Multi-task learning methods for intraday stock trading

Can news text data add a significant value in a multi-task neural network scenario? Which preprocessing method is best suitable to predict price movements? Which time granularity is the best in an intraday trading system? This work addresses these questions. In the first experiment, we have explored different news sentiment methods, both human-rated and word embedding. Provided the first result, we focus on different time granularities and different quantile labelling. The first trading system has let us choose the 2-hours granularity as this decreases the number of signals and the trading costs. Worth mentioning that it has both a selling strategy based only on the predictions and one based on technical trading signals. The third set of neural network experiments is focused on RNN cells and dense layers. The second trading system explores new time periods in the technical analysis signals, a new buy signal filter and a selling strategy based on the previous price label that shows interesting results. A permutation importance computation has been done to find whether the news feature is helping us in the prediction.