it does illustrate our point that there is some room for heuristic-based feature generation that may allow additional signals to be leveraged for better predictive outcomes. To provide some closing comments on our experiments, we also notice that the simple GRU and the stacked LSTMs both have smoother predictive curves, and are less likely to be swayed by noisy input sequences. They perform remarkably well at conserving the general trend of the stock. The out-of-set accuracy of these models (assessed with the MAE between the predicted and actual value) tells us that they perform slightly worse than the feedforward network and the simple LSTM. However, we may prefer to employ the models with the smoother...