Data & Model

VADER (Valence Aware Dictionary and Sentiment Reasoner) is a lexicon and rule-based sentiment analysis model specifically attuned to sentiments expressed in social media (Hutto & Gilbert, 2014). For a given text, the model computes the percentages of words that fall into positive, negative, and neutral. It also computes a normalized weighted composite score, where >= 0.05 indicates positive sentiment, <= -0.05 indicates negative sentiment and between -0.05 and 0.05 indicates neutral sentiment. Additional details of the model can be found in this paper and this GitHub repo. The left shows the composite scores for 3 news titles: "Thanks, bitcoin! Traders say goodbye to quiet weekends", "Should You (or Anyone) Buy Bitcoin", and "Coinbase hangover? Here's why bitcoin may be suffering its steepest slide since February". The model computes a positive, neutral, and negative sentiment in each case.

Here we choose the LSTM structure based on Li et al. (2019), which predicts stock prices without any big compute or big data techniques, so it has a long execution time and we want to improve it. The LSTM structure gives good prediction results, as shown in the figure on the bottom right, where the trend of predicted and true stock prices match well. We believe that parallel data processing and parallel computing would make the stock price predictions faster and easier to use.

To test the scalability of the models, we built and trained models on data sets of different industries and time ranges, and then used parallel model training with GPU and cuDNN toolkit on Harvard Cannon to improve run times.